The Research Of Speaker Recognition Based On Vector Quantization

Speaker recognition, as a branch of biometrics, is wildly applied in many fields. Based on the speech feature parameters that extracted from the speech waveform, which reflect the physiological and behavioral characteristics of the speaker, the speaker recognition identify the speaker’s identity automatically. The research of speaker recognition based on vector quantization is studied in this thesis.A text-independent speaker recognition system based on vector quantization is designed in this thesis, the mainly work is described as follows:(1) Establishment of speech database. The speech database is a Mandarin speech database which is recorded by myself in a background with relatively small noise. Using the voice recording software Cool Edit Pro2.1 by setting the sampling frequency as 8 kHz, the quantization precision as 16 bit, and selecting the channel mono mix, the speech database is composed of 30 male and 16 female voice, in which 10 seconds of each sample are intercepted for training and 5 seconds for testing.(2) The preliminary processing of speech signals is described. Mainly includes pre-emphasis, sub-frame, plus window and endpoint detection. Besides characteristic functions of speech signal in short time domain:short-time energy function and short-time average zero-crossing rate are analyzed. Then three methods of endpoint detection:double threshold method, spectral subtraction, spectral entropy method are introduced. The three kinds of endpoint detection method are simulated and analyzed under Matlab platform. Finally, an improved double limit endpoint detection method is proposed for the undetected problem, the undetected problem is avoided by the improved algorithm.(3) Characteristic parameters commonly used for speech signal are introduced and studied. Then the linear prediction coefficients (LPC), Linear Predictive Cepstral Coefficients (LPCC) and Mel Frequency Cepstral Coefficients (MFCC) and their extraction methods are mainly analyzed. In this thesis we choose the MFCC as the characteristic parameter, analyze the principle, procedure of the vector quantization algorithm, use the method of multistage vector quantization to improve the quantization process.(4) Finally the text independent speaker recognition system is designed under Matlab platform. By applying the multistage vector quantization, the speech database including 46 speaker’s speech are verified and a better recognition rate is obtained.