Overlapped event-note separation based on partials amplitude and phase estimation for polyphonic music transcription

We propose a discriminative model for polyphonic music transcription that deals with the well-known overlapped partial problem by taking into account the instrument envelope pattern for each note. The process to obtain the music scene-adaptive envelope patterns for each note is detailed. Firstly, spectral features are obtained individually for each note. Then, support vector machines (SVM) are trained on the notes energy. We apply a scheme of one-versus-all (OVA) SVM classifiers to make an approximation of the active frame-level note instances. Finally, amplitudes and phases are estimated by considering the envelope patterns for different notes, distributing the energy according to the note estimated envelope pattern adjustment. Also, temporal information is added by introducing Hidden Markov Models. Our approach has been tested with synthesized and real music recordings, obtaining promising results.