Reconocimiento del habla en ambientes ruidosos mediante modelos ocultos de Markov discretos

Speech recognition in noisy environments remains an unsolved problem, even in the case of isolated word recognition with small vocabularies. Recently, several techniques have been proposed to alleviate this problem. Concretely, the Short-Time Modified Coherence (SMC) parameterization and the Cepstral Projection Distortion (CPD) measure have shown excellent results when tested in a speech recognition system based on Dynamic Time Warping (DTW) and using speech contaminated by additive white noise. In this paper, a new technique based on the AR modeling of the one-sided autocorrelation sequence (OSALPC) is presented and, from a comparative study of these LPC-based techniques in the discrete Hidden Markov Model (DHMM) approach, two main conclusions are attained: 1) the slope cepstral window and a relatively high model order are preferable, and 2) the cepstral representation based on the autocorrelation modeling achieves excellent results.Peer Reviewe