A Source Reassignment Technique for Time-frequency Masking Audio Separation

A neighborhood-based source reassignment technique is proposed for being used on time-frequency masking audio source separation methods. This technique identifies all the time-frequency clusters that form the separation masks in the Short-Time Fourier Transform (STFT) domain, and labels each time-frequency bin with a value that denotes the size of their corresponding clusters. The bins corresponding to clusters of small size are reassigned to the source which has maximum likelihood in its time-frequency neighborhood. An example is described using the DUET algorithm for underdetermined mixtures, showing that this technique improves substantially the isolation of the estimated signals.