Modeling the Effect of packet Loss on Speech Quality: GP Based Symbolic Regression
-
Upload
adil-raja -
Category
Engineering
-
view
148 -
download
0
Transcript of Modeling the Effect of packet Loss on Speech Quality: GP Based Symbolic Regression
Wireless Access Research
Adil Raja June 2006
Modeling the Effect of Packet Loss on Speech Quality: GP Based Symbolic Regression
Wireless Access Research
Adil Raja June 2006
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 {A. D. Clark} Regression Using Artificial Neural Networks.
{L. F. Sun et. al. and S. Mohammed et. Al}
Wireless Access Research
Adil Raja June 2006
Speech Based Approaches
• Intrusive - ITU-T Recommendation P.862 (PESQ).
• Non-intrusive – ITU-T Recommendation P.563 (PSEAM).
• Non-intrusive PESQ – A. E. Conway.
Wireless Access Research
Adil Raja June 2006
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
Wireless Access Research
Adil Raja June 2006
The Gilbert Elliot Model
0
1
1
mmpn
i
i
p
1-q
No Loss State
1-p X=1 X=0q
Loss State
1
1
1
2
)1(1n
i
i
n
i
i imimq
Parameters of Geometrically distributed burst/gap lengths
Mean Burst length = 1/q
Variance of Burst Length Distribution = (1-q)/q2
Mean Gap Length = 1/p
Variance of Gap Length = (1-p)/p2
qp
p
1
Wireless Access Research
Adil Raja June 2006
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.
Wireless Access Research
Adil Raja June 2006
Current Approach
• Codecs – G.729 and G.723.1 and AMR-NB• Packet/frame loss simulation – Gilbert Model• Mean Loss rate ulp and clp were varied between 0-0.85 and 0-0.90
respectively.• Input Variables
mean loss rate, mean and variance of burst length distribution (VAD), mean and variance of gap length distribution (VAD), codec type and packetization interval.
VAD – Different packets have different importance {L. F. Sun | C. Hoene}.• Genetic Programming is used for Symbolic Regression.• Hyperbolic Tangent (Activation).• A total of 659 speech files out of which 35% were used for training, 15%
were used for validation and 50% were used for Speaker independent testing.
• Speech activity – 70-80%.
Wireless Access Research
Adil Raja June 2006
Genetic Programming (GP)
• GP is a Machine Learning Technique inspired by biological evolution.
Wireless Access Research
Adil Raja June 2006
Fitness: 0.0327 Test Fitness: 0.0437, r= 0.9748, r=0.9635 (tree5)
Wireless Access Research
Adil Raja June 2006
Training (t5)
Wireless Access Research
Adil Raja June 2006
Validation (t5)
Wireless Access Research
Adil Raja June 2006
Fitness: 0.0342, Test Fitness: 0.0424, r=0.9737, r=0.9642
Wireless Access Research
Adil Raja June 2006
Fitness=0.0463; Test Fitness=0.0528 r=0.9642, r=0.9566
F(X)=mysqrt(times(8,sin(X3)));
GP-MOS-LQO= -1.2442 * F(X) + 3.7511
Wireless Access Research
Adil Raja June 2006
Training tn(28)
Wireless Access Research
Adil Raja June 2006
Validation tn(28)
Wireless Access Research
Adil Raja June 2006
Fitness=0.0466; Validation Fitness=0.0527; Testing Fitness=0.0401 r=0.9640, 0.9567, r=0.9696.
F(X)=minus(X3,times(5,kozasqrt(X3)));
GP-MOS-LQO= 0.8325*F(X) + 3.8482
Wireless Access Research
Adil Raja June 2006
Training tn(25)
Wireless Access Research
Adil Raja June 2006
Validation tn(25)
Wireless Access Research
Adil Raja June 2006
Testing tn(25)