基于非连续多帧平滑的卷积混合盲分离频域算法

Multi-lnconsecutive-Frames Moving Average for the Frequency Domain Blind Source Separation Algorithm of Convolutive Mixtures

本文研究了卷积混合盲分离频域算法问题,基于短时傅立叶变换中各帧"部分卷积"的性质,提出了一种非连续多帧平滑的方法。该方法有助于降低源信号"短时谱"的瞬时混合和卷积混合信号的短时谱之间误差,从而提升分离性能。仿真实验证实了提出算法的有效性。

In this papor,a frequency-domain blind source separation algorithm for convolutive mixtures is proposed. Based on the property of partial convolution in each frame of Short Time Fourier Transform（ STFF）, an average moving method over several inconsecurive frames of convolutive mixtures is presented. Such method could reduce the error between the short-time spectra of convolutive mixtures and the instantaneous mixtures of short-time spectra of sources ,which contributes to improve the performance of separation. Simulation results verify the validity of proposed methods.