Polyphonic transcription is specially challenging in piano higher octaves due to the complexity of the spectrum of notes and therefore, chords. Besides the fundamental and second partial components, other spectral elements appears. The three peaks related to the unison as well as the second harmonic of the fundamental unison can be distinguished in most measures. Furthermore, intermodulation components are also present when non-linearity is high enough. This paper compares several methods to improve the training process that allows to synthesize the spectral patterns and masks used in transcription methods.

Download