一种基于时频分析的语音卷积信号盲分离算法

A Time-Frequency Analysis Based Blind Source Deconvolution Method

本文根据语音信号在实际环境下传播存在多径效应而造成反射信号之间产生时延差异的特点,提出了一种新的基于卷积混合矩阵模型的盲分离算法.该算法利用语音信号频谱中普遍存在的稀疏特性,通过对两路接收信号语谱图的对比分析估计出模型中时延参数和幅度系数的待定值;更进一步,从极大似然估计的思想出发,构造了一种基于时延参数和幅度系数的可信度函数,提出了通过寻找上述可信度函数的峰值以准确确定混合矩阵参数的方法.与现有的基于独立性假设的盲分离算法不同,本算法利用了语音信号频谱中普遍存在的稀疏特性,适用于求解大多数场合下的盲源分离问题.由于本算法本质上是一种非迭代的算法,且不存在发散的问题,故具有快速、稳定的特点.仿真实验和实际环境下所得到的实验结果表明,该算法能在各种信噪比的条件下准确地估计出由环境所确定的时延和幅度参数,并据此成功地分离出源语音信号,是一种面向真实环境下语音盲分离应用的有效算法.

A new blind source separation algorithm for convolved mixtures of speech singals is developed with consideration of different delayed multi-path reflections. Based on the sparseness property of speech signals ,the proposed algorithm extracted time-delay and amplitude coefficient candidates through the contrast analysis of two speech spectrograms. Motivated by maximum likelihood estimation, we defined a reliablility function constructed from the parameter candidates. The reliability analysis histogram was shown to have one peak for each pair of parameters with peak location corresponding to the real time-delay and amplitude coefficient. Instead of using the independent assumption, the proposed algorithm made use of speech spectral sparseness to perform separation and proved to be applicable in most occasions. Since it involves no iterations and does not have the convergence problem, the proposed algorithm is fast and sta ble. Simulations and real-world experiments demonstrate the proposed method can accurately estimate time-delays and amplitude coefficients in various SNR conditions. Successful recovery of source signals using estimated parameters proves algorithm feasible and can be applied to real-world speech separation applications.