Frequency domain multi-channel expectation maximization algorithm for audio background noise reduction

Abstract

In this paper we implement expectation maximization (EM) based methods in the short time Fourier transform (STFT) domain for background noise reduction in multi-channel systems. The models introduce a Wishart prior for the unknown signal covariance matrix. An EM algorithm is used to maximise the posterior probability for the clean signal, approaching a stationary point of the distribution with increasing iterations. The background noise is modelled as white and stationary in this initial work. The proposed methods are found to outperform a multi-channel Wiener filter in terms of residual noise artefacts and MSE for a small initial trial.