A Comparison of Efficient Interleaver Designs for Real Time Distributed Speech Recognition

Packet loss presents a significant problem for distributed speech recognition systems particularly when burst lengths of loss are long. This work first proposes an extension to convolutional interleavers such that bursts of packet loss are maximally dispersed to minimise the duration of bursts of loss in the received feature vector stream. This is achieved by interleaving each dimension of the feature vector stream separately. This is shown to give significant gains in recognition accuracy on a large vocabulary task, although at theexpense of increased delay. The second part of this work shows how the interleaving delay can be absorbed into the hang-over delay used to determine when a speaker has finished talking in speech recognition applications.