1
 Improved Packet Loss Recovery for
Prediction-based Speech Coders 
- Philippe GOURNAY Senior Research
Engineer VoiceAge Corporation University of
Sherbrooke François ROUSSEAU, Roch
LEFEBVRE - ICASSP 2003, Hong Kong, 6-10 April 2003
2
Error propagation (CELP decoder)
Original Coded Concealed Error
One Lost Frame !
3
The ACELP coder

• Prediction (short- and long-term quantization)
• Analysis-by-synthesis

4
Improving the Robustness (1/4)

• Concealment
• Q What to do when a (Binary) Frame is Missing ?
• A Compute a Replacement (Speech) Frame
• Recovery
• Q What to do when the Frames are Received Again
  ?
• A Control Error Propagation

5
Improving the Robustness (2/4)

• Sender-based Methods
• Forward Error Correction (FEC)
• Bandwidth
• Multiple Descriptions
• Bandwidth, Independent Transmission Paths
• Retransmission
• Delay

6
Improving the Robustness (3/4)

• Receiver-based Methods
• Frame Loss Concealment
• Limited Effectiveness
• Interpolative Update of the Decoder Internal
  State
• Very Limited Effectiveness
• Playout Buffering
• Additional Delay, Speech Rate Adaptation

7
Improving the Robustness (4/4)

• Summary
• Higher Bit Rate or Delay
• Late Packets are Considered as Lost
• Recovery Problem Largely Overlooked

8
Update (Basic Idea)

• The concealment method does not correctly update
  the Internal State (I.S.) of the decoder
• We keep a copy of the past (Good) I.S. of the
  decoder before the concealment
• We use the late frame to update the I.S.,
  starting from the past (Good) I.S.
• Smooth transition between Concealment and
  Recovery
• Best when done in the excitation domain

9
Decoder Block Diagram
10
Chronogram (1 Late Frame)
Concealment
Recovery

A)
n-1
n
n1
n2
n3

n1
n2
n3

n

B)
n-1
(i)
n
n1

C)
n-1
(ii)
(iii)
n2
n3

11
Call Sequence (1 Late Frame)
12
Complexity (1 Late Frame)

• Memory
• One Copy of the Internal State
• AMR-WB roughly 1.5 k-octet
• Processing Power
• Two Additional Decoding of the Excitation
• AMR-WB equivalent to decoding one frame

13
Sample Waveforms (1/2)
Original Coded Concealed Updated
14
Sample Waveforms (2/2)
Original Coded Concealed Updated
15
Evaluation Results (AMR-WB)

• Cond. 1 One Late Frame / 10 Frames
• Cond. 2 One Lost Frame One Late Frame / 15
  Frames
• Cond. 3 Three Consecutive Late Frames / 20 Frames

16
Demonstration Files (AMR-WB)
Original File Original File
AMR-WB (12.65 kbits/s) AMR-WB (12.65 kbits/s)
Std. AMR-WB (Cond 1) Upd. AMR-WB (Cond 1)
Std. AMR-WB (Cond 2) Upd. AMR-WB (Cond 2)
Std. AMR-WB (Cond 3) Upd. AMR-WB (Cond 3)
17
Conclusion

• Using Late Frames Substantially Improves the
  Recovery of the Decoder
• Update the Internal State of the Decoder
• Some smoothing is required between the
  Concealment and the Recovery
• In a VoIP environment
• More Robust against Jitter, with no Increase in
  Delay
• Less Delay, with no Quality Degradation