A Combination Method of Deep Learning Speech Separation and Independent Vector Analysis on Multi-channel Source Separation

Abstract

Multi-channel blind source separation problem is a difficult but commonly met task in daily life. How to separate every target signals well using several observed signals fascinates hundreds of researchers to study on it, especially when the input channels are less than output channels. In this paper, authors analyze the characteristics of different BSS methods and propose a new method to solve under-determined multi-channel source separation problem. The proposed method has a combination structure of deep learning model and independent component analysis algorithm. Comparisons and evaluations show that using proposed method in multichannel source separation boosts separation effect to a certain degree.