Online estimation of inter-channel phase differences using non-negative matrix factorization

Abstract

Estimating non-linearities in phase differences between channel pairs of a multi-channel audio recording in a reverberant environment provides more precise spatial information that yields direct improvement in signal enhancement, as we show for the case of source separation. In this study, we propose an online method for estimating inter-channel phase differences (IPDs) that do not linearly depend on frequency. For this task, we use short term cross-correlation features between the input channels and extract the non-linear IPDs as well as a measure of activation for each source using non-negative matrix factorization. Our evaluation shows that the proposed method outperforms a state-of-the-art approach based on linear phase differences by 9.4% relative improvement in signal-to-interference ratio on average. Furthermore, increasing the window length of temporal context used for the decomposition increases the source separation accuracy, which converges to the accuracy of the offline method.