改进的在线自然语音卷积混合信号时域盲分离方法

Online Time-Domain Algorithm for Blind Separation of Convolution Speech Signals

针对语音信号所具有的非平稳性和时域相关性,提出了一种新的卷积混合语音信号盲分离的在线时域算法。该算法通过利用分块处理方法和带遗忘因子更新的非完备约束条件及其推广,对于许多已有在线算法中存在的由于目标源数目随时间不断变化而产生的不稳定性问题,以及语音信号时域相关性而导致的恢复信号失真问题进行了改进,最后通过仿真,结果表明,本文方法可以有效地处理语音卷积信号的在线盲分离问题,同时在源数目变化时算法的鲁棒性较好。

Aimed at non-stationary and ttme-correlation property of new online time-domain blind separation algorithm is proposed for convolutive mixtures of natural speech. Based on block processing technique, the nonholonomic constraint updated with forgetting factor and its generalization, the algorithm provides a solution to avoid the limitations in most traditional methods, such as the severe instability problem caused by varying number of the original sources during the iteration process and the separated signals distortion resulting from the time-correlation property of speech. Experimental results confirm the efficient and robust convergence performance of the new approach.