Recognition of noisy speech

Abstract

A model-based spectral estimation algorithm is derived that improves the robustness of speech recognition systems to additive noise. The algorithm is tailored for filter-bank-based systems, where the estimation should seek to minimize the distortion as measured by the recognizer's distance metric. This estimation criterion is approximated by minimizing the Euclidean distance between spectral log-energy vectors, which is equivalent to minimizing the nonweighted, nontruncated cepstral distance. Correlations between frequency channels are incorporated in the estimation by modeling the spectral distribution of speech as a mixture of components, each representing a different speech class, and assuming that spectral energies at different frequency channels are uncorrelated within each class. The algorithm was tested with SRI's continuous-speech, speaker-independent, hidden Markov model recognition system using the large-vocabulary NIST Resource Management Task. When trained on a clean-speech database and tested with additive white Gaussian noise, the new algorithm has an error rate half of that with MMSE estimation of log spectral energies at individual frequency channels, and it achieves a level similar to that with the ideal condition of training and testing at constant SNR. The algorithm is also very efficient with additive environmental noise, recorded with a desktop microphone.