Recently, hidden Markov model(HMM) has become the predominant approach to speech recognition. The performance of an HMM-based recognizer depends on the acoustic modeling techniques adopted to represent speech characteristics. In this dissertation work, we present various methods to improve acoustic modeling capability in a discrete HMM-based speech recognition system. In order to evaluate the recognition performance for proposed methods, we established the baseline speech recognition system which is a continuous speech recognition system with a vocabulary of 1074 words.
First, we study the method to estimate robust output probabilities distributions(PD's) in considering that feature vectors within the same state have similar characteristics in their acoustical aspects and the their variation is lower than feature vectors at another state. Based on these facts, we proposed a state-dependent source-quantization(SDSQ) DHMM in which the output distribution of a state is represented as the sum of product of the frequency of a state- dependent source and its own codeword distribution. A state-dependent source is represented as an average vector, and its frequency and distribution of codewords are separately accumulated by the quantized source. The experimental result indicated that the proposed method improved the word accuracy by 2.7% as compared to the baseline system.
Second, we apply the state-dependent feature-parameter weighting for discriminative estimation of output probability at each state. In contrast to the conventional equal- weighting of output probability for each feature parameter, we adopted two kinds of weighting methods-the fixed weighting method and the varying weighted method-to get the discriminative output probability at a state. In the first approach, the fixed weighting factor for each feature parameter is derived on the training phase at once, and it is simply multiplied with the output probability of each feature parameter on the testing phase. As a second approach, the varying weighted method is adopted to reflect output probability of another feature parameter given the best feature parameter. For this, the weighting matrix is derived, and it is multiplied by the output probabilities of feature parameters at a state on the testing phase. In the comparison of weighting methods, the varying weighted method improved the performance of 2% as compared to the fixed weighting method.
Finally, we proposed the modified smoothing method based on the fuzzy mapping concept for HMM parameters. In the conventional fuzzy-based smoothing method, it dose not satisfy two criteria such that the smoothing factors are not normalized and the smoothing factor corresponding to the smoothed output probability is not always larger than other factors. To solve these drawbacks, we modified the conventional fuzzy smoothing method to get the correct smoothing factor for the target output probability. The experiment indicated that the proposed smoothing method represented the improvement of 1.4% as compared to the conventional fuzzy smoothing in an aspect of the word accuracy.
