Research And Implementation On Speaker Recognition

Speaker recognition, also known as voiceprint recognition, is a kind of technology using test voice to identify who the speaker is. As a branch of speech signal processing, this technology has been widely used in network security, speaker identification, conference record, justice verification etc. With the development of information, this biological certification technology progress to the commercial field with dropping out of the experimental stage gradually.From the point of application, speaker recognition is mainly divided into speaker identification and speaker verification, and in the content it is divided into the text-independent and the text-dependent. The purpose of this thesis is to realise a real-time system which focus on text-independent speaker identification, the main study work is as the following:(1) Analyzed of the current development of speaker recognition, mainly in two important aspects:how to extract the feature parameter and how to select the classifier model, by which this system could have good performance.(2) As for feature extraction, the Mel-Frequency cepstral coefficients(MFCC) were selected as the speech feature, and given the realization concretely. To find higher recognition rates, the difference MFCC has been combinated. One of voice activity detection(VAD) algorithm based on energy and zero crossing rate is introduced in this system.(3) Considering the trait of the feature parameter, we designed the Gaussian mixture model(GMM) for each speaker, in which the EM algorithm and the K-means methods were used the initial classification for comparision.(4) Based on Windows audio acquisition system, we design an interface to implement a real-time speaker recognition system by using MFC. Now, it can real-timely record voice and recognise in time also. The experimental result about fifty speakers is showed in this work, which means this technique is being practicably used.