加性白噪声环境下语音特征参数鲁棒性的研究

The Investigation of the Robust of Feature Extracted from Speech Signals in Additive Gaussian Noise Environments

随着说话人识别技术的发展,实用有效的说话人识别系统越来越成为研究的重点。语音特征参数的鲁棒性直接影响一个说话人识别系统的具体性能,过去主要针对移动通信环境下存在信道失真的问题,研究差分倒谱的鲁棒性。文中则主要在加性白噪声环境下研究M el倒谱参数、M el差分倒谱参数的顽健性以及它们经过倒谱系数零均值化(CMN)处理后识别性能的改进。从仿真结果可以看出:在加性白噪声环境下,差分倒谱参数具有很好的鲁棒性;倒谱系数零均值化能有效的除去加性白噪声。

With increasing demand for security in information system, the development of effective speaker recognition technologies is very important. The robust of feature extracted from speech signals has a direct influence on recognition system. In the past, under the circumstance of channel distortion, delta cepstrum has been widely studied. This paper focuses on the robust of feature in additive Gaussian noise environments. Experiments show that delta cepstrum is robust features in additive Gaussian noise environments, and that CMN（cepstral mean normalization） can effectively remove the effects of additive Gaussian noise.