talsyntes: joakim gustafson tal, musik och hörsel - · pdf filetalsyntes: joakim...

Talsyntes: Joakim Gustafson Tal, musik och hörsel

1

J. Gustafson, CTT, KTH

Speech synthesis

DT2112

Joakim Gustafson, CTT, KTH School for Computer Science and

Communication Many slides prepared by Olov Engwall (and others)

J. Gustafson, CTT, KTH 2

Text-To-Speech synthesis (TTS)

The automatic generation of synthesized sound or visual output from any phonetic string.

Our focus in this course!


Different kinds of speech synthesis

•  Recorded speech –  Words or phrases (telephone banking)

–  Fixed vocabulary – maintenance problems…

•  Concatenative speech synthesis

•  Parametric synthesis

•  Multimodal synthesis

J. Gustafson, CTT, KTH The SCALE Summer Workshop June 2012

What the voice conveys

•  The linguistic component (the words that are said)

•  The extralinguistic component (the identity of the speaker)

•  The paralinguistic component (the attitude of the speaker)

•  The dialog control component (selection of the next speaker)


Desireable synthesis features from a dialogue perspective

•  Real-time, Incremental, Interruptable •  Explicit control of prosodic parameters

–  Fundamental frequency –  Intensity –  Natural sounding lengthening, hesitation, interruptions

•  Generation of extra-linguistic sounds –  Filled pauses –  Creeks/Gargles –  Smacks/Inhalations/exhalations to give turn


The synthesis space

Intelligibility Naturalness

Bit rate

Vocabulary Units

Complexity

Processing needs

Flexibility Speech

Knowledge

Cost


2


The steps in TTS

text

Linguistic analysis

Prosodic analysis

Phonetic description

Sound generation

Morphological analysis Lexicon and rules Syntax analysis

Rules and lexicon

Rules and unit selection

Concatenation Rules

Language ident.

“abcd”


The automatic generation of synthesized sound from any text string.

From text


Preprocessor •  Sentence end detection (semicolon, period – ratio, time and

decimal point, sentence ending respectively) •  Abbreviations (e.g. – for instance) Changed to their full form with the help of lexicons •  Acronyms (I.B.M – these can be read as a sequence of

characters, or NASA which can be read following the default way) •  Numbers (Once detected, first interpreted as rational, time of the

day, dates and ordinal depending on their context) •  Idioms (e.g. “In spite of”, “as a matter of fact”– these are combined

into single FSU using a special lexicon)


Grapheme-to-phoneme conversion •  Dictionary:

–  Store a maximum of phonological knowledge into a lexicon. –  Compounding rules describe how the morphemes of dictionary items are

modified. –  Hand-corrected, expensive –  The lexicon is never complete: needs out of vocabulary pronouncer,

transcribed by rule.

•  Rules: –  A set of letter to sound (grapheme to phoneme) rules. –  Words pronounced in a such a particular way that they have their own rule

are stored in exceptions directory. –  Fast & easy, but lower accuracy

•  Machine learning: –  Cart tree –  Analogy pronunciation


11

Prosody

•  Prosody = melody, rhythm, “tone” of speech

•  Not what words are said, but how they are said •  Prosody is conveyed using:

–  Pitch –  Phone durations –  Energy

•  Human languages use prosody to convey: –  phrasing and structure (e.g. sentence boundaries)

–  disfluencies (e.g. false starts, repairs, fillers)

–  sentence mode (statement vs question)

–  emotional attitudes (urgency, surprise, anger)


Intonation: F0 contour Large pitch range (female) Authoritive (final fall) Emphasis for Finance (H*) Final has a raise – more information to come

•  Word stress and sentence intonation –  each word has at least one syllable which is spoken with higher

prominence –  in each phrase the stressed syllable can be accented depending on

the semantics and syntax of the phrase •  Prosody relies on syntax, semantics, pragmatics: personal reflection of

the reader.


3


Pitch contour modeling •  Tonetics (the British school)

–  tone groups composed of syllables {unstressed, stressed, accented or nuclear}.

–  nuclear syllables have nuclear tones {fall, rise, fall-rise, rise-fall}

•  ToBI (Tones and Break Indices) –  Phrases split into intermediate phrases composed of syllables. –  Relative tone levels: high (H) or low (L) (plus diacritics) at every

intonational or intermediate phrase boundary (%) and on every accented syllable

•  Stylization method (prosodic pattern measured from natural speech) –  Demo


The automatic generation of synthesized sound from any text string.

To Speech


Synthesis approaches

By Concatenation Elementary speech units are stored in a database and then concatenated and processed to produce the speech signal

By Rule Speech is produced by mathematical rules that describe the influence of phonemes on one another


Research trends in speech synthesis

1950 Synthesis by analysis

1960 Phonetic rules

1970 Linguistic processing

1980 Concatenation

1990 Automatic procedures

2000


Text-to-Speech Synthesis Evolution

1962 1967 1972 1982 1987 1992 1997 2002

Year

Formant Synthesis

Bell Labs; Joint Speech Research Unit; MIT (DEC-Talk); Haskins

Lab; KTH

LPC-Based Diphone/Dyad

Synthesis Bell Labs; CNET; Bellcore; Berkeley

Speech Technology

Unit Selection Synthesis

ATR in Japan; CSTR in Scotland; BT in

England; AT&T Labs (1998); L&H in

Belgium

Poor Intelligibility;

Poor Naturalness, Small footprint

Good Intelligibility;

Poor Naturalness

Good Intelligibility;

Customer Quality Naturalness

(Limited Context)

HMM ��Synthesis

HTS in Japan; CSTR in Scotland;

Multi-speaker training, speaker

adaptation; Naturalness,

generative, Small footprint


Synthesis by rule


4


Models of speech production a) humans b) machines

Carrier vocal chords, pitch, voice quality Modulation vocal tract, nose, mouth formants


The LPC Vocoder •  Vocoder is a concatenation of

the terms 'voice' and 'coding')

•  A Vocoder aims to replace the carrier with another carrier while keeping the message

•  It is used in telephony to

decrease bit rate

•  The pitch and LPC coefficients are computed on the speech

•  These are then used as input in the speech generator


The TADEM-STRAIGHT Vocoder •  Developed by Hideki Kawahara at Kyoto Univ. •  speech analysis, modification and resynthesis •  Will be used in Lab2


Articulatory Synthesis


Articulatory synthesis potential use •  Articulatory synthesis

–  Calculations directly from cross sectional areas

–  Fluid dynamics calculations

•  Visual synthesis –  Articulation training

•  Demonstrations and research


Articulatory parameters

•  Jaw opening •  Lip rounding •  Lip Protrusion •  Tongue position •  Tongue height •  Tongue tip •  Velum •  Hyoid


5


From articulation to acoustics

Transfer function

Vocal tract model

Tubes

Waveform

Cross-sections

3D air flow calculations

Area function


Benefits: •  Speech production in the same way as humans •  Can be made with few parameters •  The changes are intuitive

(raise the tongue tip, round the lips)

Disadvantages: •  Computationally demanding •  Problems with consonants •  Articulatory measurements required •  State-of-the-art articulatory synthesis still sounds bad

Summary: Articulatory Synthesis


Formant Synthesis


Formant synthesis (1959-1987)

•  Haskins, 1959 •  KTH – Stockholm, 1962 •  Bell Labs, 1973 •  MIT, 1976 •  MIT-talk, 1979 •  Speak ‘N spell, 1980 •  BELL Labs, 1985 •  Dec talk, 1987


Gunnar Fant and OVE 1962


OVE II


6


Rule-driven formant synthesis

•  Parameters are generated by rule (RULSYS , Carlson et al.)

•  Formant values are generated by interpolating between target frequencies

•  Parameters are fed to a synthesizer (GLOVE, Carlson et al.)


Summary: formant synthesis

Benefits: •  Possible to change the voice to get different:

•  speakers • emotions • voice qualities

•  Small footprint

Disadvantages: •  Hard to achieve naturalness in voice source •  Some consonant sounds are hard to model with formants (bursts)


From rule based concatenative synthesis

• Rule based sounds unnatural, while concatenative synthesis provides (piece-wise) high quality speech.

•  Certain sounds are hard to be produced by rule but easy to concatenate: –  Bursts, voiceless stops are too difficult

• Rule based had an advantage of small footprint, however storing the segment database is no longer an issue

• Change of applications: –  From reading machines for the blind to spoken dialogue systems


Synthesis by concatenation


Let’s get the terms straight

Concatenative synthesis Definition: All kinds of synthesis based on the concatenation of units, regardless of

type (sound, formant trajectories, articulatory parameters) and size (diphones, triphones, syllables, longer units). There is only one candidate per setting.

Everyday use: Concatenation of same-size sound units.

Unit selection synthesis Definition: All kinds of synthesis based on the concatenation of units where there are several candidates to choose from, regardless of if the candidates have the same, fixed size or if the size is variable. Everyday use: Concatenation of variable sized sound units.


Database preparation when building a concatenative synthesis

• Choose the speech units (Phone, Diphone, Sub-word unit, Cluster based unit selection) • Compile and record utterances • Segment signal and extract speech units • Store segment waveforms (along with context) and information in a database: Dictionary, waveform, pitch mark e.g. “ch-l r021 412.035 463.009 518.23”

diphone file Start time Middle time End •  Pitch mark file: a list of each pitch mark position in the file • Extract parameters; create parametric segment database (for data compaction and prosody matching) • Perform amplitude equalization (prevents mismatches)


7


Signal manipulations in concatenative synthesis

•  Prosodic modifications –  Possibility to modify F0 –  Possibility to lengthen or shorten segments

•  Spectral modifications –  Interpolation of spectrum at joints


Sequences of a particular sound/phone in all its environments of occurrence or all/most two-phone sequences occurring in a language: _auto_ -> _a, au, ut, to, o_ • Rationale: the center’of a phonetic realization is the most stable region, whereas the transition from one segment to another contains the most interesting phenomena, and is thus the hardest to model.

Assignment: Diphone ”synthesis”; cut and paste

Diphone Synthesis


Diphones

•  Need O(phone2) number of units –  Some combinations don’t exist (hopefully) –  ATT (Olive et al. 1998) system had 43 phones

•  1849 possible diphones •  Phonotactics ([h] only occurs before vowels), don’t need to keep

diphones across silence •  Only 1172 actual diphones

–  May include stress, consonant clusters •  So could have more

–  Lots of phonetic knowledge in design

•  Database relatively small –  Around 8 megabytes for English (16 KHz 16 bit)

Slide from Richard Sproat J. Gustafson, CTT, KTH

Building diphone schemata

•  Find list of phones in language: –  Plus interesting allophones –  Stress, tons, clusters, onset/coda, etc –  Foreign (rare) phones.

•  Build carriers for: –  Consonant-vowel, vowel-consonant –  Vowel-vowel, consonant-consonant –  Silence-phone, phone-silence –  Other special cases

•  Check the output: –  List all diphones and justify missing ones –  Every diphone list has mistakes

Slide from Richard Sproat


Designing a diphone inventory: Nonsense words

•  Build set of carrier words: –  pau t aa b aa b aa pau –  pau t aa m aa m aa pau –  pau t aa m iy m aa pau –  pau t aa m iy m aa pau –  pau t aa m ih m aa pau

•  Advantages: –  Easy to get all diphones –  Likely to be pronounced consistently

•  No lexical interference

•  Disadvantages: –  (possibly) bigger database –  Speaker becomes bored


Designing a diphone inventory: Natural words

•  Greedily select sentences/words: –  Quebecois arguments –  Brouhaha abstractions –  Arkansas arranging

•  Advantages: –  Will be pronounced naturally –  Easier for speaker to pronounce –  Smaller database? (505 pairs vs. 1345 words)

•  Disadvantages: –  May not be pronounced correctly



8


Making recordings consistent:

•  Diphone should come from mid-word –  Help ensure full articulation

•  Performed consistently –  Constant pitch (monotone), power, duration

•  Use (synthesized) prompts: –  Helps avoid pronunciation problems –  Keeps speaker consistent –  Used for alignment in labeling


Recording conditions

•  Ideal: –  Anechoic chamber –  Studio quality recording –  EGG signal

•  More likely: –  Quiet room –  Cheap microphone/sound blaster –  No EGG –  Headmounted microphone

•  What we can do: –  Repeatable conditions –  Careful setting on audio levels



Labeling Diphones •  Run an ASR in forced alignment mode

–  Forced alignment: •  In: A trained ASR system, a wavefile, a word transcription of the wavefile •  Returns: an alignment of the phones in the words to the wavefile.

•  Much easier than phonetic labeling: –  The words are defined –  The phone sequence is generally defined –  They are clearly articulated –  But sometimes speaker still pronounces wrong, so need to check.

•  Phone boundaries less important –  +- 10 ms is okay

•  Midphone boundaries important –  Where is the stable part –  Can it be automatically found?


Finding diphone boundaries

•  Stable part in phones "   For stops: one third in "   For phone-silence: one quarter in "   For other diphones: 50% in

•  In time alignment case: "   Given explicit known diphone boundaries in prompt in the label file "   Use dynamic time warping to find same stable point in new speech

•  Optimal coupling "   Taylor and Isard 1991, Conkie and Isard 1996 "   Instead of precutting the diphones

–  Wait until we are about to concatenate the diphones together –  Then take the 2 complete (uncut diphones) –  Find optimal join points by measuring cepstral distance at potential join points,

pick best



Summary: Diphone Synthesis

•  Well-understood, mature technology •  Augmentations

–  Stress –  Onset/coda –  Demi-syllables

•  Problems: –  Signal processing still necessary for modifying durations –  Source data is still not natural –  Units are just not large enough; can’t handle word-specific

effects, etc

Slide from Dan Jurafsky J. Gustafson, CTT, KTH

From diphone synthesis to Unit Selection Synthesis

•  Natural data solves problems with diphones –  Diphone databases are carefully designed but:

•  Speaker makes errors •  Speaker doesn’t speak intended dialect •  Require database design to be right

–  If it’s automatic •  Labeled with what the speaker actually said •  Coarticulation, schwas, flaps are natural

•  “There’s no data like more data” –  Lots of copies of each unit mean you can choose just the

right one for the context –  Larger units mean you can capture wider effects

Slide from Dan Jurafsky


9


Unit selection Synthesis

Slide from Tokuda J. Gustafson, CTT, KTH

Unit Selection Intuition •  Given a big database •  For each segment that we want to synthesize

–  Find the unit in the database that is the best to synthesize this target segment

•  What does “best” mean? –  Target cost: Closest match to the target description, in terms of

•  Phonetic context •  Pitch, power, duration, phrase position

–  Concatenation cost: The difference between the end of diphone 1 and the start of diphone 2: •  Matching formants + other spectral characteristics •  Matching energy •  Matching F0



The units in Unit Selection • Different types of units: e.g. diphones, phones, syllables, words, etc.

• Multiple occurrences of the units cover a wide space of the spectral and prosodic parameters

• Units nearest in this space to the targets will be chosen and will require only minor modification

• The corpus is segmented into phonetic units, indexed, and used as-is

• Selection is made on-line

•  The trend is towards longer and longer units

1999 2000 2001 2002 2003 2004 2005


•  Large databases of recorded natural speech

•  Minimal processing

•  Annotation of database – what information is needed?

•  Few cuts > maximally long units selected (but context and prosody must fit well)

•  Target and concatenation costs


Features of Unit Selection Synthesis


Database creation: a good speaker

•  Professional speakers are always better: –  Consistent style and articulation –  Although these databases are carefully labeled

•  Ideally (according to AT&T experiments): –  Record 20 professional speakers (small amounts of data) –  Build simple synthesis examples –  Get many (200?) people to listen and score them –  Take best voices

•  Correlates for human preferences: –  High power in unvoiced speech –  High power in higher frequencies –  Larger pitch range

Text from Paul Taylor and Richard Sproat J. Gustafson, CTT, KTH

Database creation: good recording conditions

•  Good script –  Application dependent helps

•  Good word coverage •  News data synthesizes as news data •  News data is bad for dialog.

–  Good phonetic coverage, especially wrt context –  Low ambiguity –  Easy to read

•  Annotate at phone level, with stress, word information, phrase breaks

Text from Paul Taylor and Richard Sproat


10


Creating database

•  Unlike diphone synthesis, prosodic variation is a good thing

•  Accurate annotation is crucial

•  Pitch annotation needs to be very accurate

•  Phone alignments can be done automatically, as described for diphones

Slide from Dan Jurafsky J. Gustafson, CTT, KTH

Practical System Issues •  Size of typical system (Rhetorical rVoice):

–  ~300M

•  Speed: –  For each diphone, average of 1000 units to choose from, so: –  1000 target costs –  1000x1000 join costs –  Each join cost, say 30x30 float point calculations –  10-15 diphones per second –  10 billion floating point calculations per second

•  But commercial systems must run ~50x faster than real time •  Heavy pruning essential:

–  1000 units -> 25 units

Slide from Paul Taylor


Summary: Unit Selection •  Advantages

–  Quality is far superior to diphones –  Natural prosody selection sounds better –  Non-linguistic features of the speakers voice built in

•  Disadvantages: –  Fixed voice –  Quality can be very bad in places

•  HCI problem: mix of very good and very bad is quite annoying –  Large footprint, itis computationally expensive –  Can’t synthesize everything you want:

•  Diphone technique can move emphasis •  Unit selection gives good (but possibly incorrect) result


From Unit selection to HMM synthesis

•  Problems with Unit Selection Synthesis –  Discontinuities: Can’t modify signal –  Hit or miss: database often doesn’t have exactly what you

want –  Fixed voice

•  Solution: HMM (Hidden Markov Model) Synthesis –  Stable, Smooth and easy to create multiple voices –  Sounds unnatural to researchers, but naïve subjects prefer it Example: Nina as unit selection and HMM synthesis voice



HMM Synthesis

Slide from Tokuda J. Gustafson, CTT, KTH 60

HMMs in synthesis


11


Relationship between unit selection and HMM synthesis

Slide from Tokuda J. Gustafson, CTT, KTH 62

Relationship between unit selection and HMM synthesis 2

Slide from Tokuda


The training part •  The training is automatic. You need:

–  The text + recordings of about 1000 sentences

•  The training of 1000 sentences –  takes 24 hours and generates a voice of less than 1 MB

•  Separate HMMs for: Spectrum, F0, Duration

•  Training in two steps: 1. Context independent models 2. Use these models to create context dependent models.

•  Clustering: –  Storing all contexts requires much space –  It may be difficult to find alternatives for missing models –  Many models are very similar = redundancy


Examples of features in HMM synthesis training

•  Segment features: –  immediate context –  position in syllable

•  Syllable features –  Stress and lexical accent type –  position in word and phrase

•  Word features –  number of syllables –  position in phrase –  morphological feature (compound or not) –  part-of-speech tag (content or function word)

•  Phrase features –  phrase length in terms of syllables and words

•  Utterance features: –  length in syllables, words and phrases

•  Speaker –  Dialect, speaking style, emotional state


Clustering •  Groups a large database into clusters •  Three decision trees: Duration, F0 and Spectrum •  Division based on yes/no questions

–  Grouping acoustic similar phonemes –  Features. –  Context.


Speaker adaptation

http://homepages.inf.ed.ac.uk/jyamagis/Demo-html/map-new.html


12


Dalarna

Norrland

Skåne

Gotland

Svealand

Götaland

vart tar universum slut? centerpar3ster och kristdemokrater menar dock a7 brudparet kan slippa betala det kan bådas föräldrar göra gula sidorna finns från och med i fredags sökbart via mobiltelefonen om man 3ll exempel tar en telefon och frågar hur den fungerar så svarar vetenskapsmannen med a7 ly@a på luren och slå numret

Exempel från Simulekt-projektet


Use of HMM synthesis

•  Various voices: –  Speaker adaptation –  Speaker interpolation

•  Security of speaker identification systems •  Very low bit rate speech coder •  Small footprint, for use in mobile phones and web

browsers –  E.g. in flash: http://www.furui.cs.titech.ac.jp/~dixonp/hts/index.html


Lab 2 synthesis Two parts: •  Manual diphone synthesis

–  Tool: Wavesurfer (download from sourceforge) –  Do it at home, or with the lab computers

•  Resynthesis experiments –  Tool: Tandem STRAIGHT Vocoder –  You will get email with link to software (PC, MAC, Linux,

but you will have to install MATLAB runtime) –  The assignments will be sent next week, together with

times to come and do/present it with the lab computers

talsyntes: joakim gustafson tal, musik och hörsel - · pdf filetalsyntes: joakim...

Documents