Recent research on Multiple Vector Quantization (MVQ) has shown the suitability of such technique to Speech Recognition. Basically, MVQ proposes the use of one separated VQ code-book for each recognition unit. Thus, a MVQ HMM model is composed of a VQ codebook and a discrete HMM model. This technique allows the incorporation in the recognition dynamics of the input sequence information wasted by discrete HMM models in the VQ process. The use of distinct codebooks also allows to train them in a discriminative manner. In this paper, we propose a new VQ codebook design method for M VQ-based systems that provides meaningful error reductions and is performed independently from the estimation of the discrete HMM part of the MVQ model. This codebook design uses a Minimum Classification Error scheme and have certain similarities with the LVQ techniques proposed by Kohonen, but overcoming any time alignment requisite.
Bibliographic reference. Peinado, Antonio M. / Rubio, Antonio J. / Segura, Josť C. / Sanchez, Victoria / Diaz, Jesus E. (1995): "MCE estimation of VQ parameters for MVQHMM speech recognition", In EUROSPEECH-1995, 533-536.