Multi-channel Itakura Saito Distance Minimization with Deep Neural Network

Abstract

A multi-channel speech source separation with a deep neural network which optimizes not only the time-varying variance of a speech source but also the multi-channel spatial covariance matrix jointly without any iterative optimization method is shown. Instead of a loss function which does not evaluate spatial characteristics of the output signal, the proposed method utilizes a loss function based on minimization of multi-channel Itakura-Saito Distance (MISD), which evaluates spatial characteristics of the output signal. The cost function based on MISD is calculated by the estimated posterior probability density function (PDF) of each speech source based on a time-varying Gaussian distribution model. The loss function of the neural network and the PDF of each speech source that is assumed in multi-channel speech source separation are consistent with each other. As a neural-network architecture, the proposed method utilizes multiple bidirectional long-short term memory (BLSTM) layers. The BLSTM layers and the successive complex-valued signal processing are jointly optimized in the training phase. Experimental results show that more accurately separated speech signal can be obtained with neural network parameters optimized based on the proposed MISD minimization than that with neural network parameters optimized based on loss functions without spatial covariance matrix evaluation.