Dual-microphone noise reduction based on semi-blind DUET

Noise reduction is a fundamental requirement of many speech applications. Sometimes the major interferences, such as music, song, cross-talking, etc., coming from loudspeakers nearby make it a more challenging problem. In this paper, by supposing the locations of interferences are fixed, a semi-blind Degenerate Unmixing Estimation Technique (DUET) approach using dual-microphone is proposed. Firstly, during system initialization, the spatial feature distribution of the interferences is estimated precisely. Then the feature distribution of unknown target speech is obtained using model adaptation. Finally, a time-frequency binary mask estimated based on likelihood comparison is used to separate the target speech. The experimental results show that the performances of our approach in adverse noise reduction are significant. 77.9% interferences are eliminated and the speech quality is preserved very well. The results are close to those of un-blind approach, where the locations of all sound sources are known.