Juan J. Ramos-Mu

1
Juan J. Ramos-Muñoz, Angel M. Gómez,Juan M.
Lopez-Soler
Intelligibility Evaluation of a VoIP Multi-flow
Block Interleaver
2
Outline

• Introduction
• Basic and Multi-flow Block Interleavers
  Algorithms
• Intelligibility Estimation
• Experimental Results
• Conclusion

3
Introduction

• Background
• Bursts degrade the perceived quality in Voice
  over IP
• VoIP services are extremely delay concerned
• Enhanced VoIP services must be subjectively
  evaluated

4
Introduction

• In this work
• We contribute to demonstrate some perceptual
  benefits that can be obtained by using active
  routers in VoIP
• We propose a new delay-aware interleaver to
  mitigate the bursty-error-prone nature of IP
• We evaluate the service performance by using
  Automatic Speech Recognition

5
Basic and Multi-flow Block Interleavers Algorithms

• Given and an interleaver is defined by
• Basic TypeI(s) to face bursts of length equal to
  s packets, with the minimum incurred delay, an (s
  x s) matrix is required
• Packets are written by rows, from left to right
  and from top to bottom.
• Packets are read by columns, from bottom to top
  and from left to right.
• Type I(s) maximum interleaver delay is given by
• Type I(s) is limited to such that
• For typical VoIP values
  and
• TypeI(s) is restricted to bursts with length s lt
  5

6
Basic and Multi-flow Block Interleavers Algorithms

• We interleave packets from different flows to
  face consecutive losses by introducing a
  tolerable delay
• We use the reference TypeII(nf) and propose
  TypeII(nf,s) where
• nf is the number of available flows and
• s is the maximum expected burst length
• Round-robin interleaver (TypeII(nf)) is suited
  for nf s
• it requires one (nf x 1) interleaver matrix
• to write the matrix each row will be assigned to
  a single flow.
• the matrix will be read from bottom to top
• if no switching delay is assumed then
• Drawback when nf lt s, TypeII(nf) interleaver
  does not isolate all the packet losses

7
Basic and Multi-flow Block Interleavers Algorithms

• The proposed TypeII(nf,s) assures the isolation
  of losses for a given (nf ,s) pair of values,
  even when nf lt s, with tolerable (bounded) delay
• Memory requirements
• If s is a multiple of nf then one (s x s) matrix
  is required.
• Otherwise, nf squared (s x s) matrices are needed.

8
Basic and Multi-flow Block Interleavers Algorithms

• Let us define
• Rij as the number of consecutive rows that the
  flow i will be assigned for matrix j.
• rem(a,b) as the remainder of the integer
  division a/b.
• int(a) largest integral value not greater than
  a.
• TypeII(nf,s) matrix writing procedure
• First matrix
• Ri1 int(s/nf), for i 1, 2, , nf
  - rem(s,nf).
• Rj1 int(s/nf) 1 for j n f
  -rem(s,nf), nf -rem(s,nf)1, , nf
• Next j 2, , nf matrices and flows i 2, , nf
• If Ri(j-1) int(s/nf)1 and
  R(i-1)(j-1) int(s/nf) then
• Rij int(s/nf) and
  R(i-1)j int(s/nf)1
• Type II(nf,s) matrix reading procedure
• Packets are read by columns, from bottom to top
  and from left to right.

9
Basic and Multi-flow Block Interleavers Algorithms

• TypeII(nf ,s) maximum delay Dmax is given by
• Dmax º
• where
• r rem(s,nf )
• d int((s-r) / nf )
• For typical VoIP values
  and
• for the best case TypeII(nf ,s) interleaver
  scatters bursts up to s lt 15
• favorably compared to TypeI(s) limited to s lt 5

10
Intelligibility Estimation

• For performance evaluation, we propose to use a
  high level end-user intelligibility estimation
  ASR rate
• Compared to MOS, ASR has lower cost and is more
  reproducible
• For end-user intelligibility estimation ASR rate
  can be more suitable than other measures like
  PESQ (P.862) or the E-model
• The Word Error Rate is defined by

11
Experimental Results

• Experimental results were obtained by simulation.
• We adopt a single error model based on a Markov
  chain
• (Yajnik et al 9).

12
Experimental Results

• For ASR evaluation we use the connected digit
  Project Aurora 2 database
• The speech recognizer is based on eleven 16-state
  continuous Hidden Markov Models (HMM)
• The HMM models are trained from a set of 8440
  noise-free sentences, while the out-of-train-test
  set comprises 4004 noise-free sentences
• More details are explained in the printed version
  of the paper

13
Experimental Results

• WER and Dmax (seconds) obtained values

14
Conclusion

• Compared to a single-flow approach, our proposed
  interleaver reduces the packet delay and makes it
  applicable under conditions where the reference
  scheme is unfeasible.
• Compared to the round-robin multi-flow
  interleaver, our proposed scheme increases the
  perceived end-user intelligibility (WER)
• With a slight penalty on the introduced delay
• We propose to consider ASR as a tool to measure
  VoIP services enhancements.
• Future work by setting up mapping functions
• for ASR rate to human intelligibility
• for ASR rate to MOS score
• and using it together with AN technology,
  enhanced VoIP services can be envisaged.
• Thank you for your attention