Score-informed pitch-wise alignment using score-driven non-negative matrix factorization

This paper aims at the task of score alignment that aligns the audio recording and its corresponding score properly. Conventional methods are mostly hard to deal with the issue: the asynchrony in the recording of simultaneous notes in the score. In this paper, we propose a note-based score alignment by means of the pitch-by-time format called piano-roll feature. One of the manners is that we develop an approach of converting audio spectrogram to piano-roll-like feature. The score-driven non-negative matrix factorization is then involved in the conversion. Secondly, the pitch-wise alignment is proposed, considering each pitch sequence (i.e., the row of pianoroll) separately. In the results, about 88% of notes match their onsets deviated from ground truth for less than 50 msbased on MIDI-Aligned Piano Sounds(MAPS) database.