Audio Blind Source Separation Based on Mutual Information Estimation

Audio Blind Source Separation aims to retrieve original independent source signals from observed mixed signals, eliminating interference and providing purer input signals for the recognition system to improve accuracy. Many studies have obtained promising performance in recent years, however, when the signal contains a high level of noise, most methods fail to achieve a satisfactory separation performance. To address the issue, this paper proposes an audio blind source separation method based on manual information estimation that separate clean audio sources from noisy signal mixtures. Specifically, it includes two modules termed Corr-Feature module and MI-Sep module. CorrFeature module enhances the correlation features by estimating the correlation between the features and the input to enhances the noise immunity; MI-Sep module optimizes mutual information between the estimated signals to improve the separation performance. The experiments are conducted using the WSJ02mix and WSJ0-3mix datasets for training. The results demonstrate that this method effectively enhances the audio separation performance and exhibits good performance in noisy audio data.