Scalar quantization of features in discrete hidden Markov models

Traditionally, discrete hidden Markov models (D-HMM) use vector quantized speech feature vectors. In this paper, we propose scalar quantization of each element of the speech feature vector in the D-HMM formulation. The alteration required in the D-HMM algorithms for this modification is discussed here. A comparison is made between the performance of D-HMM based speech recognizers using scalar and vector quantization of speech features respectively. A speaker independent TIMIT vowel classification experiment is chosen for this task. It is observed that the scalar quantization of features enhances the vowel classification accuracy by 8 to 9%, compared to VQ based D-HMM. Also, the number of HMM parameters to estimate from a given amount of training data is drastically reduced.