Predicting the Perceptual Service Quality. Using a Trace of VoIP Packets ... VoIP transmission of a telephone call ... A VoIP phone can implement any playout scheduler. How to know .... Verified software in various research projects,. e.g. voice ...
Predicting the Perceptual Service Quality Using a Trace of VoIP Packets C. Hoene, S. Wiethölter, A. Wolisz September 29th, 2004. Qofis’04, Barcelona, Spain
Technische Universität Berlin
URL: www.tkn.tu-berlin.de
Content
Introduction/Background Our Approach Playout Scheduling Listening-only Tests
Delay Spikes Non-Random Packet Losses
Conclusions
VoIP transmission of a telephone call
How to Judge the Quality of a Call? Measure the quality of a VoIP system: Gather packet loss rates, mean delay, etc.
In the end the service quality is important
Human based listening tests are extensive
ITU P.862 (PESQ algorithm) measures speech quality
Easy but inaccurate
Compares original sample with the transmitted version calculates Mean Option Score (MOS) (1=bad, 5=excellent)
ITU G.107 (E-Model) predicts quality of tel.-system
Considers echo, loudness, coding, packet loss rate, delay, … Result: R Factor (0=bad, 70=toll quality, 100=excellent)
Limits of PESQ and E-Model PESQ and E-Model cannot judge the impact of: Impact of playout scheduling cannot be predicted. Variable packet delay Ö adaptive playout time Non-random packet losses Not applicable to assess VoIP packet traces resulting from simulations or experiments
Predicting Playout Scheduling
A VoIP phone can implement any playout scheduler. How to know which is going to be used? Actually, you cannot know it. But you know the most common, published playout scheduler schemes, which are
Fixed-deadline and adaptive from Van Jacobsen, Mills, Schulzrinne, Ramjee, and Moon
Calculate packets' playout times and the mean transmission delay.
Consider only speech frames during voice activity because during silence a human cannot identify the delay.
Ö We can predict the behavior of most phones !
Assessing a Trace of VoIP Packets Our solution: Using the most common different playout schedulers and encoding schemes combine of E-Model and PESQ (MOS to R formula approved by the ITU)
PESQ calculate speech quality E-Model combines MOS rating and transmission delays
Speech Quality and Transmission Delay
Speech Recordings
Audio
Audio
PESQ Reference Degraded
Degraded Audio
Speech Coding
Experiment or Simulation
Playout Buffer
(coding distortion)
(loss or delay)
(loss or scheduling)
Decoder PLC Mean Delay
E-Model
MOS Ie Id
Result: R Factor
Equations – referring to the SPECTS’04 paper…
Content
Introduction/Background Our Approach Playout Scheduling Listening-only Tests
Delay Spikes Non-Random Packet Losses
Conclusions
PESQ not verified Öconduct listening tests
Can PESQ measure playout rescheduling? or non-random packet losses? PESQ can measure variable playout but its performance has never been verified. It has been tested only for random/bursty but not for non-random losses. Conduct formal listening-only tests!
Test Sample Design
Construct packet traces with delay spikes! Delay spikes with a height of 50 to 300ms and a width of height+10%. 1 Spike in one sample
What to do? 1. Drop late packets 2. Or delay playout?
Playout Decisions 1
2
3
4
5
6
2 2+ 2++ 3
4
5
Legend: 1...6 Frame seq. no. 2+; 2++ Conceal ed frame
Ori gi nal
1
2 2+ 2++ 5
a) Packet loss
1. 2. 3.
6
1
b) Positive adapt ation
6
1
2
5
6
c) Negative adapt ati on
DROP: Drop late packets (a) ADAPT: Delay playout to late packets (b) ADAPT&FALLBACK: Delay playout (b) and return to normal during the next silence period (c)
Listening-Only Tests Follow ITU P.830 but Multiple person at the same time Use native and foreign speakers Use linear MOS scale (no discrete scale!) Sample Sequence: For each sentence 1. 2. 3.
Original at the beginning, MNRU with 5dB (worst),15,25,35 (best) Randomly distributed: packet loss samples
Produced on CD containing all the samples
Studio...
Listening-Test and Sample Design Different encoding schemes:
G.711, G729
Produced delay spikes (only during voice activity)
Ö
with length 50,100,150,200,250,300ms with width=length*1.1
Different samples with different length 5-10s), German (using Kiel Corpus Vol. I) 220 samples (including MNRU samples as a reference) reference
Frame Analysis
speech properties
PESQ
PESQ MOS
R X
seed
packetization
Loss Generator loss rate
rate/mode
Coding algorithm
sample
speaker
language
Speech Recordings
importance
Decoding PLC
Listeningonly Tests
MOS
(Correlation)
Rating Performance
Three parts: intro, first and second half. First is best. Native speaker and foreign speaker rate as good. Some persons did not kept track with the sample no.
Human LQS-MOS vs. PESQ LQO-MOS
MOS (PESQ)
4
3
2 y = 0,9403x - 0,2139 R= 0,8316 1 1
2
3 MOS (Humans)
4
MNRU vs. MOS: Correlation: R=0.977 4
MOS
3
2
1 5
15
25
35
45
MNRU MOS
PESQ MOS
MOS cited
scaled MOS
Results for G.711 coding
MOS Variance vs. Prediction Performance Correlation between PESQ and humans
1,0 0,9 0,8
trend line: R=0.768
0,7 0,6 0,5 0,4 0,3 0,2 0
0,2 0,4 0,6 0,8 1 PESQ sample variance for different impairment patterns
Content
Introduction/Background Our Approach Playout Scheduling Listening-only Tests
Delay Spikes Non-Random Packet Losses
Conclusions
Introduction
Losing one Voice-Over-IP packet impairs the perceptual quality in a wide range, depending on
the frame speech properties the encoder/decoder/concealment algorithms decoders resynchronization time after loss (especially low-rate decoders might maintain a wrong state after loss for the following frames.) the surrounding speech.
Example: Discontinuous Transmission (DTX)
Speech frames during silence are less important Lower frame rate during silence
Packet Importance
The packet’s importance is the quality degradation that its loss would cause.
Definition: The importance of frame losses is the difference between the speech quality due to coding loss and the quality due to coding loss and frame losses, times the length of the analyzed sample. Conducted some million PESQ ratings Dropped only one VoIP packet in a sample
Just try it: www.tkn.tu-berlin.de/research/mongolia Select predefined parameter sets Compression?
chose sample Listen to it! amount of loss Choose an another loss pattern
Drop only frames with an importance between min. and max.
Sample statistics
Talking or silence? Voiced or unvoiced?
Judge the speech quality by your self!
Conclusion
Presented an approach to assess VoIP packet traces Combined PESQ, E-Model and playout scheduling because
Verified PESQ for
Speech frames differ in their importance Playout scheduler are not standardized Playout rescheduling harms speech quality. non-random packet losses (R=0.94) and delay spikes (R=0.87).
Precision performances depend on the variance of the samples’ speech quality.
Assessing VoIP: Summary
Open-source software including
Implementation of all common playout schedulers G.711 and G.729 coding Sample database including human rating results
Verified software in various research projects,
e.g. voice over WLAN, impact of handover, ad-hoc. http://www.tkn.tu-berlin.de/reseach/qofis C. Hoene, S. Wiethölter, and A. Wolisz, "Predicting the Perceptual Service Quality Using a Trace of VoIP Packets", In Proceedings of QofIS’04, Barcelona, Spain, September 2004. C. Hoene, H. Karl, and A. Wolisz, "A Perceptual Quality Model for Adaptive VoIP Applications", In Proceedings of International Symposium on Performance Evaluation of Computer and Telecommunication Systems (SPECTS'04), San Jose, California, USA, July 2004, Paper won the Best Paper Award of the conference. C. Hoene and E. Dulamsuren-Lalla, "Predicting Performance of PESQ in Case of Single Frame Losses", In Proc. MESAQIN 2004, Prague, CZ, June 2004. S. Möller and C. Hoene, "Information About a New Method For Deriving the Transmission Rating Factor R From MOS in Closed Form", ITU, May 2002, Temporary Document for the study group 12.