Multiple description transform codes with an application to packetized speech

Abstract

We consider the performance at high rates of the Multiple Description Scalar Quantizer (MDSQ) for a memoryless source, and we study its applications to sources with memory. The MDSQ is a system which uses diversity to overcome channel impairments: the information is transmitted on different channels, and in the case of two channels, which is the case considered in this thesis, the principal objective is to obtain the highest quality at the receiver when both channels are working, under a constraint on the quality received when one channel is down. We consider a memoryless source and we show that at high rates, relevant approximations allow the derivation of closed-form expressions of the distortion when both channels are working, and of the distortion when one channel is down, in the case of uniform, non-uniform and entropy-constrained quantizations. These results are then used to analyze the performance of multiple description transform codes for sources with memory and an associated optimal bit allocation problem is solved. Finally, we apply multiple description transform codes as well as multiple de scription sub-band codes to real speech. We demonstrate that for a packet switched network in which packets may be lost, significant improvements in reconstructed speech quality are obtained as compared to existing methods.