Lingfen Sun

1
Impact of Packet Loss Location on Perceived
Speech Quality

• Lingfen Sun
• Graham Wade, Benn Lines
• Emmanuel Ifeachor
• University of Plymouth, U.K.
• L.F.Sun_at_jack.see.plym.ac.uk
• j.wade,B.Lines,E.Ifeachor_at_plym.ac.uk

2
Outline

• Introduction
• Codec's internal concealment and convergence time
• Perceptual speech quality measurement
• Simulation system
• Loss location with perceived quality
• Loss location with convergence time
• Conclusions and future work

3
Introduction

• End-to-end speech transmission quality
• IP network performance (e.g. packet loss and
  jitter)
• Gateway/terminal (codec loss/jitter
  compensation)
• Impact of packet loss on perceived speech quality
  
• Loss pattern (e.g. burst/random)
• Loss location (codec's concealment)

4
Introduction (cont.)

• Previous research on loss location
• Concealment performance is speech content related
  (e.g. voiced/unvoiced)
• Analysis based on MSE or SNR for limited codec
• Perceptual objective methods only to assess
  overall quality under stochastic loss simulations
• Questions
• How does a packet loss location affect perceived
  speech quality ?
• How does a packet loss location affect codec's
  convergence time (for loss constraint)?

5
Codec's internal concealment

• What is codec's concealment?
• When a loss occurs, the decoder interpolates the
  parameters for the lost frame from parameters of
  previous frames.
• Which codec has concealment algorithm?
• G.729/G.723.1/AMR (main VoIP codecs)
• CELP analysis-by-synthesis
• What are the limitations of concealment
  algorithms?
• During unvoiced(u) or voiced(v)
• During u/v

6
Codec's convergence time

• What is convergence time?
• The time taken by decoder to resynchronize its
  state with encoder after a loss occurs. It is
  also called resynchronization time.
• For set up loss constraint distance between two
  consecutive losses for new packet loss metrics
• What is the relationship between convergence time
  with loss location, codec type and packet size?

7
Perceptual quality measurement
Reference signal
Objective perceptual quality test
Objective MOS
Degraded signal

• Transform the signal into the psychophysical
  representation approximating human perception
• Calculating their perceptual difference
• Mapping to objective MOS (Mean Opinion Score)
• Algorithms PSQM/PSQM/MNB/EMBSD/PESQ

8
Simulation System
Reference speech
Degraded speech without loss
Bitstream
decoder
encoder
convengence time analysis
Degraded speech with loss
loss simulation
decoder
perceptual quality measure
Reference speech

• Perceptual speech quality analysis with loss
  location
• Convergence time analysis with loss location

9
Speech test sentence

• Speech test sentence is about 6 seconds.
• First talkspurt (about 1.34 second, above
  waveform) is used for loss location analysis.
• Four voiced segments, V(1) to V(4), which can be
  decided by pitch delay in G.729 codec

10
Pitch delay from G.729 codec
V(2)
V(1)
V(3)
V(4)
11
Loss location with perceived quality

• Each time only one packet loss is created
• Loss position moves from left to right one frame
  by one frame
• Overall perceptual quality is measured from
  PSQM/PSQM, MNB and EMBSD
• Packet size 1 to 4 frames/packet
• Codec G.729/G.723.1/AMR
• How does a loss location affect perceived speech
  quality ?

12
Loss position with quality (1)
Loss position
reference speech
PSQM
degraded speech
PSQM
13
Loss position with quality (2)
Loss position
reference speech
PSQM
degraded speech
PSQM
14
Loss position with quality (3)
Loss position
reference speech
PSQM
degraded speech
PSQM
15
Loss position with quality (4)
reference speech
Loss position
degraded speech
PSQM
PSQM
16
Overall PSQM vs loss location (G.729)
G.729
17
Overall MNB vs loss location (G.729)
G.729
18
Overall EMBSD vs loss location (G.729)
G.729
19
Overall PSQM vs loss location (G.723.1)
G.723.1
20
Loss location with perceived quality

• Loss location affects perceived quality.
• The loss at unvoiced speech segment has no
  obvious impact on perceived quality.
• The loss at the beginning of the voiced segment
  has the most severe impact on perceived quality.
• PSQM yields the most detailed result comparing
  to MNB/EMBSD

21
Convergence time based on MSE
G.729
22
Convergence time based on PSQM
23
Convergence time based on PSQM
24
Loss location with convergence time

• Convergence time is almost the same for different
  packet size
• Convergence time for a loss at unvoiced segments
  appears stable
• Convergence time shows a good linear relationship
  for loss at the voiced segments
• maximum at the beginning
• linear descending
• Up bound to the end of voiced segments

25
Conclusions and future work

• Investigated the impact of loss locations on
  perceived speech quality
• Investigated the impact of loss locations on
  convergence time
• The results will be helpful to develop a
  perceptually relevant packet loss metric.
• Future work will focus on more extensive analysis
  of the impact of packet loss on speech content