含语音增强模块的i-向量说话人识别性能分析

Speech enhancement ini-vector speaker verification system

为解决文本无关说话人识别中训练与识别环境不同导致模式失配的问题,提出了一种采用语音增强模块进行前端预处理的i-向量说话人识别系统,从而提高系统对于环境噪声的鲁棒性.为评估不同语音增强算法的性能,利用NIST08核心测试集进行仿真实验.采用IMCRA算法对语音进行噪声估计后,分别用维纳滤波法、MMSE-LSA、传统谱减法和多频带谱减法等4种方法进行语音增强前端处理,在基于i-向量的说话人识别系统下进行实验.实验结果表明采用了语音增强的系统具有一定抗噪声性能,并且在高信噪比条件下,基于多频带的谱减法在此系统下性能最佳,而低信噪比情况下MMSE-LSA算法更有优势.

To solve the model-mismatch problem in text-independent speaker verification system when training environment differs from recognition environment,We propose a i-vector speaker verification system using speech enhancement in front-end preprocessing it can improve the system robustness to additive noise. To estimate the performance of different speech enhancement methods,we used NIST08 core test set in the experiment. Four speech enhancement methods,including wiener filtering,MMSE-LSA,traditional spectral subtraction and multi-band spectral subtraction,combining with IMCRA noise estimation,were evaluated in the speaker verification system based on i-vector. The result shows the proposed system with speech enhancement had some improvement in noise environment and that multi-band spectral subtraction method performed the best when SNR was relatively high and MMSE-LSA performed the best when SNR was low.