解卷积混合语音频域盲分离的次序问题新方法

Approach to Permutation Alignment of Blind Source Separation in Frequency-Domain for Convolutive Mixing Speech

多通道语音信号的混合往往是卷积混合,瞬时盲分离方法不能获得好的分离效果,而频域方法由于频率次序的问题使性能下降。本文采用时频掩模的方法得到各频点上具有确定次序的、但带有失真的分离信号,将其作为参考,与频域上解得的次序不定信号进行相关,从而获得精确的语音分离信号。实验表明:本文提出的方法能有效地解决频域盲分离的次序不确定性问题,得到精度更高的分离卷积混合的语音信号。

In real world, multi-channel speech sources are usually in convolutive mixing environment. Instantaneous blind source separation（BSS） cannot separate the speech sources well. In frequecy-domain the synthetical performance declines due to the permutation problem at different frequency bins. This paper uses time-frequency mask idea to find some rough separated speech sources as references, in which there are some distortions without permutation ambiguity. Then the correlation among the references and the recovered sources by independent component analysis （ICA） is used to solve the permutation problem. Experimental results show that the proposed algorithm can solve permutation problem of recovered sources in frequency-domain and separate the mixed speech sources.