A Hierarchical Bayesian Framework for Score-Informed Source Separation of Piano Music Signals

Here we propose a score-informed monaural source separation system to extract every tone from a mixture of piano tone signals. Two sinusoidal models in our earlier work are employed in the above-mentioned system to represent piano tones: the General Model and the Piano Model. The General Model, a variant of sinusoidal modeling, can represent a single tone with high modeling quality, yet it fails to separate mixtures of tones due to the overlapping partials. The Piano Model, on the other hand, is an instrument-specific model tailored for piano. Its modeling quality is lower but it can learn from training data (consisting entirely of isolated tones), resolve the overlapping partials and thus separate the mixtures. We formulate a new hierarchical Bayesian framework to run both Models in the source separation process so that the mixtures with overlapping partials can be separated with high quality. The results show that our proposed system gives robust and accurate separation of piano tone signal mixtures (including octaves) while achieving significantly better quality than those reported in related work done previously.