Improved training and recognition algorithms with VQ-based hidden Markov models

Two aspects of the application of vector quantization (VQ) in speaker-independent isolated-word recognition using discrete hidden Markov models (HMMs) are discussed. An automatic segmentation scheme using the information about the sequence of the L nearest codebook entries to a feature vector is presented. Based on the resulting segment boundaries, the parameters of the word models are estimated. A smoothing of the symbol probabilities is discussed. In the calculation of the model probabilities, the probability for the best codebook index is replaced by a weighted sum over the R best codebook indices. The algorithms were tested on a speaker-independent German database for a vocabulary with 23 words, and results are presented