Analysis and compensation of packet loss in distributed speech recognition using interleaving

The aim of this work is to improve the robustness of speech recognition systems operating in burst-like packet loss. First a set of highly artificial packet loss profiles are used to analyse their effect on both recognition performance and on the underlying feature vector stream. This indicates that the simple technique of vector repetition can make the recogniser robust to high percentages of packet loss, providing burst lengths are reasonably short. This leads to the proposal of interleaving the feature vector sequence, prior to packetisation, to disperse bursts of packet loss throughout the feature vector stream. Recognition results on the Aurora connected digits database show considerable accuracy gains across a range of packet losses and burst lengths. For example at a packet loss rate of 50% with an average burst length of 4 packets (corresponding to 8 static vectors) performance is increased from 49.4% to 88.5% with a increase in delay of 90ms.