Speech Recognition System Based On Hmm And Ann Hybrid Structure Study

Speech recognition have been attached great importance by more and more people in recent years for its wide application prospect and the theoretical value. The realization of these technologies that computers and people allow for the intelligentized exchange and have a simple interface relay on the development of the theory of speech recognitio. Speech recognition in finance, commerce, culture, educationa and other aspects, especially in computers, electronic systems, automatic control have a very wide range of applications.Many scholars have made a lot of theoretical research on speech recognition, many speech recognition system based on HMM which has srong capacity of time-domain modeling and ANN which have strong capacity of classification have made outstanding applications.In this paper, the object of study was non-specific speech recognition system, through speech signal processing research, pretreatment of voice, sub-frames, plus windows, endpoint detection and feature extraction have done a more thorough study, the double-threshold endpoint detection algorithm,LPCC and MFCC feature parameters of the extraction process made a detailed introduction. By studying the traditional theory of HMM and ANN based on applications in speech recognition and the srong time-domain modeling capacity of HMM in combination with the strong classification capacity of ANN, we propose a HMM/BP hybrid network combination. Speech signal in the HMM decoded by viterbi generate the state transition cumulative output probability as input of BP neural network and we can gain the recognition result from output of BP neural network. Hybrid network has certain degree upgrade by experiments comparing the traditionl HMM and ANN.In addition, through the experimental comparison with MFCC and LPCC as feature parameters,this paper come to the conclusion that MFCC is better than LPCC in speech recognition and found appropriate the number of hidden layer neurons by experiments.