Separation and Deconvolution of Speech Using Recurrent Neural Networks

摘要

This paper focuses on improvements of the Speech Recognition or Speech Reading (SR), due to combining multiple auditory sources. We present results obtained in the traditional Blind Signal Separation & Deconvolution (BSS) paradigm using two speaker signals from the perspective of two sources, investigating artificial linear and convolutive mixes as well as real recordings. The adaptive algorithm is based on two-dimensional (2D) system theory using recurrent neural networks (RNNs). The characteristics of convolutively mixed signals (eg. audio signals) are matched by the structure of RNNs. The feedback paths in an RNN permit the possibility of a memory of the signals at relevant delays so that better separation can be achieved. The cross-correlations of the outputs of the RNN are used as separation criterion.