Classification-based Note Tracking for Automatic Music Transcription

Abstract

Note tracking constitutes a key process in Automatic Music Transcription as it derives a note-level transcription from a frame-based pitch activation representation. While this stage is commonly performed using a set of hand-crafted rules, this work presents an approach based on supervised classification which automatically infers these policies. An initial frame-level estimation provides the necessary information for segmenting each pitch band in single instances which are later classified as active or non-active note events. Preliminary results using classic classification strategies on a subset of the MAPS piano dataset report an improvement of up to a 15% when compared to the baseline considered for both frame-level and note-level assessment.