A novel projection-based likelihood measure for noisy speech recognition

The projection-based likelihood measure, an e€ective means of reducing noise contamination in speech recognition, dynamically searches an optimal equalization factor for adapting the cepstral mean vector of hidden Markov model (HMM) to equalize the noisy observation. In this paper, we present a novel likelihood measure which extends the adaptation mechanism to the shrinkage of covariance matrix and the adaptation bias of mean vector. A set of adaptation functions is proposed for obtaining the compensation factors. Experiments indicate that the likelihood measure proposed herein can markedly elevate the recognition accuracy. Ó 1998 Elsevier Science B.V. All rights reserved.