Modeling the Effect of Packet Loss on Speech Quality: Genetic Programming Based Symbolic Regression
1 like418 views
The document discusses modeling the impact of packet loss on speech quality using genetic programming and symbolic regression techniques. It outlines previous research on packet loss, introduces the genetic programming framework, and details simulation parameters and results. The findings suggest that genetic programming may offer advantages over traditional machine learning methods for improving speech quality metrics in VoIP traffic.
![3
Packet Loss Modeling Approaches
Packet Based Approaches.
• Based on regression of packet loss parameters to MOS.
• Parameters include mean Loss rate, conditional loss probability etc.
• Some approaches include:
Markov Models [1].
Regression Using Artificial Neural Networks. [2] [3] [4]
Speech Based Approaches
• Intrusive: ITU-T Recommendation P.862 (PESQ).
• Non-intrusive: ITU-T Recommendation P.563 (PSEAM).
• Non-intrusive PESQ [5].
Adil Raja, Wireless Access research Centre University of Limerick, June 2006 3](https://image.slidesharecdn.com/dyvpxokpsnflsmxvc9aq-signature-a3f5497d26e8d92d92e05cae9e5f0706964255dfb7718b8fc6aadfe56ea06143-poli-151104144808-lva1-app6891/85/Modeling-the-Effect-of-Packet-Loss-on-Speech-Quality-Genetic-Programming-Based-Symbolic-Regression-3-320.jpg)
![7
Current Approach and Experimental Details
• Genetic Programming has been used for mapping the effect of VoIP traffic parameters on speech
quality.
• Codecs: G.729 and G.723.1 and AMR-NB.
• Packet/frame loss simulation is done using Gilbert Model.
• Mean Loss rate (ulp) was set to 10, 20, 30 and 40 %. clp was set to 10, 60, 80 and 90 %.
• Input Variables.
Mean loss rate.
Means and variances of burst and gap length distributions (VAD).
Codec type and packetization interval.
VAD: Different packets have different importance [6].
• From a total of 1408 speech files:
35% were used for training.
15% were used for validation.
50% were used for Speaker independent testing.
• Speech activity - 70-80%
Adil Raja, Wireless Access research Centre University of Limerick, June 2006 7](https://image.slidesharecdn.com/dyvpxokpsnflsmxvc9aq-signature-a3f5497d26e8d92d92e05cae9e5f0706964255dfb7718b8fc6aadfe56ea06143-poli-151104144808-lva1-app6891/85/Modeling-the-Effect-of-Packet-Loss-on-Speech-Quality-Genetic-Programming-Based-Symbolic-Regression-7-320.jpg)
![13
The GP Environment
GPLAB - A Matlab tool-box by Sara Silva (sara@dei.uc.pt).
Other GP Parameters
• Survival: Replacement (), Elitism.
• Adaptive genetic operator probabilities.
• Initial Population Size: 100.
• Generational Gap: 1.
Linear Scaling [7]
MSE (y, t) = 1/n
n
i
(ti − yi)
2
MSEs (y, t) = MSE (a + by, t) = 1/n
n
i
(ti − (a + byi))
2
a = ¯t − b ¯f(x)
b =
cov(t, y)
var(y)
Adil Raja, Wireless Access research Centre University of Limerick, June 2006 13](https://image.slidesharecdn.com/dyvpxokpsnflsmxvc9aq-signature-a3f5497d26e8d92d92e05cae9e5f0706964255dfb7718b8fc6aadfe56ea06143-poli-151104144808-lva1-app6891/85/Modeling-the-Effect-of-Packet-Loss-on-Speech-Quality-Genetic-Programming-Based-Symbolic-Regression-13-320.jpg)
![28
References
[1] A. D. Clark. Modeling the effects of burst packet loss and recency on subjective voice quality.
In 2nd IP-Telephony Workshop, Columbia University, New York, April 2001.
[2] S. Mohamed, F. Cervantes-Perez, and H. Afifi. Integrating networks measurements and speech
quality subjective scores for control purposes. In Annual Joint Conference of the IEEE Computer
and Communications Societies (INFOCOM), pages 641Ű649, 2001.
[3] S. Mohamed, G. Rubino, and M. Varela. A method for quantitative evaluation of audio quality
over packet networks and its comparison with existing techniques. In Measurement of Speech and
Audio Quality in Networks (MESAQIN), 2004.
[4] L. F. Sun and E. C. Ifeachor. Perceived speech quality prediction for voice over ip-based networks.
In IEEE International Conference on Communications (ICC), volume 4, pages 2573 -Ű 2577, 2002.
[5] A.E. Conway, Output-based method of applying PESQ to measure the perceptual quality of
framed speech signals, IEEE Communications Society, 2004.
[6] L. Sun, G.Wade, B. M. Lines, and E. C. Ifeachor. Impact of packet loss location on perceived
speech quality. In 2nd IP-Telephony Workshop, Columbia University, New York, April 2001.
[7] M. Kaijzer. Scaled Symbolic Regression. 2003.
Adil Raja, Wireless Access research Centre University of Limerick, June 2006 28](https://image.slidesharecdn.com/dyvpxokpsnflsmxvc9aq-signature-a3f5497d26e8d92d92e05cae9e5f0706964255dfb7718b8fc6aadfe56ea06143-poli-151104144808-lva1-app6891/85/Modeling-the-Effect-of-Packet-Loss-on-Speech-Quality-Genetic-Programming-Based-Symbolic-Regression-28-320.jpg)
Modeling the Effect of Packet Loss on Speech Quality: Genetic Programming Based Symbolic Regression
- 1. 1 Modeling the Effect of Packet Loss on Speech Quality: Genetic Programming Based Symbolic Regression Adil Raja adil.raja@ul.ie 1
- 2. 2 Contents • Packet Loss Modeling approaches: Related Research • Current Approach. • A Brief Intoduction to Genetic Programming (GP). • Simulation Environment and GP Parameters. • Analysis and Results. • Conclusion and Future Aspirations. 2
- 3. 3 Packet Loss Modeling Approaches Packet Based Approaches. • Based on regression of packet loss parameters to MOS. • Parameters include mean Loss rate, conditional loss probability etc. • Some approaches include: Markov Models [1]. Regression Using Artificial Neural Networks. [2] [3] [4] Speech Based Approaches • Intrusive: ITU-T Recommendation P.862 (PESQ). • Non-intrusive: ITU-T Recommendation P.563 (PSEAM). • Non-intrusive PESQ [5]. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 3
- 4. 4 Our Previous Work ANN based Regression of network traffic metrics on speech quality. • Useful Network loss Metrics. • Mean Loss Rate. • Means and Variances of Burst and Gap Length Distributions. • Codec Type and Packetization Interval. • Inter Loss Distance/Gap Length. Packet loss was modeled using a Gilbert Model. Results: - • rtraining=0.9835; • rvalidation=0.9821; • rtesting=0.9763 Adil Raja, Wireless Access research Centre University of Limerick, June 2006 4
- 5. 5 Packet Loss Simulation The Gilbert-Elliot Loss Model. p = n−1 i=1 mi/m0 /mi q = 1 − n−1 i=2 mi × (n − 1) / n−1 i=1 mi × i π1 = p p + q Adil Raja, Wireless Access research Centre University of Limerick, June 2006 5
- 6. 6 Packet Loss Simulation ... Parameters of Geometrically Distributed Burst and Gap Lengths • Mean Burst Length = 1/q • Variance of Burst Length Distribution = (1-q)/q2 • Mean Gap Length = 1/p • Variance of Gap Length Distribution = (1-p)/p2 The Gilbert Model: • Packet loss can be simulated for certain values of p and q. • During network operation bursts have to be captured for determining clp and ulp. • The Gilbert model also models the packet loss due to jitter buffer discard/overflow. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 6
- 7. 7 Current Approach and Experimental Details • Genetic Programming has been used for mapping the effect of VoIP traffic parameters on speech quality. • Codecs: G.729 and G.723.1 and AMR-NB. • Packet/frame loss simulation is done using Gilbert Model. • Mean Loss rate (ulp) was set to 10, 20, 30 and 40 %. clp was set to 10, 60, 80 and 90 %. • Input Variables. Mean loss rate. Means and variances of burst and gap length distributions (VAD). Codec type and packetization interval. VAD: Different packets have different importance [6]. • From a total of 1408 speech files: 35% were used for training. 15% were used for validation. 50% were used for Speaker independent testing. • Speech activity - 70-80% Adil Raja, Wireless Access research Centre University of Limerick, June 2006 7
- 8. 8 A Brief Introduction to Genetic Programming (GP) • GP is a Machine Learning Technique inspired by biological evolution. A branch of Evolutionary Algorithms. • Aimed at evolving program expressions/computer code. • Each individual encodes a symbolic expression. • Solution Representation. A tree structure is the most popular representation. Other representations include graphs and linear structures such as arrays. • Primary application area is modelling. Commercial Application - predicting stock index. Scientific Application - modelling physical processes. Engineering Application - reverse engineering, designing circuitry, regression, classifica- tion. Data Mining. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 8
- 9. 9 A simplified GP Breeding Cycle GP uses four steps to solve problems: • Generate an initial population of random compositions of the functions and terminals of the problem (computer programs). Functions: plus, minus, times, sin, cos, mylog, mypower, divide, sqrt, mylog2, mylog10. Terminals: Can be variables (network traffic parameters) and constants. • Execute each program in the population and assign it a fitness value according to how well it solves the problem. Minimization of χ2 error. Minimization of MSE (). Maximization of Pearson’s product moment correlation coefficient. • Copy the best existing programs (Selection). Roullete Wheel Selection - Fitness Proportinate Selection. Tournament Selection. Lexicographic Parsimony Pressure (). ... • Create new computer programs by mutation and crossover. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 9
- 10. 10 A simplified GP Breeding Cycle: A Symbolic Representation Adil Raja, Wireless Access research Centre University of Limerick, June 2006 10
- 11. 11 Mutation Two types of mutation are possible1 : • A terminal replaces a terminal or a function replaces a function. • A subtree can replace an entire subtree. 1 http://www.geneticprogramming.com/Tutorial/#anchor181526 Adil Raja, Wireless Access research Centre University of Limerick, June 2006 11
- 12. 12 Crossover Two solutions are recombined to form two new solutions or offspring. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 12
- 13. 13 The GP Environment GPLAB - A Matlab tool-box by Sara Silva (sara@dei.uc.pt). Other GP Parameters • Survival: Replacement (), Elitism. • Adaptive genetic operator probabilities. • Initial Population Size: 100. • Generational Gap: 1. Linear Scaling [7] MSE (y, t) = 1/n n i (ti − yi) 2 MSEs (y, t) = MSE (a + by, t) = 1/n n i (ti − (a + byi)) 2 a = ¯t − b ¯f(x) b = cov(t, y) var(y) Adil Raja, Wireless Access research Centre University of Limerick, June 2006 13
- 14. 14 Advantages of Linear Scaling • Bloat Control. • Faster Training. • Solutions better suited for real-time evaluations. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 14
- 15. 15 Analysis and Results A total of 50 runs were performed. Each run was spanned over 50 gener- ations. Fitness Curves: Adil Raja, Wireless Access research Centre University of Limerick, June 2006 15
- 16. 16 Analysis and Results ... Diversity and Tree Size plots Adil Raja, Wireless Access research Centre University of Limerick, June 2006 16
- 17. 17 Analysis and Results ... Parameter Representation in GP Environment GP Representation Parameter X1 codec type X2 Packetization Interval (PI) X3 Mean Loss Rate (mlr) X4 Mean Burst Length (mbl) X5 Mean Gap Length (mgl) X6 Variance of Burst Length (vbl) X7 Variance of gap length (vgl) • Talkspurt based values of all the parameters are used. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 17
- 18. 18 Analysis and Results ... • Fitness= 0.0523; Test Fitness=0.0496 rtraining=0.9074;rvalidation=0.9183; Adil Raja, Wireless Access research Centre University of Limerick, June 2006 18
- 19. 19 Analysis and Results ... Adil Raja, Wireless Access research Centre University of Limerick, June 2006 19
- 20. 20 Analysis and Results • A less bloated solution. • Fitness= 0.0670; Test Fitness=0.0555 rtraining=0.8796rvalidation= 0.9079; GP − MOS − LQO = −2.3843 sin √ X3 + sin (X3) + 3.6112 Adil Raja, Wireless Access research Centre University of Limerick, June 2006 20
- 21. 21 Analysis and Results ... Adil Raja, Wireless Access research Centre University of Limerick, June 2006 21
- 22. 22 Analysis and Results ... • Fitness= 0.0678; Validation Fitness=0.0556; Test Fitness=0.0650; rtraining=0.8780rvalidation= 0.9074; rtesting=0.8881. GP − MOS − LQO = −3.3432 √ X3 + 3.6881 Adil Raja, Wireless Access research Centre University of Limerick, June 2006 22
- 23. 23 Analysis and Results ... Adil Raja, Wireless Access research Centre University of Limerick, June 2006 23
- 24. 24 Analysis and Results ... • Fitness= 0.0678; Validation Fitness=0.0556; Test Fitness=0.0650; rtraining=0.8780rvalidation= 0.9074; rtesting=0.8881. GP − MOS − LQO = −10.0296 X3/9 + 3.6881 Adil Raja, Wireless Access research Centre University of Limerick, June 2006 24
- 25. 25 Analysis and Results ... Adil Raja, Wireless Access research Centre University of Limerick, June 2006 25
- 26. 26 Conclusions and Future Aspirations • Results obtained by GP have advantages over other machine learning algorithms (such as ANNs). Simplified results: A mathematical expression. GP searches for global minimum of the error function and is less prone to getting stuck in local minima (due to mutation property). The prelimanary results are not comparable to ANN based approaches but there is room for improvement. • Improvements: Some Speculations; The Known Knowns. Population size should be increased to 500 for the sake of having more diverse search space. Selection: Lexicographic parsimony pressure vs Tournament selection. Survival: Replacement vs some elitism criterion. • The work can be split to two parts: The Telecommunications intensive aspects. The GP intensive aspects. • Developing an in-depth understanding of GP shall be a part of future endeavors. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 26
- 27. 27 Thanks! Adil Raja, Wireless Access research Centre University of Limerick, June 2006 27
- 28. 28 References [1] A. D. Clark. Modeling the effects of burst packet loss and recency on subjective voice quality. In 2nd IP-Telephony Workshop, Columbia University, New York, April 2001. [2] S. Mohamed, F. Cervantes-Perez, and H. Afifi. Integrating networks measurements and speech quality subjective scores for control purposes. In Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), pages 641Ű649, 2001. [3] S. Mohamed, G. Rubino, and M. Varela. A method for quantitative evaluation of audio quality over packet networks and its comparison with existing techniques. In Measurement of Speech and Audio Quality in Networks (MESAQIN), 2004. [4] L. F. Sun and E. C. Ifeachor. Perceived speech quality prediction for voice over ip-based networks. In IEEE International Conference on Communications (ICC), volume 4, pages 2573 -Ű 2577, 2002. [5] A.E. Conway, Output-based method of applying PESQ to measure the perceptual quality of framed speech signals, IEEE Communications Society, 2004. [6] L. Sun, G.Wade, B. M. Lines, and E. C. Ifeachor. Impact of packet loss location on perceived speech quality. In 2nd IP-Telephony Workshop, Columbia University, New York, April 2001. [7] M. Kaijzer. Scaled Symbolic Regression. 2003. Adil Raja, Wireless Access research Centre University of Limerick, June 2006 28