Multiple Description Speech Coding

1
Multiple Description Speech Coding

• Xin Zhong
• April 12, 2002

2
A Telecomm Point of View

• Principle of separation is optimal only with
  infinite block length infinite delay and
  complexity.
• Otherwise, joint source-channel coding is better.

3
Application of MDC

• Network communication (VoIP) introduces packet
  loss.
• Asking for re-transmission
• Causes delay without guaranteed improvement.
• Low efficiency in multi-cast applications.
• Decoder-based algorithms
• Packet losses happen in bursts.
• Frame-concealment is inadequate.
• G.729 Example
• Without Packet loss
• With Packet Loss

4
A Two-Channel MDC System
Encoder 1
Encoder 2
5

Two MDC Examples

• Information Theory Example (Berger in 83)
• An i.i.d binary source symbol with equal
  probability.
• Minimum information rate for zero distortion (
  ) is
• 1 bit/symbol.
• bit/symbol gt

• DSP Example (Jayant in 81)
• Sending odd-indexed PCM speech samples over one
  
• channel, even-indexed samples over another.
• Central decoder output original PCM speech.
• Side decoder outputs construct missing samples
  by
• interpolating their neighbors.

6
Information Flow
Total Information Used by Central Coder
Packet A for Channel 1/ Side Coder 1
Packet B for Channel 2/Side Coder 2
(Usually by some obscure transformations)
7
Analysis

• Theoretical point of view available information
  is intentionally withheld from each
  encoder/channel module to make the central
  decoder relatively better than the side decoders.
• Practical point of view research effort is spent
  on how to construct the side decoder outputs, but
  majority of the time the system uses the central
  decoder.

8
Alternative MDC Design
Total information Used by Central Coder
Non-redundancy
Common information
Packet A for Side Coder 1/Channel 1 (Full
Description)
Packet B for Side Coder 2/Channel 2 (Full
Description)
9
Baseline Repetitive Coding Transmission
(Same VQ for All Encoders)

• No non-redundancy as each encoder produces the
  identical output.
• Receiving N copies is the same as receiving one.

10
Coder Diversity
(Different VQ for Each Encoder)

• Non-redundancy appears in the uncorrelated
  residual noises.
• Performance of central decoder increases by
  factor N.
• N different encoders are difficult to design
  beyond VQ.

11
Time Diversity (I)
(Telecomm.)
Time 1
Modulated Bits
Receiver
Channel condition 1
Time 2
Channel condition 2
The same information is sent each time because
binary bits allow no room for different
interpretation.
12
Time Diversity (II)
(DSP)

• Using the same coder, the same signal block
  illustrates different characteristics under
  different surroundings non-redundancy
  is preserved same performance as
  using coder diversity.

13
Time Diversity with Residual Compensation

• M of packets lost
• This is a joint source-channel coding system.

14
Summary

• To have super resolution with all information
  packets, each coder must contribute some
  non-redundant information about the same signal
  block.
• Non-redundancy can be generated via diversities
  time, frequency, coders.
• One coders weakness (residual) can be studied,
  absorbed, and compensated by the next to improve
  the overall system.
• These techniques can be applied to MDC of any
  signals for any applications.

15
15
Time Diversity with Residual Computation (I)
Twice-Coded LSF with Residual Quantization
Spectral Envelopes by Unquantized LSF.
Single Quantization. Two
quantizations with residual
compensation.