Research On Vector Quantization In Speech Recognition

The basic goal of speech recognition is studying out a kind of machine with hearing function which can recognize phonetic message and perform human's intention in any condition.Speech endpoint detection is very important in speech recognition. Short-time energy and zero-crossing rate can not locate efficiently endpoint for speech in lower SNR. According to speech inter-frame correlation, endpoint detection of finite-state vector quantization is used to detect speech. The traditional algorithm with double threshold and algorithm with finite-state vector quantization are applied to speech endpoint detection together. The implementation procedure is given in detail. Experimental results show the efficiency of the new algorithm, especially in condition of lower SNR.Vector quantization is one of model training and pattern matching technologies with good performance which are applied to speech recognition at present. In the process of codebook design in vector quantization, traditional LBG(Linde-Buzo-Gray) algorithm owns the advantage of fast convergence, but it is easy to get the local optimal result and be influenced by initial codebook. Because genetic algorithm has the capability of getting global optimal result, this paper proposed a new clustering algorithm GA-L which based on genetic algorithm and LBG to improve the quality of codebook. This paper applied GA-L algorithm to mandarin continuous digit speech recognition, the experiments show it is more effective than traditional LBG algorithm.