In this thesis work, we study optimization of codebook for speech recognition using discrete hidden Markov models. We propose a new codebook merging algorithm that merges codewords having similar statistical properties. A tree structured classfier constructed by iteratively growing and pruning process is adopted for classifying codewords.
The characteristics of this algorithm is to combine the codebook design process and the HMM training process. It is pseudo supervised vector quantization by Viterbi alignment. The proposed algorithm is used to reduce the codebook size and results in a more efficient system. Also, it enhances the recognition accuracy by suppressing confusible codewords.
To confirm our algorithm, we simulate using 74 Korean words spoken by 7 male speakers. Experimental results show both codewords reduction and performance improvement.