The Research Of Speaker Recognition Under Noisy Environment

Speaker recognition is one of important research branches of speech signal processing field. Now, most speaker recognition systems have high performance in low noise and low distort environment. Once the system applies in noisy environment, the performance is degraded dynamically. So, the research of speaker recognition under noisy environment, specially reducing the effect of noise, is key to the practicable speaker recognition system, is one focus of current research work too. This paper investigates how to improve the performance of speaker recognition system under noisy environment, the main work in this paper is listed as follow:In the speech pre-processing, the principle of wavelet-based speech enhancement is investigated, and an improved arithmetic is presented. Experiments show that the improved arithmetic not only has more excellent effects on speech enhancement but also has potential to improve robustness of a speaker recognition system under noisy environment.In the module of feature extraction, the robustness of MFCC and delta cepstrum is focused on. The experimental results clearly demonstrate that higher order of MFCC has better robustness than lower order, delta cepstrum is robust and cepstral mean normalization(CMN) can effectively remove the effects of noise.Gaussian Mixture Models(GMM) for speaker recognition is investigated, and a modification is proposed to the standard diagonal GMM. Experiments show that the OGMM always gives a better performance.