Isolated word recognition in noisy environment by vector quantization of the HMM and noise distributions

Abstract

The author reports on the application of vector quantization (VQ) of the continuous HMM (hidden Markov model) and noise distributions for isolated word recognition in a noisy environment. Separate codebooks and HMMs for noise-free speech and noise are constructed. An input frame is associated with the speech or noise codebook according to the minimum distortion criterion, and the probability density function of the combined speech-noise signal is obtained from the vector-quantized PDFs of speech and noise, and their respective cumulative distribution functions. The methodology was applied to an isolated digital recognition task in an office environment, in a car environment, and with additive white and impulsive noise. The VQ approach resulted in much improved recognition performance. >