基于联合对角化的声信号深度卷积混合盲分离方法

Deep Convolution Blind Separation of Acoustic Signals Based on Joint Diagonalization

声信号在空间中的传播具有较强的多径效应,在接收端往往以卷积形式相互叠加,尤其在海洋、剧场等强混响条件下,混合滤波器冲激响应的长度会显著增加,现有的频域卷积盲分离算法将失效。为了消除长脉冲响应导致解混合模型失效的问题,该文对观测信号进行两次短时傅里叶变换(STFT),第1次STFT缩短了脉冲响应长度,第2次STFT将信号模型转化为瞬时盲分离,最终利用联合对角化(JD)技术估计出分离矩阵。与现有方法相比,所提方法解决了深度卷积混合下模型失效的问题,并且当源信号数较多或存在加性噪声时,可以得到更好的分离性能。仿真结果验证了方法的有效性和性能优势。

The propagation of acoustic signal in space has a strong multipath effect,and the receiver often overlaps in the form of convolution.Especially in strong reverberation conditions such as ocean and theatre,where the length of impulse response of hybrid filter increases significantly.In order to eliminate the problem that long impulse response leads to the failure of the frequency domain convolution blind separation algorithm,two Short-Time Fourier Transforms(STFT)are applied to the observed signal.The first STFT shortens the length of the hybrid filter.The second STFT converts the signal model into instantaneous blind separation.Finally,the separation matrix is estimated by Joint Diagonalization(JD)technique.Compared with the existing methods,this method solves the problem of model failure under deep convolution mixing,and can obtain better separation performance when the number of source signals is large or additive noise exists.The simulation results verify the effectiveness and performance advantages of the proposed method.