ICASSP 2005 - Poster M. Chibani

1
INCREASING THE ROBUSTNESS OF CELP-BASED CODERS BY
CONSTRAINED OPTIMIZATION
Mohamed Chibani, Philippe Gournay and Roch
Lefebvre
University of Sherbrooke Sherbrooke, Quebec,
Canada

• INTRODUCTION
• The use of the adaptive codebook (ACB) in
  CELP-like speech coders allows the achievement of
  high quality speech, especially for voiced
  segments.
• The problem
• The dependency of the ACB on the history of the
  signal makes it vulnerable to frame erasures.
  When a frame is lost, it affects not only the
  corresponding segment, but also the following
  frames due to error propagation, and the decoder
  may take several frames to recover.
• The proposed solution
• To reduce the error propagation when a frame is
  lost, we propose to lessen the inter-frame
  dependency by limiting the energy of the ACB
  contribution to the synthesized speech. The ACB
  gain by itself is not limited.

4. EXPERIMENTAL RESULTS Constrained encoding
improves the recovery time of the decoder after a
frame erasure
6. CONSTRAINED JOINT QUANTIZATION OF THE ACB AND
ICB GAINS The ACB contribution can be controlled
more precisely during the gains quantization
process, since at this stage both ACB and ICB
excitation shapes are known. The
constrained optimal solution can be found
analytically using the Kuhn-Tucker conditions.
3. CONSTRAINED SEARCH ALGORITHM If the ACB
energy contribution is high, a fraction of this
contribution is forced to be left in the new
target. The ACB contribution is estimated by
comparing, in the perceptual domain, the energy
due to the ACB excitation to the target energy.
The quadratic error relative to the quantized ACB
and ICB gains

• 2. EXCITATION SEARCH IN THE AMR-WB CODEC
• The standard search for the excitation parameters
  is done sequentially. It consists first in
  finding the gain and the delay for the ACB, then
  the shape (pulse positions) for the innovative
  codebook (ICB). Finally, the ACB and ICB gains
  are jointly quantized. The procedure may be
  summarized as follows
• Perform the LP analysis on the speech signal,
  extract the residual signal and build the target.
• Search for the ACB parameters that match the
  target (in an analysis-by-synthesis loop),
  according to the error criterion
• Update the target by subtracting the ACB
  contribution.
• Search for the ICB parameters using the updated
  target.
• Jointly quantize the ACB and ICB gains (quadratic
  error minimization in the perceptual domain).

A value of Rth1 as low as 0.55 preserves the
overall quality in clear channel while
significantly improving the robustness against
frame erasures.
NO PREFERENCE IN CLEAR CHANNEL
STRONG PREFERENCE WITH A FRAME LOSS RATE OF 10

• 7. CONCLUSION
• Limiting the ACB contribution, significantly
  improves the recovery time of the decoder after a
  frame loss.
• In clear channel, the quality degradation due to
  the constraint is very small.
• The modified coder is completely interoperable
  with the standard, it requires no extra bits and
  adds no extra delay.
• Although the ACB contribution can be controlled
  more precisely at the gains quantization stage,
  modifying only the target for the ICB search is
  the simplest and most effective approach.

• 5. SUBJECTIVE TESTS
• Two subjective tests were carried out in order to
  compare the standard and modified codecs in the
  clear channel condition (no frame erasures), and
  with a simulated frame loss rate of 10.
• Test features
• 20 sentences for each condition.
• 12 listeners participated in the tests.
• Listening using binaural headphones.
• AMR-WB in mode 2 (12.65 kb/s).
• Audio files on http//www.gel.usherb.ca/gournay/