Efficient channel coding for CELP using source information

Abstract

Since simultaneous optimization of source and channel coding is often not practical for speech-coding algorithms, it is useful to develop channel codes optimal for a particular speech coding algorithm operating under a specified range of channel-error conditions. Such source-dependent channel-error codes can be obtained by minimizing an appropriate speech distortion criterion. To carry out this minimization, we use a simulated-annealing procedure. The resulting channel codes are efficient, since they provide error correction of non-uniform accuracy (highly probable quantization levels receive more accurate correction) and/or non-uniform error detection (errors which greatly impact the speech quality are more likely to be detected). An optimal trade-off between error correction and error detection can be obtained. Source-dependent channel codes aimed at counteracting low, random error rates (up to 2%) are applied to the CELP algorithm and the resulting performance is reported. It is found that a small allocation of codewords (equivalent to less than one bit) for the protection of a particular parameter often results in a large performance improvement.