Improved speech recognition in digital mobile communication environments

Abstract

This work considers the problems of speech recognition in digital mobile communication environments, and presents several methods to improve the speech recognition performance. In digital mobile communication networks, speech recognition systems conventionally first reconstruct speech and then extract feature parameters. In this work, we introduce an efficient approach of incorporating speech coding parameters into the speech recognizer and show the advantages of this approach by the measures of spectral distortion and recognition accuracy. Most speech coders employed in modern digital mobile communications represent line spectrum pairs (LSPs) as spectral parameters. We introduce two ways to improve the recognition performance of the LSP-based speech recognizer. One is to devise weighted distance measures of LSPs based on spectral sensitivity and mel-frequency warping. The other is to transform LSPs into cepstral domain features including pseudo-cepstrum (PCEP). The speech recognition experiments are performed for several databases including connected Korean digit and phonetically-balanced isolated word databases. The recognition results show that the proposed LSP weighting methods provide recognition accuracies considerably higher than the unweighted ones do. Also, the cepstral features converted from LSPs give more improved performances. Among the proposed methods, the mel-scale PCEP gives the best performance in view of recognition accuracy and complexity. Moreover, we present the effects of several standard speech coders on speech recognition performance under the adverse conditions such as tandem, frame erasure and background noise. The recognition results showed that the speech recognition performances are much affected by bit-rates and some optional schemes of speech coders. These comparative results can provide a guideline for selecting and/or designing a speech coder when a speech recognition service is needed in digital communication networks.``