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 di erent 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 systemÕs 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. Ó

Instantaneous-distortion based weighted acoustic modeling for robust recognition of coded speech

“…The basic idea of the WAM method [4,5] is related to the observation described in Section 2 that not all segments of speech in a coded corpus are distorted to the same extent. As noted above, when the speech codec performs a short-term and long-term analysis of the speech signal, the level of distortion introduced by the long-term predictive analysis of the shortterm residual can be associated with the predictability of the speech signal.…”

“…In this work we focus on reducing the effect of this distortion on recognition accuracy by making use of the Weighted Acoustic Modeling method employing different distortion estimates. In the WAM technique [4,5] a set of acoustic models, each representing a certain distortion condition, is employed during decoding and its contribution to the overall likelihood is weighted by a running estimate of the distortion observed by each observation frame. In this work we estimate the instantaneous distortion in four different ways: using measured cepstral distortion, the long-term gain (adaptive codebook gain), the long-term predictability, and recoding sensitivity.…”

Single-channel noise reduction using unified joint diagonalization and optimal filtering

“…However, the purpose of noise reduction is sometimes as a preprocessor to, e.g., a speech recognition algorithm. Here, the word error rate increases when the SNR decreases [29,30], but on the other hand, the algorithms are also sensible to distortion of the speech signal [31,32]. In such cases, it might, therefore, be optimal with another relationship between SNR and speech distortion than the one having the best perceptual performance.…”

The influence of speech coding on recognition performance in telecommunication networks

“…Investigations have been carried out to determine the influence of speech coding on the performance of speech recognition systems [1,2] and to improve recognition performance in such situations [3,4,5]. Most of this work has a focus on the GSM full-rate coding scheme that was introduced as first coding technique in GSM mobile networks.…”