DISCRIMINATIVE CODEBOOK DESIGN USING MULTIPLE VECTOR QUANTIZATION IN HMM-BASED SPEECH RECOGNIZERS

Recent research on multiple vector quantization (MVQ) has shown the su itability of such technique for speech recognition. Basically, MVQ pro poses the use of one separated VQ codebook for each recognition unit. Thus, a MVQHMM model is composed of a VQ codebook and a discrete HMM m odel. This technique allows the incorporation in the recognition dynam ics of the input sequence information wasted by discrete HMM models in the VQ process. The use of distinct codebooks also allows to train th em in a discriminative manner. In this paper, we propose a new VQ code book design method for MVQ-based systems, obtained from a modified max imum mutual information estimation. This method provides meaningful er ror reductions and is performed independently from the estimation of t he discrete HMM part of the MVQ model. The results show that the propo sed discriminative design turns the MVQHMM technique into a powerful a coustic modeling tool in comparison with other classical methods as di screte or semicontinuous HMM's.