Research Of Speaker Recognition Technology Based Text-Independent

In actual life, identity authentication is applied to every field and there are many ways of identity authentication, such as fingerprint recognition, iris recognition, face recognition, and so on. Voiceprint recognition is one of it and also known as speaker recognition. It can recognize speaker by speaker's voice. It can be divided into two kinds:text-dependent speaker recognition and text-independent speaker identification. The latter is more useful in our life. In this paper, to improve the accurate rate of text-independent speaker identification, studied on its key technology.In this paper, on the basis of text-independent speaker identification existing technology, combined with phonetics, phonology and the feature of voice signal, studied text-independent speaker identification key technology. The primary research contents includes:Distinguish between unvoiced and voiced. On the basis of voice signal of removing silence, in view of the question that zero-crossing rate way can't distinguish between unvoiced and voiced when it deals with the voice signal whose mean amplitude keep away from horizontal axis for a long time, proposed the method of effective flip rate; In addition, for the unvoiced and voiced signal which is similar in effective flip rate, distinguished them by spectrum mean way.Pitch detection. On the basis of the methodology that categorizes speech signal into three types, silence, voiceless sound, voiced sound, in view of the random distribution of obvious periodic property speech, the improved algorithm of length varied average magnitude difference function (LVAMDF) and comprehensive multi-factor for pitch frequency detection was put forward to categorize voiced sound into two types, one was obvious periodic property speech, the other one was unobvious periodic property speech. At the same time, the starting and ending points of all accurate pitch period in the obvious periodic property speech was achieved. For a few pitch periods divided into frequency doubling or half frequency, the recognition and correction method was proposed which has a high recognition and correction rate.The text-independent speaker identification system. According to the theory of the system, by MATLAB and C++ mixed programming, have finished the system and system testing and the EER (Equal Error Rate) achieved 0.4762%.The theory study of speaker recognition based on phoneme Category. On the basis of manually annotated phoneme of TIMIT corpus, using the theory of confusion matrix, researched the similarity and otherness of the different pronunciation of different speaker, to improve the system performance by emphasizing the otherness and avoiding the similarity. This part have finished several theoretical research and several function.In the paper, improved two key technology performance, and finished the text-independent speaker identification system. Finally, came up with the improvement project of the identification theory of the system. But because the workload is huge, the improvement project need to perfection and verification, the system performance need to further enhance in phoneme recognition, feature extraction and the system theory.