噪声环境下基于稀疏表示的说话人识别

Sparse representation based speaker recognition in noisy environment

在噪声环境下,稀疏表示方法并没有表现出它出色的区分性能,反而由于特征的分散导致性能的大幅下降。根据语音特征参数之间的相关性,提出了一种适用于稀疏表示说话人识别的全局补偿方法。该方法对不同阶特征参数进行逐一分析,目的是为了找出被噪声影响最严重的一阶参数并去除之,以此增强测试语音与训练语音之间的相关性。理论分析和实验结果表明,该方法具有很好的抗噪性能,在信噪比为5d B时,带有白噪声的语句识别率达到了85.7%,而在高信噪比时,其识别率能够达到97.5%,几乎等同于干净语音的识别率。

Sparse representation based speaker recognition cannot perform well where environment noise exists and even decreases sharply because of the scatter of features. Utilizing the correlation of utterances, this paper proposes a universal compensation method for sparse representation based speaker recognition. The method is used to analyze each feature vector members one by one and to find the most corrupted one and remove it, thus the correlation of the test utterances and training utterances is enhanced. According to theoretical analysis and simulation results, the method can improve the robustness for environment noise of the speaker recognition system based on sparse representation. The accuracy is 85.7% when the SNR equals 5dB. And the accuracy can even reach 97.5% in a high SNR environment equaling to the accuracy of recognizing a clean utterance.