A selection function for pitched instrument source separation

Yukai Gong, Longquan Dai, Jinhui Tang

Abstract: There exist a large number of methods for pitched instrument source separation. The core problem is to separate the time-frequency overlapping harmonics. To yield better results, we propose a function to select fine harmonic separation results from existing methods. Our strategy is based on the discovery that a fine separation result usually has a low total amplitude fluctuation. For source harmonics overlapped in a frequency band, each method produces a separation result. Employing the harmonics separated by each method, we can estimate the total amplitude fluctuation of each group of overlapping harmonics. Our selection function maps the band index to the method index by selecting the method with the minimum total amplitude fluctuation. Experiments are conducted on sample mixtures from the University of Iowa Musical Instrument Sample Database. Three advanced separation techniques are compared, including common amplitude modulation (CAM), harmonic bandwidth companding (HBW-comp) and ideal binary mask (IBM) filtering. Experiment results indicate that the proposed selection function is able to boost the separation performance significantly.