Distortion-class modeling for robust speech recognition under GSM RPE-LTP coding

Abstract

We present a method to reduce the degradation in recognition accuracy introduced by full-rate GSM RPE-LTP coding by combining sets of acoustic models trained under different distortion conditions. During recognition, the a posteriori probabilities of an utterance are calculated as a weighted sum of the posteriors corresponding to the individual models. The phonemes used by the system’s word pronunciations are grouped into classes according to amount of distortion they undergo in coding. The acoustic model used in the decoding process is a weighted combination of models derived from clean speech and models derived from speech that had been degraded by GSM coding (the source models), with the relative combination of the two sources depending on the extent to which each class of phonemes is degraded by the coding process. To determine the distortion class membership, and hence the weights, we measure the spectral distortion introduced to the quantized long-term residual by the RPE-LTP codec. We discuss how this distortion varies according to phonetic class. The method described reduces the degradation in recognition accuracy introduced by GSM coding of sentences in the TIMIT database by more than 70% relative to the baseline accuracy obtained in matched training and testing conditions with respect to a system using the source acoustic models, and up to 60% relative to the best baseline systems regardless of the number of Gaussians.