Transcript www.rtc-conference.com

Effect of FEC
mechanisms in the
Performance of Low Bit
Rate Codecs in Lossy
Mobile Environments
Rolando Herrero, PhD
Northeastern University, Boston, MA, U.S.A.
Martin Cadirola, Director/VP Business Development
Ecotronics Ventures LLC, Clarksburg, MD, U.S.A.
Background

Developers of Kapanga (Voice/Video/Fax SIP client)

Field work with OEM, QA and military environments

Collaboration with academia
© 2014 - IPT Communications Conference - September 29, 2014
Motivation
© 2014 - IPT Communications Conference - September 29, 2014
Motivation
Voyager 1
- Launched in 1977
- 130 AU from the Sun
- 14 hours to Tx/Rx
- 38,000 mph = 61,000 kph
© 2014 - IPT Communications Conference - September 29, 2014
Motivation
HOW?
© 2014 - IPT Communications Conference - September 29, 2014
FEC!
Mobile Networks
- Multipath loss
- Slow/Fast fading channels
- Applicable in RT over 3G/4G/LTE
Source: http://www.ice.rwth-aachen.de/
Use of Forward Error Correction (FEC) techniques
in media transmission are critical here too!
© 2014 - IPT Communications Conference - September 29, 2014
Motivation

Mobile networks

Interest in creating real-time QA tools

Room for innovation in improving performance
© 2014 - IPT Communications Conference - September 29, 2014
Motivation

Work based on Barton, Lemberg, Sarraf, Hamilton,
“Performance analysis of packet loss concealment in
mobile environments”, in Communications Quality and
Reliability (CQR), 2010 IEEE International Workshop
Technical Committee.

Network impairments in mobile networks are mostly
due to fading, especially in high multipath environments

The characteristics of fading channels are a well known
phenomenon and their statistics have been modeled
mathematically by a two-state Markov process
How do LBR codecs perform in such a mobile network?
How can we assess quality of speech?
© 2014 - IPT Communications Conference - September 29, 2014
Theoretical Model: Markov Process

2-state Markov Process models a fading channel mobile network
environment
•
Model has 2 states:
• High Packet Loss State (H)
• Low Packet Loss State (L)
•
Model has 5 parameters:
• Transition probability from low to high loss states
(p)
• Probability that the channel stays in a high loss
state (α)
• Distance between an RTP packet and the
transmission of the corresponding FEC RTP packet
in packet times (D)
• Packet loss probability when the channel is in the
low loss state (β)
• Packet loss probability when the channel is in the
high loss state (γ)
© 2014 - IPT Communications Conference - September 29, 2014
Experimental Framework

We implemented 2 kinds of FEC techniques in the UA

XOR-based


Repetition-based


Optimal for lower bandwidths
Best quality for higher bandwidths
Assumptions wrt Markov process: β = 0 and γ = 1

All RTP packets are dropped in the high loss state

No RTP packets are lost in the low loss state
© 2014 - IPT Communications Conference - September 29, 2014
A little FEC background
160 bytes
Speech Packet
Pn
160 bytes
Repetition-based FEC
XOR-based FEC
320 bytes
320 bytes
P1
P1
P2
P2
P3
160 bytes
160 bytes
160 bytes
160 bytes
160 bytes
P1
P2
P2-P1
P3
For each packet (on each case)
RTP Packet
RTP
Header
© 2014 - IPT Communications Conference - September 29, 2014
RTP Payload
…
P3-P2 …
Experimental Framework

Modify a SIP UA to implement FEC
mechanisms under study

Setup UA with Linux box running
Netem

Use ITU-based speech waveforms
*no tones nor signals*

Use of G.723.1, G.729A, AMR-NB,
AMR-WB, EVRC, SILK, Opus
© 2014 - IPT Communications Conference - September 29, 2014
Experimental Framework

Netem takes the transition probabilities of the two-state
Markov model, namely α and p, as input parameters

In order to validate different bursty scenarios the following
combination of parameters is considered; α = 0.01, 0.05, 0.1
and 0.15 and p = 0.25, 0.5 and 0.75

In order to minimize latency, in part due to the restrictions
imposed by playout buffers in regards to tolerable speech
quality, the packet distance between the transmitted RTP and
the FEC RTP is D = 2
© 2014 - IPT Communications Conference - September 29, 2014
Comparative Analysis
Bursty Packet Loss (Repetition-based FEC)
α
p
0.01
0.01
0.01
0.05
0.05
0.05
0.10
0.10
0.10
0.15
0.15
0.15
0.25
0.50
0.75
0.25
0.50
0.75
0.25
0.50
0.75
0.25
0.50
0.75
Success %
Theoretical Experimental
95.01
94.79
83.39
81.36
67.99
72.62
95.00
93.21
83.53
80.78
68.46
71.86
94.89
91.05
83.57
79.29
68.86
70.42
94.66
90.72
83.43
78.59
69.06
69.64
PESQ
G.723.1
3.91
3.46
2.87
3.83
3.38
2.81
3.75
3.32
2.73
3.68
3.24
2.62
G.729A
3.84
3.10
2.61
3.78
3.01
2.53
3.72
2.91
2.45
3.66
2.83
2.39
© 2014 - IPT Communications Conference - September 29, 2014
AMR-NB
3.89
3.17
2.43
3.81
3.09
2.35
3.74
3.01
2.29
3.67
2.94
2.23
EVRC
3.82
3.23
2.68
3.74
3.12
2.59
3.67
3.04
2.52
3.61
2.95
2.46
AMR-WB
3.75
3.29
2.70
3.68
3.20
2.61
3.61
3.11
2.54
3.52
3.04
2.47
PESQ-WB
SILK
3.66
3.12
2.64
3.55
3.07
2.56
3.47
2.93
2.48
3.41
2.83
2.41
Opus
3.65
3.27
2.71
3.59
3.18
2.64
3.51
3.09
2.53
3.44
3.01
2.45
Comparative Analysis
Bursty Packet Loss (XOR-based FEC)
α
p
0.01
0.01
0.01
0.05
0.05
0.05
0.10
0.10
0.10
0.15
0.15
0.15
0.25
0.50
0.75
0.25
0.50
0.75
0.25
0.50
0.75
0.25
0.50
0.75
Success %
Theoretical Experimental
95.01
94.79
83.39
81.36
67.99
72.62
95.00
93.21
83.53
80.78
68.46
71.86
94.89
91.05
83.57
79.29
68.86
70.42
94.66
90.72
83.43
78.59
69.06
69.64
PESQ
G.723.1
3.77
3.32
2.73
3.71
3.27
2.63
3.64
3.21
2.54
3.57
3.15
2.46
© 2014 - IPT Communications Conference - September 29, 2014
G.729A
3.69
2.97
2.44
3.63
2.90
2.84
3.54
2.82
2.77
3.48
2.75
2.71
AMR-NB
3.73
3.04
2.28
3.66
2.95
2.89
3.58
2.88
2.81
3.52
2.80
2.74
EVRC
3.66
3.07
2.52
3.60
3.02
2.93
3.55
2.96
2.86
3.49
2.91
2.81
AMR-WB
3.59
3.14
2.54
3.52
3.08
3.01
3.46
3.01
2.96
3.40
2.97
2.90
PESQ-WB
SILK
3.41
2.97
2.48
3.35
2.92
2.84
3.29
2.87
2.76
3.22
2.82
2.69
Opus
3.42
3.10
2.57
3.37
3.01
2.93
3.31
2.93
2.88
3.24
2.89
2.81
Comparative Analysis
Rate-Distortion (for a set of α and p)
© 2014 - IPT Communications Conference - September 29, 2014
Conclusions

We presented a framework to evaluate and compare the
experimental and theoretical performance of RTP FEC
applied to a set of narrowband and wideband codecs in a
lossy mobile environment

It can be verified that the experimental probability of success
of an endpoint receiving an RTP packet is within 15% of the
analytical value (error) obtained by modeling the system
with a two-state Markov process

Speech quality scores, based on PESQ and PESQ-WB
metrics, follow the probability of decoding success for
both when repetition FEC and XOR FEC schemes are used
© 2014 - IPT Communications Conference - September 29, 2014
Conclusions

When put in the context of rate-distortion plots narrowband
codec G.723.1 and wideband codec Opus provide better
performance by decoding same quality speech at lower
transmission rates

Although the probability of success (from the theoretical
model) does not take into account the codec type nor the
characteristics of the playout mechanism it provides a good
approximation to estimate the overall quality of the
communication

When comparing XOR-based vs Repetition-based FEC it is
clear that the latter gives better media quality at the cost of
significantly increasing the transmission rate

FEC and other error correcting techniques offer
improvements to the quality of *any* real-time media
application: RTC, any rtp-based communications
© 2014 - IPT Communications Conference - September 29, 2014
Next Steps

Challenges (as any new technology) -> standarization,
unless everybody uses one kind of UA

Continue research on new FEC mechanisms
(TurboCodes, Reed Salomon)

Utilization of these techniques in:

UAV communications for civilian and military fields

LTE deployments overseas

Space applications
Drones/UAV/UAS
Earth Science
© 2014 - IPT Communications Conference - September 29, 2014
Thank you! Q & A!
© 2014 - IPT Communications Conference - September 29, 2014