Researches On Speech Feature Extraction And Implementation Of Speaker Recognition System

Speaker recognition is considered to be the most natural biometrics, which authenticates people according to their physiological and behavioral characters in speech. Feature extraction acts an significant role in speaker recognition. However, it is difficult to extract a simple and reliable feature. Especially when the recognition process is executed in noise conditions, the accuracy would decline sharply. Researches on pitch, Linear Prediction Cepstrum Coefficients(LPCC) and Mel Frequency Cepstral Coefficients(MFCC) have been made. Then based on it, the main contribution of our work in feature extraction aiming at its stability, veracity and robustness are as following:(1) In traditional feature extraction, fixed length windowing could depress the stability of speech, and cause harmonic leakage. Pitch synchronous flexible length windowing preprocessing method for feature extraction is proposed in this paper. First, the synchronous pitch period of speech is extracted. The purely periodic portion of speech is reserved as the analysis frame based on pitch period. Then, the feature of the analysis frame is extracted. Experiment result shows the rise of recognition rate if this preprocessing is done before LPCC and MFCC feature extraction in the speaker recognition algorithm.(2) Since the preprocessing stated above can save more high frequency information and shorten the distortion between spectrum from two sessions of speeches, and the sound pressure level would decline with the increase of frequency, high frequency weighted process for the extraction of MFCC is proposed and applied to the speech with pitch synchronous preprocessing method. This can count in the raise of recognition rate in noise conditions.(3) After the validation proof of proposed methods in this paper, a proto-type system of real time speaker recognition is realized, and pitch synchronous flexible windowing preprocessing high frequency weighted MFCC extraction method is applied as well as traditional MFCC extraction. The speaker recognition system, with high accuracy and robustness, can fulfill the requirement of the application in office and home noise conditions with a small number of users.