Independent Component Analysis And Its Applications In Speech Feature Extraction

Speech is the most direct and convenient method in the communication between different person。In rencent years, we always earnestly long for achieving audio communication between human and comuter, in which,speech recognition is one the key point technique. Generalized speech recognition is technique,by which the computer can make right response to the human voice. It is manily divided into speech recognition, speaker recognition, language recognition, etc.Speaker recognition is a technique to recognize identity of the speaker with speech parameter reflecting their physiological and behavior feature. The core procedure of speaker recognition is obtaining speech features through speech samples of speakers, then saving them into database, confirming the identity of speaker by matching the target speech to the feature in database at last.Essentially, speaker recognition can be divided into two parts on the whole, feature extraction and recognition model. Mel-frequency cepstral coefficients is the main feature in traditional speaker recognition task. But the robustness of this feature is not enough in noisy environment. Scarcity of reliable speech feature has been an important handicap to development of speaker recognition technique.Blind source sparation (BSS) recover independent source signals from the observed mixtures only by the statistic characteristics, without any preknowledge of mixing pattern and source signals. Independent component analysis (ICA) is a fire-new data analysis and signal processing method which appears during the research of BSS. Since its appearance, ICA has become a hot topic in signal processing, data analysis, statistics and neural networks, etc. And it has been widely used in speech processing, biomedical signal processing, pattern recognition, feature extraction, data compression, image processing, and telecommunications, etc. Recently, ICA was applied to log filter-bank-energies, by which could get speech features with similar characteristic to mel-frequency cepstral coefficients (MFCC) and better performance. It has great sense to the research on speaker recognition technique.In this dissertation, we introduce the basic theoretics of ICA, as well as component analysis (SCA); then, we apply ICA to the feature extraction procedure in speaker recognition task, do research as follows:1. A signal frames based speech feature extraction system, which can get the MFCC features, is introduced. The ICA-based features are obtained by apply ICA to the linear transformation procedure in this system, instead of the discrete cosine transform.2. ICA-based features obtained by training the clean speech signals, is applied to feature extraction procedure in clean and guassian noisy environment. The effectiveness of ICA features obtained from clean speech is proved by academic analysis based on kurtosis. In experiment simulation, ICA features obtained from clean speeh and MFCC features are both applied to speaker recognition task for clean and guassian noisy corrupted speech. The experimental results prove that, ICA-based features have better performance to recognition.3.The shortage of ICA features obtained from clean speech is analyzed in theoretic point of view in non-guassian noisy environment. A new signal representation model is proposed, in which ICA-based features are obtained by training the non-guassian corrupted speech. In experiment simulation, ICA features obtained from noisy speech is applied to speaker recognition task in non-guassian noisy speech. The experimental results prove that, ICA features based on noisy speech have obvious better performance to recognition than MFCC features and ICA features based on clean speeh.However, the robustness of speech features refered in thie dissertation still need to be improved in noisy environment. At the end of this dissertation, the future directions of our research are summarized and prospected.