基于盲源分离的单通道语音信号增强

Single Channel Speech Enhancement Based on Blind Source Separation

在运用基于独立分量分析(ICA)的盲源分离法进行语音增强时,要求观测信号(含噪语音)的个数不少于源信号(纯净语音和噪声)的个数。由于含噪语音通常是单通道的,所以必须合理地生成另一路的虚拟观测信号,以实现纯净语音和噪声的分离是个关键。介绍了一种基于盲源分离和谱减法的单通道语音信号增强的方法。首先运用谱减法对语音进行部分去噪,产生了ICA其中的一路观测信号,并产生了对噪声的估计值。用语音和噪声估计值的帧平均能量构成了加权函数,将噪声的估计值与原始含噪语音进行加权组合,生成另一路的虚拟观测信号。由于虚拟观测信号很好地再现了实际的观测信号,所以运用ICA可以较好地实现了噪声和语音的分离。同时,盲源分离和谱减法相互结合,使语音增强的性能提高。实验证明了算法可以在信噪比很小的情况下实现对噪声的去除,其效果要优于传统的去噪算法。

In the speech enhancement,when using Blind Source Separation based on Independent Component Analysis,the number of observed signals（noised speech signals） must not be less than the number of source signals（pure speech and noise）.As the noised speech is usually a single channel signal,so how to make another hypothetical observed signal and to separate the pure speech and noise is an important thing.This paper proposes a method of single channel speech enhancement based on blind source separation and spectral subtraction.At first,by denoising the speech with the method of Spectral Subtraction,an observed signal and the estimated noise are obtained.And another hypothetical observed signal is generated by using a weighted sum of the estimated noise and the original noised speech,where the weighted function is obtained from the average power of the speech and the estimated noise.As the hypothetical observed signal can better represent the real observed signal,the noise and pure speech can be separated better using independent component analysis.And combining the Blind Source Separation with Spectral Subtraction,the system gets a better performance.Experimental result shows that the algorithm is more efficient than other conventional algorithms even in a low signal-noise ratio.