噪声环境下的鲁棒性说话人识别系统研究

In practical applications, noise or channel interference resulted in great degradation in the performance of Speaker Recognition (SR) system,to solve this problem, we proposed a system design which use Wiener filter and cepstral mean normalization (CMN) in the front-end of the SR system. Regard MFCC and first&second order MFCC as different features of voice,use the Gaussian mixture model (GMM) train them separately and combined the final score to identify the speaker. The experimental results show that the Wiener filter and cepstral mean normalizationcan effectively weaked the impact of noise, hybrid modeling of the MFCC parameters can help distinguish the speakers better and increase the robustness of the system under noisy environment , the overall performance has been greatly improved.