Reviving discrete HMMs: the myth about the superiority of continuous HMMs

Despite what is generally believed, we have recently shown that discrete-distribution HMMs can outperform continuousdensity HMMs at significantly faster decoding speeds. Recognition performance and decoding speed of the discrete HMMs are improved by using product-code Vector Quantization (VQ) and mixtures of discrete distributions. In this paper, we present efficient training and decoding algorithms for the discrete-mixture HMMs (DMHMMs). Our experimental results show that the high-level of recognition accuracy of continuous mixture-density HMMs (CDHMMs) can be maintained at significantly faster decoding speeds.

[1]  Vassilios Digalakis,et al.  Quantization of cepstral parameters for speech recognition over the World Wide Web , 1999, IEEE J. Sel. Areas Commun..

[2]  Vassilios Digalakis,et al.  Genones: optimizing the degree of mixture tying in a large vocabulary hidden Markov model based speech recognizer , 1994, Proceedings of ICASSP '94. IEEE International Conference on Acoustics, Speech and Signal Processing.

[3]  Satoshi Takahashi,et al.  Discrete mixture HMM , 1997, 1997 IEEE International Conference on Acoustics, Speech, and Signal Processing.

[4]  Vassilios Digalakis,et al.  Product-code vector quantization of cepstral parameters for speech recognition over the WWW , 1998, ICSLP.

[5]  Vassilios Digalakis,et al.  Efficient speech recognition using subvector quantization and discrete-mixture HMMS , 2000, Comput. Speech Lang..

[6]  P. J. Price,et al.  Evaluation of Spoken Language Systems: the ATIS Domain , 1990, HLT.