a methodology for deriving voip equipment impairment factors … · 2008-07-20 · a methodology...
TRANSCRIPT
A Methodology for Deriving VoIP EquipmentImpairment Factors for a mixed NB/WB Context
Adil Raja1 R. Muhammad Atif Azad2 Colin Flanagan1
Conor Ryan2
1Wireless Access Research CentreDepartment of Electronic and Computer Engineering
2Bio-Computing and Developmental SystemsDepartment of Computer Science and Information Sysmtems
University of Limerick, Ireland
5th Annual "Humies" Awards for Human-Competitive ResultsMonday, 14th July, 2008
Outline
1 Background and Motivation
2 Our Goal
3 Our Approach
4 Our Results
5 Our Achievements
Background
VoIP enablespacket-basedtelecommunication.
IP NETWORK Packetizer Encoder Depacketizer Decoder
It is cheap and cost effective.
SupportsWB telephony conveniently.
QoS provisioning remains a problem.
Speech quality estimation is important.
Background
VoIP enablespacket-basedtelecommunication.
IP NETWORK Packetizer Encoder Depacketizer Decoder
It is cheap and cost effective.
SupportsWB telephony conveniently.
QoS provisioning remains a problem.
Speech quality estimation is important.
BackgroundSpeech Quality
Speech Quality Measurement
Subjective Methods Instrumental Methods
Figure:Various categories of speech quality assessment methods.
BackgroundSpeech Quality
Speech Quality Measurement
Subjective Methods Instrumental Methods
Intrusive Non-Intrusive
Signal-based
Figure:Various categories of speech quality assessment methods.
BackgroundSpeech Quality
Speech Quality Measurement
Subjective Methods Instrumental Methods
Intrusive Non-Intrusive
Signal-based Parametric
Figure:Various categories of speech quality assessment methods.
BackgroundSpeech Quality Standardization
International Telecommunications Union –ITU-T – deals withthe standards.
Intrusive: ITU-T P.862.* – NB-PESQ andWB-PESQ.Nonintrusive:
1 Parametric: ITU-T G.107 – TheE-Model.2 Signal-based: ITU-T P.563.
Our GoalEquipment Impairment Factor
To derive equipment impairment factors forWB VoIP as described bytheE-Model:
SpeechQuality = f (Impairments) (1)
Ie,WB,eff = f (codec, loss, ...) (2)
Our Approach
To evolveIe,WB,eff using GP.
Table:Input domain parameters used in evolutionary modeling
No. Input Parameter Description1 Ie,WB Impairments due to codecs2 mlr mean loss rate3 PI packetization interval (ms)4 mbl mean burst length5 grad Coarse estimate of codec specific loss
robustness factor
Our ApproachSimulation Environment
Speech Database
Decoder Loss
Simulator Encoder
Reference speech signal
Figure:Simulation system for derivation ofIe,WB,eff
Our ApproachSimulation Environment
Speech Database
Paremeter Extraction
Decoder Loss
Simulator Encoder
Reference speech signal
Figure:Simulation system for derivation ofIe,WB,eff
Our ApproachSimulation Environment
Speech Database
Paremeter Extraction
Decoder Loss
Simulator Encoder
Reference speech signal
Degraded speech
I e,WB,eff WB-PESQ
Figure:Simulation system for derivation ofIe,WB,eff
Our ApproachSimulation Environment
Speech Database
Model Evolved By GP
Paremeter Extraction
Decoder Loss
Simulator Encoder
Reference speech signal
Evolved I e,WB,eff
Degraded speech
Target I e,WB,eff
VoIP Traffic parameters
I e,WB,eff WB-PESQ
Figure:Simulation system for derivation ofIe,WB,eff
Our ResultsThe Derived Models
Ie,WB,eff = (3)˘
11− mbl + ln(grad) + grad × mlr + Ie,WB
−2.log2(PI)} × 0.8619+ 9
Ie,WB,eff = (4){
ln
(
9× (Ie,WB + mlr × grad2)
mbl5 − mlr
)
+ mlr + Ie,WB
+grad × mlr} × 0.8303+ 8.9977
Ie,WB,eff = (5)`
log10(log10(log2(Ie,WB − 2 × mbl) + mlr))´
×321.7017+ 95.3708
Our ResultsComparison with E-Model
Table:Prediction gain over E-Model in various network operating conditions
Loss Type Train TestGeneral (Bursty) 14.54 16.36
Specific (Random) 36.72 35.89
Network conditions:
Fulfilled by ITU-T G.1050.
11,280 input/output patterns.
training data – 70%.
test data – 30%.
Our ResultsComparison with E-Model
Table:Prediction gain over E-Model in various network operating conditions
Loss Type Train TestGeneral (Bursty) 14.54 16.36
Specific (Random) 36.72 35.89
Network conditions:
Fulfilled by ITU-T G.1050.
11,280 input/output patterns.
training data – 70%.
test data – 30%.
Our ResultsComparison with E-Model
Table:Prediction gain over E-Model in various network operating conditions
Loss Type Train TestGeneral (Bursty) 14.54 16.36
Specific (Random) 36.72 35.89
Network conditions:
Fulfilled by ITU-T G.1050.
11,280 input/output patterns.
training data – 70%.
test data – 30%.
Our ResultsCodecs
Table:The following codecs were employed
Codec Type ModesITU-T G.722.1 WB 2ITU-T G.722.2 WB 9ITU-T G.729 NB 1ITU-T G.723.1 NB 1AMR-NB NB 2Total – 15
Our Achievements
Our model outperforms ITU-T recommendation G.107 i.e. TheE-Model– Criterion B.
Demonstrates superior performance for a wide range of ITU-Trecommended NB/WB codecs.
TheE-Modelis downloadable fromwww.itu.int . CriterionC.
Our Achievements
The results have been accepted for publication indomain specificIEEE Transactions on Multimedia. Criterion D.
Successful andparsimonioususe of a richer feature space.
Suitable for real-time evaluation.
Our Achievements
0 5 10 15 20 25 30 35 4020
30
40
50
60
70
80
90
100
110
120
mean loss rate (mlr) %
I e,W
B,e
ff
G.729 (8)G.723.1(6.3)AMR−NB (7.4)G.722.2 (19.85)G.722.1 (32)
Figure:Ie,WB,eff as a function ofmlr for various NB/WBcodecs. Past research proposed codec specific proposals.
Past approaches:
Logarithmic – codecspecific.
Quadratic – codecspecific.
The E-Model–unified, superior.
Our model:
Unified.
Superior to all thepast models.
Criterion E.
Our Achievements
0 5 10 15 20 25 30 35 4020
30
40
50
60
70
80
90
100
110
120
mean loss rate (mlr) %
I e,W
B,e
ff
G.729 (8)G.723.1(6.3)AMR−NB (7.4)G.722.2 (19.85)G.722.1 (32)
Figure:Ie,WB,eff as a function ofmlr for various NB/WBcodecs. Past research proposed codec specific proposals.
Past approaches:
Logarithmic – codecspecific.
Quadratic – codecspecific.
The E-Model–unified, superior.
Our model:
Unified.
Superior to all thepast models.
Criterion E.
Our Achievements
GP evolved model is better than a series of enhancements of theE-Model (G.107).Criterion F.
E-Model: used in transmission planning and live networkmonitoring.
Clients: ISPs, network operators, telcos.
Our Achievements
GP evolved model is better than a series of enhancements of theE-Model (G.107).Criterion F.
E-Model: used in transmission planning and live networkmonitoring.
Clients: ISPs, network operators, telcos.
Thank You