Speech Recognition System Based On Independent Component Analysis

With the development of Human-Computer Interaction(HCI) technology, speech recognition has already become the key point of research in the field of artificial intelligence and pattern recognition. Speech is not only the most important and convenient method of information transmission, but also one of the most direct ways of the HCI realization. It is of great significance to make machines execute corresponding commands after recognizing the speech contents accurately. The related research has expansive application prospects in the fields of medicine, military and industry. However, due to the existence of background noise in real environment, the recognition rate of the existing speech recognition techniques degrades dramatically in practice. Therefore, how to improve the stability of the speech recognition system has become a priority.This thesis focuses on the research of fast voice activity detection(VAD) based on sliding window, speech enhancement methods based on independent component analysis(ICA). The main research work is shown as follows:(1) Front-end processing and voice activity detection. The methods of speech front-end processing have been introduced in this part, which include digitization, pre-emphasis and framing of speech signal. Besides, a recursive calculating algorithm for higher-order cumulants over sliding window is proposed which has been applied to voice activity detection. Furthermore, the VAD method based on dynamic higher-order cumulants is proposed. The principle and detection process of the proposed VAD algorithm have been described in detail. Finally, a contrast experiment about the detection performance of the proposed algorithm and the G.729b VAD algorithm has been carried out in different noisy at different signal-to-noise ratios(SNRs) environments. Experimental results reveal that the proposed VAD algorithm has better robustness and higher calculation efficiency.(2) The research of independent component analysis algorithms. Firstly, the principle of ICA algorithms has been introduced, which includes the mixed models, construction of objective functions and optimization algorithm.Secondly, the thesis deeply studies the ICA algorithms in convolution mixed model and states the principle and realization process of the convolution ICA algorithms in detail. On this basis, three kinds of instantaneous ICA algorithms in complex field are introduced: Jointly Approximate Diagonalisation Eigenmatrics(JADE) algorithm, ICA algorithm based on kurtosis maximum and information maximum(Infomax) respectively. Moreover, the thesis analyzes the inherent fuzzy problems of ICA algorithms and proposes the corresponding solutions. Finally, a series of emulation experiments have been carried out on MATLAB platform. Experimental results show that the permutation alignment algorithm based on envelope coefficient employed in this thesis can effectively decrease the probability of incorrect permutation to improve separation effect and the instantaneous ICA algorithms in complex field have acquired good separation performance.(3) Speech feature extraction based on ICA. First of all, some feature parameters of speech signal have been presented, covering short-time energy, linear predictive cepstrum coefficients(LPCC) and Mel frequency cepstrum coefficients(MFCC). Then the principle and realization process of the speech feature extraction algorithm based on ICA have been stated in detail.(4) Realization of speech recognition system based on ICA. Two speech recognition techniques have been introduced, including the technique based on dynamic time wrapping(DTW) and hidden makov model(HMM). The thesis focuses on the research of the principle of the technique based on HMM. After the analysis of the composition of the speech recognition system, each module of the system has been realized in MATLAB. Finally, speech recognition experiments are conducted in the emulational and real environments separately. Experimental results indicate that the combination of ICA algorithms and the sliding window technique makes the speech recognition system obtain good robustness in noisy environment.