Minuet: Musical Interference Unmixing Estimation Technique

We propose a noise cancellation technique that performs robustly in the presence of poor channel estimates and channel synchronization errors. The technique is based on the assumption that the signals have a sparse representation in a chosen signal basis, in this case, the time-frequency domain. Moreover, we assume the components of the signal of interest that contain a majority of its power overlap with components of the interference signal containing negligible power. In case of speech mixed with music, this occurs because in the time-frequency domain both music and speech are sparse and the large magnitude coefficients rarely overlap. The robustness of the technique to channel estimation and synchronization errors is demonstrated experimentally on speech/music mixtures.