Multiple Description Quantization for Recognizing Voice Over Packet Networks

Abstract

The practical design ofmultiple description vector quantizers for robust distributed speech recognition over packet networks is investigated. In the proposed system, speech parameters are quantized and mapped to multiple descriptions for transmission over independent channels. A new approach to the index assignment optimization is presented on the basis of a linear programming framework. Also, a fast local search algorithm is proposed to find the optimal index assignment without compromising the speech recognition accuracy. Experiments with random packet loss in a range of loss conditions are conducted on the Mandarin digit string recognition task. Simulation results indicate that the proposed multiple description scheme provides more robust performance than the ETSI standardized split vector quantization scheme with a single description.