Chinese Text Speech Recognition Derived from VQ-LBG Algorithm

Abstract

Speech recognition has been drawn extensive attentions for identity verification, man-machine interactive control, security appliance and so on, while the recognition rate of the speech text is essentially affected by the pronounce mode of different language. In order to recognize the Chinese text, which is syllabic in nature with severe word boundary uncertainty problem, a combined algorithm has been established based on Linde–Buzo–Gray (LBG) method to achieve vector quantization (VQ) for the MFCC processed Chinese speech text. Simulation indicates that the proposed algorithm is effective to construct codebooks for each solicited Chinese text word constricted by the minimum average distortion measure criterion, and the speech recognition is achieved through comparing between the codebooks and the unknown speech signal. Results also show that the codebook template dimension plays important role to affect the recognition rate, with 32 dimensions as the optimal size which can result in higher recognition rate and few computational resources occupancy. The proposed algorithm can be integrated in Chinese speech analyzing software to achieve voice signal recognition.