噪声鲁棒性说话人识别语音高频加权MFCC提取  

High frequency weighted MFCC extraction for noise robust speaker verification

作　　者：陈迪[1] 龚卫国[1] 李波[2] 

机构地区：[1]重庆大学光电技术及系统教育部重点实验室,重庆400044 [2]重庆大学现代物理中心材料科学与工程博士后流动站,重庆400044

出　　处：《仪器仪表学报》2008年第3期668-672,共5页Chinese Journal of Scientific Instrument

摘　　要：本文提出了一种可提高噪声环境下的说话人确认识别率的语音MFCC参数高频加权方法。由于Mel频率与线性频率成对数关系,频谱能量在高频部分分辨率逐减,而语音经过基音同步可变窗长加窗后的语音会在一定程度上避免语音信号的谐波泄露,从而保留更多高次谐波信息。将语音频谱能量高频部分进行加权,则可使语音增强,提高语音鲁棒性。该方法被用于基音同步预处理MFCC参数提取中,并进行了说话人确认实验。实验结果表明,即使在信噪比较低的情况下,该方法都会在一定程度上提高多种噪声环境下的说话人确认识别率。This paper proposes a high frequency weighted MFCC extraction method to improve the performance of speaker verification in noise conditions. As the Mel frequency has a logarithmic relationship with linear frequency, spectral resolution in high frequency domain would decline. Frames of purely periodic speech signal can avoid harmonic leakage, and more high frequency information would be reserved. To get speech enhancement, high frequency energy amplitude weighted method is proposed. This method was applied in pitch synchronous preproeessing MFCC feature extraction, and speaker verification experiments were conducted. The results show that the recognition rates are improved in several kinds of noise environments even when the SNR is low.

关 键 词：高频加权 说话人确认 基音同步 鲁棒性 MFCC 

分 类 号：TP192.3[自动化与计算机技术—控制理论与控制工程]