In this dissertation, we propose several methods to improve recognition accuracy of a hidden Markov model(HMM)-based speech recognition system by using segmental information of speech signal. As the segmental information, we use the HMM state segments which possess common stochastic characteristics of speech signal. Using the segmental information, we propose a modified corrective training algorithm which could improve the discrimination ability of HMMs. Then a new HMM parameter estimation algorithm and a new post-processor are proposed to reduce training and recognition time as well as to improve recognition accuracy. In order to obtain benchmark performances of the proposed algorithms, we implemented two baseline speech recognition systems based on phoneme-like units: one is a speaker-dependent system for 100 phonetically-balanced Korean words and the other is a speaker-independent system for 75 phonetically-balanced Korean words.
First, we present a modified corrective training algorithm using HMM state segment information. The modified corrective training method corrects the HMM parameters using the segmental k-means algorithm instead of the forward-backward algorithm used in the conventional corrective training method. It is motivated from the fact that the segmental k-means algorithm has more emphasis on the model state segment information. Applying this method to the speaker-dependent baseline system, we observe that the proposed method results in higher recognition accuracy than the conventional method. That is, the phoneme and word, recognition accuracies in the conventional method are 72.5% and 89%, respectively, and those in the proposed method are 74.9% and 93%, respectively. Also, the proposed method requires much less computation time than the conventional method in training process.
Second, a fuzzy segmental k-means(FSKM) algorithm for the HMM parameter re-estimation is proposed. A fuzzy vector quantization(FVQ)-based HMM (FVQ/HMM) scheme requir...