Study On Speaker Recognition Technology

Speaker recognition is a branch of speech processing technology, it has a very broad research and application background. As a kind of biometric systems,it has unique advantage.To sound as a feature, its data acquisition equipment is non-contact, user acceptable.The process is that it extract the speaker's personality from a speaker's voice.Through the analysis and recognition these personality characteristics, so as to achieve the purposes of speaker recognition or confirmation.At first,this paper introduced the pre-processing of speech signals, that is you must standardized the speech signal before extraction thefeature . Speech point detection is a very critical step in the pre-processing of speech signals . This paper use short energy-frequency value as parameters for noisy speech point detection signal. This method are better to the traditional double-gate point detection.Then,it analyses the speech feature extraction process, the common feature are include LPC, LPCC, MFCC. One of the most characteristic parameter is MFCC.The pitch frequency will affect the MFCC parameters to accurately describe the vacal track characteristics,and then impact the performance of the Speaker recognition system. In order to reduce the impact of its pitch frequency, this paper proposes a improved MFCC parameters which is based on smoothing amplitude spectrum envelope. The results of experiment show that, the improved MFCC performance parameters are better than the MFCC parameters.Finally,in speaker recognition models, vector quantization model plays an important role in speech signal processing.It is commonly used in speaker recognition.This paper has modified the vector quantization model and proposed a recognition model that is based on variance and standard deviation weight distortion measure vector quantization model. Using Matlab program to experiment the improved feature extraction methods and recognition model show that the recognition system to achieve a high recognition rate.