A Divide and Conquer Approach to Automatic Music Transcription Using Neural Networks

摘要

This paper describes a new approach for the automatic music transcription problem. We take advantage of the divide and conquer design paradigm and create several artificial neural networks, each one responsible for transcribing one musical note. This way, we depart from the traditional approach which resorts to a single classifier for transcribing all musical notes. To further improve results, an additional post-processing stage using artificial neural networks with the same design paradigm is also proposed. This last stage comprises three main steps: (1) fix notes duration, (2) fix notes duration regarding onsets and (3) fix onsets. The obtained results show that these steps were essential to improve the final transcription. We also compare our results with existing neural network-based approaches. Our approach is able to surpass current state-of-the-art works in frame-based results and, at the same time, reach similar results in onset only, thus demonstrating its viability.