INTERLEAVING AND ESTIMATION OF LOST VECTORS FOR ROBUST SPEECH RECOGNITION IN BURST-LIKE PACKET LOSS (FriAmOR3)

Author(s) : Alastair James (University of East Anglia, UK) Ben Milner (University of East Anglia, UK)

Abstract : Analysis into the effect of packet loss on speech recognition performance shows that both the burst length and the overall proportion of packets lost contribute to a deterioration in accuracy. To combat this burst-like packet loss several methods are compared for estimating the value of missing feature vectors. Three forms of interleaver are then compared which distribute long duration bursts of packet loss into a series of smaller bursts in the feature vector stream. Experimental results are presented on a range of channel conditions and demonstrate that substantial accuracy gains can be achieved using estimation techniques provided burst lengths are short. For longer burst lengths interleaving is necessary to maintain performance. For example at a packet loss rate of 50% and average burst length 20 packets (which represents 40 feature vectors or 400ms) performance is increased from 49.6% with no compensation to 86% with interleaving and cubic interpolation.

Menu