Score-Informed Source Separation Based on Real-Time Polyphonic Score-to-Audio Alignment and Bayesian Harmonic Model

This paper proposes a system on the basis of guidance information from music score and Bayesian harmonic model and a two-dimensional Hidden Markov (2D-HMM) states model with particle filtering to address the separation of single-channel polyphonic music source. It is showed in a large number of experiments that in recording and synthetic polyphonic music material, the informed separation method performs well in objective performance and subjective listening experience.