The Research Of Small Vocabulary Speaker-independent Isolated Word Speech Recognition System

Speech recognition is a cross-disciplinary which involves digital signal processing, artificial intelligence, computer science,mathematical models and many other disciplines, and is one of the fastest growing field of information research now. With the continuous development of people's needs and the technology of the embedded chip and mobile terminal equipment, speech recognition system, in particular, the speaker-independent isolated word speech recognition system is applied more and more into our daily lives (such as voice dialing of cell phones, car voice navigation, etc.),which brings us a very great convenience.This paper makes in-depth research for small vocabulary, speaker-independent speech recognition of isolated word, major work includes:An improved dual-threshold speech endpoint detection algorithm based on short-time average magnitude increment and short-time average zero-crossing rate is presented, which uses the continuous rise of speech short-time average magnitude as the first-level decision threshold and short-time average zero-crossing rate as the second-level decision threshold.Experimental results show that this algorithm is accurate, easy and reliable in the case of ideal signal to noise ratio(SNR).Several feature extraction algorithms of speech recognition are researched:linear predictive coding (LPC) based on the model of human pronunciation, linear prediction cepstrum coefficient (LPCC) and mel frequency cepstral coefficient (MFCC), and several methods that can improve the performance of mel frequency cepstral coefficient are listed. The paper introduces the algorithms of dynamic time warping (DTW) and hidden markov model (HMM) for isolated word speech recognition. For the DTW algorithm, dynamic programming (DP) technique and relaxation endpoint DTW algorithm are introduced. When introducing hidden markov model, the paper starts from the concept, introduces its three basic problems,their solution and continuous hidden markov model. On the basis of the research for the speech recognition technology, a algorithm program of building a small vocabulary speaker-independent isolated word speech recognition system is proposed, and the simulation of the speech recognition system is implemented in Matlab environment. Describe the problems that are encountered in the process of realization of the system and problem-solving methods in detail, and verify the performance of the system by experiments in the end.The speech recognition system includes:pre-processing of speech signal (includ pre-emphasis, normalization, framing and endpoint detection algorithm proposed in this paper), improved MFCC extraction algorithm and continuous hidden markov model.