Korean speech vector quantization using a continuous hidden Markov model

The classification error of vector quantization (VQ) is a major factor which affects the performance of speech recognition. The most common VQ algorithms are the Linde-Buzo-Gray (LBG) algorithm and the K-means algorithm, proposed by Linde et al. in 1980, which have the advantages of being simple in concept and implementation with low computational costs. However the quantization error of VQ using these algorithm degrades the performance of speech recognizer. We propose an alternative VQ method for Korean speech using a continuous hidden Markov model (CHMM). The CHMM classifies the signal space into clusters of which each cluster is represented by a Gaussian function in the state of HMM. The results show that VQ using a CHMM classifies Korean speech space more effectively.