Monaural Source Separation from Musical Mixtures Based on Time-Frequency Timbre Models

We present a system for source separation from monaural musical mixtures based on sinusoidal modeling and on a library of timbre models trained a priori. The models, which rely on Principal Component Analysis, serve as time-frequency probabilistic templates of the spectral envelope. They are used to match groups of sinusoidal tracks and assign them to a source, as well as to reconstruct overlapping partials. The proposed method does not make any assumptions on the harmonicity of the sources, and does not require a previous multipitch estimation stage. Since the timbre matching stage detects the instruments present on the mixture, the system can also be used for classification and segmentation.