The Research On Feature Extract And Classifying Method Of Text-Independent Speaker Recognition

Automatic Speaker Recognition is a biometric characterization process aimed at automatically recognizing who is speaking based on unique information inherent in speaker signal. There are two key steps in this technology. The first is how to extract the feature parameter which can discriminate different speaker from original voice signal. The second is to design a classifier which could yield good performance.In order to solve problems with which this technology confronts, some deep research has been made, it main includes the following two aspects:As for feature extraction, the most popular feature parameter is the MFCC at present, which is extracted based on Short-Time Fourier Transform, assuming the voice signal is invariant in short time. In fact voice signal is a variant signal typically, but short-time analysis cannot alter time-frequency property. On the other hand, wavelet transform is a signal process based on time-scale representation, in which the time and frequency resolution basis function change with a scale factor. So based on study the MFCC's extraction theory and wavelet packet decomposition to speech signal processing, a new feature parameter named WPDC(wavelet packet decomposition coefficient) is proposed. In this way, we describe a frequency bands division of the signal by combing the nodes selected from the wavelet packet tree to secure a mel-like scale without overlapping.In classifier design, we have concentrated on the application of the neural network techniques to the task of speaker recognition. Firstly, a speaker recognition system using RBF network is established. Based on studying traditional VQ model and Neutral network model, another speaker recognition system using VQ combined with neutral network is proposal, in which a self-adaptation is took into the account to avoid the problem that performance degradation with the time passing.