频域盲解卷积局限性分析及一种改进算法

Limitations of Frequency-Domain Blind Deconvolution and an Improved Algorithm

在频域盲解卷积问题中,时域信号的卷积混合转化为频域信号在有限频点的瞬时混合,使算法复杂度大大降低。但这种算法的局限是分离结果存在次序和幅度上的不确定性,并且窗函数长度和信号非平稳性之间存在相互制约的关系。文中对语音信号频域盲解卷积算法存在的制约因素进行分析并提出一种改进的基于包络相关性的排序方法。在分裂谱法的基础上,通过"分裂"后的多路信号求得"总包络",再依据"总包络"进行排序,从而克服传统的直接依据输出信号包络相关性进行排序的不足。实验结果表明,采用本方法可获得较高的分离质量。

In frequency-domain blind deconvolution, the convolutive mixture of time-domain signals is converted to instantaneous mixtures on several frequent bins, which makes the algorithms much less computational. But this kind of approach has its basic limitations that the results of separation are indeterminate in their scaling and permutation, further more, the length of window function is added to the observation signals and the non-stationary characteristic of speech signal are restricted with each other. Analyzes the limitations of 1713-blind deconvolution algorithms and proposes an improved permutation algorithm based on the correlation of envelops on each frequent bin. On the basis of ＂split spectrogram＂ algorithm,get the ＂general envelop＂ from the splited signals and remove the permutation ambiguity by measure the ＂general envelop＂. This overcomes the deficiency of traditional algorithms which measure the correlation of envelops of output signals directly. The experiment results show that the algorithm proposed in this paper can get a relatively high separation quality.