(A) study on the use of perceptual information for speech recognition

Abstract

Speech recognition by machines has applications in many areas, but it has been achieved with only a limited success. This is due to the fact that the current template-matching-based speech recognition approach relies heavily on the general-purpose pattern recognition algorithms and utilizes little of speech-specific knowledge. The main objective of this dissertation work is the development of a speech recognition system which can yield improved recognition accuracy by attending to constraints imposed by the knowledge of human speech perception, while maintaining the advantages of the template matching approach. For this purpose, we first propose a spectral representation of the speech signal based on the human peripheral auditory system. In this representation, a bank of band-pass filters with filter characteristics that produce cochlear-like frequency mapping is used for the spectral analysis. The frequency characteristics of filter bank are designed so that they may reflect faithfully the knowledge of peripheral auditory processing and psychophysical relations. We compare the performances of various filter-bank-oriented features including the proposed filter bank feature with respect to the recognition accuracy. From the isolated word recognition experiments, we show that the proposed feature outperforms other existing features, especially for the speaker-independent case. Second, we propose a modified distance measure that is insensitive to perceptually irrelevant spectral variations. The proposed method may be realized simply by applying a nonlinearity to the conventional distance measure which computes the acoustic dissimilarity between two spectra. We examine several realization schemes for the modified distance measure, and show that the discriminability of phonetically similar words is significantly improved even by a very simple, threshold-type nonlinearity applied to the conventional distance measure. Since the optimal choice of the threshold is indep...