Maximum likelihood approach for blind audio source separation using time-frequency Gaussian source models

In this paper we propose a simple time-frequency Gaussian model of audio signals that allows for separation of possibly underdetermined and noisy linear instantaneous mixtures. An efficient EM algorithm is proposed to estimate the mixing matrix, the noise covariance and covariances of the source t-f coefficients over a chosen frame/subband tiling of the time-frequency domain. Results are given on 4 × 4 and 3 × 4 noisy mixtures of audio sources.