一种基于多判据综合决策的语音信号频域盲解卷积排序算法

A Synthetical Algorithm Based on Multiple Criterions to Solve Permutation Ambiguity in Audio Signal Frequency-Domain Blind Deconvolution

在基于子带ICA的语音信号频域盲解卷积问题中,通过短时傅立叶变换把信号从时域卷积混合转化为各频点的瞬时混合将导致分离结果出现次序和幅度上的不确定性.在传统排序方法的基础上提出一种改进的基于多判据综合决策的排序算法,同时通过实验探讨了窗函数长度和信号非平稳性之间的关系.仿真实验表明,多判据决策有效克服了各种传统单一方法的局限,使得排序精度和最终分离质量获得提高.

In audio signal frequency-domain blind deconvolution based on subband ICA,the transformation of signals from convolutive mixture in time-domain to instantaneous mixtures in frequency-domain may cause indeterminacies in their scaling and permutation.Based on traditional permutation alignment algorithms,this paper proposes a synthetical permutation alignment algorithm based on multiple criterions.The relationship between the length of window function and the non-stationary characteristic of speech signal is explored by experiments as well.The simulation shows that the deficiency of traditional algorithms based on single criterion can be overcame by this approach and the quality of permutation accuracy and separation results is improved greatly.