改进最大信噪比的独立成分分析单通道语音增强算法

针对现有基于独立成分分析(ICA)的盲源分离算法在单通道语音增强中的不稳定性和信噪比低的问题,提出了新的基于最大信噪比的ICA语音增强算法.该算法首先用带噪语音直接乘以二维向量,并经过列满秩的转换,得到既具有源信号特性、又不会引入新噪声的二路观测信号,保证了系统的稳定性;再结合用小波系数改进的最大信噪比的ICA算法来实现,为增强的效果和提高信噪比提供了依据.实验结果分析表明,该算法是稳定的,且能有效地提高信噪比的值. 

An Improved Single-Channel Speech Enhancement Algorithm with Maximum SNR Based on Modified Independent Component Analysis

Considering the shortcomings of unstable and low SNR of basic ICA (independent component analysis) in single-channel speech enhancement algorithm, a new ICA algorithm with maximum SNR for speech enhancement is presented in this article. By directly multiplying noisy speech with two-dimensional vectors and converting it into full column rank matrix, the two observed signals could be obtained. The obtained two observed signals have the properties of source signal without adding new noise so that it could ensure the system stability. ICA algorithm, which is modified by wavelet coefficients based on maximum SNR, is combined with the new two observed signals to realize the high efficiency and maximum SNR of the system. The experiments show that the presented algorithm is stable and it improves the SNR obviously.