analysis and classification of sounds produced by asian

147
Portland State University Portland State University PDXScholar PDXScholar Dissertations and Theses Dissertations and Theses 5-5-2009 Analysis and Classification of Sounds Produced by Analysis and Classification of Sounds Produced by Asian Elephants ( Asian Elephants ( Elephas Maximus Elephas Maximus) ) Sharon Stuart Glaeser Portland State University Follow this and additional works at: https://pdxscholar.library.pdx.edu/open_access_etds Part of the Animal Sciences Commons, and the Biology Commons Let us know how access to this document benefits you. Recommended Citation Recommended Citation Glaeser, Sharon Stuart, "Analysis and Classification of Sounds Produced by Asian Elephants (Elephas Maximus)" (2009). Dissertations and Theses. Paper 4066. https://doi.org/10.15760/etd.5951 This Thesis is brought to you for free and open access. It has been accepted for inclusion in Dissertations and Theses by an authorized administrator of PDXScholar. Please contact us if we can make this document more accessible: [email protected].

Upload: others

Post on 04-Jul-2022

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Analysis and Classification of Sounds Produced by Asian

Portland State University Portland State University

PDXScholar PDXScholar

Dissertations and Theses Dissertations and Theses

5-5-2009

Analysis and Classification of Sounds Produced by Analysis and Classification of Sounds Produced by

Asian Elephants (Asian Elephants (Elephas MaximusElephas Maximus) )

Sharon Stuart Glaeser Portland State University

Follow this and additional works at: https://pdxscholar.library.pdx.edu/open_access_etds

Part of the Animal Sciences Commons, and the Biology Commons

Let us know how access to this document benefits you.

Recommended Citation Recommended Citation Glaeser, Sharon Stuart, "Analysis and Classification of Sounds Produced by Asian Elephants (Elephas Maximus)" (2009). Dissertations and Theses. Paper 4066. https://doi.org/10.15760/etd.5951

This Thesis is brought to you for free and open access. It has been accepted for inclusion in Dissertations and Theses by an authorized administrator of PDXScholar. Please contact us if we can make this document more accessible: [email protected].

Page 2: Analysis and Classification of Sounds Produced by Asian

THESIS APPROVAL

The abstract and thesis of Sharon Stuart Glaeser for the Master of Science in

Biology were presented May 5, 2009, and accepted by the thesis committee and the

department.

COMMITTEE APPROVALS:

RandyZeli~

Luis Ruedas

<::;;>

Thomas Dolan

David Shepherdson

DEPARTMENT APPROVAL:

Michael Murphy, Chair Department of Biology

Page 3: Analysis and Classification of Sounds Produced by Asian

ABSTRACT

An abstract of the thesis of Sharon Stuart Glaeser for the Master of Science in

Biology presented May 5, 2009.

Title: Analysis and classification of sounds produced by

Asian elephants (Elephas maximus)

Relatively little is known about the vocal repertoire of Asian

elephants (Elephas maximus), and a categorization of basic call types and

modifications of these call types by quantitative acoustic parameters is needed to

examine acoustic variability within and among call types, to examine individuality,

to determine communicative function of calls via playback, to compare species and

populations, and to develop rigorous call recognition algorithms for monitoring

populations.

This study defines an acoustic repertoire of Asian elephants based on

acoustic parameters, compares repertoire usage among groups and individuals, and

validates structural distinction among call types through comparison of manual and

automated classification methods. Recordings were made of captive elephants at

the Oregon Zoo in Portland, OR, USA, and of domesticated elephants in Thailand.

Acoustic and behavioral data were collected in a variety of social contexts and

environmental noise conditions. Calls were classified using perceptual aural cues

Page 4: Analysis and Classification of Sounds Produced by Asian

plus visual inspection of spectrograms, then acoustic features were measured, then

automated classification was run. The final repertoire was defined by six basic call

types (Bark, Roar, Rumble, Bark, Squeal, Squeal, and Trumpet), five call

combinations and modifications with these basic calls forming their constituent

parts (Roar-Rumble, Squeal-Squeak, Squeak train, Squeak-Bark, and Trumpet­

Roar), and the Blow. Given the consistency of classifications results for calls from

geographically and socially disparate subject groups, it seems possible that

automated call detection algorithms could be developed for acoustic monitoring of

Asian elephants.

2

Page 5: Analysis and Classification of Sounds Produced by Asian

1ANALYSIS AND CLASSIFICATION OF SOUNDS PRODUCED BY

ASIAN ELEPHANTS (ELEPHAS MAX/MUS)

by

SHARON STUART QLAESER

A thesis submitted in partial fulfillment of the requirements for the degree of

MASTER of SCIENCE in

BIOLOGY

Portland State University 2009

Page 6: Analysis and Classification of Sounds Produced by Asian

Dedication

This thesis is dedicated to my dear husband, Bobby, who

offered his love and support throughout this adventure and

encouraged me to share my heart with elephants.

Page 7: Analysis and Classification of Sounds Produced by Asian

Acknowledgements

First and foremost I would like to thank my thesis committee for their

expertise and guidance: Dr. Randy Zelick, Dr. Luis Ruedas, Dr. Thomas Dolan,

Dr. David Shepherdson, and Dr. David Mellinger. I am especially thankful to my

advisor, Dr. Zelick, who met the challenges of data collection with enthusiasm and

was so patient with me. I am eternally grateful to the Oregon Zoo elephant handlers,

without whom this study would not have been possible. I especially appreciate the

support from those most involved, April Yoder, Bob Lee, Joe Sebastiani, Jeb Barsh,

and Dimas Dominguez. Dr. Mellinger and Dr. Bolger Klinck, with the OSU/NOAA

Cooperative Institute for Marine Resource Studies, provided invaluable supervision

and support with acoustic analysis at a critical time. I appreciate the support and

encouragement from the Oregon Zoo Conservation Program staff, Anne W amer,

David Shepherdson, and Karen Lewis, and Deputy Director Mike Keele. I thank Dr.

Mandy Cook for insight and guidance with early analysis, and Jean Lea Hofheimer

and Bobbi Estabrook for scoring calls. I thank the managers at the Royal Elephant

Kraal and Elephant Nature Park for the opportunity to record domesticated elephants

in Thailand. Travel funding was granted by Portland State University Marie Brown

and Academically-Controlled Auxiliary Activities Fund. I thank my loving family,

Meghan Martin, and other dear friends who embraced this project with me. Finally, I

am eternally grateful to Nancy Scott and the late Dr. LEL "Bets" Rasmussen for

welcoming me into the elephant research community and paving the road, and to the

many elephant handlers, veterinarians, and elephants who have taught me so much.

ii

Page 8: Analysis and Classification of Sounds Produced by Asian

TABLE OF CONTENTS

ACKNOWLEDGEMENTS ........................................................................................... ii

LIST OF TABLES ......................................................................................................... v

LIST OF FIGURES ....................................................................................................... vi

CHAPTER I: INTRODUCTION ................................................................................... 1

TAXONOMY AND CONSERVATION STATUS ................................................................... 2

ECOLOGY AND COMMUNICATION ................................................................................ 4

VOCAL PRODUCTION AND INDIVIDUALITY IN ACOUSTIC SIGNALS ................................ 6

Low-FREQUENCY SOUNDS ........................................................................................... 9

HEARING .................................................................................................................... 11

PURPOSE AND SIGNIFICANCE OF THIS THESIS RESEARCH ............................................ 12

CHAPTER II: PUTATIVE CALL TYPES AND MANUAL CALL CLASSIFICATION ...................................................................................................... 16

METHODS .................................................................................................................. 16

RESULTS .................................................................................................................... 21

CHAPTER III: CALL DISTRIBUTION AND CALL RA TE - HOW THE REPERTOIRE IS USED BY GROUPS AND INDIVIDUALS .................................. 27

METHODS .................................................................................................................. 27

RESULTS .................................................................................................................... 29

CHAPTER IV: ACOUSTIC PARAMETERS AND AUTOMATIC CLASSIFICATION OF CALLS .................................................................................. 33

METHODS .................................................................................................................. 33

RESULTS .................................................................................................................... 53

lll

Page 9: Analysis and Classification of Sounds Produced by Asian

CHAPTER V: DISCUSSION ...................................................................................... 78

REFERENCES ............................................................................................................. 82

APPENDIX A: DISTRIBUTION PLOTS OF PARAMETERS .................................. 87

HISTOGRAMS AND BOXPLOTS OF ORIGINAL DATA ...................................................... 87

Q-Q PLOTS OF SCALED DAT A .................................................................................... 104

Box PLOTS OF SCALED v ARIABLES FOR IDENTIFYING OUTLIERS .............................. 111

APPENDIX B: CLASSIFICATION ON TREE RAW DAT A OUTPUT ................. 118

APPENDIX C: PRINCIPAL COMPONENT ANALYSIS DAT A - RAW DATA OUTPUT .................................................................................................................... 123

APPENDIX D: ANALYSIS OF SIMILARITY (ANOSIM) DATA OUTPUT ........ 124

APPENDIX E: POWER SPECTRA OF RUMBLES AND NOISE .......................... 127

lV

Page 10: Analysis and Classification of Sounds Produced by Asian

LIST OF TABLES

Table 1: Acoustic ethogram of putative call types showing single calls, trains of calls, and call combinations ........................................................................................... 22

Table 2: Social contexts for call comparison ................................................................ 28

Table 3: Acoustic parameters measured by Osprey ...................................................... 38

Table 4: Calls in dataset for statistical analysis (N=927) ............................................ .42

Table 5: Acoustic parameters that were used in statistical analysis ............................ .45

Table 6: Calls in dataset (N=908) after outliers removed ............................................ .49

Table 7: Confusion matrix for tree ................................................................................ 59

Table 8: Classification rate(%) for tree ........................................................................ 59

Table 9: Summary of pair-wise ANOS IM showing significant difference between all call types ............................................................................................................... 66

Table 10: Median values of acoustic parameters shown by the classification tree and PCA to differentiate basic call types ..................................................................... 67

Table 11: Final ethogram for Bark, Roar, Rumble, Squeak, Squeal, Trumpet, and Blow ...................................................................................................................... 68

v

Page 11: Analysis and Classification of Sounds Produced by Asian

LIST OF FIGURES

Figure 1: Call distribution of the Oregon Zoo herd versus domesticated elephants in Thailand ................................................................................................................ 29

Figure 2: Call distribution of females in the Oregon Zoo herd .................................... 30

Figure 3: Call distribution of one male (Tusko) in the Oregon Zoo herd .................... 31

Figure 4: Mean call rate as a function of social context ...................... ." ........................ 32

Figure 5: Osprey window of Squeak with annotation box and feature box ................. 34

Figure 6: Osprey window of Trumpet with annotation box and feature box ............... 35

Figure 7: Acoustic parameters measured by Osprey that can be visualized ................ 3 7

Figure 8: Boxplot of scaled variables across all call types (blow, bark, roar, rumble, squeak, squeal, trumpet) ....................................................................................... 43

Figure 9: Scaled variables for Roar showing outliers and gradients ..................... .48

Figure 10: Example histograms and boxplots of acoustic parameters for each call type show distribution and variability .......................................................................... 55

Figure 11: Final classification tree model .................................................................... 58

Figure 12: Principal Component Analysis plot for basic call types .......................... 61

Figure 13: Global ANOSIM test for difference in parameters among the call types ... 65

Figure 14: Power spectra of Rumble and background noise at the Oregon Zoo ........ 75

Figure 15: Power spectra of Rumble and background noise at Elephant Nature Park ................................................................................................. 76

Figure 16: Power spectra of rumbles and background noise at the Royal Elephant Kraal ......................................................................................... 77

VI

Page 12: Analysis and Classification of Sounds Produced by Asian

CHAPTER I: INTRODUCTION

Relatively little is known about the vocal repertoire of Asian elephants

(Elephas maximus). Our knowledge of acoustic communication in elephants is based

primarily on research on three wild African elephant (Loxodonta spp.) populations

(Amboseli National Park in Kenya, Etosha National Park in Namibia, central African

forests), three captive African elephant herds (Disney's Animal Kingdom in Lake

Buena Vista, Florida, U.S.A., Vienna Zoo in Austria, and Daphne Sheldrick's

orphanage in Nairobi National Park in Kenya), one wild Asian elephant population

(Gal Oya National Park in Sri Lanka), and one captive Asian elephant herd (Oregon

Zoo in Portland, OR, USA, formerly Washington Park Zoo). Additional elephant

populations are currently under study, but publications are limited.

Over 30 calls have been published for African elephants (Loxodonta spp.)

(Olson, 2004; Stoeger-Horwath et al., 2007), and nine have been published for Asian

elephants (Elephas maximus) (Olson, 2004; McKay, 1973); but most are described

only by context or function and not by acoustic features. Among the published African

elephant calls, six infant calls have been described by both spectral and temporal

features in addition to context and function, so there is advancement in our

understanding of vocal ontogeny (Stoeger-Horwath et al., 2007). However, less than

16 calls produced by sub-adults and adults have been described by temporal or

spectral features, and most of these appear to be variations of the low frequency

rumble that differ in social context and function (Olson, 2004; Langbauer, 2000).

Among the published Asian elephant calls, only the low frequency rumble has been

1

Page 13: Analysis and Classification of Sounds Produced by Asian

described by spectral and temporal features. African elephants are capable of

producing calls ranging from 5 Hz to 9 kHz (Poole, in prep). Infrasonic calls (below

20 Hz) have been recorded from captive Asian elephants (Payne et al., 1986), but the

frequency range of Asian elephant vocalizations is unknown.

TAXONOMY AND CONSERVATION STATUS

Elephants belong to the order Proboscidea and family Elephantidae. The two

extant genera of the family Elephantidae include the African elephants (Loxodonta

africana, Blumenbach 1769) and the Asian elephants (Elephas maximus, Linnaeus

1758).

Subspecies taxonomy of Elephas maximus varies among authors, but the

International Union for Conservation of Nature and Natural Resources (IUCN)

follows Shoshani and Eiesenberg ( 1982) in recognizing three subspecies: E. M.

indicus on the Asian mainland (India and Indochina), E. m. maximus on Sri Lanka,

and E. m. sumatranus on the Indonesian island of Sumatra (IUCN, 2008). These

subspecies designations were based primarily on morphological differences, but

mtDNA variation suggests that E. m. sumatranus is monophyletic (Fleischer et al.,

2001). Borneo's elephants are not currently listed as a subspecies and their origin is

controversial; however, comparison of mtDNA of Borneo elephants to that of Asian

elephants across their range suggest that Borneo's elephants are genetically distinct,

with divergence indicative of a Pleistocene colonization of Borneo (Fernando et al.,

2003). Given genetic distinction, the IUCN could define these taxons as evolutionarily

significant units (ESU) (IUCN, 2008).

2

Page 14: Analysis and Classification of Sounds Produced by Asian

The IUCN uses Sukumar's (2003) estimate of 41,410-52,345 wild elephants

for the global population, but acknowledges the argument by Blake & Hedges (2004)

that these estimates may be inaccurate due to the challenges of surveying forest­

dwelling populations. A more recent estimate lists a wild population of 38,534-52,566

animals (Sukumar, 2006) and a global captive population of approximately 16,000

(Sukumar, 2006; Fischer, 2004). The Asian elephant is listed as Endangered by the

IUCN (Choudhury et al., 2008. Elephas maximus. In: IUCN 2008) and is listed in

Appendix I of the Convention on International Trade in Endangered Species of Wild

Fauna and Flora (CITES) (CITES, 2008). The greatest threat to the Asian elephant is

habitat loss, degradation, and fragmentation driven by an expanding human

population, which leads to increasing conflicts between human and elephant. Poaching

for ivory and other body parts also pose a major threat to the long-term survival of

some populations. Large-scale hunting has reduced populations significantly over

wide areas; but even selective removal of tusked males can result in skewed adult sex

ratios and reduced genetic variation, especially in areas where populations are isolated

(Choudhury et al., 2008. Elephas maximus. In: IUCN 2008).

Although there is genetic evidence that suggests there may be at least two

species of African elephants, namely the African savanna elephant (loxodonta

africana) and the African forest elephant (Loxodonta cyclotis) (Roca et al., 2001), both

the IUCN and CITES classify all African elephants as a single species encompassing

both savanna and forest populations (IUCN, 2008; CITES, 2008). Many members of

the international.elephant management and medical community currently recognize

3

Page 15: Analysis and Classification of Sounds Produced by Asian

the reclassification of Loxodonta africana into two separate species, Loxodonta

africana and Loxodonta cyclotis, with the following subspecies of Loxodonta

africana, L.a. africana (South African bush elephant), L.a. Knochenhaueri (East

African bush elephant), and L.a. oxyotis (West African bush elephant).

The African elephant is listed as Near Threatened by the IUCN (Blanc, 2008.

Loxodonta ajricana. In: IUCN 2008) and is in Appendix I of CITES, with four

populations down-listed to Appendix II (CITES, 2008). Poaching for ivory and meat

remains a significant threat, but the most important perceived threat is habitat loss and

fragmentation caused by an expanding human population and land conversion, which

in tum leads to an increase in human-elephant conflict (Blanc, 2008. Loxodonta

africana. In: IUCN 2008).

ECOLOGY AND COMMUNICATION

Elephants are intelligent, long-lived animals that live in a complex and fluid

society in which several modes of communication play a role in maintaining group

cohesion and social order, and in locating and assessing reproductive state of potential

mates (Langbauer, 2000; Eisenberg et al., 1971). Asian and African elephants have a

similar family structure. Females live in matriarchal herds consisting of other female

relatives and their young. Elephant society is multi-tiered, with family units joining to

form bond groups or larger clans that share a home range and coordinate movements

(Douglas-Hamilton, 1972; Charif et al., 2005). Males are excluded from familial units

at a young age and live as solitary elephants or in temporary association with other

males in "bachelor herds" at the fringes of the female herds (Eisenberg et al., 1971;

4

Page 16: Analysis and Classification of Sounds Produced by Asian

Poole, 1987; Vidya & Sukumar, 2005). Recent studies suggest that bull society may

be more structured than previously thought, and that young bulls may seek the

company of mature bulls (Kate Evans, pers. comms). There is evidence that larger

clans may not form in Asian elephant society, so ranging behavior may be influenced

to a lesser degree by extended kin compared to African savannah elephants (Fowler &

Mikota, 2006).

Data on Asian elephant mating behavior in the wild are scarce and incomplete,

but studies of African savannah elephants suggest that mate choice by females and

mate guarding by males appear to play crucial roles in the reproductive behavior of at

least the African savannah species (Moss, 1983; Poole, 1989).

Infrasonic vocalizations and chemical signaling are considered primary

modalities for long-distance communication (Payne et al., 1986; Poole, 1999);

whereas visual, tactile, auditory, and chemical signals are all important at close range

(Schulte & Rasmussen, 1999). Recent findings with seismic detection and

discrimination suggest that vocalizations traveling in the seismic channel could be

used for both short- and long-distance communication (0' Connell-Rodwell et al.,

2006; O'Connell-Rodwell et al., 2007). Acoustic signals are temporally short-lived and

both sender and receiver must be present for the signal to be communicated, and thus

these signals provide information only on an immediate situation, for example,

location (Langbauer, 2000). Chemical signals are temporally long-lived and the sender

and receiver need not both be present for the signal to be communicated, and thus ·

5

Page 17: Analysis and Classification of Sounds Produced by Asian

these signals provide information on a constant state, for example, reproductive state

(Langbauer, 2000).

VOCAL PRODUCTION AND INDIVIDUALITY IN ACOUSTIC SIGNALS

Mammals can modify certain basic sounds by changing the amplitude,

temporal patterning, or stressing of overtones (Muckenhim in McKay, 1973). Vocal

individuality has been found in other mammalian species, for example bottlenose

dolphins (Tursiops truncatus) (Janik et al., 2006), rhesus macaques (Macaca mulatta)

(Rendall et al., 1998), fallow deer (Dama dama) bucks (Reby et al., 1998), swift fox

(Vulpes velox) (Darden et al., 2003), but it is unknown how or if individual identity is

communicated via the acoustic channel in Asian elephants. Strnctural variation within

a call may serve a communicative function (Green, 1975 & Bayart et al., 1990 in

Soltis et al., 2005), but call structure must first be characterized.

Characteristics of vocalizations that arise from inherent properties of vocal fold

vibration in the larynx (the source) vary independently from properties that arise from

vocal tract resonance (the filter), so either the source or filter may provide the auditory

characteristics a receiver needs to determine identity (McComb et al., 2003 ). Air

driven from the lungs sets the vocal folds in motion. Vibration of the vocal folds in the

larynx has afundamentalfrequency and harmonics. In terrestrial mammals, the

fundamental frequency is determined by how many times the vocal folds vibrate in

one second, measured in cycles per second or Hertz; and the fundamental frequency is

negatively correlated to vocal cord mass (Fitch, 1997). The fundamental frequency

determines the pitch that we hear. Harmonics of sound waves are integer multiples of

6

Page 18: Analysis and Classification of Sounds Produced by Asian

the fundamental frequency; that is, if the fundamental frequency is 25 Hz, then the

harmonics have frequencies of 25 Hz, 50 Hz, 75 Hz, etc. The human auditory system

does not perceive separate harmonics in sound, but harmonics provide the tone quality

or richness of the sound.

The source sound from the larynx passes through the vocal tract, which

selectively amplifies certain frequencies, and thus filters the spectral envelope to

produce peaks calledformants (McComb et al., 2002). As vocal tract changes shape,

its resonance frequencies change, and different formants are produced. Articulators,

t.he tongue and lips, produce more complex configurations of the vocal tract and

hence more complicated formant patterns (Baken, 1996). The average spacing

between consecutive formants, or formant dispersion, can provide information on the

length of the vocal tract in mammals (McComb et al., 2003), with an increase in vocal

tract length resulting in a decrease in formant spacing in many mammalian species

(Sanvito et al., 2007; McComb et al., 2003; Fitch & Hauser, 2002).

The sound energy received is a factor of sound production intensity and

acoustic propagation loss in the sound channel, which is influenced by topography,

boundary conditions (surface and vegetation), and atmospheric structure (temperature,

humidity, pressure) (Garstang, 2004).

The vocal tract of an elephant is the nasal passages of the skull, the trunk, and

the pharyngeal pouch (Garstang, 2004). The pharyngeal pouch is at the posterior one­

third of the nasopharynx (Fowler & Mikota, 2006). The elephant tongue is unable to

protrude from the mouth due to the attachment from the tip of the tongue to the floor

7

Page 19: Analysis and Classification of Sounds Produced by Asian

of the mouth (Fowler & Mikota, 2006), so the tongue is limited in the articulation of

sounds.

Adult Asian elephants have a size range of 2,000-5,500 kg (4,410-12,125 lbs)

and 2-3.5 meters (6'7"-11'6") at the shoulder (Shoshani, 2000), and there is a visible

difference between individuals in the size of the head and trunk, or the vocal tract.

Center frequencies of formants depend on the size and posture of the caller, and

different individual African savannah elephants have been found to overlap in these

frequencies (McComb et al., 2002). However, there has been no comparative work

done on the size of the vocal folds of elephants.

Soltis et al. 2005 (2005) found that structural variation in the rumble from

adult female African elephants housed at Disney Animal Kingdom (Lake Buena Vista,

Florida, U.S.A.) reflected individual identity and negative emotional arousal of caller.

In the presence of dominant females, subordinate females produced rumbles with low

tonality and unstable pitch compared to rumbles produced outside the presence of

dominant females. Results of acoustic playback studies with African savannah

elephants on female recognition of the contact call (McComb et al., 2003) and on

signal assessment of the musth rumble (Poole, 1999) suggest individual- and size­

related differences in acoustic production, which elicit questions regarding

individualism, social recognition, and size assessment. Adult females were able to

discriminate familiar and unfamiliar contact calls, and it appeared that the fundamental

frequency contour extracted from the harmonics was the key characteristic in the

signal that could be used for distinguishing individual calls at a distance (McComb et

8

·--1

Page 20: Analysis and Classification of Sounds Produced by Asian

al., 2000; McComb et al., 2003). Adult males responded to playback of musth rumbles

in a way that suggests they are able to assess characteristics of the caller in this

acoustic signal, but it is unclear whether they assess size in the signal or whether they

recognize the caller. The musth rumble is a low-frequency vocalization produced only

by males in musth, and is shown to advertise the musth state (Poole, 1987; Poole,

1989). The musth rumble has been studied extensively in the African elephant but not

in the Asian elephant. In response to the musth rumble of high-ranking males, other

musth males approached aggressively, while non-musth males walked away (Poole,

1999), which was consistent with prior studies of this population on the state of musth,

dominance, and a willingness to contest access to females (Poole, 1989). Elephants

are long-lived, intelligent animals, and they meet and interact over a period of

decades, so it is reasonable to hypothesize that males also recognize individuals by

their calls (Poole, 1999), and that they assess size based on prior interactions rather

than assessing the size of an individual by their call. To further investigate the

potential for size assessment in acoustic signals, the acoustic features that have the

potential to code for size (fundamental frequency and formant dispersion) need to be

measured for calls used over distances or barriers where visual access is limited.

LoW-FREQUENCYSOUNDS

lnfrasound is an anthropocentric term for describing frequencies below the

human hearing range, which is nominally 20 Hz to 20 kHz. The long wavelengths of

low-frequency sound are more likely to hit only large objects and be reflected, and are

thus resilient to atmospheric attenuation with distance (Langbauer et al., 1991; Larom

9

Page 21: Analysis and Classification of Sounds Produced by Asian

et al., 1997). Low-frequency sound below 100 Hz shows little attenuation in forest

environments (Marter et al., 1977; von Muggenthaler et al., 2001). fufrasonic

vocalizations of elephants have fundamental frequencies typically in the range

14 Hz to 35 Hz, with harmonics that extend into the audible range (Poole et al., 1988;

Payne et al., 1986; Langbauer et al., 1989; Langbauer et al., 1991). Elephants can

vocalize in the infrasonic range up to levels of 103 +/- 3 dB at 5m from the source

(Garstang et al., 1995). The role and production mechanism of infrasound varies

among species. Other species known to produce infrasonic sounds include the okapi,

Bengal and Siberian tigers, giraffe, and Sumatran rhino (von Muggenthaler, 1992; von

Muggenthaler, 2000; von Muggenthaler et al., 200 I; von Muggenthaler et al., 2003 ).

The hyoid apparatus of elephants differs from the basic mammalian scheme

and allows a greater flexibility of the larynx, which may aid in the production of

infrasound (Langbauer, 2000). This difference in the hyoid apparatus presents a

potential proximate cause for production of infrasound, and the need for long-distance

communication presents an ultimate cause. McComb (2002) suggested that elephants

produce fundamental frequencies in the infrasonic range simply because of their large

size rather than an evolved mechanism for long-distance communication·. McComb

(2002) analyzed the source- and filter-related acoustic features of the African elephant

female contact call in the Amboseli National Park population, measured degradation

of the spectral structure of this call with distance, and performed playback to measure

signal perception and assessment. Harmonics peaks around 115 kHz were most

prominent and had the highest persistence with distances of 0.5 km to 2.5 km. Female

10

Page 22: Analysis and Classification of Sounds Produced by Asian

groups showed signs of detecting calls from distances of 2 km to 2.5 km, but did not

respond as though they were familiar or unfamiliar until the distance narrowed to 1.0

or 1.5 km. These results suggest that although lower frequencies have the potential to

travel further, the most important frequency components for long distance

communication of social identity in African elephants may be in the harmonics around

115 Hz rather than in the infrasonic range of the fundamental frequency of the contact

call (mean 16.8 Hz) (McComb et al., 2002).

HEARING

There is some evidence that elephants may be better adapted for perceiving

acoustic signals in the 100 Hz to 4 kHz range than in the range of the lowest

frequencies produced. Heffner & Heffner (1982) measured the hearing sensitivity of a

captive Asian elephant, and found the elephant to have an audibility curve similar to

that of other mammals, but with a greater sensitivity to low frequencies and lower

sensitivity to high frequencies than any other mammal tested prior to this study. The

elephant's absolute threshold was 16 Hz (at 65 dB) to 12 kHz (at 72 dB), with a 17

Hz to 10.5 kHz hearing range by the 60 dB criterion. Although the elephant had a low

frequency threshold of 16 Hz, it was considerably less sensitive to frequencies below

100 Hz than to those between 100 Hz and 4 kHz, and the maximum sensitivity was

1 kHz (at 8 dB). From 31.5 Hz to 2 kHz the thresholds were close to background noise

level, so it is possible that the animal's actual sensitivity in this region was masked by

background noise. Frequency discrimination tests indicated the elephant's frequency

discrimination was best below 1 kHz.

11

Page 23: Analysis and Classification of Sounds Produced by Asian

Localization of sound relies on the time of arrival to the ear and the intensity of

sound reaching the ear. Because the interaural distance in the elephant is large, sound

reaches one ear long before the other, and the intensity of the sound on the ear closest

to the sou_nd is greater than on the opposite ear due to the buffering effects of the head.

The larger the interaural distance, the greater the ability to localize low frequency

sounds because the greater interaural distance means a larger phase and amplitude and

difference of the incoming sound wave between the two ears. Heffner & Heffner

(1982) conducted localization experiments for a frequency range of 125 Hz to 8 kHz,

source angles of 0° to 60°, and various stimuli. The elephant performed best at

localizing sound below 300 Hz and was virtually unable to localize 4 kHz to 8 kHz.

Pinnae extension appeared to play a role in localizing sound in this study

(Heffner et al., 1982).

PURPOSE AND SIGNIFICANCE OF THIS THESIS RESEARCH

The purpose of this research is to contribute to the basic science of elephant

communication and to conservation efforts. The specific goals are to (1) define an

acoustic repertoire of Asian elephants based on acoustic parameters, (2) investigate

how the repertoire is used by groups and individuals, (3) compare manual and

automated classification to validate structural distinction among call types, (4) provide

a basis for comparing acoustic communication among elephant species and

populations, and (5) explore the potential for using call parameters to develop an

automatic detector of Asian elephant calls for acoustic monitoring applications.

12

-----,

Page 24: Analysis and Classification of Sounds Produced by Asian

Communication is the intended transfer of information (a signal) from a sender

to either an intended or unintended receiver. Communication is intended to confer

some advantage to the sender, receiver or both via natural or sexual selection. This

study analyzes sounds produced by elephants, but there was no examination of an

intended transfer of information by measuring the response of conspecifics, so it is not

implied that these sounds are signals with communicative value. Data were collected

for future investigation of the communicative function of these sounds, but this topic

was not included in this analysis.

This study provides a basis for future research. A categorization of basic call

types and modifications of these call types by quantitative acoustic parameters is

needed to examine acoustic variability within and among call types, to examine

individuality, to determine communicative function of calls via playback, to compare

species and populations, and to develop rigorous call recognition algorithms for

monitoring wild and managed populations.

The task of vocalization classification and speaker identification is common in

bioacoustic analysis. Automated methods are particularly useful with large datasets or

when the repertoire is being compared between individuals or social groups (Deecke

& Janik, 2006). Detection algorithms are used extensively in the passive acoustic

monitoring of marine mammals, and there are many published classification methods

and tools that could be used with other vocalization datasets. Examples of methods

referenced for this current study include classification of African elephant and marine

mammals vocalizations. Campbell et al. (2002) used artificial neural network to

13

Page 25: Analysis and Classification of Sounds Produced by Asian

identify individual female Steller sea lions (Eumetopias jubatus). Clemins et al.

(2005) developed a Hidden Markov Model based on human speech recognition

techniques to automatically classify African elephant vocalizations for call type and

individual.

Classification of sounds by acoustic parameters is the first step in developing

call recognition algorithms for call detection and acoustic monitoring or census. Payne

et al. (2003) provides evidence that elephant calling patterns can be reliable indicators

of group size and composition, both of which are important for acoustic monitoring.

Calls were divided into three structures, single-caller low-frequency, single-caller high

frequency, and multiple-callers low-frequency. The rate of calling increased with

increasing numbers of elephants, and the distribution of these call categories changed

with group composition. Vocalizations may provide another tool for mitigating

human-elephant conflict, which often results in injury or death to both humans and

elephants (Kemf & Santiapillai, 2000). In Sri Lanka, researchers are investigating the

use of speech recognition techniques as a means of remotely identifying individual

elephants as they approach crops (Doluweera et al., 2003).

A library of vocalizations from known individuals in known reproductive

states will facilitate a more rigorous categorization of sounds by acoustic parameters,

and will allow an examination of sources of variability. An understanding of this

variability is needed to reliably determine the meaning of various calls via playback

(Langbauer, 2000). A database of acoustic communication of African savannah

elephants (Loxodonta africana) is currently being developed by the Savannah

14

Page 26: Analysis and Classification of Sounds Produced by Asian

Elephant Vocalization Project (SEVP) in Amboseli National Park (SEVP, 2008) and

by Disney's Animal Kingdom (John Lehnhardt, Joseph Soltis, pers comms). SEVP has

collected more than 70 different elephant call types of African savannah elephants and

linked them to observations of elephant behavior (SEVP, 2008), but the descriptions

of only a subset of these calls have been published. There is no similar database

known to be developed for wild or captive Asian elephants (Elephas maximus), so one

goal of this current study is to contribute to an animal communications database for

future research.

This study aims to provide a basis for comparisons of acoustic communication

in wild and captive Asian elephants, and could potentially serve as a basis for

comparisons between Asian and African elephants. These comparisons may provide

insights into the ecological role such communication plays and the requirements of

counterpart populations. As elephants lose habitat and find themselves under varying

degrees of management, the need to understand their requirements in captivity will

become even more important to the survival of the species (Riddle et al., 2003).

15

Page 27: Analysis and Classification of Sounds Produced by Asian

CHAPTER II: PUTATIVE CALL TYPES AND MANUAL CALL

CLASSIFICATION

METHODS

Study site and study subjects

Data were collected at the Oregon Zoo in Portland, Oregon, U.S.A. and in

Thailand. Data collected at the Oregon Zoo were from a captive community of three

male and four female Asian elephants. During the course of this study, the females

were managed as a single herd with only temporary separations. The males were

housed alone with no visual or physical access to other males; however, they did have

acoustic and olfactory access given their close proximity and the movement of

elephants between enclosures. Males and females were housed together during social

introductions and breeding. The Oregon Zoo elephant exhibit is designed with two

outside yards totaling 3 I 40 m2, and seven inside rooms. The Oregon Zoo is in an

urban setting located near a major interstate. Sources of anthropogenic noise include

highway traffic, air traffic, hydraulics, water, electric fences, zoo construction, a zoo

train, and visitors.

Data collected in Thailand were from two herds of domesticated elephants, one

herd of approximately 80 elephants at the Royal Elephant Kraal in urban Ayutthaya,

and one herd of approximately 30 elephants at the Elephant Nature Park in the rural

Mae Taeng Valley north of Chiang Mai. The Royal Elephant Kraal is a working

elephant village with adult and semi-adult bulls and cows, geriatric cows, and a

nursery of 10-12 calves. The Elephant Nature Park manages injured and abused

16

Page 28: Analysis and Classification of Sounds Produced by Asian

animals in a semi-captive setting where elephants are allowed to roam during the day

within the grounds and chained at night in social groups. At the time of recording,

there were three adult bulls, two sub-adult males, two calves, and females of varying

ages. Sources of anthropogenic noise in both Thailand sites were sporadic road noise,

machinery, water, and faint talking of visitors.

Acoustic data collection

· Data were collected at the Oregon Zoo from February 2005 to March 2009,

and consisted of over 80 hours of observations and recordings, with 56 hours of usable

data. Data were collected continuously for one hour (occasionally 30 min to 2 hours)

during social introductions, breeding events, temporary separations, arrival of a bull,

the death of a matriarch, novel events, and routine husbandry. TQe following data were

collected during each session: social context. behavior, vocalizations, and visual signs

of musth. Reproductive state measured by hormone levels was provided by the Oregon

Zoo. The caller, if identified during observation by sound localization or visual cues,

was dictated into the recorder or noted on a checksheet. One challenge was detecting

and localizing vocalizations at low frequency; however the harmonics often extended

into the audible range, making it possible to detect and localize at close proximity to

the caller.

Data were collected in Thailand from November to December 2009, and

consisted of approximately 6 hours of data. Data were collected continuously for one

hour sessions (or opportunistically for short sessions) during the morning release to

pasture, morning routines, night feedings, greetings between human and elephant, and

17

Page 29: Analysis and Classification of Sounds Produced by Asian

elephant painting. The following data were collected during each session: social

context, vocalizations, and focal animal if there was a focal.

Acoustic data were collected with four systems during the course of the study.

The frequency responses given for these systems are those provided by the

manufacturer, unless otherwise specified. The systems used were: (1) M-Audio

MicroTrack II (20 Hz -20 kHz,+/- 0.5 dB, recorder tested down to 10 Hz+/- 3 dB),

(2) Bruel & Kjaer 4145 condenser microphone (2.6 Hz -18 kHz+/- 2 dB), an ACO

type 012 preamp (flat to 0.5Hz), an ACO PS2000 power supply, and a Racal V-Store

24 Instrumentation Recorder (analog, DC-45.5 kHz), (3) Edirol R-09 mp3 recorder

(20 Hz - 40 kHz+/- 2 dB), (4) acoustic extraction from digital video (Panasonic PV­

DV700, 20 Hz - 20 kHz) that is being used as part of an ongoing behavioral study.

Only the M-Audio and digital video were used to collect data in Thailand. Preliminary

data were collected with an loTech Wavebook 512 12-bit IMHz Data Acquisition

System, USBGear USB Sound card modified to record to DC, and Vetter 820 analog

recorder.

During recording sessions, the microphone was fixed in position with a tripod

to minimize extraneous noise from holding it directly, as per McComb (1996).The

distance to source was variable as subjects moved throughout the recording session, so

distance to source was only approximated for each recording session and there was no

attempt to measure absolute intensity of calls.

A Cambridge Electronic Design (CED) Micro1401 rnkII data acquisition unit

running Spike2 software (v5.12) was used to upload analog acoustic data. Adobe

18

Page 30: Analysis and Classification of Sounds Produced by Asian

Audition (vl.5) was used to convert file formats and split recording channels. Files

were first down-sampled to 25kHz to reduce the file size. The sampling rate was

selected based on the published frequency range of African elephant calls, 5 Hz to

9 kHz (Poole, in prep) which gives a minimum sampling rate at the Nyquist

frequency of 18 kHz. However, energy was found to extend above 12 kHz, so the

original sampling rate of either 32 kHz or 44 kHz was used for final measurements of

acoustic parameters.

Scoring calls and classifying calls into putative call types

The ethogram used for preliminary data collection included vocalizations

described by McKay (1973) and those in the Elephant Husbandry Resource Guide

(Olson, 2004). The ethogram of elephant behaviors in the Elephant Husbandry

Resource Guide is a compilation of ethograms from approximately 30 publications

and manuscripts, so it is a relatively comprehensive source. Ad libitum sampling was

employed to incorporate vocalizations not yet described. Six basic call types were

defined (trumpets, squeaks, squeals, roars, rumbles, and barks), and these call types

and modifications of these call types were fit into published nomenclature where

possible. Combination calls were also defined (trumpet-rumble, trumpet-roar, squeak­

rumble, squeak-squeal, roar-rumble). Call descriptions were compared to those

recorded in an ongoing study in Uda Walawe National Park in southern Sri Lanka, and

a joint ethogram was developed for calls that were common to the captive herd at the

Oregon Zoo and the free-ranging elephants in Sri Lanka (Shermin de Silva, pers

comms.)

19

Page 31: Analysis and Classification of Sounds Produced by Asian

Exemplar calls were presented to elephant handlers and university students to

gather descriptions of how the sounds were perceived compared to familiar non­

elephant sounds, with the aim of providing a description that would allow elephant

handlers and researchers to discuss elephant calls in terms of aural cues; for example,

a Squeal sounds like rubber shoes on a wood floor.

A signal or sound can be viewed as a distribution of energy in time and

frequency. These distributions can be represented several ways. A spectrogram

displays a three-dimensional plot that shows how the sound varies over time, with

frequency on the y-axis, time on the x-axis, and the relative power (or the logarithm of

relative power) at a given point in frequency and time represented as a darkness value.

A power spectrum displays average energy of the signal over a period of time, with

relative power on the y-axis and frequency on the x-axis (frequency domain). A

waveform displays amplitude on y-axis and time on x-axis (time domain).

Praat software ( v4.5.16, Institute of Phonetic Sciences of the University of

Amsterdam, The Netherlands) was used to annotate (score) sounds by adding

boundary markers and typing call notations (call type, caller ID, quality) in an

annotation field within the boundary. Calls were categorized into the putative call

types using perceptual aural cues and visual inspection of spectrograms for

differentiation of the following acoustic parameters: fundamental frequency contour

(start and end frequency, maximum and minimum frequency, inflection), tonality, and

signal duration. Categorizing by aural cues and visual inspection is considered a

manual method of classification. With the exception of blow sounds, every detectable

20

Page 32: Analysis and Classification of Sounds Produced by Asian

occurrence of every call type was scored. Environmental sounds were also scored for

future analyses with call detection. Half of the recordings were first scored by two

trained but inexperienced reviewers, then verified and corrected. Of 1243 sounds

scored by inexperienced reviewers using aural cues and spectrogram inspection, 90

required correction, for an interobserver reliability of 93%. The caller (if known) was

annotated for each call. A quality score of 1 to 4 was assigned to each call using

methods by Cambell et al. (2002), which were based on presence/absence of sporadic

sound overlap (elephant, human, environmental) and a subjective measurement of

degradation by ambient noise (e.g., highway, wind, water, electricity, faint sounds of

visitors, insects).

RESULTS

The ethogram of putative call types with descriptions and exemplar

spectrograms is provided in Table 1. The aural cue and spectrogram descriptions were

used for categorizing sounds into call types. The visual cue was used to help identify

the caller. The published descriptions include only those for Asian elephants.

21

Page 33: Analysis and Classification of Sounds Produced by Asian

Table I: Acoustic ethogram of putative call types showing single calls, trains of calls, and call combinations.

Call type and spectrogram

Basic Sounds

RUMBLE

ROAR "'

'""' ~ ~ 3000

~ Jt

2000

1000

00 05 10 1 5 20 lime ~

BARK "'

dOOO

6000

"""'

2000 - . ...

' c.~-·?l- - "-=-

0 1 O? 0.3 1} 4 05 06 time. s

Definition

Aurul cuf:' : Sounds like a diesel engine. motorboat. distant helicopter: PuJsatmg with a t1a1 pitch. Perceived as low frequency.

Spnrrogram inspt'Crion: Flar fundamental frequency with narrow bandwidth. Pulsating in spectrogram associated w/ amplitude modulation in waveform. Harmonics present.

Visuul me ufcaller: Mouth visibly open. forehead flutter.

Puhlished Jescriprion: A resonant growl (Olson. 2004). Growl modified by resonance in trunk. used in contex1 of mild arousal (McKay. l 97J). Call with fundamental frequency in range l 4 Hz to 30 Hz. lasting 10-1:5 sec.

Aural cuf:': Sounds like a lion roar with changing pitch Pitch sounds like it changes because of amplitude modulation. Perceived as low frequency.

Spt'crrogram inspecrion: Flat or slightly modulated fundamental frequency. Amplitude modulated. with greatest amplitude somewhere during the call rather than at beginning or end. Tonal. noisy. or mixed. Sometimes ends in rumble. Broader bandwidth than rumbles. but inconsistent. Harmonics present.

Visuul cue of caller: Mouth wide open or widens during call. lower abdomen contracts.

Puhlished description : Growl modified by amplitude increase. used in context of long-distance contact. Used primarily by juveniles who have been separated from their group (McKay. 1973 ). Pulsating sound. Loud growl (Olson. 2004 ).

Aural cuf:': Sounds like a short loud grunt. clearing throat. cape buffalo. chuff of cow. cheetah. Ends ahruptly with no roll off. Noisy compared to rumble.

Spt!crrogram inspn-rion: Flat hmdamentaJ frequency. Short duration. Ca!J is usuaJ!y noisy. Harmonics not always visible (or difficult to see with noise).

Visual cue of caller: Mouth open.

Puhlished description: None for A~ian elephants. but may be similar to the bark African elephant calves (Stoeger-Horwath et al .. 2007).

22

Page 34: Analysis and Classification of Sounds Produced by Asian

~ ~ 8000

i 4000

"""'

10000

aooo

0.IJ

Call type and spectrogram

aooo

6000

4000

2000

SQUEAK

0 0 0.2 o.3 o~

SQUEAK TRAIN

.,

SQUEAL

0 4 0.6 ume '>

o.e 1.0

Definition

Aural cue: Sounds like a wet finger across tight rubber surface. audible pedestrian crossing signal. baby alligator. gibbon caJI. High-pitched sound that falls in pitch, powerful initial sound.

Spectrogram inspection: Descending fundamental frequency that is either straight or curved in shape. Small shon blip may precede pitch fall. Harmonics present. Modified by repetition.

Visual cue of caller: Cheeks depressed. mouth open slightly or closed. tail sometimes erect.

Published description: Not heard in wild (McKay. l 973).

Aural cue: Train of repeated Squeaks with short breaks between Squeak sounds of about l sec. Rate of squeaking may change.

Spectrogram inspection: Squeaks are in a train if time between edges (end of one/beginning of next) of Squeaks is less than l sec. Ot was easier to perceive a break of l sec than 0.5 sec. so l sec was used to define a train versus a bout of individual calls.) Amplitude sometimes shows decrease of subsequent Squeaks _towards the end of the train. Harmonics present.

Visual cue of caller: Cheeks depressed, mouth open slight] y or closed. tail sometimes erect.

Published description: Chirp=multiple short squeaks. Possibly in context of conflict (McKay. l 973).

Sounds like rubber shoes on a gym floor, release of air out of a baJloon, windshield wiper on a dry window. a dog whine. Short utterance.

Acoustic description: Flat or modulated pitch. Short duration (less 1-J .5sec from aural cues), usually flat in pitch or difficult to tell (because it looks like just a point). Harmonics present. Modified by repetition.

Visual cue of caller: Cheeks depressed, mouth open slightly or closed.

Published description: none

23

Page 35: Analysis and Classification of Sounds Produced by Asian

0000

~

r~ 4000

2000

aooo

~ "'6000

J .... 2000

6000

~ "'6000

! """' 2000

Call type and spectrogram

SQUEAL TRAIN

IO

SQUEAL LONG

0 0 02 0 4 0 .6 08 I 0 .., '""' '

Definition

Aural cue: Train of repeated Squeal with shon breaks between sounds. Sounds like squeals in staccato. intennittent release of air from a balloon. Each squeal sounds flat in pitch but they jump up and down.

Specrrogram inspecrion: Squeals are in a train if time between edges (end of one/beginning of next ) of Squeals is less than l sec. Flat fundamental frequency of each squeal. with frequency shifts or jumps between Squeals in the train. Harmonics present.

Visual cue of cu.ller: Cheeks depressed. mouth open slightly or closed.

Published description: none

Aural cue: Sounds like continual release of air from a balloon. fingers on a wet baJloon. sliding pitch. sounds Jong.

Specrrogram inspection: Modulated fundamental frequency with many inflection points. Longer than Squeal but duration variable. clearly longer than l .5-2sec. Sound is continuous with no apparent breaks and pitch slides rather than jumps. Frequency values similar to Squeal Trains but with continuous energy rather than breaks. Harmonics present.

Visual cue of caller: Cheeks depressed. mouth open slightly or closed.

Published description: none

Aural cue: Sounds like a trumpet blast. Perceived as a higher frequency call.

Specrrogram inspection: Flat fundamental frequency. Narrowband. Sometimes has broadband energy at higher frequencies (top hat of noise). Harmonics present.

Visual cue of caller: maybe a lifting at base of trunk. trunk sometimes extended. tail sometimes erect sometimes (any behavior indicating extreme arousal).

Published description: Squeak modified by increased amplitude and duration of call. Used in context of extreme arousal. Pulsating sound (McKay. l 973).

Aural cue: Forced exhalation through end of trunk. Not a vocalization.

Specrrogram inspection: Noisy broadband. Looks like a foot scrape. but lower frequency than foot scrape. Harmonics not clearly visible.

Visual cue of caller: air blast from trunk

Published description: May be same as snort in McKay (1973).

24

Page 36: Analysis and Classification of Sounds Produced by Asian

Call type and spectrogram

BLOW-TONAL

250C

1000 rl

~ 1000 !!1

; I

lw'° ii

""' o:i os 10 15 20 2.5 30 35

'""'.

Combination calls

Definition

Aural cue: Sounds like blowing into a glass bottle or rubbing water on the rim of a crystal glass. Very clear.

Specrrogram in~pection: Tonal. Flat fundamental frequency. Harmonics not clearly visible.

Visual cue of caller: Forehead protrudes enough to see at a distance. Produced only by one animal in study (Oregon Zoo male. Tusko).

Published description: none

Calls in which constituent parts are basic calls types. but with a continuous frequency contour and no apparent break in sound between the constituent parts . The criterion of continuity differentiate combination calls produced by a single individual from overlapping calls produced by multiple individuals.

'"" 3500

:woo

~ ;.~

Ii ~ 2000

/! 1500

,,..

"'°

•0000

.,,., ~

rff.Ol'IO

~

1-"""'

'""" '""' ""'

·10

ROAR/BARK + RUMBLE

., ,, •• .. " "

A Roar or Bark ending in a Rumble.

Same acoustic structure as singles calls. but with continuous sound and frequency contour.

May just be a variance of Roar. Many Roars have at least a slight fall off at the end that sounds like a Rumble. In (Stoeger-Horwath er al.. 2007). Roars are broken down into mixed Roars of various combinations.

Any combination of Squeals and Squeaks occurring in a train with less than l sec between the edges of the calls.

Same acoustic structure as singles calls.

A Squeak ending with any variation of Bark. Rumble. or Roar. Usually ends with a Bark.

Same acoustic structure as singles calls. bur with continuous sound and frequency contour.

25

Page 37: Analysis and Classification of Sounds Produced by Asian

Call type and spectrogram

TRUMPET+ ROAR/RUMBLE

.

:;.:; >.

Definition

A Trumper endmg. in a Roar or Rumble. Trumpet ending. in Rumble' is rare .

Same acoustic structure as singles call s. bul with continuous wund and frequency contour.

26

Page 38: Analysis and Classification of Sounds Produced by Asian

CHAPTER III: CALL DISTRIBUTION AND CALL RATE - HOW THE

REPERTOIRE IS USED BY GROUPS AND INDIVIDUALS

METHODS

Comparing call distribution of groups and individuals

To compare incidence of call types among groups and individuals, the number

of calls of each putative call type (Table 1) was totaled for the study sites and for

individuals in the Oregon Zoo herd. For this comparison, the following calls were

removed from the dataset: duplicates of calls recorded with multiple systems, Blows,

and recordings of one female elephant who vocalized on command. All combination

calls with basic call types as the constituent parts were combined into one group called

"combo."

The call rate was not weighted to account for differences in recording time

among sessions. For comparison of the Oregon Zoo herd to domesticated elephants in

Thailand, the recording duration, sample size of calls, and number of subjects were

very different, but the contexts in both sites offered situations of mild to extreme

arousal. For comparing individuals within the Oregon Zoo herd, the call rate was not

weighted because the total recording time for the cows was similar. Also, the caller

has not yet been identified for every call, so these call compositions provide only a

snapshot of how each cow uses the repertoire. Only one male, Tusko, was included in

this comparison because of limited data for the other males.

27

Page 39: Analysis and Classification of Sounds Produced by Asian

Comparing call rate in various social contexts

The rate of calling, regardless of the caller, was compared among the various

contexts in which data were collected at the Oregon Zoo. For this comparison, the

following calls were removed from the dataset: duplicates of calls recorded with

multiple systems, Blows, recordings of one female elephant who vocalized on

command, and males recorded when by themselves. The total number of calls was

2147. The number of calls of each call type per recording session were calculated,

then adjusted to a 60 minute recording duration. The social contexts in which

recordings were made are shown in Table 2. Most of the recordings were scheduled

during social introductions, with some opportunistic recordings.

Table 2: Social contexts for call comparison

Context Description #of sessions

Routine No specific event, normal training, husbandry, shifting 6

Enrichment Novel enrichment in yard (not just browse, maybe new tires or 3 something rare)

Introduction 1 male in with female group (breeding, intro), includes howdy 39

Reunion Females join after separation of group for an Introduction event 1 after introduction

Separation Focal animal is separated and by themselves (not with male as in intro) 2

Reunion Females join after a Separation event 1 after separation

Transfer Arrival of bu11 Tusko 1

Death Euthanasia of matriarch Pet, including time during euthanasia, visitation 1 of herd members to the body, other herd members after leaving the body

28

Page 40: Analysis and Classification of Sounds Produced by Asian

RESULTS

Call distribution between sites

With the exception of the Squeak call. the domesticated elephants recorded in

Thailand produced the same repertoire as the Oregon Zoo elephants. but with a

different distribution of calls, as shown in Figure 1. The Squeak call was made by

Thai elephants. but only when asked to speak or when engaging tourists. Some factors

that potentially account for the difference in call distribution are a difference in social

context. difference in number of animals in the herd. environmental noise, and

individuality.

A.

Corrbo Rurrble

6%

25%

B.

37%

Combo

3%

Squeal

23%

Rumble

Roar 12%

ark

Figure J : Call distribution of the Oregon Zoo herd versus domesticated elephants in Thailand Panel A shows the call distribution of the Oregon Zoo herd (N=2066 calls). Panel B shows the call distribution of domesticated elephants in Thailand (N=279 calls) .

Call distribution among individuals

Individuals within the Oregon Zoo herd use the repertoire quite differently. as

shown in Figure 2 and Figure 3. For most of the recording sessions, the cows were

together as a group, so the difference in calls produced could be a function of

29

Page 41: Analysis and Classification of Sounds Produced by Asian

individuality. differences in roles within the herd. or differences in level of arousal

within the same social context. Investigating these differences is planned with future

analyses. Only one of the three bulls was recorded consistently during the course of

this study, and this bull produced primarily one sound. a tonal Blow.

A.

Trurrpe\

0%

c.

RuntJle Roar ContJo 1% 1%

Bark

0% Squeak

0%

B.

RuntJle ContJo 1% Roar Bark

3°/o . . .'.· ·. ~

Squeal 27%

D.

Figure 2: Call distribution of females in the Oregon Zoo herd.

Squeak 54%

Panel A shows Pet. Panel B shows Sung-Surin (Per's daughter). Panel C shows Rose Tu. Panel D shows Chendra. Panel E shows Tusko (male).

30

Page 42: Analysis and Classification of Sounds Produced by Asian

Run"ble 0% Roar

6% Squeak

0% Squeal

0%

2% Con"bo

Blow-tonal

81 %

0%

Figure 3: Call distribution of one male (Tusko) in the Oregon Zoo herd.

Call rate in various social contexts

Figure 4 shows a comparison of call rates as a function of social context. The

highest rate was in temporary separatjons and reunion after those separatjons. Not

surprising is the low rate of calling during routine husbandry. What is surprising is the

relatively low rate of calling during the death of the matriarch, the transfer in of a bull.

and social introductions. Given that elephants use many modes of communication, it is

possible that the rate of acoustic production for this herd is not a reliable indicator of

arousal, or that what we perceive as a situation that warrants increased arousal does

not actually do so.

31

Page 43: Analysis and Classification of Sounds Produced by Asian

70

60

.. 50 :::J 0 ;i: GI 40 'tii a:

~ 30 c Ill

~ 20

10

0

-~0 ~ «-0

~ ~0

!...°"'"" i..'O'

«,.~

Call Rate by Context

.,p ~

~-s

-o~

J>-'>CJ

'~ ·t:l -o~'b' ~

«-0"

~o ,~

~o~ . ~'lf c-;,0~

<!::-0' 'lf

~<§:' 'b-{b'

c-;,0~

-o~ ~ «-0-s

Context

Figure 4: Mean call rate as a function of social context

~~ ~<:<)

"{b'

!§' ()0

32

Page 44: Analysis and Classification of Sounds Produced by Asian

CHAPTER IV: ACOUSTIC PARAMETERS AND AUTOMATIC

CLASSIFICATION OF CALLS

METHODS

Measuring acoustic parameters

Praat scripts were run to make a separate sound file for each vocalizations.

Only sounds of basic call types (Bark, Roar, Rumble, Squeak, Squeals, Trumpet) and

Blows with no overlapping sporadic sounds were selected for measurement. All calls

were measured at the original sample rate of either 32 kHz or 44 kHz.

Osprey (vl.7) on MATLAB (v7.7, Mathworks, Inc.) was used to measure the

acoustic parameters in order to characterize the signal structure of each call type.

Osprey's measurement system was developed for characterizing the marine sounds in

the Macaulay Library of Natural Sounds. Osprey measurements are based primarily on

Fristrup and Watkin's (1993) AcouStat approach, with additional measurements and

modifications that use estimators of central tendency and dispersion that are robust to

outliers, namely the median, interquartile range, and quartile skewness. Measurements

that use these estimators of central tendency have more consistent values at variable

noise levels than measurements that rely on manual selection of signal extremes

(Mellinger & Bradbury, 2007; Cortopassi, 2006). Signal extremes are sensitive to

noise and outliers. In addition, manually measuring signal extremes requires

assessment of signal onset and offset in both time and frequency, which can be

affected by display settings and can be biased by researcher expectation and

experience (Cortopassi, 2006).

33

Page 45: Analysis and Classification of Sounds Produced by Asian

Calls of the same type and sampling rate were merged together into a sequence

of calls for quicker and more consistent measuring. Spacers were added between calls

in the merged file for Rumbles and Squeals to help distinguish separate call s. The

spectrogram parameters in Osprey were set to show separation of harmonics and good

detail in both time and frequency. and were set the same for each call type (window

type=Hamming, hop size=14 , zero padding=lx. frame size (digital samples per FFf)

=512 to 2048, for a filter bandwidth of 63 to 349 Hz). An annotarion box (or selection

box) containing the entire sound in a region of time and frequency was drawn liberally

to contain the entire sound plus some background noise, as shown in the two example

calls in Figure 5 and Figure 6. Because only calls with no overlap of other sporadic

sounds were used for measuring acoustic parameters. the annotation box included only

the ambient noise and the focal sound.

8000

.. ::i:. 6000 >. 0 c .. :::J er .. .:: 4000

2000

0 000 0.05 0 10 015 020 025 030 0 35 0 40 0 45 050

time . s

Figure 5: Osprey window of Squeak with annotation box and feature box. X-axis is time. Y axis is frequency. Coloration is intensity. Annotation box encompasses the entire signal. Feature box encompasses 90% of the signal energy.

34

Page 46: Analysis and Classification of Sounds Produced by Asian

12000

10000

:r 8000

~ c::

" " ~ 6000

4000

2000

0.2 0.4 0.6 0.8 1.0 1.2 1 4 1 8 20 time. s

Figure 6: Osprey window of Trumpet with annotation box and feature box. X-axis is time. Y axis is frequency. Coloration is intensity. Annotation box encompasses the entire signal. Feature box encompasses 90% of the signal energy.

Osprey creates a spectrogram of the annotation box and applies a de-noising

algorithm to the entire spectrogram to remove general ambient noise (Mellinger &

Bradbury. 2007). It then creates a feature box that encompasses the inner 90% of the

signal energy in time and frequency, with the strongest 90% of the signal represented

and the weakest 10% excluded. In the time domain (horizontal axis in the

spectrogram), energy in each column is summed, with the sequence of sums forming a

row vector known as the time envelope. The row vector is then ranked (sorted) highest

to Lowest in energy. and the cumulative sum beginning at the high end is computed

until 90% of the total energy is reached. The earliest and latest time indices included

in this 90% cumulative sum define the time bounds of the feature box. The inner 90%

is the upper 90% after the values in the time envelope are sorted into ascending order.

An analogous process happens in the frequency domain (vertical axis in the

35

Page 47: Analysis and Classification of Sounds Produced by Asian

spectrogram), with the sequence of row sums forming the frequency envelope (power

spectrum), and the 90th percentile of the sorted energies being used to define the

frequency bounds of the feature box.

After the feature box is established, Osprey extracts 28 acoustic features within

this feature box and calculates the signal to noise ratio within the annotation box.

Osprey weights measurements at each instant by the normalized intensity of the signal

at that instant. This weighting means that louder parts, which are least affected by

background noise, have the strongest influence on the measurement value (Mellinger

& Bradbury, 2007).

For calls that have low frequency energy around 20 Hz (Rumbles, Barks, and

Roars), the frequency of the fundamental was verified by measuring the harmonic

spacing of the call in Praat. The annotation box was then drawn so the boundary

matched the fundamental frequency rather than drawing it more liberally to include

energy below the fundamental frequency. The reason for this was twofold: (1) one

goal of this study was to determine the frequency range of elephant calls, so an

accurate measure of the minimum frequency for low frequency calls was needed, and

(2) drawing the annotation box below the fundamental frequency sometimes resulted

in the inner 90% of the signal energy encompassing energy below the frequency that

was verified as the fundamental frequency by harmonic spacing.

36

Page 48: Analysis and Classification of Sounds Produced by Asian

Acoustic parameter definitions

The acoustic parameters measured by Osprey are illustrated in Figure 7 and

described in Table 5. A complete descrjption of the calculation of these parameters is

given by Mellinger & Bradbury (2007).

A.

10000

8000

... ~ 6000 :>. u i :;J

e-- 4000

2000

M11JM20

0

B. M17 M7 time. s

i~)

1· 1 . M25

/

v-Ji\ · 1~ ID \ V\ M27 l. I) I •

~ ~J \ _: · M6 f\ g \ ~ _ii F'I ' \. /\-//\r_J '- ·-J,

~ J ,_,, \t'h (' u~ I I <I u -- !()YI X('·l <·i'l 41l'1 C;)' -I t;f('1 7(11"1 fjf{f ~Yf

M20 ' ''°lt.ielr(\ ti :

Figure 7: Acoustic parameters measured by Osprey that can be visualized Panel A shows the spectrogram. Panel B shows the power spectrum.

37

Page 49: Analysis and Classification of Sounds Produced by Asian

Table3: A dbvO

Parameter Units Description Type

Ml-M6: Feature box

Start Time sec Lowest time index in bounds that encompass the temporal (Ml) inner 90% of the signal strength in the time

envelope.

End Time sec Highest time index in bounds that encompass the temporal (M2) inner 90% of the signal strength in the time

envelope.

Lower Frequency Hz Lowest frequency index in bounds that encompass frequency (M3) the inner 90% of the signal strength in the frequency

envelope.

Upper Frequency Hz Highest frequency index in bounds that encompass frequency (M4) the inner 90% of the signal strength in the frequency

envelope.

Duration sec Width of feature box: M2-Ml temporal (M5)

Bandwidth Hz Height of feature box: M4-M3 frequency (M6)

M7-M14: Central values and variation

Uses measures that do not assume normality: median, quartile ranges, quartile skewness, concentration.

Median Time sec Time at which 50% cumulative signal energy is temporal (M7) reached.

(Measured relative to start of file, so M7 was calculated as M7new=M7-Ml)

Temporal Interquartile sec Concentration of a call around the median time (M7) temporal Range measured as the duration of the interquartile range of (M8) signal energy (Q3-Ql).

Counts energy going forward and back from the median time (M7). Q3=median + 25% of signal energy Q l=median-25% of signal energy

Temporal sec Concentration of a call measured as the time span temporal Concentration encompassing loudest 50% of time envelope values. (M9) Counts energy from the loudest parts down towards

the smallest parts regardless of where the parts occur in time.

Temporal Asymmetry none Skewness of energy along time axis within temporal (MlO) interquartile range (-1.0 to 1.0)

38

Page 50: Analysis and Classification of Sounds Produced by Asian

Parameter Units Description Type

Median Frequency Hz Frequency at with 50% cumulative signal energy is frequency (Mll) reached.

More stable than extreme values of LowerFreq and UpperFreq in varying noise conditions.

Spectral Interquartile Hz Concentration of a call around the median frequency frequency Range (M 11) measured as frequency range of interquartile (M12) range of signal energy (Q3-Q 1 ).

Counts energy going forward and back from the median frequency (Ml I). Q3=median + 25% of signal energy Ql=median-25% of signal energy

Spectral Concentration Hz Concentration of a call measured as the frequency frequency (M13) span encompassing loudest 50% of frequency

envelope values. Counts energy from the loudest parts down towards the smallest parts regardless of where the parts occur in time.

Frequency Asymmetry none Skewness of energy along frequency axis within frequency (Ml4) interquartile range (-1.0 to 1.0)

M15-M20: Peak intensity

Time of Peak Cell sec Time of single loudest spectrogram cell. temporal Intensity Time of the cell containing the peak intensity. (Ml5) (Measured relative to start of file, so MIS was

calculated as M15new=M15-Ml)

Relative Time of Peak % Relative time of peak intensity (M15/M5) temporal Cell Intensity (M16)

Time of Peak Overall sec Largest value in time envelope, which is the largest temporal Intensity vertical sum of the spectrogram over all frequencies. (M17) Time of the peak intensity in the trimmed time

envelope. (Measured relative to start of file, so M7 was calculated as Ml 7new=Ml 7-Ml)

Relative Time of Peak % Relative time of peak intensity (M 17 /M5) temporal Overall Intensity (M18)

Frequency of Peak Hz Frequency of cell containing the peak intensity. frequency Cell Intensity (M19)

Frequency of Peak Hz Frequency of peak intensity in the trimmed frequency Overall Intensity frequency envelope. (M20)

39

Page 51: Analysis and Classification of Sounds Produced by Asian

Parameter Units Description Type

M21-M24: Amplitude and frequency modulation (variation of amplitude and frequency over time)

AM Rate Hz Dominant rate of amplitude modulation. amplitude (M21) Frequency of the maximum rate in the power

spectrum of the trimmed time envelope.

AM Rate Variation Hz Variability of amplitude modulation measured as the amplitude (M22) width of peak at M21-6 dB.

Values are discretized because at 6dB down from the peak, the widths may be a only a few bins wide so the values are integer multiples of the bin width.

FM Rate . Hz Dominant rate of frequency modulation. frequency (M23) Frequency of the maximum rate in the power

spectrum of the trimmed frequency envelope.

FM Rate Variation Hz Variability of frequency modulation measured as the frequency (M24) width of peak at M23-6 dB.

(How much the rate of change varies, may be related to inflections and steepness of upsweeps and downsweeps)

M25-M28: Fine features of harmonic structure, shifts in periodicity, direction of frequency change, rate of change in frequency

Cepstrum Peak Width Hz Harmonic structure structure (M25) Average width of peaks (harmonics) in power

spectrum. Peak width is measured at 6 dB down from maximum value. At 6 dB down from the peak, the widths may be a only a few bins wide (like M22 and M24), but M25 is an average of integers so the values are not discretized. Narrow peaks means narrowband/tonal harmonics.

Overall Entropy Hz Entropy, shifts in periodicity structure (M26) Distribution of energy across frequency blocks in a

given time block. Shift from periodicity and linearity to chaos. Change in noisiness v. tonality.

Upsweep Mean Hz Direction of frequency change frequency (M27) Measures how much the frequency increases.

Average change in median frequency between successive time blocks, weighted by total energy in the block. Inflection points with rising and falling frequencies throughout call result in a low M28 (closer to 0) compared to a consistent directional change. Measure is weighted to emphasize contribution oflouder signal components. M27<0 means frequency is decreasing.

40

Page 52: Analysis and Classification of Sounds Produced by Asian

Parameter Units Description Type

Upsweep Fraction % Rate of directional frequency change frequency (M28) Counts the number of times the frequency content

increases. Fraction of time in which the median frequency in one block is greater than in the preceding block, weighted by total energy in the block. Indicates how much of the call has a directional change in the frequency. Inflection points with rising and falling frequencies throughout call result in a high M28, just as a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M28 always positive.

M29: Signal strength

Signal-to-Noise Ratio dB Signal to noise ratio within the annotation box. amplitude (M29) Ratio of the signal power (loudest cell) to the noise

power (power of cell at 25th percentile). Cells are ranked low to high and the cell at the 25th percentile represents noise. (25th percentile is used because the animal call likely takes up less than 75% of the total spectrogram cells.) Measurement assumes that the within the annotation box at least 25% of the cells are without a focal signal.

Dataset for measuring acoustic parameters

The original dataset contained 2791 calls representing all call types, including

combination calls. Acoustic parameters were measured for basic call types (Bark,

Roar, Rumble, Squeak, Squeals, Trumpet) and Blows having no overlapping sporadic

sounds and no echo (N=lOl 1). Calls that met the following criteria were removed

from this dataset for statistical analysis: duplicate calls using different recording

systems (N=83) and calls with a low signal-to-noise ratio (SNR< 10 dB) (N=l). The

final dataset used for statistical analysis is shown in Table 4. It is important to note

that the number of samples per call type was not balanced, so the statistical power is

41

Page 53: Analysis and Classification of Sounds Produced by Asian

low. Also, Blow sounds were not scored for every occurrence, so the sample size was

very low relative to the number of occurrence.

Table 4: Calls in dataset for statistical analysis (N=927)

Call type Oregon Zoo (N) Thailand (N) Total (N)

Blow 73 5 78

Bark 55 8 63

Roar 98 10 108

Rumble 38 21 59

Squeak 193 5 198

Squeal 211 31 242 (includes Squeal, Squeal train, and Squeal long,)

Trumpet 123 56 179

TOTAL 791 136 927

Process for statistical analysis for call classification

The goal of the statistical analysis was to determine if the data support

automatic classification of calls that were previously categorized using perceptual

aural cues and visual inspection of spectrograms, and to determine which acoustic

parameters differentiate the call types. The process for the analysis was as follows:

1) screen data to investigate distributions and variability, 2) standardize data to

account for different units of measure, 3) determine parameters to use for statistical

analysis, 4) remove outliers, 5) determine parametric or non-parametric statistical

analyses, and 6) run statistical analysis for automatic classification. All statistics were

run using R (v2.8.1).

Screening data for statistical analysis

Histograms and boxplots were created to determine the distribution and

variability among call types, to do a preliminary assessment of variables that

42

Page 54: Analysis and Classification of Sounds Produced by Asian

differentiate call types, and to learn more about how these variables reflect perceptual

aural cues.

Standardizing the data for statistical analysis

The acoustic parameters are measured in different units, so the data must be

standardized, or scaled, for the variables to be dimensionally homogeneous. A

boxplot of all variables was created for the scaled dataset, as shown in Figure 8.

"' Q)

~ >

-0 Q)

(;3

:ro

15

;o

<.) rf) I

-6

J

0

0

0

0

0 0 0

0

0

0

0

0

0

0 0

0

0 0 0 8 0

0 0 0 ° @ 0 0

o e 0 o 0 0 8 ~ i § @ .o 0 0 0 0

T ' I ' J I I ~ ' I : ' I T I T ~ ~ i I I ! • I i ' aa~$s8¢f e$1~~a~sss~!~1e~~$ . . . . t i i • • i • I t

0

0 0 . 0

0 0

9 I I I I I I I I I I I I I I I

i ... ~ ~ ~ ~ ~ 0 "' "' ~ "'

., ~ "' "' fl N ,., j "' ~

~

~ ::; i ~ ~ ~ ~ i i ~ ~ i ~ I .,. "' t: 1: ::; ::; ::; ::;

e !! ~ '!! E 8 E tr " g E .... .... .. ~ ~ i ti ~ .. !! !! 8 ~ ~ .. 'j ~ ~ 'li 'li .. ,:: g ;. 8 ~ ! a i 0 0: tJ "' ~ ~ ~

... " ~ "' 8 it it ~

,,_ .. ... 0 ,;; " l .,. " if. 'i J if. ~ ~ .s !" .l! '6 .. -g I! u 0 0 "' "' l "' "' :!l: ....

~ .. ... ,,.

t t "" ~ ::; ! !:. ~ ::; E 0 ... ... (j :::> ...

t :::>

Figure 8: Boxplot of scaled variables across all call types (blow, bark, roar, rumble, squeak, squeal, trumpet) Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis.

43

Page 55: Analysis and Classification of Sounds Produced by Asian

Removing parameters from statistical analysis

Some parameters were removed from the statistical analysis because they were

highly correlated with other parameters for this dataset (6 parameters) or were

absolute measures that could be represented by relative measures (4 parameters).

A correlation matrix was created using the scaled dataset of the basic call types

(Bark, Roar, Rumble, Squeak, Squeal, Trumpet) and Blow to determine potential

candidates for removal from further analyses. The correlation matrix measures the

strength and direction of the linear relationship between pair wise combinations of

variables (given by the Pearson correlation coefficient). Pairs of acoustic parameters

having high correlation coefficients were considered candidates for removal. Because

a correlation measures only the strength of linear relationships, it is possible that

additional pair wise combinations had a strong relationship, but the relationship was

not linear.

Correlations were greater than 0.5 for 14 pairs of variables. Although high

correlations could indicate redundancy and therefore suggest candidates for removal

from further analyses, not all candidates were removed because some relationships

may be biologically relevant for investigating acoustic communication of this species

as compared to others. For example, Overall Entropy (M26) was highly positively

correlated with three variables related to frequency range, Upper Frequency (M4 ),

Bandwidth (M6), and Median Frequency (M 11 ), so M26 was a candidate for removal.

However, the relationship of entropy to frequency range may provide insight into

sound production and perception. Frequency of Peak Overall Intensity (M20) was

highly correlated with Low Frequency (M3 ), so M20 was a candidate for removal.

44

Page 56: Analysis and Classification of Sounds Produced by Asian

However, this relationship confirms that the peak intensity occurs at or close to the

fundamental frequency of the call. Other high correlations were related to central

values and variation and their relationship to frequency and duration. These

correlations may indicate greater stereotypy in elephant calls compared to calls of

other species, so the relationship between these variables may be of biological interest.

Further analysis may be needed to determine measurements that are most important

given the study subjects and recording environment.

A total of 18 acoustic parameters were used in the statistical analyses, which

are shown in Table 5 with notation for each parameter that appears in the data output.

-Parameter Notation in Output Description

Lower Frequency (Hz) LowerFreqM3 Lowest frequency index in bounds that (M3) encompass the inner 90% of the signal strength

in the frequency envelope.

Upper Frequency (Hz) UpperFreqM4 Highest frequency index in bounds that (M4) encompass the inner 90% of the signal strength

in the frequency envelope.

Duration (sec) DurationM5 Width of feature box: M2-Ml (MS)

-

Bandwidth (Hz) BandwidthM6 Height of feature box: M4-M3 (M6)

Median Time (sec) MedianTimeM7 Time at which 50% cumulative signal energy is (M7) reached.

(Measured relative to start of file, so M7 was calculated as M7new=M7-Ml)

Temporal Concentration TimeConcentM9 Concentration of a call is measured as the time (sec) span encompassing loudest 50% of time (M9) envelope values.

Counts energy from the loudest parts down towards the smallest parts regardless of where the parts occur in time.

Temporal Asymmetry TimeAsymmM 10 Skewness of energy along time axis within (MIO) interquartile range (-1.0 to 1.0)

45

Page 57: Analysis and Classification of Sounds Produced by Asian

Parameter Notation in Output Description

Median Frequency (Hz) MedianFreqMl 1 Frequency at with 50% cumulative signal (M11) energy is reached.

More stable than extreme values of LowerFreq and UpperFreq in varying noise conditions.

Frequency Asymmetry FreqAsymmMl 4 Skewness of energy along frequency axis within (M14) interquartile range (-1.0 to 1.0)

Relative Time of Peak PkOverallRelTM 18 Relative time of peak intensity (Ml 7 /MS) Overall Intensity (%) (M18)

Frequency of Peak Pk0verallFM20 Frequency of peak intensity in the trimmed Overall Intensity (Hz) frequency envelope. (M20)

AM Rate (Hz) AMRateM21 Dominant rate of amplitude modulation. (M21) Frequency of the maximum rate in the power

spectrum of the trimmed time envelope.

AM Rate Variation (Hz) AMRate V arM22 Variability of amplitude modulation measured (M22) as the width of peak at M21-6 dB.

Values are discretized because at 6dB down from the peak, the widths may be a only a few bins wide so the values are integer multiples of the bin width.

FM Rate (Hz) FMRateM23 Dominant rate of frequency modulation. (M23) Frequency of the maximum rate in the power

spectrum of the trimmed frequency envelope.

FM Rate Variation (Hz) FMRate V arM24 Variability of frequency modulation measured (M24) as the width of peak at M23-6 dB.

(How much the rate of change varies, may be related to inflections and steepness of upsweeps and downsweeps)

Overall Entropy (Hz) EntropyM26 Entropy, shifts in periodicity (M26) Distribution of energy across frequency blocks

in a given time block. Shift from periodicity and linearity to chaos. Change in noisiness v. tonality.

Upsweep Mean (Hz) UpswpMeanM27 Direction of frequency change (M27) Measures how much the frequency increases.

Average change in median frequency between successive time blocks, weighted by total energy in the block. Inflection points with rising and falling frequencies throughout call result in a low M28 (closer to 0) compared to a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M27<0 means frequency is decreasing.

46

Page 58: Analysis and Classification of Sounds Produced by Asian

Parameter Notation in Output Description

Upsweep Fraction (Hz) UpswpFracM28 Rate of directional frequency change (M28) Counts the number of times the frequency

content increases. Fraction of time in which the median frequency in one block is greater than in the preceding block, weighted by total energy in the block. Indicates how much of the call has a directional change in the frequency. Inflection points with rising and falling frequencies throughout call result in a high M28, just as a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M28 always positive.

Removing outliers

Boxplots of all variables in the scaled dataset were created for each call type to

determine potential outliers within each call type. The criteria for determining outliers

was very conservative in order to preserve data that may have biological relevance.

Given that the gradients of variability may have biological relevance, data points were

considered outliers only if they deviated from the overall pattern of distribution and

did not appear to belong to a long tail (gradient), as shown in Figure 9. Calls were

considered outliers only if they were multivaried outliers. Because variance across two

variables could be a result of covariance alone, calls were considered outliers and

removed from the dataset only if they were outliers across at least three variables.

Because calls recorded in Thailand could represent variability in call structure not

found in the Oregon Zoo call repertoire, and the Oregon Zoo data is better represented

in the dataset, outlier Thailand calls were removed only if they sounded like they did

not meet the definition of the calls as established by aural cues.

47

Page 59: Analysis and Classification of Sounds Produced by Asian

8

6

4

CJ'>

g 2

~ > ]

c<:I u

(/J ()

-2

-4 -I

0

0 0 0

0 0

0

8

0

0

0

outliers not outliers

0 0 0 0

0

0

0

0

0 \:) 0 ow; 0 8 Q 0 0 Q 0 I 0 0 -, 0 ~ ~ 8 I

I 0 I I -, 0 I 1-r"T" I 0 I I 0 -, I I I I I 0

0

I -, 0 I -, I I I I I 0 I

A~8~~~taQB~QO~cig~~ _J_~-'-~ .....1....1 I I _LQ I

: ~ ~ 0 : I I

~ 0 ~

0 0

Q

8 8 I I I

"' .. "' ~ ~ "' c ... "' 0 ;;;. "' M .. "' ;:. "' ::;; ::;; ::!: ::;: i "' N "' ,..

"' N

"' r:r c ~ t: ::!: ::;: ::;: ::;: ::;: ~ ::;: ~ ~ ~ ::;:

E E !! '§ § .,-[

,_ U- .. ~ " .. E ;; 2 '!;j ' ;IS c. .. E

~ "- ~ ;:: u lo " c: ~ "- a:: "' a:: e ..

U-

" "O " 0 c -e ~ ~ ::;: ~ ai ::;: "-3 c. 0 c: "' u ~ "' t if !:I.

0. "' "i 0 a:: "- "' " .9 ;;; "' ~ .. -a <T ~ "' .. .. "' ~ ::;: .. c. ::!: E ::;: E 0 "- .... " ;;; f-- ;:: ~

"' ;;;

"-

Figure 9: Scaled variables for Roar showing outliers and gradients Small circles in the plot indicate points that are outside of the interquartile range. Ellipses show examples of data points considered as outliers and data points not considered outliers.

Out of 927 calls, only 19 calls (Blow N=6, Bark N=4, Roar N=3, Rumble N=2,

Squeak N=l, Squeal N=3, Trumpet N=O) met the conservative criteria for outliers.

Table 6 shows the final dataset with outliers removed (N=908).

48

Page 60: Analysis and Classification of Sounds Produced by Asian

Table 6: Calls in dataset (N=908) after outliers removed

Call type Oregon Zoo (N) Thailand (N) Total (N)

Blow 68 4 72

Bark , 51 8 59

Roar 97 8 105

Rumble 36 21 57

Squeak 193 4 198

Squeal 209 30 239 (includes Squeal, Squeal train, and Squeal long,)

Trumpet 123 56 179

TOTAL 777 131 908

Determining parametric or non-parametric statistical analysis

After scaling and removal of outliers, normal probability plots were created for

the scaled dataset (N=908) to determine how well the normal distribution describes the

data. The normal probability plot is a quantile-quantile (Q-Q) plot against the standard

normal distribution. The observed data is ranked and the quantile is calculated. If the

observed data set matches the theoretical distribution, the shape of the plot will be a

straight line where y = x.

Based on inspection of the Q-Q plots, approximately half of the variables were

normally distributed and only slightly skewed, and half were either very skewed or not

normal. Therefore, this dataset does not meet the assumptions of normality, linearity,

and constant variance. By not meeting these assumptions, only non-parametric

statistics can be used.

49

Page 61: Analysis and Classification of Sounds Produced by Asian

Running statistical analyses for automatic classification

A classification tree was run with the original data (non-standardized) as a

preliminary analysis for determining the most important predictors for classification

into the basic call types (Bark, Roar, Rumble, Squeak, Squeals, Trumpet), and to

explore its potential as a predictive model for classifying new calls into the pre­

defined call types.

A Principal Component Analysis (PCA) of the scaled dataset was run to

determine the parameters that explain most of the variance among call types. The main

objective of PCA is to reduce the dimensionality of the dataset. Ideally, PCA should

be multinormal and reasonably unskewed. Although the data did not meet these

assumptions, the clustering of data along principal component axes can still provide

insight into the variables that group and separate call types.

Global and pair-wise Analysis of Similarity (ANOS IM) tests of the scaled

dataset were run to determine if there were significant differences in the acoustic

parameters among the pre-defined call types and between pairs of call types. If two

call types are significantly different in their parameter values, then the dissimilarities

between the call types will be greater than those within each call type. In order to

evaluate the dissimilarity within and between call types as a measure of true distance

between parameter values, the Ward's linkage method was used with the Euclidian

distance. The number of iterations for assessing significance was 1000. A Bonferroni

correction was calculated to reduce the family-wise Type I error from 15% to 4.8%.

Other analyses considered were a Discriminant Function Analysis (DFA) and

unsupervised clustering. DFA is better than PCA for discriminating existing groups

50

Page 62: Analysis and Classification of Sounds Produced by Asian

and classifying into pre-defined categories; however, even the scaled data did not meet

the assumption of multinormality required by DF A. A hierarchical agglomerative

cluster analysis of the scaled dataset was run on the Oregon Zoo data using R.

Although there appeared to be at least a weak group structure in support of the pre­

defined call types, the output plot was illegible given the number of data points so it

was not possible to determine other potential call groupings from inspection.

Unsupervised clustering should be considered for future comparison of call

classification methods.

The Blow sound was included in the distributions, but not the classification

tree, PCA, or ANOSIM tests. This non-vocal sound is made frequent~y by both males

and females and in most situations, so characterizing this sound by its acoustic

structure may be useful for acoustic monitoring. However, preliminary runs of the

classification tree and PCA showed that it complicated the classification and grouping

of the vocal sounds, so it was handled separately from the basic call types.

Comparing low frequency calls to background noise

Although low frequency vocalizations have the potential for communication in

forested environments and over long distances, they serve a communicative function

only if they can be detected by a receiver. Power spectra of low-frequency Rumble

calls and background noise were compared to investigate bandwidth and intensity of

Rumbles in relation to spectral characteristics and intensity of the background noise.

Only data collected with equipment capable of recording below 20 Hz were

used. Paired Rumble and noise samples were selected from the same recording to

51

Page 63: Analysis and Classification of Sounds Produced by Asian

ensure that the recording levels were equal. Noise segments were selected as close as

possible to the Rumble to take into account that a caller may adjust the intensity of

sound production depending on background noise. The noise segments were selected

with a duration approximately equal to the rumble. A frequency range of 14 Hz to

1.2 kHz was used to compare bandpass characteristics because rumbles have a range

of 14 Hz to 1.2 kHz for the inner 90% of the call energy. Thailand recordings were

separated by site because one site is rural (Elephant Nature Park) and the other is

urban (Royal Elephant Kraal).

Osprey uses an arbitrary reference to generate power spectra, where a sample

value of 1 is effectively 0 dB. Hence, comparisons of signal and noise can be made

only if the recording levels are equal so that the power is referenced to the same value.

With this reference to the same value, the signal to noise ratio at any frequency value

is the difference of the signal power (dB) and noise power (dB). To make these

comparison, the signal and noise were not separated within each sound segment;

rather, the spectrums for the paired Rumble and noise segments were compared

visually to estimate the difference between the signal power and the noise power at

various frequency values.

The sample size used in these analysis was small (Oregon Zoo N=3, Elephant

Nature Park N=4, Royal Elephant Kraal N=3), so this does not constitute a complete

noise analysis, but it does offer some insight into the potential of Rumble calls to be

buried in noise in various recording environments. Future analyses include a more

52

Page 64: Analysis and Classification of Sounds Produced by Asian

quantitative analysis of the signal to noise ratio at various frequency values for an

increased number of call and noise samples.

RESULTS

Call types

Six basic vocalization types were identified as structurally different, and were

labeled Bark, Roar, Rumble, Squeak, Squeal, Trumpet. An additional non-vocal

sound, Blow, was also found.

Frequency range, bandwidth, and duration of acoustic repertoire

The frequency range of the basic call types (Bark, Roar, Rumble, Squeak,

Squeal, Trumpet) and Blow encompassing all of the signal energy was 14 Hz to

18 kHz. The frequency range encompassing only 90% of the energy was 14 Hz to

9 kHz. The bandwidth encompassing 90% of the signal energy for individual calls

ranged from 54 Hz to 9 kHz (median 1680 Hz). The duration encompassing 90% of

the signal energy varied from 0.1 sec to 14 seconds (median 0.7 sec).

Variability of acoustic parameters among call types

Only a subset of the graphical results are included here. All plots are provided

in the appendices. By comparing the histograms and boxplots, it appears that the

distribution of each variable differs among call types, as shown in Figure 10. From the

distributions of commonly used measurements (e.g., fundamental frequency and

duration), one can see how the acoustic parameters reflect aural perception. For

example, the fundamental frequency, as measured by Lower Frequency (M3), is

53

Page 65: Analysis and Classification of Sounds Produced by Asian

consistent or stereotypic for Bark (BRK), Roar (ROR), Rumble (RUM) and Squeak

(SQK), but varies more widely for Trumpet (TMP) and even more so for Squeal

(SQL). Blow shows the greatest variance, but this may be because it is a non-vocal

sound that is mostly broadband noise. Consistent with perception by aural cues, Bark,

Roar, and Rumble are lower in frequency than Squeak, Squeal, and Trumpet. Most of

the energy in Rumbles is low, whereas Roars have energy that extends higher. The

Duration (M5) is stereotypic for Trumpet, Squeak, Blow, and Bark, but his highly

variable for Rumble and Squeal. Squeal include the longest calls, followed by Rumble

then Roar.

54

Page 66: Analysis and Classification of Sounds Produced by Asian

A.

~ " 1"

!

c.

:s ~ 0 c .. i: " Q_

''° .. ..

'°1~~~1100 ' ~ .. "" .. I :

100 BRK 61.W

~~,.JL1tn-o 5(10 1000 1500 2000 0 500 1000 \500 21:){10

d!a.6Slowe<freoM3

TMP

100L[ 80

60

• I ~ 1 """ '°" I ""'

ROR l~IL_!. ff

BRK

'~1LJl_/L_1 0 5 10 15 o s •o 1s

dla.6SOurationM5

B.

3000

2500

i 2000

~ 1500 ~ _J 1000

500

0

D.

15

!{)

~ 10 0

:z:;

~ :J

0 5

,-~

~ ~ ~ ··r

BRK ROR RUM SOK

Gal

"

'

~ q::i $ ..L

BRK ROR RUM SOK

Call

--,--

~]

SOL

.l. '

' '

ci::J r TMP

~

8 ~1-SOL TMP

Figure 10: Example histograms and boxplots of acoustic parameters for each call type show distribution and variability. TMP=Trumpet, RUM=Rumble, SQK=Squeak, SQL=Squeal, BLW=Blow, BRK=Bark, ROR=Roar. Panel A and B show the Low Frequency (M3) parameter values or each call type. Panel C and D show the Duration (M5) parameter values for each call type.

55

Page 67: Analysis and Classification of Sounds Produced by Asian

Classification tree

A classification tree was run on the non-standardized data (N=836), with

outliers removed and excluding Blows. The tree served as a preliminary analysis for

. determining the most important predictors for classification, and may have potential as

a predictive model for classifying new calls into the pre-defined call types (Bark,

Roar, Rumble, Squeak, Squeal, Trumpet). The final classification tree model is shown

in Figure 10. Tree pruning was used as cross-validation for determining the final

model. The goal of cross-validation was to evaluate alternative models that reduce the

number of branches, but when the number of branches was reduced , two call types

(Bark and Rumble) were left out and the misclassification rate doubled. The final

model explained 78.4% of the variation among call types using decision rules with

only six variables: Lower Frequency (M3), Upper Frequency (M4), Duration (M5),

Frequency of Peak Overall Intensity (M20), FM Rate Variation (M24), and Upsweep

Fraction (M28).

The overall misclassification rate was 12.2%, which means successful

classification of 87 .8% of the calls into the predicted call type. Table 7 and Table 8

show the confusion matrix and the classification rate for each call type. The Squeak

had the highest successful classification rate (95 .9% ), followed by Squeal (91.2% ),

Trumpet (91.1 % ), Rumble (78.9% ), Roar (73.3% ), then Bark (71.1 % ). Rumbles were

most often misclassified as Roars. Roars were most often misclassified as Rumbles.

Barks were most often misclassified as Trumpets. The sample sizes of Bark and

56

Page 68: Analysis and Classification of Sounds Produced by Asian

Rumble were low, which may impact the decision rules and thus successful

classification of these call types.

The tree was most successful in classifying the group of higher frequency calls

(Squeak, Squeal, and Trumpet), but classifying Squeaks and Squeals required multiple

decision rules. Squeals were best differentiated by having a Lower Frequency (M3)

above 675 Hz, but for those Squeal calls that did not meet this criteria, it appears that

Upper Frequency (M4), Duration (M5), and FM Rate Variation (M24) differentiated

this call type. The Squeal call type is highly variable, and it appears from the decision

rules that misclassifications are primarily with the short Squeal call, which are often

low intensity compared to Squeal trains and long Squeals. Squeaks were best

differentiated by a low value for FM Rate Variation (M24), which means the

frequency modulation rate was relatively consistent throughout the call. For those

Squeaks that did not meet this criteria, it appears that Upper Frequency (M4),

Upsweep Fraction (M28); Duration, and Lower Frequency (M3) differentiated this call

type.

The group of lower frequency calls (Bark, Roar, and Rumble) appears to be

primarily differentiated by low values of Lower Frequency (M3) and Upper Frequency

(M4), then by Duration (M5). Among these call types, the Bark is separated from the

Roar and Rumble by its short duration. The Roar and Rumble are separated by the

Frequency of Peak Overall Intensity (M20), with the Rumble being below 170 Hz.

Given that Roars and Rumbles were misclassified most often as each other, this single

parameter may not be sufficient for separating these call types.

57

Page 69: Analysis and Classification of Sounds Produced by Asian

Vt

00

r ,_~.,.,

·--·--

----~v

.i .. ., .. ,, _

_ _

I l I I I -·

----

-·--

~~Fr

~M<.

:_8!

>! __

__ . __

______

_ _

Dut

lltiO

Oi

< 0

867

5

LO~fll

~l< 325

JK

f~24>"0011

s&

;&.

42/1

1/0I

Of1

/1

OJO

l(l/1

91'0.

'0 l)I

WQ

,"()/3

11

Pk0.11ta

ll~ >"

' 170

1

l--~-~~---.,

R

Rl.I

Gi1

1'11

!\li2

1Z

4113

14S

i()l

()IO

Fig

ure

11:

Fin

al c

lass

ific

atio

n tr

ee m

odel

. (T

MP=

Tru

mpe

t, R

UM

=Rum

ble,

SQ

K=S

quea

k, S

QL

=Squ

eal,

BL

W=B

low

, BR

K=B

ark,

R

OR

=Roa

r)

t2/.l

l1i!

il13!

16J

sL !I

QIO

ll6t/6

10

O/O

ll'lfl/2

1111

9

Page 70: Analysis and Classification of Sounds Produced by Asian

Table 7: Confusion matrix for tree Diagonal cells (in grey) show correct classification into the predicted class. Off-diagonal cells show misclassifications. Columns represent the distribution of calls into each leaf (call type). Rows --r------- --- ------------ -- ------ ---- - .-- ,----------- -- - -- - . .,.. __ ,,_

Bark Roar Rumble Squeak Squeal Trumpet

Bark 42 11 0 0 1 1 Bark class contains 42 barks, 11 roars, 1 squeal, 1 trumpet

Roar 0 77 11 0 2 2 Rumble 4 13 45 0 0 0 Squeak 1 0 0 189 5 3 Squeal 0 0 0 3 218 10 Trumpet 12 4 1 5 13 163

I•~ 59 105 57 197 239 179 # calls of each call type

59 bark calls recognized by tree as bark (42), rumble (4), squeak (1) trumpet (12)

Table 8: Classification rate ( % ) for tree Diagonal cells (in grey) show correct classification into the predicted class. Off-diagonal cells show misclassifications

Bark Roar Rumble Squeak Squeal Trumpet Bark 71.1 10.5 0 0 0.4 0.6 Roar 0 73.3 19.3 0 0.8 1.2 Rumble 6.7 12.4 78.9 0 0 0 Squeak 1.7 0 0 95.9 2.1 1.7 Squeal 0 0 0 1.3 91.2 5.6 Trumpet 20.3 3.8 1.8 2.1 5.4 91.1

t Bark calls classified correctly at 71.1 % with a 28.9% misclassification rate

Parameters measured by Osprey are numerous, and one could limit the

parameters used in the model to those commonly measured by spectrographic analysis

tools to design a tree that could be used widely by researchers of Asian elephant calls.

Principal Component Analysis

A Principal Component Analysis (PCA) of the scaled dataset (N=836) with

outliers removed and excluding Blows was run to determine the parameters that

account for the variance within and among call types. In accordance with broken.stick

validation, which determines whether the observed pattern is significantly different

from a random pattern, only PCl and PC2 should be kept. PCl and PC2 explained

59

Page 71: Analysis and Classification of Sounds Produced by Asian

only 42.8% of the variance, with PCl explaining 24.6% of the variance and PC2

explaining 18.2%.

The PCA plot in Figure 12 shows some support for the grouping of call types,

but with mixing of call types. The relative contribution of each variable to the

principle components, or loading of principal components, highlights a subset of the

variables as being important for separating call types and grouping calls within each

type. The sign of the numerical loading value indicates whether the value of that

parameters increases or decreases along the principal component axes. The loading is

expressed in the eigenvectors of PC 1 and PC2 below:

PCl= (-0.29)LowerFreqM3 + (-0.42)UpperFreqM4 + (-0.38)BandwidthM6 + (-0.44)MedianFreqM11 + (-0.38)Pk0veral1FM20 + (0.23)AMRateVarM22 + (0.21)FMRateVarM24 + (-0.38)EntropyM26

PC2= (-0.14)LowerFreqM3 + (-0.52)DurationM5 + (-0.51)MedianTimeM7 + (-0.51)TimeConcentM9 + (-0.10)Pk0verallFM20 + (0.24)AMRateM21 + (0.24)FMRateM23 + (-0.19)UpswpFracM28

60

-----i i

Page 72: Analysis and Classification of Sounds Produced by Asian

.. !)

0 I ~

' f"~:r~· ~' '..('

,,.-~· ~ -

~-~---r

~ .. c

"" ~· ~ . "' c . Q_

<> BRK -+ . ROR 0 RUM

SOK 0 _J I SOL

TMP

i.J.- j

-5 0

oc 1

Figure 12: Principal Component Analysis plot for basic call types

(~

Oo 0 c

0 ( •

~;'

:::j

0 c•

0

TMP=Trumpet. RVM=Rumble. SQK=Squeak. SQL=Squeal. BL W=Blow. BRK=Bark. ROR=Roar PC I is represented primarily by spectral parameters. so clustering along PC l results primarily from variability in spectral structure. PC2 is represented primarily by temporal parametc-rs. so clustering along PC I results primarily from variability in temporal patterns.

Along PC l axis. the variability is explained by AM Rate Variation ( M22 ).

Overall Entropy (M26 ). and frequency parameters. namely Lower Frequency (M3 ).

Upper Frequency (M4). Bandwidth (M6). Median Frequency (Ml J ). Frequency of

Peak Overall Intensity (M20J. and FM Rate Variation (M24). The loading of AM Rate

V aria ti on (M22) and FM Rate V aria ti on (M24) is positive. so the value of these

61

Page 73: Analysis and Classification of Sounds Produced by Asian

parameters increase along the PCl axis. All other variables decrease along the PCl

axis. Starting at the most positive end of the PCl axis, Rumbles have the lowest values

for Lower Frequency (M3), Upper Frequency (M4), Bandwidth (M6), Median

Frequency (Mll), Frequency of Peak Overall Intensity (M20), and Overall Entropy

(M26), and the highest values for FM Rate Variation (M24) and AM Rate Variation

(M22). ·Moving from right to left along the PCl axis are Rumbles, Roars, Barks,

Squeaks, then Trumpets and Squeals with the highest values for Lower Frequency

(M3), Upper Frequency (M4), Bandwidth (M6), Median Frequency (Ml 1), Frequency

of Peak Overall Intensity (M20), and Overall Entropy (M26) , and the lowest values

for Freq1:1ency modulation (M24) and Amplitude modulation (M22).

Squeaks show the least within-call variation along the PCl axis, so this call

type is the most stereotypic in its spectral structure. Trumpets show the greatest degree

of variation along the PC 1 axis, so this call° type is highly variable in its spectral

structure. These results are consistent with perceptual aural cues of lower and higher

frequency calls, and of stereotypy in the Squeak call. It is difficult to perceive rate of

change in modulation rates, so the results here provide differentiation that would

otherwise be missed. Frequency jumps were perceived in the Squeal and changes

tonality and noisiness were perceived in the Trumpet, so the results here for the

variability in Entropy are consistent with aural cues for these call types.

Along the PC2 axis, the variability is explained primarily by AM Rate (M21 ),

FM Rate (M23), and temporal parameters, namely Duration (M5), Median Time (M7),

Temporal Concentration (M9). There is some contribution by Lower Frequency (M3),

62

Page 74: Analysis and Classification of Sounds Produced by Asian

Frequency of Peak Overall Intensity (M20), and Upsweep Fraction (M28), but the

contribution is small compared to the other parameters, so these were not included in

the interpretation of the plot. The loading of AM Rate (M2 l) and FM Rate (M23) is

positive, so the value of these parameters increase along the PC2 axis. All other

parameters decrease along the PC2 axis. Moving from top to bottom along the PC2

axis are Squeaks with the lowest values for the temporal parameters, followed by

Barks, Trumpets and Roars, then Rumbles, then Squeals with the highest values for

temporal parameters.

Trumpets and Barks show the least within-call variation along the PC2 axis, so

these call types are most stereotypic in their temporal pattern Squeaks and Barks are

the shortest calls. Squeals and Rumbles include the longest calls and show the greatest

degree of variation along the PC2 axis, so these call types are highly variable in their

temporal pattern These results are consistent with perceptual aural cues of at least

duration.

Five of the 6 parameters used as decision rule in the classification were

contributors to the principal components. Only Up Sweep Fraction (M28) was not a

contributor, and this parameter was used only to differentiate a small number of

Squeak calls. In the classification tree, Lower Frequency (M3), Upper Frequency

(M4 ), and FM Rate Variation (M24) separated most of the higher frequency calls

(Squeak, Squeal, Trumpet) from the lower frequency calls (Bark, Roar, Rumble),

which is consistent with the separation along the PCI axis. In the classification tree,

Duration (M5) separated Barks from the other low frequency calls, Roars and

63

Page 75: Analysis and Classification of Sounds Produced by Asian

Rumbles. Roars and Rumbles were separated by Frequency of Peak Overall Intensity

(M20), which is supported by the separation along the PCl axis.

Analysis of Similarity (ANOSIM)

The classification showed a high rate of successful classification, and the

parameters that separated call types were consistent with the parameters that

contributed to the variance in the Principal Component Analysis, but neither of these

analyses determine if the difference among call types is significant. The null

hypothesis states that there are no differences in acoustic parameters among the six

basic call types (Bark, Roar, Rumble, Squeak, Squeal, Trumpet). A test statistic of

differences among groups was computed using an Analysis of Similarity (ANOSJM).

Global and pair-wise ANOSIM tests of the scaled dataset were run to determine if

there were significant differences in the acoustic parameters among the pre-defined

call types and between pairs of call types. Squeak and Squeal and Roar and Rumble

were sometimes challenging to differentiate using perceptual aural cues, and the

grouping in the PCA showed overlap of these call types. ANOSJM tests were used to

determine if these call types were significantly different or if they were gradations of a

composite call type.

The global ANOSJM test showed significant differences among the call types

(R=0.432, p<0.001), as shown in Figure 13. The pair-wise ANOSJM tests with a

Bonferroni-corrected alpha (a=0.0033) indicated a significant difference between

every call type, as shown in Table 9. Not surprisingly, the dissimilarity was greatest

between the Rumble and the higher frequency calls, Squeak (R=0.807, p<0.001),

64

Page 76: Analysis and Classification of Sounds Produced by Asian

Trumpet (R=0.731, p<0.001) and Squeal (R=0.675, p<0.001). The call types that are

challenging to distinguish by aural cues had mid-range R values, namely Squeal and

Squeak (R=0.425, p<0.001) and Roar and Rumble (R=0.417, p<0.001), so it appears

that these call types are indeed distinct. The most similar pairs were Bark and Roar

(R=0.227, p<0.001), Bark and Rumble (R=0.278, p<0.001), and Squeal and Trumpet

(R=0.230, p<0.001), which is consistent with misclassification rates in the

classification trees and the clustering overlap in the PCA plots. These calls were easy

to distinguish by aural cues, so it may be that the acoustic structure is more similar

across acoustic features that are difficult to perceive.

It is important to note that the group sizes were not balanced, so the statistical

power is low, and the pairs with low R values may not be significantly different in a

dataset with balanced group sizes.

R= 0432 P= <QQ01

i~ ~ ' '

I ' ' ' ' ' ' ' ' ' '

I I ' ' '

1 '

~~ I ' : ' ' ' .!.

' ' ' ' '

~~ lj ' :· ' ' '

~ ' ' ' ' ' ' '

~ '

~ ! ' '

~~ ' ' ' ' ' ' ' ' '

' ' ' ' ' ' ' ' : . ' ' ' ' 0 ----'--- , , ~ ~ ~

Between BRK ROR RUM SOK SQL TMP

Figure 13: Global ANOSIM test for difference in parameters among the call types. TMP:::Trumpet, RUM==Rumble, SQK=Squeak, SQL=Squeal, BLW=Blow, BRK==Bark, ROR==Roar Plot shows that the difference between call types is larger than the difference within each call type.

65

Page 77: Analysis and Classification of Sounds Produced by Asian

Table 9: Summary of pair-wise ANOSIM showing significant difference between all call types

Call pair R value Significance Significant Difference (Bonferroni-corrected a=0.0033)

BRKv.ROR 0.227 <0.001 Yes

BRKv. RUM 0.278 <0.001 Yes

BRK v. SQK 0.539 <0.001 Yes

BRK v. SQL 0.448 <0.001 Yes

BRKv. TMP 0.342 =0.001 Yes

RORv.RUM 0.417 <0.001 Yes

ROR v. SQK 0.685 <0.001 Yes

ROR v. SQL 0.453 <0.001 Yes

ROR v. TMP 0.468 <0.001 Yes

RUMv.SQK 0.807 <0.001 Yes RUMv.SQL 0.675 <0.001 Yes

RUMv. TMP 0.731 <0.001 Yes

SQKv. SQL 0.390 <0.001 Yes

SQKv. TMP 0.422 <0.001 Yes

SQLv. TMP 0:230 <0.001 Yes

Descriptions of call types by acoustic parameters

Acoustic measurements using Osprey yielded comparative measurements of

the basic call types (Bark, Roar, Rumble, Squeak, Squeal, Trumpet) and Blow. The

parameters selected for describing the calls types were those that were shown to be

most important in separating call types by the classification tree and PCA. Out of the

18 parameters that· were included in the analyses, 14 were shown to be important in

differentiating call types. Four of the parameters (Bandwidth (M6), Median Frequency

(Ml 1), Median Time (M7), and Temporal Concentration (M9)) were highly correlated

with other parameters and were not used by the classification tree, but were included

in the final description for completeness. Table 10 lists median values for these 14

parameters across the basic call types for quick comparison of acoustic structure.

66

Page 78: Analysis and Classification of Sounds Produced by Asian

Table 10: Median values of acoustic parameters shown by the classification tree and PCA to differentiate basic call types

for the inner 90% of the signal energy rather than th ~, - ----- - --r-.----

Bark Roar Rumble Squeak Squeal Trumpet

Lower Frequency (Hz) 75 124 26 495 1031 484

(M3)

Upper Frequency (Hz) 1065 777 665 1830 2896 4620

(M4)

Duration (sec) 0.5 1.5 2.3 0.2 1.9 0.5

(MS)

Bandwidth (Hz) 991 657 583 1318 1873 4091

(M6)

Median Time (sec) 0.22 0.66 0.98 0.10 0.75 0.24 (M7)

Temporal Concentration (sec) 0.22. 0.74 1.21 0.10 1.07 0.27

(M9)

Median Frequency (Hz) 408 322 165 912 1509 1305

(Mll)

Frequency of Peak Overall Intensity (Hz) 172 237 133 775 1443 1098 (M20)

AM Rate (Hz) 2.0 0.9 0.5 4.7 2.3 2.2

(M21)

AM Rate Variation (Hz) 0.029 0.046 0.372 0.012 0.023 0.023 (M22)

FM Rate (Hz) 2.3 0.8 0.5 4.9 1.5 2.7

(M23)

FM Rate Variation (Hz) 0.006 0.012 0.093 0.003 0.006 0.006

(M24)

Overall Entropy (Hz) 91 42 25 136 145 184

(M26)

Upsweep Fraction (%) 46 50 44 34 47 49

(M28)

Table 11 provides the final acoustic ethogram for the basic call types (Bark,

Roar, Rumble, Squeak, Squeal, Trumpet) plus Blow. This ethogram describes calls

both by aural cues and acoustic parameters. Values for Median Time (M7) and

Temporal Concentration (M9) in relation to Duration (M5) showed that energy was

67

Page 79: Analysis and Classification of Sounds Produced by Asian

distributed fairly evenly around the center of the call for all call types, so these values

were not included in the final ethogram. What we can say instead is that the energy is

distributed around the center of the call for all of the basic call types in the repertoire.

Upsweep Fraction (M28) was not included in the description because it differentiates

only a small number of Squeaks.

All Osprey measurements are made on the upper 90% of the call energy, so

range values may be greater for calls measured with signal processing tools that

measure extreme values.

Table 11: Final ethogram for Bark, Roar, Rumble, Squeak, Squeal, Trumpet, and Blow Measurements are for the inner 90% of the signal energy rather than the entire signal. The spectrograms are not the same frequency range or duration. The images were captures to show the ranQ:e of the oarticular call rvoe.

Call type and spectrogram

BARK Hz

8000

Definition

Aural cue: Sounds like a short loud grunt, clearing throat. cape buffalo. chuff of cow. cheetah. Ends abruptly with no roll off. Noisy compared to rumble.

Visual cue of caller: Mouth open.

Measurements of90o/o signal energy

Duration 0.1 - 1.3 sec (median 0.5)

26 - 240 Hz (median 75)

6000

4000

Lower Frequency

Upper Frequency

Bandwidth

Median Frequency

248 - 4963 Hz (median 1065)

129 - 4910 Hz (median 991)

101- 1182 Hz (median 408)

2000

0.0 0.1 0.2 0.3 0.4 0.5 0.6 times

Freq of Peak Intensity 37 - 840 Hz (median 172)

AM Rate 0 - 21.5 Hz (median 2.0)

AM Rate Variation

FM Rate

FM Rate V aria ti on

Overall Entropy

0- 0.186 Hz (median 0.0.29)

0 - 24.8 Hz (median 2.3)

0 - 0.046 Hz (median 0.006)

17 - 468 (median 91)

68

Page 80: Analysis and Classification of Sounds Produced by Asian

"'

4000

3000

2000

1000

00

GOO

500

""'

""' i '

?00 ~

l 100

I

Call type and spectrogram

0.5

0.ti l 0

ROAR

1.0 1.5 hme s

RUMBLE

'' ,,,.. ,

2.0

'l.t 3.0 J .~ I

Definition

Aural cue: Sounds like a lion roar with changing pitch Pitch sounds like it changes because of amplitude modulation. Perceived as low frequency.

Visual cue of' caller: Mouth wide open or widens during call. lower abdomen contracts.

Measurements of90% signal energy

Duration 0.5 - 3. 7 sec (median 1.5)

Lower Frequency 23 - 361 Hz (median 124)

Upper Frequency 210 - 3763 Hz (median 777)

Bandwidth

Median Frequency

Freq of Peak Intensity

AM Rate

AM Rate Variation

54 - 3650 Hz (median 657)

122 - 1214 Hz (median 322)

40 - 1292 Hz (median 237)

0.3 - 4.3 Hz (median 0.9)

0.012 - 0.464 Hz

FM Rate

FM Rate Variation

Overall Entropy

(median 0.046)

0.3 - 6.8 Hz (median 0.8)

0.012- 0.232 Hz (median 0.012)

11 - 208 (median 42 )

Aural cue: Sounds like a diesel engine. motorboat. distant helicopter. Pulsating with a flat pitch. Perceived as low frequency .

Visual cue of caller: Mouth visibly open, forehead flutter.

Measurements of90% signal energv

Duration 0.9 - 10.0 sec (median 2.3)

Lower Frequency 14 - J 20 Hz (median 26)

Upper Frequency 104 - 2498 Hz (median 665)

Bandwidth 86 - 2422 Hz (median 583)

Median Frequency 26 - 590 Hz (median 165)

Freq of Peak Intensity 17 - 750 Hz (median 133)

AM Rate 0.1 - 3.6 Hz (median 0.5)

AM Rate Variation 0.023 - 2.972 Hz (median 0.372)

FM Rate 0.1 - 2. 7 Hz (median 0.5)

FM Rate Variation 0.023 - 0.929 Hz (median 0.093)

Overall Entropy 4 - 133 Hz (median 25)

69

Page 81: Analysis and Classification of Sounds Produced by Asian

Call type and spectrogram

SQUEAK

8000

6000

4000

2000

0.0 0.1 0.2 0.3 0.4 11me. s

Definition

Aural cue: Sounds like a wet finger across tight rubber surface. audible pedestrian crossing signal. baby alligator. gibbon call. High-pitched sound that falls in pitch. powerful initial sound.

Visual cue of caller: Cheeks depressed. mouth open slightly or closed. tail sometimes erect.

Measurements of 90% signal energv

Duration

Lower Frequency

Upper Frequency

Bandwidth

Median Frequency

Freq of Peak Intensity

AM Rate

AM Rate Variation

FM Rate

FM Rate Variation

Overall Entropy

0. l - l.2 sec (median 0.2)

194 - 883 Hz (median 495)

797 - 3 768 Hz (median l 830)

301 - 3273 Hz (median 1318)

538 - 1905 Hz (median 912)

537- 2024 Hz (median 775)

0.8 - 44.9 Hz (median 4. 7)

0.003 - 0.026 Hz (median 0.0.012)

l.l - 70.5 Hz (median 4.9)

0.003 - 0.032 Hz (median 0.003)

58 - 294 Hz (median 136)

70

Page 82: Analysis and Classification of Sounds Produced by Asian

Call type and spectrogram

SQUEAL

8000

~ /;' 6000

• ! """' 2000

0000

~

/;' 6000

l """' 2000

0.0 05 LO " 2.0 time!>

Definition

Highly variable in strucrure and duration. Variations numbered in aural cues.

Aural cue: ( l) Short utterance . Sounds like rubber shoes on a gym floor. release of air out of a balloon. windsrueld wiper on a dry window. a dog whine. (2) Train of repeated Squeal with short breaks between sounds. Sounds like squeals in staccato. intermittent release of air from a balloon. Each squeal sounds flat in pitch but they jump up and down. (3) Sounds like continual release of air from a balloon. fingers on a wet balloon. sliding pitch. sounds long.

Visual cue of caller: Cheeks depressed. mouth open slightly or closed.

Measurements of 90% signal energ'' ·

Duration 0.1 - 16.5 sec (median 1.9)

Lower Frequency 226 - 1927 Hz (median 1031)

Upper Frequency 851 - 8193 Hz (median 2896)

Bandwidth 129 - 7816 Hz (median 1873)

Median Frequency 575 - 2384 Hz (median 1509)

Freq of Peak Intensity 422 - 2218 Hz (median 1443)

AM Rate

AM Rate Variation

FM Rate

FM Rate Variation

Overall Entropy

0.1 - l 2.3 Hz (median 2.3)

0.006 - 0.087 Hz (median 0.023)

0.1 - 23.6 (median 1.5)

0.006- 0.041 (median 0.006)

35 - 479 (median 145 )

71

Page 83: Analysis and Classification of Sounds Produced by Asian

Call type and spectrogram

TRUMPET

8000

6000

4000

"""'

0.0 02 " 0.6 0.8 10 .., '-'

htne ~

BLOW

"'°"" " 15000 '

'J

"' ~10000 1

! 5000

- - . -- -

0 - -· -QC 02 0.4 06 °' 1.0

"

1 2

Definition

Aural cue: Sounds like a trumpet blast. Perceived as a higher frequency call.

Visual cue of caller: maybe a lifting at base of trunk. trunk sometimes extended. tail sometimes erect sometimes (any behavior indicating extreme arousal).

Measurements of'90o/o signal energv

Duration 0. 1 - 2.4 sec (median 0.5 )

Lower Frequency 54 - 980 Hz (median 485)

Upper Frequency 128 1 - 8926 Hz (median 4620)

Bandwidth

Median Frequency

Freq of Peak Intensity

AM Rate

AM Rate Variation

FM Rate

FM Rate Variation

Overall Entropy

1012 - 8742 Hz (median 4091)

146 - 4454 Hz (median 1305)

129 - 353 1 Hz (median 1098)

0.4 - 27.6 Hz (median 2.2)

0.006- 0.061 Hz (median 0.023)

0.4 - 23.l Hz (median 2.7)

0.006 - 0.244 Hz (median 0.006)

29 - 83 l Hz (median 184)

Aural cue: Forced exhalation through end of trunk. Not a vocalization.

Visual cue of caller: air blast from trunk

Measurements of'90o/o signal energy

Duration 0.3 - 1.2 sec (median 0. 7)

Lower Frequency

Upper Frequency

Bandwidth

Median Frequency

81 - 1512 Hz (median 372)

1448 - 9318 Hz (median 3790)

1217 - 8990 Hz (median 3273)

531 - 4021 Hz (median 1156)

Freq of Peak Intensity 97 - 3295 Hz (median 894)

AM Rate 0.8 - 3.6 Hz (median 1.4)

AM Rate Variation 0.012 - 0.07 Hz (median 0.046)

FM Rate

FM Rate Variation

Overall Entropy

0.8 - 25.1 Hz (median 1.6)

0.012 - 0.046 Hz (median 0.012)

102 - 875 Hz (median 286)

72

Page 84: Analysis and Classification of Sounds Produced by Asian

Nonlinear dynamics of vocal production

These basic call types exhibited some nonlinear dynamics of vocal production,

which include subharmonics, deterministic chaos, bifurcations, biphonation, and

frequency jumps (Mann et al., 2006). Subharmonics are sounds at frequencies other

than the fundamental frequency or integer harmonic of the fundamental that can be

generated by coupled oscillators. Deterministic chaos comprises noisy signals that are

not random noise, but generated by chaotic process. Bifurcation is the rapid shifts

between tonal harmonics and chaos. Biphonation is the generation of two independent

frequencies simultaneously. Frequency jumps are abrupt up or down transitions

between two frequencies or harmonics that are unpredictable (Mann et al., 2006). The

Trumpet exhibited clear evidence of bifurcation, as measured by the Overall Entropy

(M26) parameter and by visual inspection of Trumpet spectrograms. The Squeak and

Squeal also exhibited evidence of bifurcation as measured by the Overall Entropy

(M26) parameter, but it was not evident from visual inspection of the spectrogram.

The Roar and Bark appear to be noisier sounds, but the shift from linearity to non­

linearity is not evident. The Squeal exhibited clear frequency jumps when produced in

a train of Squeals. The degree to which these non-linearities vary with perceived level

of arousal is being investigated, but is not included here.

Comparison of low frequency calls to background noise

Power spectra of Paired Rumbles and noise for the three study sites were

compared to investigate the potential for Rumbles being buried in background noise.

At the Oregon Zoo (Figure 14) the Rumble had multiple peaks below 200 Hz. The

73

Page 85: Analysis and Classification of Sounds Produced by Asian

noise floor started to rise below 300 Hz, with the peak below 1 OOHz. From visual

inspection, the signal-to-noise ratio was less than 10 dB below 100 Hz, but increased

in the 100-300 Hz range. Given this ratio, it appears that signal energy below 100 Hz

could become inaudible due to noise in this environment. In the entire range of 15 Hz

to 20 kHz, noise declined more sharply from lower frequencies to approximately 2

kHz, then declined gradually or flattened. In one recording there was a slight increase

in noise in the 6-8 kHz range.

At the Elephant Nature Park in rural Thailand (Figure 15), the Rumble had

multiple peaks below 200 Hz. The noise floor started to rise sharply below 100 Hz,

with the peak below lOOHz. Noise had the highest power at low frequencies (below

approximately 30 Hz). From visual inspection, the signal-to-noise ratio was less than

10 dB below 100 Hz, but increased in the 100-300 Hz range. Given this ratio, it

appears that the signal energy below 100 Hz could become inaudible due to noise in

this environmentas well. Wind gusts were a factor at this site. In the entire range of 15

Hz to 20 kHz, noise declined sharply from low frequency to approximately 400 Hz,

then declined gradually across the range. In all selected recordings there was a notable

increase in noise in the 2-4 kHz and 6-11 kHz ranges. Based on aural cues at the time

of observation, noise in 6-11 kHz band may be sounds of insects or other arthropods.

At the Royal Elephant Kraal in urban Thailand (Figure 16), the Rumble had

multiple peaks below 200 Hz. The noise floor started to rise below 300 Hz, with the

peak below lOOHz. From visual inspection, the signal-to-noise ratio was less than 10

dB below 200 Hz (except for recording at close proximity), and remained fairly low to

74

Page 86: Analysis and Classification of Sounds Produced by Asian

about 300 Hz. Given this ratio, it appears that signal energy below 200 or 300 Hz

could become inaudible due to background noise in this environment as well. Wind

gusts were a factor at this site. In the entire range of 15 Hz to 20 kHz, noise declined

sharply from 15 Hz to approximately 400Hz, then declined gradually across the range.

In all selected recordings there was a notable increase in noise in the 6-11 kHz range.

With the exception of the increase in the 6-11 kHz range, the noise profile across the

entire range of 14 Hz to 20 kHz looked more similar to the Oregon Zoo than to the

Elephant Nature Park, so the frequency band of the background noise maybe

primarily a result of the degree of urbanization.

.,, ~

iii 80

i 75

fH" 70

65

60

Rumble: 14 Hz to 1.2 Hz

V'v·A~

55 1m 2m 300 4m 500 60J 700 ooo 900 1000 1100

frequency, Hz

Noise: 14Hzto1.2 kHz 100

95

90

85 .,, g

80

j 75

fli 70

65

60 A •. ~r'V-w.,.,A 1 ~W\./W •V'v\r1;.

55 200 400 600 800 1000 1200

frequency, Hz

.,, ~

90

g 85 ~

~ g 80

fli 75

70

Rumble: call bandwidth

ss~~-~-~-~-~-~-~-~-~-~ 60 m ~ ~ B D ~ 400 ~ 500

frequency, Hz

Noise: 14 Hz to 20 kHz 100.---.,.--~--.,.--~----~-.,.--~-~-~

00

BO

20 2oo:i 4000 6000 8000 10000 12000 14000 16000 18000

frequency, Hz

Figure 14: Power spectra of Rumble and background noise at the Oregon Zoo Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

75

Page 87: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz fllr-~~r-~~~~~~~~~~~~~~~~

75

" g BJ ~

~ 55 §

~ 50

45

40

35 ~ ~ 200 400 500 ~ ~ 1200

frequency, Hz

Noise: 14 Hz to 1.2 kHz lllr-~~~~~~~~~~~~~~r-~~~---.

75

70

66

40

~" 35

:J:J 100 200 m 400 500 500 100 soo oo:i 1000 1100 frequency. Hz

Rumble: call bandwidth BOr-~~~~~~~~~~~~~~~~~~--,

'll g 65

" ~ 5_ 60 lg

/\ 55

50

45 ~~40~-,60~~00~~,oo".:-~,20~~,40~~,60~-,-,eo".:--=-200".:--=-220._,_~

frequency, Hz

Noise: 14 Hz to 20 kHz 80r-~~~r-~r-~~~~~~~~~~~~--,

70

60

50

m = == ~1=1~1=1=1~ frequsncy, Hz

Figure 15: Power spectra of Rumble and background noise at Elephant Nature Park Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

76

Page 88: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz 85~~~~~~~~~~~~~~~~~~~~---,

80

70 .., g Ei5 ~.

j ro

i55 50

45

40

35 200 400 600 800 1000 1200

frequency, Hz

Noise: 14 Hz to 1.2 kHz 10~~~~~~~~~~~~~~~~~~~~

60

l 55 ~

-! ro §

~ 45

40

35 v~ 3J 100 200 30'.l 400 500 600 700 000 900 1000 1100

frequency, Hz

.,, . 75

70

f 65

i § 60

!li 55

50

45

Rumble: call bandwidth

40 ~

.,, . ~ c ~

100 200 300 400 500 600 frequency, Hz

Noise: 14 Hz to 20 kHz 10~~~~~~~~~~~~~~~~~~~---,

60

50

i 40

ai .., 30

20

10 = ~~ ~1~1=1~1~1~

frequency, Hz

Figure 16: Power spectra of rumbles and background noise at the Royal Elephant Kraal Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

77

Page 89: Analysis and Classification of Sounds Produced by Asian

CHAPTER V: DISCUSSION

The goals of contributing to the basic science of elephant communication were

met by the results of this research. An acoustic repertoire of Asian elephants based on

acoustic parameters was defined. A comparison of how the repertoire was used by

groups of elephants and individuals was made. The manual methods of classification

based on perceptual aural cues and visual inspection of spectrograms was compared to

automated classification, and structural distinction among call types was validated.

The fact that a limited number of acoustic parameters differentiated call types suggests

that these parameters could be used for detection of elephant calls. The next steps

would be to determine if the parameter set could be reduced further, and to examine

differentiation of elephant sounds from non-elephants, the data of which has already

been collected and is ready to analyze. Finally, the descriptions of elephant sounds

presented in this thesis serve as a basis for comparisons among captive and wild Asian

elephants and between Asian and African elephants.

The final repertoire was defined by 6 basic call types (Bark, Roar, Rumble,

Bark, Squeal, Squeal, and Trumpet), 5 call combinations and modifications with these

basic types as their constituent parts (Roar-Rumble, Squeal-Squeak, Squeak train,

Squeak-Bark, and Trumpet-Roar), and a sound that was produced frequently by many

elephants, the Blow. Results suggest these call types are differentiated by 11 temporal

and spectral parameters, and with future analyses this feature set may be able to be

reduced further.

78

Page 90: Analysis and Classification of Sounds Produced by Asian

Given the high success rate of the classification tree using only 6 parameters as

decision rules, it appears that this tree does have potential as a predictive model for

classifying new calls into pre-defined call types. Consistent with aural cues was the

separation of the higher frequency calls (Squeak, Squeal, and Trumpet) from the lower

frequency calls (Bark, Roar, Rumble), and the confusion between Roars and Rumbles.

Separation of the call types with the Principal Component Analysis appeared to be in

agreement with the separation by the classification tree. Finally, analysis of similarity

tests showed significant difference among the 6 basic call types and between all pairs

of these call types, so these call types are structurally distinct.

Data used in this study were collected in both urban and rural settings with

many sources of noise and recorded with multiple recording systems, and results

suggest that automated call detection is possible in varying recording situations, with

various noise sources and using various recording systems.

Biological sources of variability within call types may include individuality of

the caller, social context, level or arousal and state of motivation, and potentially the

size of the animals and size of the head space. Some call types appear to be stereotypic

in either temporal or spectral structure and other vary widely. The Squeak is a

stereotypic call, the sound production of which includes manipulation of the cheek or

lips. The Squeal is highly variable, but the sound production appears to be the same.

Little is known about sound production in elephants, but it may be that these sounds

are produced by s.ome source other than the vibration of vocal cords alone.

79

Page 91: Analysis and Classification of Sounds Produced by Asian

Results of repertoire usage agreed with (Langbauer, 2000) in that there Was a

difference in calling pattern of males and females, with male elephants producing

fewer call types than their female counterparts (Langbauer, 2000). The repertoire

defined by this study suggests some differences between African elephants and Asian

elephants. The most common vocalization of African elephants is the Rumble, which

is highly variable and is used in multiple contexts and serves multiple functions (Soltis

et al., 2005; Olson, 2004). The Rumble is produced by Asian elephants, but it was not

the most common sound produced in this study. The Asian elephants in this study

were captive or domesticated and were not wild, but the published calls of wild

elephants by McKay (1973) suggest that at least some of the call types defined in this

study are also produced by wild Asian elephants. Ecological differences in populations

of African and Asian elephants include predator pressures, resource availability, group

size, and human encroachment. Home range size may be a function of these

differences, and communication modalities and distance may be a function of home

range and group size. One could hypothesize that a reason for the Rumble being the

primary call of African elephants is a greater need for long distance communication in

order to maintain separation while remaining in contact for the purpose of resource

utilization.

Future analyses for this dataset include completing the caller identification,

investigating the communicative function of these sounds based on behavioral data

already collected, investigating potential explanations for differences in call

distribution among individuals, running unsupervised call classification as another

80

Page 92: Analysis and Classification of Sounds Produced by Asian

validation of established call types, doing a more rigorous analysis of signal-to-noise

ratio at various frequency values, and running non-elephant sounds through a

classification tree as a preliminary test for call detection potential.

81

Page 93: Analysis and Classification of Sounds Produced by Asian

REFERENCES

Baken, R. J. 1996. Clinical Measurement of Speech and Voice. San Diego: Singular Publishing Group, Inc.

Blake, S. & Hedges, S. 2004. Sinking the Flagship: the Case of Forest Elephants in Asia and Africa. Conser1lation Biology, 18, 1191-1202.

Blanc, J. 2008. Loxodonta africana. In: IUCN 2008. 2008 IUCN Red List of Threatened Species. <www.iucnredlist.org>. Downloaded on 11 May 2009.

Campbell, G. S., Gisiner, R. C., Helweg, D. A. & Milette, L. L. 2002. Acoustic identification of female Steller sea lions (Eumetopiasjubatus). The Journal of the Acoustical Society of America, 111, 2920-2928.

Charif, R. A., Ramey, R. R., Langbauer, W. R., Payne, K. B., Martin, R. B. & Brown, L. M. 2005. Spatial relationships and matrilineal kinship in African savanna elephant (Loxodonta africana) clans. Behavioral Ecology and Sociobiology, V57, 327-338.

Choudhury, A., Lahiri Choudhury, D. K., Desai, A., Duckworth, J. W., Basa, P. S., Johnsingh, A. J. T., Fernando, P., Hedges, S., Gunawardena, M., Kurt, F., Karanth, U., Lister, A., Menon, V., Riddle, H., Rubel, A. & Wikramanayake, E. 2008. Elephas maximus. In: IUCN 2008. 2008 IUCN Red List of Threatened Species. <www.iucnredlist.org>. Downloaded on 11May2009.

CITES. 2008. Appendices I, II, and III to Convention on International Trade in Endangered Species of Wild Fauna and Flora <http://www.cites.org>. Downloaded on 11 May 2009. Geneva.

Clemins, P. J., Johnson, M. T., Leong, K. M. & Savage, A. 2005. Automatic classification and speaker identification of African elephant (Loxodonta africana) vocalizations. The Journal of the Acoustical Society of America, 117, 956-963.

Cortopassi, K. A. 2006. Automated and Robust Measurement of Signal Features. Bioacoustics Research Program, Cornell Laboratory of Ornithology, Cornell University.

Darden, S. K., Dabelsteen, T. & Pedersen, S. B. 2003. A potential tool for swift fox (Vulpes velox) conservation: Individuality of long-range barking sequences. Journal of Mammalogy, 84, 1417-1427.

Deecke, V. B. & Janik, V. M. 2006. Automated categorization ofbioacoustic signals: A voiding perceptual pitfalls. The Journal of the Acoustical Society of America, 119, 645-653.

Doluweera, G., Fernando, S., Gunasekera, C., Madanayake, A., Rossel, G., Seneviratne, L., Wi jetunga, S. & Y apa, S. 2003. Infrasound - Can }t be more than a means of communication between Elephants? In: Symposium on Human-Elephant Relationships and Conflicts. Sri Lanka.

Douglas-Hamilton, I. 1972. On the ecology and behaviour of the African elephant. In: Oriel College, pp. 257: Oxford.

82

Page 94: Analysis and Classification of Sounds Produced by Asian

Eisenberg, J. F., McKay, G. M. & Jainudeen, M. R. 1971. Reproductive behavior of the Asiatic elephant (Elephas maximus maximus L.). Behaviour, 38, 193-225.

Fernando, P., Vidya, T. N. C., Payne, J., Stuewe, M., Davison, G., Alfred, R. J., Andau, P., Bosi, E., Kilbourn, A. & Melnick, D. J. 2003. DNA analysis indicates that Asian elephants are native to Borneo and are therefore a high priority for conservation. Plos Biology, 1, 110-115.

Fischer, M. 2004. AZA Elephant TAG/SSP Regional Collection Plan. Saint Louis: Saint Louis Zoo.

Fitch, W. T. 1997. Vocal tract length and formant frequency dispersion correlate with body size in rhesus macaques. The Journal of the Acoustical Society of America, 102, 1213-1222.

Fitch, W. T. & Hauser, M. D. 2002. Unpacking "Honesty": Vertebrate Vocal Production and the Evolution of Acoustic Signals. Springer Handbook of Auditory Research, 16: Acoustic Communication, 65-137.

Fleischer, R. C., Perry, E. A., Muralidharan, K., Stevens, E. E. & Wemmer, C. M. 2001. PHYLOGEOGRAPHY OF THE ASIAN ELEPHANT (ELEPHAS MAX/MUS) BASED ON MITOCHONDRIAL DNA. Evolution, 55, 1882-1892.

Fowler, M. E. & Mikota, S. K. 2006. Biology, Medicine, and Surgery of Elephants. Ames: Blackwell Publishing.

Fristrup, K. M. & Watkins, W. A. 1993. Marine Animal Sound Classification. pp. 29: Woods Hole Oceanographic Institution Technical Report WHOI-94-13.

Garstang, M. 2004. Long-distance, low-frequency elephant communication. Journal of Comparative Physiology A: Sensory, Neural, and Behavioral Physiology, 190, 791 - 805.

Garstang, M., Larom, D., Raspet, R. & Lindeque, M. 1995. Atmospheric controls on elephant communication. The Journal of Experimental Biology, 198, 939-951.

Heffner, R., Heffner, H. & Stichman, N. 1982. Role of the elephant pinna in sound localization. Animal Behaviour, 30, 628-630.

Heffner, S. R. & Heffner, H. E. 1982. Hearing in the elephant (Elephas maximus): absolute sensitivity, frequency discrimination, and sound localization. Journal of Comparative and Physiological Psychology, 96, 926-44.

IUCN. 2008. 2008 IUCN Red List of Threatened Species. <www.iucnredlist.org>. Downloaded on 11 May 2009.

Janik, V. M., Sayigh, L. S. & Wells, R. S. 2006. Signature whistle shape conveys identity information to bottlenose dolphins. Proc Natl Acad Sci US A, 103, 8293-7.

Kemf, E. & Santiapillai, C. 2000. Asian Elephants in the Wild. In: WWF Species Status Report (Ed. by Wilson, A.). Gland, Switzerland: WWF-World Wildlife Fund for Nature.

Langbauer, J., William R., Payne, K. B., Charif, R. A., Rapaport, L. & Osborn, F. 1991. African elephants respond to distant playbacks of low-frequency conspecific calls. The Journal of Experimental Biology, 157, 35-46.

83

Page 95: Analysis and Classification of Sounds Produced by Asian

Langbauer, J., William R., Payne, K. B., Charif, R. A. & Thomas, E. M. 1989. Responses of captive African elephants to playback of low-frequency calls. Canadian Journal of Zoology, 67, 2604-2607.

Langbauer, W.R., Jr. 2000. Elephant Communication. Zoo Biology, 19, 425-445. Larom, D., Garstang, M., Lindeque, M., Raspet, R., Zunckel, M., Hong, Y., Brassel,

K., O'Beime, S. & Sokolic, F. 1997. Meteorology and elephant infrasound at Etosha National Park, Namibia. J. Acoust. Soc. Am., 101, 1710-1717.

Mann, D. A., O' Shea, T. J. & Nowacek, D. P. 2006. Nonlinear Dynamics in Manatee Vocalizations. Marine Mammal Science, 22, 548-555.

Marter, K., Quine, D. & Marler, P. 1977. Sound transmission and its significance for animal communication II: tropical forest habits. Behavioral Ecology and Sociobiology, 2, 291-302.

McComb, K. 1996. Studying Vocal Communication in Elephants. In: Studying Elephants (Ed. by Kangwana, K.), pp. 112-119. Nairobi, Kenya: African Wildlife Foundation (A WF).

McComb, K., Baker, D.R. L., Moss, C. & Sayialel, S. 2002. Long-distance communication of social identity in African elephants. Animal Behaviour, 65, 317-329.

McComb, K., Moss, C., Sayialel, S. & Baker, L. 2000. Unusually extensive networks of vocal recognition in African elephants. Animal Behaviour, 59, 1103-9.

McComb, K., Reby, D., Baker, L., Moss, C. & Sayialel, S. 2003. Long-distance communication of acoustic cues to social identity in African elephants. Animal Behaviour, 66, 195-195.

McKay, G. M. 1973. Behavior and ecology of the Asiatic elephant in southeastern Ceylon. In: Smithsonian Contributions to Zoology, pp. 1-113. Washington, DC: Smithsonian Institution Press.

Mellinger, D. K. & Bradbury, J. W. 2007. Acoustic Measurement of Marine Mammal Sounds in Noisy Environments. In: Second International Conference on Underwater Acoustic Measurements: Technologies and Results. Heraklion, Greece.

Moss, C. J. 1983. Oestrous behaviour and female choice in the African elephant. Behaviour, 86, 167-196.

O'Connell-Rodwell, C., Wood, J., Rodwell, T., Puria, S., Partan, S., Keefe, R., Shriver, D., Amason, B. & Hart, L. 2006. Wild elephant ( Loxodonta africana ) breeding herds respond to artificially transmitted seismic stimuli. Behavioral Ecology and Sociobiology, 59, 842-850.

O'Connell-Rodwell, C. E., Wood, J. D., Kinzley, C., Rodwell, T. C., Poole, J. H. & Puria, S. 2007. Wild African elephants (Loxodonta africana) discriminate between familiar and unfamiliar conspecific seismic alarm calls. The Journal of the Acoustical Society of America, 122, 823-830.

Olson, D. 2004. Elephant Husbandry Resource Guide. pp. 280: International Elephant Foundation.

84

Page 96: Analysis and Classification of Sounds Produced by Asian

Payne, K. B., Jr., W.R. L. & Thomas, E. M. 1986. fufrasonic calls of the Asian elephant (Elephas maximus). Behavioral Ecology and Sociobiology, 18, 297-301.

Payne, K. B., Thompson, M. & Kramer, L. 2003. Elephant calling patterns as indicators of group size and composition: The basis for an acoustic monitoring system. African Journal of Ecology, 41, 99-107.

Poole, J. H. 1987. Rutting behavior in African elephants: the phenomenon of musth. Behaviour, 102, 283-316.

Poole, J. H. 1989. Announcing intent: the aggressive state of musth in African elephants. Animal Behaviour, 37, 140-152.

Poole, J. H. 1999. Signals and Assessment in African Elephants: Evidence from playback experiments. Animal Behaviour, 58, 185-193.

Poole, J. H. in prep. Elephant vocal communication. In: The Amboseli Elephants: A long-term perspective on a long-lived mammal. Chicago: University of Chicago Press.

Poole, J. H., Payne, K. B., Jr., W.R. L. & Moss, C. J. 1988. The social contexts of some very low frequency calls of African elephants. Behavioral Ecology and Sociobiology, 22, 385-392. .

Reby, D., Joachim, J., Lauga, L., Lek, L. & Auglagnier, S. 1998. Individuality in the groans of fallow deer (Dama dama) bucks. Journal of Zoology (London), 245; 79-84.

Rendall, D., Owren, M. J. & Rodman, P. S. 1998. The role of vocal tract filtering in identity cueing in rhesus monkey (Macaca mulatta) vocalizations. The Journal of the Acoustical Society of America, 103, 602-614.

Riddle, H. S., Rasmussen, B. & Schmitt, D. L. 2003. Are captive elephants important to conservation? Gajah, 22, 57-60.

Roca, A. L., Georgiadis, N., Pecan-Slattery, J. & O'Brien, S. J. 2001. Genetic Evidence for Two Species of Elephant in Africa. Science, 293, 1473-1477.

Sanvito, S., Galimberti, F. & Miller, E. H. 2007. Vocal signalling of male southern elephant seals is honest but imprecise. Animal Behaviour, 73, 287-299.

Schulte, B. & Rasmussen, L. 1999. Signal-receiver interplay in the communication of male condition by Asian elephants. Animal Behaviour, 51, 1265-1274.

SEVP. 2008. Savanna Elephant Vocalization Project <http://www.elephantvoices.org>.

Shoshani, J. 2000. Elephants: majestic creatures of the wild. New York: Checkmark Books.

Shoshani, J. & Eisenberg, J. F. 1982. Elephas maximus. Mammalian Species, 1-8. Soltis, J., Leong, K. & Savage, A. 2005. African elephant vocal communication II:

rumble variation reflects the individual identity and emotional state of callers. Animal Behaviour, 70, 589-599.

Stoeger~Horwath, A. S., Stoeger, S., Schwammer, H. M. & Kratochvil, H. 2007. Call repertoire of infant African elephants: First insights into the early vocal ontogeny. The Journal of the Acoustical Society of America, 121, 3922-3931.

85

Page 97: Analysis and Classification of Sounds Produced by Asian

Sukumar, R. 2003. The living elephants: evolutionary ecology, behavior, and conservation. New York: Oxford University Press.

Sukumar, R. 2006. A brief review of the status, distribution and biology of wild Asian elephants Elephas maximus. International Zoo Yearbook, 40, 1-8.

Vidya, T. N. C. & Sukumar, R. 2005. Social organization of the Asian elephant (Elephas maximus) in southern India inferred from microsatellite DNA. Journal of Ethology, 23, 205-210.

von Muggenthaler, E. 1992. Infrasound from the okapi. In: American Association for the Advancement of Science (A.A.A.S) 158th conference.

von Muggenthaler, E. 2000. Infrasonic and low-frequency vocalizations from Siberian and Bengal tigers. The Journal of the Acoustical Society of America, 108, 2541.

von Muggenthaler, E., Baes, C., Hill, D., Fulk, R. & Lee, A. 2001. Infrasound and low frequency vocalizations from the giraffe; Helmholtz resonance in biology. In: Acoustical Society of America, regional conference. South Carolina: Proceedings of Riverbanks Consortium.

von Muggenthaler, E., Reinhart, P., Lympany, B. & Craft, R. B. 2003. Songlike vocalizations from the Sumatran Rhinoceros (Dicerorhinus sumatrensis). Acoustics Research Letters Online (ARLO ), Acoustical Society of America, 4, 83-88.

86

Page 98: Analysis and Classification of Sounds Produced by Asian

APPENDIX A: DISTRIBUTION PLOTS OF PARAMETERS

IDSTOGRAMS AND BOXPLOTS OF ORIGINAL DATA

All of the histograms and boxplots below include. outliers. The summary statistics exclude outliers removed. Plots are included only for the 18 parameters that were used in the statistical analysis.

The figure panels are as follows: A) Histogram of parameter for all call types combined. B) Histogram of parameter for each call type. C) Summary statistics for all call types combined. D) Box plot of parameter for all call types combined. E) Boxplot of parameter for each call type.

Abbreviations for the call types are as follow: TMP (Trumpet), RUM (Rumble), SQK (Squeak), SQL (Squeal), BL W (Blow), BRK (Bark), ROR (Roar).

A.

20

~ 15 0 l-o i t-0

~ ll.

500 1000 15&D 2000

dta.6$LowerFreqM3

B.

"" .. " 40

21)

100

" .. " "

0 SOO 100tl 1SOG :2000 o soe 1000 1soo 2ooe

D.

c ;;;

. ,,.. "

"' 1000

Lowt'f'FreqM3

is~

E.

.., ;;

"' ~ ., !: 0 -'

""

d!a6$LowerF"'qM3

2000

1500 ·--i

' 1000 '

D 500 0

' __ @ __

~ c.tJ BLW BRK ROR

Figure 17: Histograms and boxplots of Lower Frequency (M3) TMP and SQL more variable, other call types more stereotypic.

a::;

RUM

cau

·c.

LowerFreqM3 Min. : 14 .13 lstQu.: 188.42 Median : 495.26 Mean : 532.77 3rd Qu.: 721.36

'"' 1 Max. : 1927. 22

" " " "

• . 0

-~!"--

4=J 0

0

SQK

:~-] I•

L~-·-'

[..-: ' rj' ' '

__ J._ __

' '

SQL TMP

87

Page 99: Analysis and Classification of Sounds Produced by Asian

A.

"

" ! 1~ .. ~

! "

D.

~

"

.<000 JOOD 5000

dta 6Sl/ppe!'FreqM4

.

2CIOO 400Q ~

UpperFieqM4

B.

~ l­o t' .. e " a.

60

40

20

60

... OIR.• : 'U :>~soli<:' "Y~:lE:;H: u:sai:;y;: · ·

LL1_.1_. BlW BRK ~i!' ROft

""'' ::J_L..1-11. L

.-....,.~. > ,;.-, ,.

"'""

a 2000 6COO 10000

dta.6SUpperFreqM4

E.

8000

" :::. 6000 O' ., ~ ., ! 4000

2000

,.........., i I

I • r

1 I y

Bl.W

~ __ .!.~-

BRK

0 21JOO 6000 10000

·····~·-· c:p ~ ROR RUM

cau

Figure 18: Histograms and boxplots of Upper Frequency (M4)

60

40

20

c.

UpperFreqM4 Min. : 103.6 1st Qu. :1341.8 Median :2261. 0 Mean :2780.7 3rd Qu.:3434.5 Max. : 9318. 5

' ' ' ' '

rLJ i

1 • L,.._J

Q •

__ , __ !_ __

SQ)( SOL TMP

ROR and RUM have narrow range. TMP, SQL, and BL W highly variable, which may be partially a function of signal strength (distance to caller).

88

Page 100: Analysis and Classification of Sounds Produced by Asian

A.

"" 11111

l:JL

D.

J ~

1~

a:a.6$0ISabonM5

n, __ .. DuJauonM5

B.

100

80

60

H 40 20

0

I 1U 0 f-0 c .. ~

" Q_

100

80

60

" I 4l)

20

.. "

;.{H:t'if;:~'ii'MP.?:t>:\q;.~

L :i; RIJM\''::::v: ) SOK ,,., . ., :::

L LIL f BlW BRK :1 ROR

L L L 5 10 15 5 10 1'5

dta6$DurationM5

E.

15

"' :::< 10 c 0

~

e ~ [~] ~ ~~

__ ., __

BLW BRK ROR RUM

Call

Figure 19: Histograms and boxplots of Duration (MS)

100

80

60

40

20

.J..,.

SOK

c.

DurationM5 Min. : 0.067 1st Qu.: 0.331 Median : 0.737 Mean : 1. 292 3rd Qu. : 1. 5 7 9 Max. : 14. 820

_J :

Q :.l i::;;:Q.

SOL TMP

ROR and RUM overlap in perception. SQL variable in duration. SQK, BRK, and TMP stereotypic.

89

Page 101: Analysis and Classification of Sounds Produced by Asian

A.

! " " j

"

D.

l3 ~

""'

200<)

dta,6SBa.'IC1Wialh~

'''" BandwidthM6 ""

B.

1il 0 t-0 1: .. ~

" "-

"""'

6-0 50 40 30 20

1D i ~ 0 ,,,, ----- y,_ ----~·~•'-'" _,, , , ~~" · ·, , I

60 50 40 30 2ll 1-0 0

l Btw

0 2000 6-000

E.

8000

"' 6-000 ::> ~ ~ ~ 4000 .. al

2000

dta. 6$8andwid1hM6

;

r-i , I ; !

! • ! i I ! I -~---

c --r-G.J '

' .... J. •• _

BtW BRK

0 2000 6-000

,--•:::i

ROR

[tJ

RUM

can

Figure 20: Histograms and boxplots of Bandwidth (M6)

6-0 50

c.

BandwidthM6 Min. : 53.83 1st Qu.: 904.39 Median :1682.08 Mean :2247.91 3rd Qu. :2802.61 Max. : 8990-11

'

J_ --~--

n y~

'

' ,---:--,

~~~T~-'. 8

SQl< SQL TMP

RUM mostly low frequency (as expected). BLW and TMP have very high bandwidth and variability, which may partially be due to distance to source. Squeal has relatively high bandwidth and very high variability. SQK has relatively wide bandwidth, but it is stereotypic.

90

Page 102: Analysis and Classification of Sounds Produced by Asian

A.

.. ! ~ .. j "

D.

. ~

"

d:a.6SMedlaniimeM7

........ , ~

1;· r •M~-a·~~ L~.,

MectanT1meM7

B.·

100

BO

60

40

20

0

~ 1--0 c: "' ~ ., ll.

100

80

60

40 I

20

0

" "

C'' TMe:!cA'''!l

L FIT RUM SOK:;::_: '.;o;Jh <·/"'l;-.rw:·.··,

,. 1L1t. fr Blil'VL: '.'.'}; BRK !:i JIDR

LILI~ E.

12

10

:i 8 .. E

~ .. '5 ., ::;; 4

0

10 10

dta.6$MedianTimeM7

"''i''"-

_J __ ct 0 g::: Si~.?

__ ... __ ~

BLW BRK ROR RUM

Call

Figure 21: Histograms and boxplots of Median Time (M7)

10

80

60 40

20

c.

MedianTimeM7 Min. 0.0210 1st Qu.: 0.1340 Median : 0.3245 Mean : 0.5950 3rd Qu.: 0.6773 Max. :10.5830

• • --l-

.1. ~ e i:iik~

SQK SQL IMP

91

Page 103: Analysis and Classification of Sounds Produced by Asian

A. B.

100

80

60

40

20 0

1 .. i I r I 2 c >-0 c .. e ..

Q.

100

60

60

40 ~ 6STtmeConcer.1M9 I 20

™"".'''''

L • RUMii>i"''' : SOK

IL L i BLW BRK

L L

SOL

kn_ '.; ROR

L

c.

TimeConcentM9 Min. : 0. 0000 1st Qu~: 0.1340 Median : 0.3370 Mean : 0.7946 3rd Qu. : 0. 8590

100. Max. :13.8450

80

60

40

20

10 15 10 15

D.

s j[ ! ~· ~---~--"" <>->-»->.>

~

~L

" TimeConcentM9

E.

"' 10

~ ~ c 0 0

.~ 5 >-

dta. 65TimeConcentM9

--~--

-···y .. •·

[;] ~ ~~·;".J ~ BLW 8RK ROR RUM

Gall

Figure 22: Histograms and boxplots of Temporal Concentration (M9) SQL highly variable. RUM more variable than others.

~ --~--

'

.4- LJ --~-­~

SQK SQL TMP

92

Page 104: Analysis and Classification of Sounds Produced by Asian

A.

! 60

" l

D.

!!

~ .:

..

....

-ea a..~

ala.€STimeAs)'mrr.M10

I , --·~ ~ i

~:J.l! '·' 0.2

TimeAsymmM10

B.

100

80

6(}

40

20

0

~ I-0 c .. e .. "-

100

60

60

40

20

... "

c.

%',:·:: ':TMP TimeAsymmMlO

l_ Min. :0.39400 1st Qu. :0.00000 Median :0.00000 Mean :0.0006586 3rd Qu. :0.00100 'liir= ,070600

\&:\fr i BlW H''.lE::.\BRK ROR

1-l~ll_ I I i I I I I I I I I I I I I I I ( l I

-0.4 O.Q 0.20.40.60.B -0.4 0.00.20.40.60.8

dta.6$TimeAsymmM10

E.

0.6

0 0.4

~ E

! 02 ~ .. .

0 ~ 0 . -+ + -t-E o.o -+- + ~ .........

i= 8 >

-0.2

-0.4

SlW SRK ROR RUM SOK SOL TMP

Caff

Figure 23: Histograms and boxplots of Temporal Asymmetry (MlO) Stereotypic with minimal variability.

93

Page 105: Analysis and Classification of Sounds Produced by Asian

A.

"

! ,, " i Q_"

D.

~

"

1000 2!:-00 ::000

dl:t 6.SMediatlF reqM 11

~- -- ---1 ' t-------lLJ · ---·-·-··--····L. 0

: l : : ~ :

1000 """ '""' MediarFreqM11

B.

'"'

~ l-o c " re ..

Q_

60

40

20

60

40

20

'IEi'~;

;~;;.··

LIL 0 1000 3000 0 1000 3000

dta6$MedianFreqM11

E.

4000

i 3000

i-j 0

.. -r .... 0

1000 . s . c!J ~ ~;:~

BtW BRK ROR RUM

Call

Figure 24: Histograms and boxplots of Median Frequency (Mll)

60

40

20

3 .

c.

MedianFreqMll Min. : 26. 24 1st Qu.: 663.81 Median :1046.76 Mean :1105.67 3rd Qu. :1489.14 Max. :4454.33

. . . . 8 [J i •

cb Ll --g-----L-.

SOK SOL ThlP

94

Page 106: Analysis and Classification of Sounds Produced by Asian

A. B.

6()

40

20

0 .....

i')'.t!ii:\•iRVM ·-· ! JO >;

! !T."

~

I f-

_l_ 0 1:

aL .. ~ Cl.

" r.; .. : .. •.: BLW BRK

" eta 6$F'"&!'.;AsyrnmM14

60

_lj 40

~,___._ 2D

I I ! I I I I

-0.5 0.0 05 -0.S 0.0 0.5

dta. 6$FreqAsymrnM 14

D. E.

g 0.5 4 --:--

' ' ' ' ' ... 0 ' '

:1 ··-1··

GJ 8 E n ~ I I I E GJ L!J . -1·----· . «<•<>:•!!<>< .f o.n

0- . ~ ' u. ' '

··:.

-0.5 -G.5 00 " "eqAsymmM14

' BLW BRK ROR RUM

Cal

Figure 25: Histograms and boxplots of Frequency Asymmetry (M14)

[ 60 40

2D

--?--

(j-1 . ; .~

--6--.

SQK

c.

FreqAsymmM14 Min. :-0.63400 1st Qu.:-.00600 Median :0.02800 Mean :0.05637 3rd'Qu. :0.12400 Max. : 0. 66600

,

~ I --~ -·-

4::J --~--

0 __ ._ __

SOL TMP

95

Page 107: Analysis and Classification of Sounds Produced by Asian

A.

! 10

j

tll.6$?1(QveralR~18

D.

~

"

,. .. .. PkOverallReffM18

"

B.

~ 0 c ., l: ..

Cl.

25 20 15

10

5

25 20 15 10 5

, ..

,,,lifV !TMF':0',S'.1;1

0 20 40 60 80 100 0 20 40 60 so 100

dta,6$Pk0vera11RelTM18

E. 100

--i--

' BO

' "' ' :ii

'

LJ !:: 60 ..

~ Q:; ~ ! • .,

40 > 0

"" Cl.

20 ' ' ' ' ' ' ' 0

--i--

Bl'IV 8RK ROR RUM

Gaff

25 20

15 10

5

c.

PkOverallRelTM18 Min. : 0.00 1st Qu.:23.50 Median : 41. 50 Mean : 41. 89 3rd Qu. : 5 7 . 7 3 Max. : 99. 63

--;--' ' ' ' '

D~ SOK SOL

1MP

Figure 26: Histograms and boxplots of Relative Time of Peak Overall Intensity (M18) High variability within call type, but similar among call types.

96

Page 108: Analysis and Classification of Sounds Produced by Asian

A.

% ~ $

i .r"

D.

<'l ~

B.

~ I-0 c .. ~

" Q.

dta.6$PkOYeralFM2o

1000 ""' 3000

PkO\'el'allFM20

00

60

40

20

0

ao 60

=~ ill 0 1000 2000 3000

E.

3000

§ !!,, 2000 ~

§ Q.

1000

K sot'"·

LIL d1a.6SPkOVeraHFM20

L:J

BtW

--1-· GJ BRK

0 1000 2000 3000

e ~~-~~ w

ROR RUM

Galt

60

60

40

2\)

0

'

c.

PkOverallFM20 Min. : 17. 50 1st Qu.: 481.71 Median : 861. 33 Mean : 931.14 3rd Qu. :1346.23 Max. :3531.45

' '

Q ~ lTJ __ i.._.

SQK SQL TMP

Figure 27: Histograms and boxplots of Frequency of Peak Overall Intensity (M20)

97

Page 109: Analysis and Classification of Sounds Produced by Asian

A.

"

! ~JO

~"

D.

~ ;;;

11·1

lO

dla 6SAMRa!et.121

" f ' w,---.--»-~

" " AMRa!eM21 " ..

B.

~ r-0 c ., ~ ..

CL

100

80

~~'

~~L_

c.

AMRateM21

LfLILt~ ROl'Ui: -

Min. : 0.000 1st Qu.: 1.118 Median : 2.207 Mean : 3.053 3rd Qu. : 3 . 7 4 5 Max. :44.939

BLW BRK

~lLLL 010203040

E.

~ ~

4V

30

ll'. 21l

~

10

dla.6$AMRateM21

_J __ ~

Bt.W

.L r.::::P

BRK

010203040

. __ _. __ c::b

ROR

~ RUM

Call

i __ L

C!J ' ~~~ L-'--. ..i

. --V-- --'--

SOK SOL TMP

Figure 28: Histograms and boxplots of AM Rate (M21)

98

Page 110: Analysis and Classification of Sounds Produced by Asian

A.

~ ~

j~

D.

~

"

llilllll

dd.6SAMRateVa11-R2

f i il~ . .,, ><·<'

ii ~ i

'·' B " 1.5

A.\1RafeVarM22

B.

100

80

6-0

40

20

0

~ I-Ci c "' ~ ., 0.

100

80

60

40

20 0

" " 3.0

c RUM t'. SOK k' <::ru

L[L/L![ I~ .. · . 61.W ) . · SRI(;;'.::!:;:>:: ROR

LILIL 2

dta.6SAMRateVarM22

E.

3.0

2.5

~ 2.tl ~ ~ ~ 1.5

Cl'.

~ 1.0

OS 1 0

00 i -- ~

Bl.W BRK

0

' ~

RQR

' r.:= RUM

Call

Figure 29: Histograms and boxplots of AM Rate Variation (M22)

c.

AMRateVarM22 Min. :0.00000 1st Qu. :0.01200 Median :0.02300 Mean :0.05642 3rd Qu. :0.04600 Max. : 2. 97200

--- ~· _. __

SOK SOL Th'IP

99

Page 111: Analysis and Classification of Sounds Produced by Asian

A.

~ ~ 4~

i

D.

~ ~

:ta.6$FMR.l!eY2l

:11 ; ' :J· -- +--0 ~. 'LJ :

"' " FMRateM23

B.

100

80

60

40 20

~ f-0 c ., ~ ., "-

1-00

80

60

40

20

..

c.

i~~ea~71:~~~"~ FMRateM23

L Li LtE

Min. : 0.0000 lstQu.: 0.9653 Median : 2.2160 Mean : 3.7613 3rd Qu.: 4.4170 Max. :70.4720

BtW BRJ'<k~:::::.;;: ROR

LILIL 0204060

E.

"' N

60

i 40

C! a: 20

0

dla.6$FMRateM23

0 0 0

~ 0 --1'--

~ ± BLW BRK

0 204Q60

' ·---~--- Ml~ ~

ROR RlJM

Call

i --~-- 0 •

-l- _ _; __

~ ci;:J c:t::i SOK SQL TMP

Figure 30: Histograms and boxplots of FM Rate (M23)

100

Page 112: Analysis and Classification of Sounds Produced by Asian

A. B. c.

Hi TMP FMRateVarM24 100

L Min. :0.00000

so 1st Qu.:0.00600 60 Median :0.00600 40 Mean :0.02094 20 .. 0 3rd Qu. :0.01200

,. RUM ll! "rn Max. :0.92900

~ L L 100 .... 60 i' t-

~ 0 60 ~ ~ ~ c

~ 40 ..

e 20 Q)

"' a. 0

i/ Bl..W BRK .. ROR

- 100

L L ' so

L ... ., " .. . .. ,, 60

~.GSFMR.i.!e'lart'.424 40 20

0

0.0 02 0.4 0.6 0.8 1.0 0.0 0.2 04 0.6 0.8 1 0

dta .6$FMRateVarM24

D. E. ¢

o.s 0

" ~ 0.6 Ill

~ - .• ' '0 ·• .; ~ . < ~ 0.4 0

t: ::;, LL ··-r--,, .

0.2 ., <> ii

--T-- I • ; D.D {i.2 '·' o.e "

8 ~ c~:J

l.._.,.. _ _I _._ --1-- _j_ 0.0 -·--FMR8teVarM24

BlW BRK ROR RUM SOK SOL IMP

Call

Figure 31: Histograms and boxplots of FM Rate Variation (M24)

101

Page 113: Analysis and Classification of Sounds Produced by Asian

A.

,, ! " J :0

D.

.. ~ ;;:

"' tia.6$Eri!ropyt.126

; ·-----~-1·~·><r..,.<-«flN>~·;<o

200 4C-O "' EnD"opyM26 "'

B.

~ I­C;

c "' !,>

&'.

80

&ll

4l)

20

b'ff'.z".~1MF:'n ·.

_, . .,,, I -:j _L.__ I

1 I 1 I I I I I I I

0 200400600 800 I I

dta 6$EntropyM26

E.

1000

• =i ' ' ' ' '

l 400 [] 200 -1

' __ ... __ '

G.J "'" -BLW BRK

rL I I I l I l

0 200400600 800

--7--

4;1

ROR

. --~-­l±:J

RUM

Call

Figure 32: Histograms and boxplots of Overall Entropy (M26)

80

60

40

2{)

c.

EntropyM26 Min. : 3. 968 1st Qu.: 79.487 Median :136.609 Mean :159.961 3rd Qu. :194.310 Max. :1000. 792

--,--

g --~--

GJ a; cb --1.-- .... .:. ...

SQK SQL TMP

Highly variable among call types. TMP has highest variability (tonal-noisy variability), which may reflect level of arousal in switch to noise. Wider bandwidth calls have higher entropy. Chaotic noise has been perceived in the calls with higher frequency ranges (roar, trumpet, squeals), so this is consistent with perceptual aural cues.

102

Page 114: Analysis and Classification of Sounds Produced by Asian

A.

! . " j

D.

~ 'ii:

dta..65UoswiiFracM28

; ! :

B.

]! 0 f-0 c "' " "' Q_

; Di''···,,,., ... ! .:•<> -----r---------1 . • ---------r·<. .JN~ :;.;.<

' I ' ; 1 :

"' " .. " UpswpFracM28

,ii{1!ki~ll'~Pi :v,::,ij

RL ;>r,,:;SQL

_lJ~

lLl_tJ .~ .. 1 0 ' '

"'

0 2Q 4() 60 80 100 o 20 40 60 ao 100

ctta,6$UpswpFracM28

E. 100

SG

a:!

"' ~ ·1 ~ u.. C,

~ 40 0,

=>

20

--7--

~ [5 ..... l

'

BLW BRK ROR

--r-C;J

'

-+-RUM

Call

Figure 33: Histograms and boxplots ofUpsweep Fraction (M28)

c.

UpswpFracM28 Min. : 6 .196 1st Qu.: 40.340 Median : 46.450 Mean : 44.931 3rd Qu.: 51.132 Max. :100.000

--:--r-i I • I ! ! L-,-1

SQK

--r-lfJ -··i--

SQL

--T--

' [~]

TMP

103

Page 115: Analysis and Classification of Sounds Produced by Asian

Q-Q PLOTS OF SCALED DATA

In a Q-Q plot, the x-axis is the expected value for a normal distribution and the x-axis is observed. The data are ranked and the quartile is calculated. If data are distributed normally then the shape will be a straight line where y=x.

M 2 ::>

l ~ 0 ' i ~

.,

·2

<JlO!l11

;,>Cc' ~ .. ,,.

Figure 34: Q-Q plot of Lower Frequency (M3)

1 ~

i' ~

-1

·2

. '

/ I

(JlOrm

Figure 35: Q-Q plots of Upper Frequency (M4)

~ ~

l \:: U> J

~

i ~

.,

~ i 3

~ '

-1

-3 -2 ·1 G 1 2 J

"J

"I /l.L/ .. ,.· -~ f--1 ..,.,,,,-- . . ' . , <' ... :BRK:';·!'.<'.".'"dli".'J'.ROR·::;·.:-->:Tc• !'.'.'hlUJl,F

.,~· .,.~~· ~·· -3 -2 -1 G ,1 2 3 .J -2 -1 0 1 2 3

qnurm

·3: -2 -1 0 1 2 3

b'::. •-:t::.S:OJ< !) ·;"L'..·•<''.•JMFl <.d '.·

-1

(" /,•' .~··1 .. / ,,

:oo<"·~ ~+(t~~ b~/-~f~'.:'::1u1u~·,._, ,,,,.,'T

,_,.!' <00___.. .~ -3 -2 -1 0 1 2 -3 -2 -~ I) l 2 3

qnorm

104

Page 116: Analysis and Classification of Sounds Produced by Asian

"' " g .!! •

i ~

I

·2

C/l(lml

Figure 36: Q-Q plots of Duration (M5)

"' ' i ~ :. Ill' Ji 4i

-1

-2

qnorm

<.; ·~

<P'~

I

Figure 37: Q-Q plots of Bandwidth (M6)

10

~

i ~ 6

I i 4

·2

--'-~~~~~~~~~~

qnorm

,, sn

! "'

Figure 38: Q-Q plots of Median Time (M7)

~ ! 1·.:u.,:;>E!f\1(.;;.: .:; . ~

,.,-«''

.3 -2 -1 I) 1 2 3

.~' /

.,/ -3 ·2 -1 c- 1 2 3 .3 -2 -1 0 1 2 :J

I ! el Ji 3

ii

~

j ~ i ~ 1Q

ii '

·;·· ;:;·-:;;SIJKi "· ...

/' . ,.,..,,.,,

+i :;:;:i'!iF!~

,ff

,,_.,,/ -3 -2 -1 0 1 2 '3

qnorm

.3 ·2 ·1 0 1 2 3

; TM"ff';

./'/ •· ''.f\OJ<:.': · '"l' < :c<RUM :

,,__..-8 ,,,~-·

.J ·2 -1 0 1 2 3

qnorm

-3 -2 -1 0 1 2 3

,,•1 / QQ~ ')'.~"P<;. :~'.; J~.f1C';'' :;<'>--;-

~~'<.~'"' ,,___.,.,,-- ,/'

,_..../ -J -2 -1 0 1 2 J -J -2 -1 0 1 2 J

qnorm

-I

10

105

Page 117: Analysis and Classification of Sounds Produced by Asian

.3 ·2 ·1 0 1 2 l

" ·::-£-,./'.~ :S.-Qt<.e;:;_ ·!'<FW•:d• '{{>]:-·.:,: :·0.··JMP ,.

i 5 § u <J'

.j!

~ ii

! 10

/ .J ,____.,."f-o

~ i 4 / ~-,ti~K ~· .. ~·~''

~

,__/ ~ a

,oo_/j~ ,-"' "~

·2 .) -2 -1 0 1 2 3 .3 -2 -1 i) 1 2 J

qnorm qnoon

Figure 39: Q-Q plots of Temporal Concentration (M9)

20

15

0 ,.

~ §..

i 5

~ "' v ~

~

-10 ') ,)·'

~( <'> ,, ______ ../ r

-2

qnonn

0

1 ':i E 20

~ 15

~ 10

-5

-10

Figure 40: Q-Q plots of Temporal Asymmetry (MlO)

" { 'O a <ll ~

'"°"n.:•

'!{' . -~

•"

I LL

Ii ~ ::; .

i :i

..J -2 -1 0 1 2 3

ii>YiU'-.SOW ·., 1.·sot ;. :) ..:-:it:\:: .!::ndf!:' Y·'·,.

"' ___ .. ,,--I

:.:.u i:;imKin1r .• :hF"'··· -,.RCIF: · ·:: .i.• : ·:::m1W•:' ··

,,"'~~<;? <f~Q

-3 -2 -1 0 1 2 3 .J -2 ·1 -0 1 2

qnoon

..J ·2 ·I O 1 2 3

,,, '"'" :: '"'' ;;:,::so , ;~ l '.,_TMP~',';_ ··"

.:'"'

i

-~·: ,__.,,,-· ./·1,/ •"

•. ,.·<:i!lRttc>• · :·Roo: •. · .]7j;'. ·f'RDi.I

_ _,.../ ·•· I ___,,t••P

(-'>-~ 6~· .__...,.,.· ·2 -3 ·2 ·1 0 1 2 .3 -2 -1 0 1 2 3

qnorm qnorm

Figure 41: Q-Q plots of Median Frequency (Mll)

"' 15

10

.. -10

106

Page 118: Analysis and Classification of Sounds Produced by Asian

,/'<>c

ff'

;t

" .,.

i i ll

i g i:;'

l ,s ~ -'2

-4

·2

qnorm

·2

..

.J -2 ·1 0 l 2 3

:

/l.,-/.1/ ~"o

/I./·· -3 -2 -1 () 1 2 3

qnorm

r" /

.:19>'

•• <:i

-3 -2 _, 0 1 2 3

Figure 42: Q-Q plots of Frequency Asymmetry (M14)

.J ·2 ·1 0 1 2 3

""' p-,··

-2

-4

"' i ,_ ;!i =;

"' § ~ ~

/')""·' /,,~,

./ .. / i'i i a

~

§ i 2

m '5 !

I .;·~

li·~\W/>BRI¢>:: :n}.< :1 :'/llofii'i ,, ... <·RUM'.·· ..,,..,.

./ '·'

I ,, ,,;·

I • ,,,,.!

??

·I

·2 -3 -2 -1 "-0 1 2 3 .J ·2 -1 0 1 2 3

qnorm qnorm

Figure 43: Q-Q plots of Relative Time of Peak Overall Intensity (M18)

~ u. 15 .2

§ l1 ;'l

(}<~

·2

qnorm

l'l ~ ~ §

! ll J

.3 -1 ·1 0 1 2 3 J......,.J....,-- i l l

,, .·,:'.;sr. ·1r.tP~<(~

'

,___,,l/./~I /' .i·:·::.·:·aRK"''''' •·'l :.•ii .:1R0R· ·.:·r-- :1 :RUMY.

-·" ,~~"'?" I .,__,-· I . __.,,r· -J ·2 ·1 0 1 2 3 -J -2 -1 G 1 2 3

qnorm

Figure 44: Q-Q plots of Frequency of Peak Overall Intensity (M20)

-1

107

Page 119: Analysis and Classification of Sounds Produced by Asian

IO

"' "' t & (!'.

~· ': "' ' ~

"'

.2

qnorm

Figure 45: Q-Q plots of AM Rate (M21)

2tl

15

I ~" i ~

oil

~------.r -2

qnorm

Figure 46: Q-Q plots of AM Rate Variation (M22)

12

10

"' ~

1 ti: 6

~ IS ~ 4 ,.

g

.. ____/ -2

qnorm

Figure 47: Q-Q plots of FM Rate (M23)

~

to ~

;:: :;;;

i (!'.

! 20 IS ~ ,5

.., 51 ~

10

~ 12

c.:i: ~o ,!j

.) •2 •t I} 1 2 J ! '

i______L__L '(, 'TMP

.J , ____ ,,r I .. _...r ,..'!llllt•• '•'!·-'.Ci!ibiL "'JjJJM.

,_

,__.,,,,.~·' I ,,__,,..,~.- ~~~;-:-.~

-3 -2 ·1 0 1 2 ) -J -2 -1 0 1 2 )

Qn<lf!TI

..J -2 -1 0 1 2 J '-'--L_L_'

0 !'SOti! +· -·TMP"' ·'"-'

.')!'.T!ii::!ll!!K;'.;!'{• .,•'.lt!T'71::;r-R0R •'· ., ·. "!'!/· . 'Rini!!" ·

, ,~

-3 -2 -I 0 1 2 3

_,,' ,---Ql1<Jf!TI

..J -2 -1 I) 1 2 3

J. ,,,,,,,,,.,...

-3 -2 -1 0 1 2 3

iK".TC0'1;E+ ;·;+: 'S• "'JMIFi'C. •;;

•'

_ ___; __ ___.,,,,,;:- ,'

,__,/ · -.,,J"iBRK' ·'' ·· i:'!':i'.! '.\ROR·· :C.';i,E".ii\' ('lIDM!f·

, ~..---""'"" ,,...-""'' I ·---··· .3 ·2 -1 0 1 2 J .) ·2 -1 (I 1 2 )

qnorm

-10

2tl

15

10

12

10

108

Page 120: Analysis and Classification of Sounds Produced by Asian

-3 -2 -1 0 1 2 J

+-<;/;: -SOK-• ~~P'.-

15

~ ~ 10

~

.. ~ ~

~ ·-s>''> .,..

i~~'.,:"tSPk'' "'' -"'-ROO:"'- ;-,.-----~"-----~ .. ;; . i 15

<I ;

-'" ~

-2

Cl'OfTTl

Figure 48: Q-Q plots of FM Rate Variation (M24)

•''

"' ~ /' g ~ 2

' i -r

{l

-2

qrorm

Figure 49: Q-Q plots of Overall Entropy (M26)

~ i ~

i i

i'-:>-'

,.p:

.. -2

<JlO!lTI

Figure 50: Q-Q plots of Upsweep Mean (M27)

"'

l :;, ~.

0---

~ ~

! i ~

-5

10

o~''C ,«¢.~- , . .,.,,-,,,,,,,.:

-3 -2 -1 0 I 2 3 -3 -2 -1 0 1 2 3

qoorm

..J -2 -1 0 1 2 J

0 I I ,~'l/'-/ ~~H?Sji;:•:· ----:"RORL: c-• f· --•;-.;,'ROM"•_..,_ -

,_/ ,~~ ,_,~-·

-J -2 -1 0 1 2 3 .3 -2 -1 0 1 2 3

qoorm

-l -2 -1 0 1 2 3

'''.U!:SllK''." i!_E,V ·::;·soc.,~::, ,_, .:' ~f'..;_ ~~~ ·>::ThlP·~~: :~<- !~·

.. --1 /.

,,/\~/ •:Uiit:'<aRKG?')::;v: ??t:ROR - f'' -----:q3!J~

~ ,•

-3 -2 -1 0 1 2 3

-~" ,,,,.,----

qno<m

~~,-

.3 -2 -t ti 1 2 3

" "

-5

109

Page 121: Analysis and Classification of Sounds Produced by Asian

~ 2

{ ~

i 0

ll

-2

).·.

•'

/

qnorm

Figure 51: Q-Q plots ofUpsweep Fraction (M28)

~ ~ £ ~ a <1l. ll

-2

-3 -2 -1 0 1 2 3

I "o',AAK ,A:<;:,~nn TMP'

/I./' / ,,, -2

>··liRil::·' '<1100' '''<•l -·-,-·< iliJlli;,·

/ j,/''" /,•' ..,,,,e

-3 -2 -1 1l 1 2 3 -3 -2 -~ 0 1 2 3

qnorm

110

Page 122: Analysis and Classification of Sounds Produced by Asian

Box PLOTS OF SCALED VARIABLES FOR IDENTIFYING OUTLIERS

0

ll -

0

II -

0

0 0 0

4 -0

0 0 0 0

€10 0 0 0 0 .- 0 0 0 0 ,--,0 I ,- 0 0,-

,- I I I I 0 . I 2-i I-., I I 'T' I -r--, I 8 0 I -.,

I I I I I I I I 1-r-r 0 I I I I I I I I I I I I I 0

8 1 I

I I I I I I I I I I I I 0 I

H~RBR8B+Q8888-~-Q+$ _i_ ...L. : -'- : I -L : : : -'- _!_ 8 :

I I I 0 I I ..J.... 0 I I ~ ..J.... ..J.... I I

-2--t ~ J_ €1 ..J....

0 0

0

-4 I 0

0

I

"' ,.. ...,

"' ... <:":I 0 ... <D· c ;:; "' M ... "' "'" "" .::;: ::;: ::;: ::;: :< :< i i N "' "' "' N N '"' ,,. <r c: .::: ... 'E ::::;. ::;: ::;: ::;: ~ :< :e ~ ::;: ::;: ~ e 0 ~ g E er E 'i "- .. .. c: <>

~ .. e 1i ' 1ii " <I. .. e .... " ~ ~ 2 ~ i; .. I- c .... "' "' "' "' .. l! "' "-.. ::i .., c 0 "' ~ 1ii i£

:; Q. 3 0. 0 c "' u ~ ~ ~ e ~ ::;: m a. 3 .s a. "' 'i5

"' 0 "' "- "' .. ::> m. .. ~

.. .. ::r !I .... ~ ::;: .. "' ::;: ,g ::;: e 0 "- "- a. :3" I- ..... .... ::J

, 12.

Figure 52: Boxplot of scaled variables for Blow AcoustiC parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

111

Page 123: Analysis and Classification of Sounds Produced by Asian

0

0

6 - 0

0 0 0

0 4 - 0 0 0 0

2 -

0 -

0 @ 0

0 GI @ 0 0 -r -. 1::1 I I 0 A t t 6-rT" " -rt t Qt t 0

I I I ,.- I I Qg.,-0 t t 8 t t t t 0 0 -1L 1-r-1 I I I I I ,-

:T:::: : : : : § :-r "Tit-;

$$9$8Bf B~WO~Q~o~$$ -L t t t -Lt t..L. 0 t

I -1..-'-8 ..L. t 0-1... -'- -'-

0

-2 -I 0 8 0 0 0

0 -4 -j

0

-6 -I

0

I I I I I I I I I I I I I I I I I I

"' ... ""' (<I .... "' 0 ... "' 0 ;; "' C'J· ... "" .... a:> :! :! :! :ii: :ii: :! N N ..... "' N "' N CT D" c: .c "' '!: :! :ii: :! :! :ii: :! E :ii: E ~ ::< :ii: e. !! 0 'S E E i:r E t ..... e "' c <>

~ ..

E ti , 1ii 'j <>. m E ... ... 3 i= "' i. §. .,

i! E ., ~ & <: ... a: a: "' a: ., ..._ .. ::l "CJ c: 0 <:::

~ ~ ~ '!i'l :! '15 "' :! <>. 3 Q,. c c: .. u t .!!! if Ill 0.

~· .s <>. .. '5 " 0 a: ... a: " :::> m "' "' "' er ~ ... il :a .. :a E ~

.. E 0. :ii 0 c... .... Cl. :::> i= "-

"" :::>

"-

Figure 53: Boxplot of scaled variables for Bark Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

112

Page 124: Analysis and Classification of Sounds Produced by Asian

£II

s~xc-J\ atp uo a.rn s1a+aur1:utid alp Jo san1t?A (paZfit?UIJou) pa1ti:>S ·s~xti-x atp uo a1ti s1a+aurti.rnd ::i9sno::iy .nm-H JOJ sa1qt:pt:A paJt:;Js 10 lOJdxog :ps a.1n~!.i

c: c: '"Cl .., .,,. .. s:: "'

>E s IJ;l ;o .. !!!. .,, ll: ... .... a Q

iil "' '

.. "' " :::t

~ l: s: s:: ..., ., ..., "' "'

..... "' ...

8

e 0 0

0 0 0

0

0

.,, s: ::0 !!le .. s:: "' c.>

0 0 0 0

~ ::0 ~ ai'

;; ;o

' !a

~ ., s::

"' ~ "'

0

0

" ,.. . .., ..,, 0 ;;; "" a .<> 0

ii. ~ a ii ~ il. .. :;; ::; 3' s:: s: s: ..., a; 0 ...

0

~ ~

s:: 3 ,l!: .. .. .. tD c: b ·9' .. a. !' "'

"'O

" n ... ::r 0 "t:r s

" 3 2

" a. t::

~ ~ ..... s i iil a ~ .,, ...... 3' <> 3 !ij; c;· ii <i .a .. s: s:: ... :::r :::! ,,, D

s: s: s: s: s:: s:: 0 "' .... "' "' ... '"'

8

0

0 0 0

0 0 8 0 0

0

0 0 0 0

0

0

Page 125: Analysis and Classification of Sounds Produced by Asian

6 -

4 -

2 -

!) -

-'2 -

-4 ~

-!! -

0

0 0

0

0 0

0

0 0

0

0

0 0 0 0

0

0 0

0 ~ 0 "'T" 0 I 0 ~ I Q I 0 ~0: :~ 0 Q I -,- I I I I I "'T""'T" I I I I IR fl "'T" I I 1-r-1 I I I I~ ~ 0 I

Oa~aga+Q8B~$$8e~+$ I -L ....1.... I I I ..J.... -L I

I

51 ~ 't: .. " .3

....L. ....L. -L I I I I I 8 I

I

• ::;; tT l! 't: .. & :::>

I

"' i !! ::> c

I

~ ~ 'ij; -g ~

I ,.,_ ::;; "' ~

I

I

°' ::;; l! g l5 0 .. E f=:

0

0

0

I 0

i E ~ if ... E ;::

I

i [ ... c .. '5 .. ::;;

I ....1.... I

l_ §

I "'It

j ~ t CT

2 ....

I <O·

~· <;

"' I x

I

~ ::;; .... E ~ 0 ::.:. "-

I I

~ ' ..

N

:a: .. 1ii

"' 15

"" ~ ~

I

"' N :a: .. 1i

"' :a: ....

I "'It

~

' ~ "' ~

I

"' i l

0

0

I

"" "' ~ .. "' ::;; Q.

3 ., J:I. ::>

I

"" N

~ e .... Q.

" ~ ::>

Figure 55: Boxplot of scaled variables for Rumble Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

114

Page 126: Analysis and Classification of Sounds Produced by Asian

s 1 6 ~

4 -I

2 -

() -

-2 -

-4 -I

0 0

0 0

0 0

0

0 0

0

0

0 0

0 0

0

0 0 0

0 0 0

0 0

1§1 0 e 0

0 0 0 0 0 0 0 0 0 0 ... 8 8 0 0 -r § 0 0 0

i~: o® -9-: 0 0 •o

0 : ~ : I 1§1 g -:- : : i J 0 ! : -. ... I I -Y- 0 I I I I 0 I I I -r- I ,- I I I I I I I I I

$8$8~$f $~8~$D~i$$8 0

I I -L I I I I I I I I I I I : I I I I ~~ ~ ~~~ ~I I

e o v : 0 § ~

0 8 0 0

0

I I I I I I I I I I I I I I I I I I

"' ... It> "" .... al 0 ~ :a :a ~ ~ ~

~ r:r C' c: ~ .. 'C e l'! 3 :!2 E .. i ~ ...... e s ;:: " ti c .. ::I .., <::: 0 s .Cl. 0 <: "' (J t .:9 "'- .. '5

.l ::J "" " .,

~ ~· I- I-

... "" 0 ;:; N "' ... "' .... "' "' N N N N "' "' ~ ~ ~ :a :a ~ :a ~ ~ :a :a C" E I- "- ... .. <::: 0 e Ci e 15 ~ 10 ~ Cl. "' J! .... ~ c: c: .. c: ., E .. ... c: 'E ~ ~ 1i :a 1i ~

:a ... .. t Cl.

"' '5 0 c: .... c: ,. II>

"' t:r '!l .>< i :a .. Cl. :a e 0

a_ .... Cl. ::J u.

""" ::J

CL

Figure 56: Boxplot of scaled variables for Squeak Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

115

Page 127: Analysis and Classification of Sounds Produced by Asian

to

~

0

-!i

0

0

0

0 0

0 0

0

0

0

8 0

0 0 0

0 0 0 9 0 0 0 0

0 0 8 0 0 8 0

0 0

0 o0 I 0 .. e 0 0 ~ 0 I 0

I • @ ~ I 0 0 0 0 e ! I • I Jo 1 8 T T .- ~- ! ~ o 8 lo

0

I I 0 I t I I 1-? T "T" I -r I I I -r I I I I I I I

8$8~~8f $9898B~1~~$ l_.L-'- I l...L..1-L Ill I I ...J.... I -L I I I 181 I ...J....

I * 0 ...J.... S ~ ~ 0 8 0 0

0 0 0 0

g 0

I I I I

"' "" .,,

"' ,... 0 0 ., CD· c

N "' "' "" ., ..... "' :e ::;: :e ::;: :e :e i i N "' "' N "' "' N

"'" .,. c .::: .. :e :e :e :e ::;: ~ ::;: ~ ~ ::;: ::

I!? l!! t :2 .s E er E ..... ... .. .. c u

"' e <i ,. ~ 1ii ' Q. "' l!! ... ... <.> i. ~ e :;; ~ 3' ..... c: ... c:: c:: "' ct: .. e .. ...

"' ::i "' c 0 c ~ ~ ::;: Q.

3' <>. 0 c ~· 0 f "' Jf I! ~ ,. ::;: ,,.

Q. s .s Q. "' .. '6 0 c:: .... ct: 3' ::>· m .,, .. ,,. ~

.. ::;: _g . s .. e "" ~ ::;: .. a.

:a 0 c.. ... <>. ::> I- ..... I;..

"" :::>

0..

Figure 57: Boxplot of scaled variables for Squeal Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

116

Page 128: Analysis and Classification of Sounds Produced by Asian

lII

spro-A: aip uo a.m s1aiaurn.md aipJO sanp:?A (paznmurnu) pap:?;)S ·s~xu-x aip uo a1u s1aiaurn.md ;)nsno;)y 1adwn.1i .IOJ sa1q'Bp'BA p<lJ'B:>s JO io1dxog :ss a.1nlj!.!I

"'D

c ,.. ..., .... "::! c -a ..., ~

"'D 0 ;;J s: 3 s: .., O> s: "'" a .0 .. ~ 3 .. m c .Ei" rA

" "' ..., ::lJ

~ 0

i!L ~ ~ .. "' CL

"' "O

" ""C

~ !!i. s: l!i. a ("j ... ::i 0 .,, " ""C

s: ii ::r 0 ~ . Q. c: -~ .. ..., d "' "' .. ::0

~ 3 .,, 3 ::r

~ iil .. '

!!i.

' l!i. "'

;;J "

-I "

..., ~ .. -a =i 3 3 .. 3 j;j; ii ii 0 :> 1: ~ .. ~ .. .,, .0

.. ., s: :s: :s: :s: s: :s: s: s: :s: ... ::r ::J ..Q ..Q

""' .., ...,

'"' "' "' ~ "' c :s: s: :s: s: s: s: "'

..... "' ... "' "'

0 "' ... "'

_, . "' "' ... "' I I I I I I I I I I I I I I I I I I

0

0 0

8 0

~ 1 § ~ ~ ~ 0 ~ ~ I I """I I I 8 ~ I I T I:: I ,-..,-II"::

$$~-~9~~~~~t~~8~S~ I l I ~ ~ () ii-I : I I : l_ :· I I I I I 0~ t!l 1 1--1.....L I U I I I I

--1... 0 t 0 @ 8 ~ ..L 0 ; 8 8 --1... s --1... g o e o o o 8 @ o o S

0 0

0 0 0

0 0 0 0

0

-0

0

0

0

0

Page 129: Analysis and Classification of Sounds Produced by Asian

APPENDIX B: CLASSIFICITON TREE RAW DATA OUTPUT

vFQC11F...;.w .. 1m

, __ ["~ ·-+;; r~ ... ~ . .r- 'i;T';:;i.,""!' l

.tt'MJGi~\r. ~ IUol ®':l!1"'111 t.'\l.-0.:C.1<1 tvnm!artll .£rln~'tl-ll

lolHIF~~'7;1

$QI( 1''\ln.'161!\W

"" CIClfliJ;l1C.~

Figure 59: Classification tree final model - plots

~

Table 12: Classification tree final model - raw data outout [l] 0.1220096 > surnmary(mod) Call: rpart(formula calltype - ., data params)

n= 836

CP nsplit rel error xerror xstd 1 0.27638191 0 1.0000000 1.0000000 0.02188309 2 0.22948074 1 0.7236181 0.7236181 0.02420221 3 0.16247906 2 0.4941374 0.4941374 0.02314367 4 0.05276382 3 0.3316583 0.3400335 0.02076694 5 0.03182580 5 0.2261307 0.2428811 0.01833773 6 0.01340034 6 0.1943049 0.2211055 0.01766021 7 0.01005025 7 0.1809045 0.2093802 0.01727080 8 0.01000000 8 0.1708543 0.2043551 0.01709825

Node number 1: 836 observations, complexity param=0.2763819 predicted class=SQL expected loss=0.7141148

class counts: 59 105 57 197 239 179 probabilities: 0.071 0.126 0.068 0.236 0.286 0.214

left son=2 (614 obs) right son=3 (222 obs) Primary splits:

LowerFreqM3 < 675.0855 to the left, improve=l63.3921, FMRateVarM24 < 0.0055 to the left, improve=l49.4871, MedianFreqMll < 561.782 to the left, improve=l00.2804, PkOverallFM20 < 530.3175 to the left, improve= 99.7010, DurationM5 < 0.4395 to the right, improve= 92.1828,

Surrogate splits:

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

PkOverallFM20 < 1302.758 to the left, agree=O. 868, adj=0.505, (0 split) MedianFreqMll < 1318.912 to the left, agree=0.859, adj=0.468, DurationM5 < 2.4235 to the left, agree=0.787, adj=0.198, AMRateVarM22 < 0.007 to the right, agree=0.785, adj=0.189, TimeConcentM9 < 1.6775 to the left, agree=0.782, adj=0.180,

Node number 2: 614 observations, complexity param=0.2294807 predicted ciass=SQK expected loss=0.6840391

class counts: 59 105 57 194 29 170 probabilities: 0.096 0.171 0.093 0.316 0.047 0.277

left son=4 (452 obs) right son=5 (162 obs) Primary splits:

(0 split) (0 split) (0 split) (0 split)

FMRateVarM24 < 0.0055 to the right, improve=128.59100, (0 missing) AMRateVarM22 < 0.0215 to the right, improve=l20.66540, (0 missing) PkOverallFM20 < 530.3175 to the left, improve= 97.40044, (0 missing)

118

Page 130: Analysis and Classification of Sounds Produced by Asian

MedianFreqMll < 561.782 to the left, improve= 97.01046, (0 missing) BandwidthM6 < 2616.284 to the left, improve= 94.27497, (0 missing)

Surrogate splits: AMRateVarM22 < 0.0215 to the right, agree=0.879, adj=0.543, (0 split) DurationM5 < 0.287 to the right, agree=0.824, adj=0.333, (0 split) UpswpFracM28 < 34.3755 to the right, agree=0.800, adj=0.241, (0 split) AMRateM21 < 4.0295 to the left, agree=O. 796, adj=0.228, (0 split) TimeConcentM9 < 0.1265 to the right, agree=0.787, adj=0.191, (0 split)

Node number 3: 222 observations predicted class=SQL expected loss=0.05405405

class counts: 0 0 0 3 210 9 probabilities: 0.000 0.000 0.000 0.014 0.946 0.041

Node number 4: 452 observations, complexity param=0.1624791 predicted class=TMP expected loss=0.6238938

class counts: 58 105 57 33 29 170 probabilities: 0.128 0.232 0.126 0.073 0.064 0.376

left son=8 (237 obs) right son=9 (215 obs) Primary splits:

UpperFreqM4 < 1866.577 to the left, improve=91.40859, MedianFreqMll < 642.902 to the left, improve=89.47726, BandwidthM6 < 1361.89 to the left, improve=87.85287, PkOverallFM20 < 468.347 to the left, improve=85.20569, FMRateVarM24 < 0. 011 to the right, improve=77.18512,

Surrogate splits:

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

BandwidthM6 < 1361.89 to the left, agree=0.985, adj=0.967, (0 split) MedianFreqMll < 680.2285 to the left, agree=0.889, adj=0.767, (0 split) EntropyM26 < 95.3805 to the left, agree=0.863, adj=0.712, (0 split) PkOverallFM20 < 468.347 to the left, agree=0.858, adj=0.702, (0 split) LowerFreqM3 < 241.5455 to the left, agree=O. 827, adj=0.637, (0 split)

Node number 5: 162 observations predicted class=SQK expected loss=0.00617284

class coilllts: 1 0 0 161 0 0 probabilities: 0.006 0.000 0.000 0.994 0.000 0.000

Node number 8: 237 observations, complexity param=0.05276382 predicted class=ROR expected loss=0.5738397

class counts: 46 101 56 19 11 4 probabilities: 0.194 0.426 0.236 0.080 0.046 0.017

left son=l6 (83 obs) right son=l7 (154 obs) Primary splits: ·

DurationM5 < 0.8675 MedianTimeM7 < 0.37 PkOverallFM20 < 170.0515 LowerFreqM3 < 384.906 AMRateVarM22 < 0.3255

Surrogate splits:

to the left, improve=33.89742, to the right, improve=28.32799, to the left, improve=23.31349, to the left, improve=22.73504, to the left, improve=22.52987,

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

MedianTimeM7 < 0.4005 to the left, agree=0.920, adj=0.771, (0 split) AMRateM21 < 1.1595 to the right, agree=O. 869, adj=0.627, (0 split) FMRateM23 < 1.1525 to the right, agree=0.865, adj=0.614, (0 split) TimeConcentM9 < 0.2685 to the left, agree=0.857, adj=0.590, (0 split) UpperFreqM4 < 1184.326 to the right, agree=O. 7 85, adj=0.386, (0 split)

Node number 9: 215 observations, complexity param=0.01005025 predicted class=TMP expected loss=0.227907

class counts: 12 4 1 14 18 166 probabilities: 0.056 0.019 0.005 0.065 0.084 0.772

left son=18 (17 obs) right son=19 (198 obs) Primary splits:

UpswpFracM28 < 36.3945 FMRateVarM24 < 0.011 UpperFreqM4 < 3016.089 TimeConcentM9 < 1.318 BandwidthM6 < 2616.284

Surrogate splits:

to the left, improve=ll.40962, to the right, improve=ll.36771, to the left, i~prove=ll.33500,

to the right, improve=l0.79894, to the left, improve=l0.19102,

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

TimeConcentM9 < 0.0305 to the left, agree=0.930, adj=0.118, (0 split)

119

Page 131: Analysis and Classification of Sounds Produced by Asian

TimeAsyrnmMlO DurationM5 FMRateM23

< -0 .2115 < 0.1325 < 15.4305

to the left, agree=O .930, adj=O .118, (0 split) to the left, agree=0.926, adj=0.059, (0 split) to the right, agree=0.926, adj=0.059, (0 split)

Node number 16: 83 observations, complexity param=0.0318258 predicted class=BRK expected loss=0.4939759

class counts: 42 11 0 19 9 2 probabilities: 0.506 0.133 0.000 0.229 0.108 0.024

left son=32 (55 obs) right son=33 (28 obs) Primary splits:

LowerFreqM3 < 325.017 PkOverallFM20 < 539.676 AMRateVarM22 < 0.0215 MedianFreqMll < 669.679 DurationM5 < 0.3615

Surrogate splits:

to the left, to the left, to the right, to the left, to the right,

improve=21.43904, (0 improve=l8.54354, (0 improve=l8.52093, (0 improve=l8.18897, (0 improve=13. 89051, ( O

missing) missing) missing) missing) missing)

MedianFreqMll < 662.2645 to the left, PkOverallFM20 < 539.676 to the left, DurationM5 < 0.3615 to the right, UpperFreqM4 < 1372.742 to the left, AMRateM21 < 2.7815 to the left,

agree=0.940, agree=0.940, agree=0.904, agree=0.867, agree=0.867,

adj=0.821, adj=0.821, adj=0.714, adj=0.607, adj=0.607,

(0 split) (0 split) (0 split) (0 split) (0 split)

Node number 17: 154 observations, complexity predicted class=ROR expected loss=0.4155844

class counts: 4 90 56 0 2 probabilities: 0.026 0.584 0.364 0.000 0.013

left son=34 (92 obs) right son=35 (62 obs) Primary splits:

param=0.05276382

2 0.013

PkOverallFM20 < 170.0515 to the right, improve=28.37610, (0 missing) AMRateVarM22 < 0.3255 to the left, improve=24.03463, (0 missing) FMRateVarM24 < 0.0695 to the left, improve=22.60592, (0 missing) LowerFreqM3 < 22.5425 to the right, improve=l8.17635, (0 missing) EntropyM26 < 25.117 to the right, improve=l5.06470, (0 missing)

Surrogate splits: MedianFreqMll < 181.817 to the right, agree=0.857, adj=0.645, (0 split) LowerFreqM3 < 35.6645 to the right, agree=O. 786, adj=0.468, (0 split) AMRateVarM22 < 0.087 to the left, agree=0.753, adj=0.387, (0 split) FMRateVarM24 < 0.0195 to the left, agree=O. 740, adj=0.355,. (0 split) UpperFreqM4 < 425.1125 to the right, agree=0.714, adj=0.290, (0 split)

Node number 18: 1.7 observations predicted class=SQK expected loss=0.4705882

class counts: 0 0 0 9 5 3 probabilities: 0.000 0.000 0.000 0.529 0.294 0.176

Node number 19: 198 observations predicted class=TMP expected loss=0.1767677

class counts: 12 4 1 5 13 probabilities: 0.061 0.020 0.005 0.025 0.066

Node number 32: 55 observations predicted class=BRK expected loss=0.2363636

class counts: 42 11 0 0 1 probabilities: 0.764 0.200 0.000 0.000 0.018

163 0.823

1 0.018

Node number 33: 28 observations, complexity param=0.01340034 predicted class=SQK expected loss=0.3214286

class counts: 0 0 0 19 8 1 probabilities: 0.000 0.000 0.000 0.679 0.286 0.036

left son=66 (19 obs) right son=67 (9 obs) Primary splits:

FMRateVarM24 AMRateVarM22 EntropyM26 BandwidthM6 PkOverallRel TM18

Surrogate splits:

< 0. 011 < 0.0215 < 97.3315 < 1012.061 < 56.2775

to the right, improve=ll.007940, to the left, improve= 9.385714, to the right, improve= 8.058442, to the right, improve= 4.919048, to the left, improve= 4.500000,

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

AMRateVarM22 < 0.0215 to the left, agree=0.964, adj=0.889, (0 split)

120

Page 132: Analysis and Classification of Sounds Produced by Asian

EntropyM26 < 80.201 to the right, agree=0.929, PkOverallRelTM18 < 56.2775 to the left, agree=0.857, BandwidthM6 < 973.779 to the right, agree=0.821, MedianFreqMll < 693.84 to the right, agree=0.821,

Node number 34: 92 observations predicted class=ROR expected loss=0.1630435

class counts: 0 77 11 0 2 2 probabilities: 0.000 0.837 0.120 0.000 0.022 0.022

Node number 35: 62 observations predicted class=RUM expected loss=0.2741935

class counts: 4 13 45 0 0 0 probabilities: 0.065 0.210 0.726 0.000 0.000 0.000

Node number 66: 19 observations predicted class=SQK expected loss=O

class counts: 0 0 0 19 0 0 probabilities: 0.000 0.000 0.000 1.000 0.000 0.000

Node number 67: 9 observations predicted class=SQL expected loss=0.1111111

class counts: 0 0 0 0 8 1 probabilities: 0.000 0.000 0.000 0.000 0.889 0.111

·--rt.•

--··» ....... 10.

adj=0.778, (0 split) adj=0.556, (0 split) adj=0.444, ( 0 split) adj=0.444, (0 split)

_[J_ • ..*-.w

!,j~ -~ ·------- ---·----~~:::o""""'i .... L

Figure 60: Classification tree validation model - plots

Table 13: Classification tree validation model - raw data output (1] 0.2368421 > summary(new) Call: rpart(formula calltype - ., data= parru~s)

n= 836

CP nsplit rel error xerror xstd 1 0.2763819 0 1.0000000 1.0000000 0.02188309 2 0.2294807 1 0.7236181 0.7252931 0.02420020 3 0.1624791 2 0.4941374 0.4974874 0.02317902 4 0.1000000 3 0.3316583 0.3400335 0.02076694

Node number 1: 836 observations, complexity param=0.2763819 predicted class=SQL expected loss=0.7141148

class counts: 59 105 57 197 239 179 probabilities: 0.071 0.126 0.068 0.236 0.286 0.214

left son=2 (614 obs) right son=3 (222 obs) Primary splits:

LowerFreqM3 < 675.0855 to the left, improve=163.3921, FMRateVarM24 < 0.0055 to the left, improve=149.4871, MedianFreqMll < 561.782 to the left, improve=l00.2804,

(0 missing) (0 missing) (0 missing)

121

Page 133: Analysis and Classification of Sounds Produced by Asian

PkOverallFM20 < 530.3175 to the left, improve= 99.7010, (0 missing) DurationM5 < 0.4395 to the right, improve= 92.1828, (0 missing)

Surrogate splits: PkOverallFM20 < 1302.758 MedianFreqMll < 1318.912 DurationM5 < 2.4235 AMRateVarM22 < 0.007 TimeConcentM9 < 1.6775

to the left, to the left, to the left, to the right, to the left,

agree=0.868, agree=0.859, agree=0.787, agree=0.785, agree=0.782,

adj=0.505, adj=0.468, adj=0.198, adj=0.189, adj=0.180,

Node number 2: 614 observations, complexity param=0.2294807 predicted class=SQK expected loss=0.6840391

class counts: 59 105 57 194 29 170 probabilities: 0.096 0.171 0.093 0.316 0.047 0.277

left son=4 (452 obs) right son=5 (162 obs) Primary splits:

(0 split) (0 split) (0 split) (0 split) (0 split)

FMRateVarM24 < 0.0055 AMRateVarM22 < 0.0215 PkOverallFM20 < 530.3175 MedianFreqMll < 561.782 BandwidthM6 < 2616.284

to the right, to the right, to the left, to the left, to the left,

improve=128.59100, improve=120.66540, improve= 97.40044, improve= 97.01046, improve= 94.27497,

(0 missing) (0 missing) (0 missing) (0 missing) (0 missing)

Surrogate splits: AMRateVarM22 < 0.0215 DurationM5 < 0.287 UpswpFracM28 < 34.3755 AMRateM21 < 4.0295 TimeConcentM9 < 0.1265

Node number 3: 222 observations

to the right, to the right, to the right, to the left, to the right,

agree=0.879, agree=0.824, agree=0.800, agree=0.796, agree=0.787,

predicted class=SQL expected loss=0.05405405 class counts: 0 0 0 3 210 9

probabilities: 0.000 0.000 0.000 0.014 0.946 0.041

adj=0.543, adj=0.333, adj=0.241, adj=0.228, adj=0.191,

Node number 4: 452 observations, complexity param=0.1624791 predicted class=TMP expected loss=0.6238938

class counts: 58 105 57 33 29 170 probabilities: 0.128 0.232 0.126 0.073 0.064 0.376

left son=8 (237 obs) right son=9 (215 obs) Primary splits:

(0 split) (0 split) (0 split) (0 split) (0 split)

UpperFreqM4 < 1866.577 to the left, improve=91.40859, (0 missing) MedianFreqMll < 642.902 to the left, improve=89.47726, (0 missing) BandwidthM6 < 1361. 89 to the left, improve=87.85287, (0 missing) PkOverallFM20 < 468.347 to the left, improve=85.20569, (0 missing) FMRateVarM24 < 0.011 to the right, improve=77.18512, (0 missing)

Surrogate splits: BandwidthM6 < 1361. 89 to the left, .agree=O. 985, adj=0.967, (0 split) MedianFreqMll < 680.2285 to the left, agree=0.889, adj=0.767, (0 split) EntropyM26 < 95.3805 to the left, agree=0.863, adj=0.712, (0 split) PkOverallFM20 < 468.347 to the left, agree=0.858, adj=0.702, (0 split) LowerFreqM3 < 241.5455 to the left, agree=0.827, adj=0.637, (0 split)

Node number 5: 162 observations predicted class=SQK expected loss=0.00617284

class counts: 1 0 0 161 0 0 probabilities: 0.006 0.000 0.000 0.994 0.000 0.000

Node number 8: 237 observations predicted class=ROR expected loss=0.5738397

class counts: 46 101 56 19 11 4 probabilities: 0.194 0.426 0.236 0.080 0.046 0.017

Node number 9: 215 observations predi~ted class=TMP expected loss=0.227907

class counts: 12, 4 1 14 18 166 probabilities: 0.056 0.019 0.005 0.065 0.084 0.772

122

Page 134: Analysis and Classification of Sounds Produced by Asian

APPENDIX C: PRINCIPAL COMPONENT ANALYSIS DATA - RAW DATA

OUTPUT

Table 14: PrinciEal ComEonent Anall'.sis - raw data outEut Standard deviation 2.1032890 1.8106105 1.31698516 1.15100919 1.11924231 Proportion of Variance 0.2460624 0.1823465 0.09647373 0.07368938 0.06967798 Cumulative Proportion 0.2460624 0.4284088 0.52488258 0.59857195 0.66824993

Loadings: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5

LowerFreqM3 -0.29 -0.14 -0.35 -0.17 0.14 UpperFreqM4 -0.42 0.25 0.15 DurationM5 -0.52 BandwidthM6 -0. 38 0. 35 0. 20 MedianTimeM7 -0.51 0.18 TimeConcentM9 -0.51 -0.11 0.12 TimeAsyrnmM10 -0.32 0. 35 MedianFreqMll -0.44 0.15 FreqAsyrnmM14 0.33 0.20 -0.23 PkOverallRelTM18 0.28 -0.47 PkOverallFM20 -0.38 -0.10 -0.24 -0.13 0.14 AMRateM21 0.24 -0.34 0.17 AMRateVarM22 0.23 0.28 0.12 0.42 FMRateM23 0.24 -0.25 0.10 FMRateVarM24 0.21 0.24 0.46 EntropyM26 -0. 38 0.24 0.15 UpswpMeanM27 0.17 -0.54 -0.29 UpswpFracM28 -0.16 0. 31 -0.48 -0.25

broken.stick(18) j E (j)

[1, l 1 0.194172671 [2' l 2 0.138617115 [3' l 3 0.110839338 [4, l 4 0.092320819 [5, l 5 0.078431930 [6, l 6 0.067320819 [7, l 7 0.058061560 [8, l 8 0.050125052 [9'} 9 0.043180608

[10,] 10 0.037007768 [11, l 11 0.031452212 [12,] 12 0.026401707 [13,] 13 0.021772078 [14,] 14 0.017498574 [15, l 15 0.013530320

'[16,] 16 0.009826616 [17,] 17 0.006354394 [18,J 18 0.003086420

Comp.l Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Proportion of Variance 0.2460624 0.1823465 0.09647373 0.07368938 0.06967798 0.05750879

123

Page 135: Analysis and Classification of Sounds Produced by Asian

APPENDIX D: ANALYSIS OF SIMILARITY (ANOSIM) DAT A OUTPUT

Table 15: Pair-wise ANOSIM plots for 15 pair-wise tests Pairs are noted above each plot -

BRKv.ROR BRKv.RUM

~ R:: 0??7 P:: <nntH R= n">7Q P- <nnn1

' ' ~ : : ;

or 0 §

~ §

s g

18 iii 0

~

0 0 ~

Between SRK ROR RUM SQK SOL TMP Between BRK ROR RUM SQK SOL TMP

BRKv.SQK BRKv.SQL 0

R=0~1<1 P=<•"'nn1 R::::fllAll P::<ilfl(l~

~

y '

§

8 g ... ~

~

i B 0

9 ~ §

0 ~ § + 0 0

Between BRK ROR RUM SQK SOL TMP Between BR~ ROR RUM SQK SQL TMP

BRKv.TMP RORv.RUM

D: ""'' P: <111101

~ ~= n.H7 P= <nt"\n1

g

~ ~

~ g 6 '

8 ~

0

§ ~ g j ~ ~

0 0 ~ i Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SQL TMP

124

Page 136: Analysis and Classification of Sounds Produced by Asian

RORv.SQK RORv.SQL

~

9 §

r l ~

~

s §

8 8 ~ l'l

~ ~ ~ 0

0 0

Between BRK ROR RUM SOK SOL TMP Bef'.Neen BRK ROR RUM SQK SQL TMP

RORv.TMP RUMv.SQK

~ ~ 1 g

t r Ii

T 0

! i

0

R ~

~ g

8 ·@

~ ~ -T

' § 0 0 t

Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL TMP

RUMv.SQL RUMv.TMP

O: n;:qs:.. P= c:"nn•

g

~ ~ 8

! ~ ~

~ ~ ~ 9 ~ r ~ t 0 0

Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL TMP

125

Page 137: Analysis and Classification of Sounds Produced by Asian

SQKv. SQL SQKv. TMP

i g - ____n~ 0

18 :::

i y g y H ;;\ :il i

f g

8 "' g

i 8 l ~ ~ 0

.'! -.-Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL 1MP

SQLv.TMP

!J 8 g~ ~

BetWeen BRK ROR RUM SQK SQL TMP

126

Page 138: Analysis and Classification of Sounds Produced by Asian

APPENDIX E: POWER SPECTRA OF RUMBLES AND NOISE

Rumble: 14 Hz to 1.2 Hz 105,--~~..--~~~~~--,~~~,.-~~-,.-~~----.

Rumble: call bandwidth 105~~~~~~~~~~~~~1

100

95

~ 00

~ ,! 851-

~ i;g" 00

75

70

v 65 200 400 600

frequency, Hz 800

Noise: 14 Hz to 1.2 kHz

100

95

v 90

85

1000 1200 80

100 200 3lO 400 500 600 700 800 frequency, Hz

Noise: 14 Hz to 20 kHz 95.-~~..--~~~~~--,~~~,.-~~-,.-~~----. 100~~..--~,.-~~~~~~~~~~~~..--~~~

70

65

ffi 200 400 600 800 ~ = frequency, Hz

~ ~

g ~

i ig"

90

80

30 2000 4000 GOOD 8000 10000 12000 14000 16000 18000 frequency, Hz

Figure 61: Power spectra of Rumble and background noise at the Oregon Zoo (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

127

Page 139: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz 100

95~ n I~ 90 ~

~ . ~ a5

i " 80 rg"

75

70

65

90

~

~ 80 ~

i !g- 75

70

~ 100 200 300 400 500 600 700 800 900 1000 1100

frequency, Hz

Noise: 14 Hz to 1.2 kHz

65'--~~'--~~-'-~~--'-~~~'--~~-'-~~--' JJO 400 600 BOO 1000 1JJO

frequency, Hz

Rumble: call bandwidth 100

95

~ . 90 ~

! !g- 85

00

75'-~'--~--'-~~-'-~~-'-~~'--~--'-~~-'-__J 100 200 300 400 500 600 700

frequency. Hz

Noise: 14 Hz to 20 kHz 90~~~~~~~~~~~~~~~~~~~~~-

~ . 0 c ~

00

~ 60 §

rg" so

40

3)'-~-'-~-'-~-'-~-'-~--'-~~'-~'--~-'-~-'-~~ JJOO 4000 6000 8000 10000 12000 14000 16000 18000

frequency, Hz.

Figure 62: Power spectra of Rumble and background noise at the Oregon Zoo (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

128

Page 140: Analysis and Classification of Sounds Produced by Asian

~ ~ ~ 5 !fi

70

65

60

55 100

100

95

90

85

80

75

70

65

60

55i

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

Ell

~~

~ 75

70

200 300 400 500 600 700 800 65~~~~~~~~~~~~~~~~~-~~~~

OCIJ 1000 1100 5IJ M ~ D B m B a s 500 frequency, Hz frequency. Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz 100~~~~~~~~~~~~~~~~~~~~

90

Ell

70

:m 400 500 800 1000 1200 20

2000 4000 6000 8000 10000 12000 14000 16000 1 BOOJ frequency, Hz frequency, Hz

Figure 63: Power spectra of Rumble and background noise at the Oregon Zoo (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

129

Page 141: Analysis and Classification of Sounds Produced by Asian

"" ~ u

65

~ 60

i § 55 aj

"" 50

45

40

Rumble: 14 Hz to 1.2 Hz

35 ~ 100 200 300 400 500 600 700 800 900. 1000 1100

frequency, Hz

Noise: 14 Hz to 1.2 kHz

65f

60f

55'

SOf

45

40

Rumble: call bandwidth

~\ ~

50 100 150 200 250 300 350 400 frequency.Hz

Noise: 14 Hz to 20 kHz 70~~~~~~~~~~~~~~~~~~~~ 10~~~~~--~~~~~~~~~~~~~~

65

-g 55 g

j 50

rg" 45

40

35

3J 200 400 600 frequency, Hz

800 1000 1200

65

60

55

"" g 50

i 45 § !g- 40

25

20 2000 4000 6DOO 8000 10000 12000 14000 16000 18000 frequency, Hz

Figure 64: Power spectra of Rumble and background noise at the Elephant Nature Park (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

130

Page 142: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth oo~~~~~~~~~~~~~~~~~~~~~ so~~~~~~~~~~~~~~~~~~~~~

75

70

65

~

~ 60 ~ i 55 § !fl. 50

45

40

35

30 200 400 ---

frequency, Hz

Noise: 14 Hz to 1.2 kHz

65

60

55

50

45 40 60 BO 100 120 140 160 180 200 220

frequency, Hz

Noise: 14 Hz to 20 kHz oo~~~~~~~~~~~~~~~~~~~---, oo~~~~~~~~~~~~~~~~~~~---,

75

70

65

45

40

35 ~:

30'---'-~-'-~-'----''-----'~-'-~-'-~-'--~'-----'~--'--~~ 100 200 300 400 500 600 700 800 900 1000 1100

frequency, Hz

70

60

50

10 ~ ~ = ~1~1=1~1~1~ frequency. Hz

Figure 65: Power spectra of Rumble and background noise at the Elephant Nature Park (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

131

Page 143: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz BO

75

70

65

... g 60 ~

j 55

~· 51)

45

40

35

30 100 200 300 400 500 600 700 800 900 100J 1100

frequency, Hz

Noise: 14 Hz to 1.2 kHz 70.----~~~~~~~~~~~~~~.----~~~~

65

~ e so

i 45

iii 40

35

30

25 100 200 300 400 500 600 700 800 900 1000 1100

frequency, Hz

Rumble: call bandwidth so~~~~~~~~~~~~~~~~~~~~~~

] 65

~ 60

~ ~· 55

50

45

40 50 100 150 200 250 300

frequency, Hz

Noise: 14 Hz to 20 kHz 70.----~~~~~~~~~~~~~~~~~~~

iii

60

50

10~~~~~~~~~~~~~~~~~~~~~~

2000 4000 6000 8000 10000 12000 14000 1600) 18000 frequency, Hz

Figure 66: Power spectra of Rumble and background noise at the Elephant Nature Park (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

132

Page 144: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth Ill Ill

75 75

70

65 70

~ ~

~ 60 ~

w g 65 ~

i 55

; ro i § 60 !ll.

45 55

40 50

35

~ 100 200 300 400 500 600 700 800 900 1000 11 DO

45 50 100 150 200 250 300 350

frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz Ill.-~~.-~~~~~~~~~~~~~~~--, BO.-~.-~.-~.-~.-~.-~.-~.-~.-~.---,

75

70

~~~--'~~~~~~-'-~~~~~---'-~~--'

200 400 600 frequency, Hz

800 1000 1200

70

60

10 =~ ~~1=1=1~1~1~ frequency, Hz

Figure 67: Power spectra of Rumble and background noise at the Elephant Nature Park (example 4) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

133

Page 145: Analysis and Classification of Sounds Produced by Asian

.,, 65

i i SJ

~.55

50

45

Rumble: 14 Hz to 1.2 Hz

.,, . ~ 65

i " 60 ill-

55

50

Rumble: call bandwidth

40 100 200 300 400 500 600 700 800 9()0 1000 1100

frequency, Hz

45L-LC.---'-.---'-.--~.---'-.---'-.---L.----'.--.--L-.---'-J 50 100 150 200 250 300 350 400 450 500

frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz 80.--.--.--.--.--.--~.--.--~.--.--.--~.--.--~.--.-----. 80.--.--~.--~.--~.--~.--~.--~.--~.--~.--~~

'g 65 ~ I!! -s 60 § ~- 55

50

45

40 200 400 600 800 1000 1200

frequency, Hz

.,, .

70

60

~ 50

i 40 ill-

30

20

10 2000 4000 6000 8000 10000 12000 14000 16000 18000

frequency, Hz

Figure 68: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

134

Page 146: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth 85.---~~~~~~~~~~~~~~~~~~~~

\f\~ 65

60

55

50

45

1000 1200 Ellll 400 600 35

200 40 -

200 300 400 500 600 10U frequency, Hz frequency, Hz

Nmse. . . 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz 70 m~~~~~~~~~~~~~~~~~~~~~

~ 55 e i 50 5 !g. 45

40

35

30 100 200 DJ 400 500 600 700 Ellll 900 1000 1100

frequency, Hz

~ ~ u

60

50

~ i 40

~ 30

20

w ~ ~ =~1~1=uooo1=1~ 1requem:y, Hz

Figure 69: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

135

Page 147: Analysis and Classification of Sounds Produced by Asian

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth 85~~~~~~~~~~~~~~-....~~~~-.... as~-....~~~~~~~~~~~~~~~~~-....

00

.,, g 65 !!

~ ro § !g. 55

50

45

40

35 100 200 300 400 500 600 700 800 900 1000 1100

frequency, Hz

Noise: 14 Hz to 1.2 kHz

60

55

50

45

40

35 100

~ I./

200 300 400 500 600 700 800 900 frequency, Hz

Noise: 14 Hz to 20 kHz 65~·~~~~~~~~~~~~~~~~~~

i: 50

~ 'i 45 §

~ 40

35

30

1200 25'-----x~~~o;!;n-----'--_J 800 11XXJ 200 400 600 frequency, Hz

60

55

50

45

20 21XXJ 4000 6000 8000 10000 12000 14000 16000 18000

frequency, Hz

Figure 70: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

136