On the fuzzy vector quantization based hidden Markov model

There is a mathematical inconsistency in conventional FVQ/HMMs proposed by Tseng et al. (1987). This inconsistency appears to affect the recognition performance. We formulate two new types of FVQ/HMM to remove this inconsistency: multiplication type FVQ/HMM and addition type FVQ/HMM. According to experimental results, the MT-FVQ/HMM shows the best performance among the VQ type HMMs for a wide range of code-book size. It is also shown that the MT-FVQ/HMM can be derived on the basis of the Kullback-Leibler divergence between the a priori probability distribution of clusters defined at each state of a given model whose likelihood of yielding a given observation sequence y₁, ..., y_T is to be calculated, and the a posteriori probability distribution of the clusters for given y_t