An application of the multiple description scalar quantizer to speech coding on correlated fading channels

Abstract

utilizing multiple description scalar quantizers (MDSQ) for correlated Rayleigh fading channels. In this thesis, we construct two different types of MDSQ-based systems: a fixed length coding system and a variable length coding system. Both employ the discrete cosine transform (DCT) to decorrelate blocks of sampled speech. For the fixed length coding MDSQ system, DCT coefficient distributions are modeled and Lloyd-Max quantizers are constructed. We consider the performance tradeoff between fixed and adaptive bit allocations. We also discuss various strategies for exploiting the structure and redundancy of the multiple description quantizer. In the variable length coding system, we build a set of entropy-constrained, uniform threshold quantizers with optimized reconstruction levels. We accomplish the variable length coding of the DCT coefficients using Shannon-Fano-Elias (arithmetic) coding modified to accommodate the distinct distributions of the coefficients. We obtain performance improvements in this system by employing error detection that takes advantage of the residual redundancy in arithmetic codes. We measure the performance of both speech coding systems using objective measures such as the segmental signal-to-noise ratio (SEGSNR) as well as subjective listening tests. Additionally, we offer comparisons between our MDSQ system and another proposed system that uses dynamic bit allocation subband speech coding and rate-compatible punctured convolutional channel coding.