Score-informed leading voice separation from monaural audio

Separating the leading voice from a musical recording seems to be natural to the human ear. Yet, it remains a difficult problem for automatic systems, in particular in the blind case, where no information is known about the signal. However, in the case where a musical score is available, one can take advantage of this additional information. In this paper, we present a novel application of this idea for leading voice separation exploiting a temporally-aligned MIDI Score. The model used is based on Nonnegative Matrix Factorization (NMF), whose solo part is represented by a source-filter model. We exploit the score information by constraining the source activations to conform to the aligned MIDI file. Experiments run on a database of real popular songs show that the use of these constraints can significantly improve the separation quality, in terms of both signal-based and perceptual evaluation metrics.