Vector quantization of the HMM distributions: further results

The statistical vector quantization approach is extended to cepstral features and cepstral features augmented by differenced cepstral coefficients. The use of binary tree search codebooks is investigated. A suboptimal algorithm for obtaining binary tree search codebooks is presented. The new techniques are evaluated in terms of their recognition performance with the digits vocabulary. Experimental results indicate that the degradation introduced by this technique is much smaller than the degradation caused by conventional binary search codebooks