Speaker Recognition Technology In Noise Environment

Voice is one of the most direct and effective way to communicate. Voice not only contains information exchange, but also includes the speaker’s personality traits. Speaker recognition technology is used to identify the identity of the speaker by using the parameters which represent the speaker’s personality. In the past few decades, speaker recognition technology has made great progress and development. The main obstacle to further speaker recognition technology to practical is the noise in a variety of practical application scenarios. These noise greatly reduces the performance of the recognition system and affects the application and promotion of speaker recognition technology. This paper studies the problem of speaker recognition in noisy environments. This paper studies the Voice Activity Detector technology, anti-noise signal space technology, anti-noise feature space technology.The main work of this paper includes the following aspects:(1) Study on the Voice Activity Detector (VAD) of low SNR:Firstly, it introduces two VAD methods which have good performance in low SNR environment. The two VAD methods are the method based on the CO complexity and the method based on the MFCC similarity. Secondly, a new VAD method is proposed on the basis of the previous two methods. Finally, experiments are conducted to compare the performance of the three VAD methods. The result shows that the proposed method works well and has higher stability.(2) Study on the anti-noise signal space technology:Firstly, it introduces several conventional anti-noise signal space methods:Spectral subtraction method, Wiener filtering method and Wiener filtering method based on a priori SNR. Secondly, a new filtering method is proposed by adding MFCC similarity VAD to the winner filtering method based on a priori SNR. Finally, experiments are conducted to compare the performance of the four methods. The result shows that the proposed method has a better recognition performance than the others.(3) Study on the anti-noise feature space technology:Firstly, it introduces several conventional robust features:Perceptual Linear Prediction Coefficients and Mel-frequency cepstral coefficients. And it improves each of them to get two new features. Secondly, the two new features are fused to a new feature. Finally, experiments are conducted to compare the three features. And the result shows that the new feature has a better recognition performance than the others..