Design And Realization Of Uyghur Isolated-word And Connected Digit Speech Recognition System

Presently, speech recognition research mainly used the statistical model (especially was Hidden Markov Model) method as well as the artificial intelligence (particularly neural networks) method as a research way.In this article we have done some research on Uyghur small-vocabulary speaker-dependent isolated-word speech recognition (used DTW model), basically realized a small speech recognition engine; then, we emulate training and recognizing algorithm of HMM model used in Uyghur connected digit recognition with Matlab6.2;It mainly includes two parts of work. The first part is small- vocabulary speaker-dependent isolated-word Uyghur speech recognition system. Small-scale Uyghur recognition engine used DTW model, its universality is better. Extracting MFCC as parameter, using LBG algorithm to make codebook of centroids.It is the key to construct engine corresponding to dictionary defined in this paper. Training process produces VQ file of each word in dictionary and uses them as reference model. Recognition model adopts DTW (Dynamic Time Warping) method, computes distortion measure of all centroids and returns the smallest one. Then compare and search in dictionary, finally display it as a result.In second part, do some research on Uyghur connected digit recognition based on isolated digit speech recognition. First of all, construct phoneme model to ten Uyghur numerals. Use HMM model in connected digit recognition, chose clustering algorithm to train reference pattern and use LB+Viterbi algorithm to recognize in this paper. Then, directly analyze phone-type string from HMM model and get the result of recognition. Finally, we emulate record function, training and recognizing algorithm of HMM model used in Uyghur connected digit recognition with Matlab6.2;It can recognize digit string of unpredictable length(length of 2,3,4 tested in this paper).