Integration and Adaptation of Harmonic and Inharmonic Models for Separating Polyphonic Musical Signals

摘要: This paper describes a sound source separation method for polyphonic mixtures of music to build an instrument equalizer remixing multiple tracks separated from compact-disc recordings by changing the volume level each track. Although such usually include both harmonic and inharmonic sounds, difficulties in dealing with types sounds together have not been addressed most previous methods that focused on either two separately. We therefore developed integrated weighted-mixture model consisting harmonic-structure inharmonic-structure tone models (generative power spectrogram). On basis MAP estimation using EM algorithm, we estimated all parameters this under several original constraints preventing over-training maintaining intra-instrument consistency. Using standard MIDI files as prior information parameters, applied achieved equalizer.