In-depth performance analysis of the state-of-the-art algorithm for automatic drum transcription. MZ

Описание: In-depth performance analysis of the state-of-the-art algorithm for automatic drum transcription.

Authors: Mickaël Zehren (Umea Universitet, Sweden), Marco Alunno (Universidad EAFIT, Colombia) and Paolo Bientinesi (Umea Universitet, Sweden)

DMRN+18, QMUL December 2023

Abstract—In this work, we assess the most common sources of errors in a recent drum transcription algorithm.

In music information retrieval, the task of Automatic Music Transcription (AMT) is especially important because the results it produces—i.e., the notes played by the instruments—help estimating many high-level features of a
musical track, such as structure, melody, and rhythm. A subtask of AMT is automatic drum transcription in the presence of melodic instruments (DTM), which focuses on the estimation of the notes’ onsets and their corresponding drum instrument in multi-instrument tracks.
Recently, we presented a new DTM algorithm based on large supervised learning from crowdsourced annotations; thanks to the size and diversity of the datasets curated, we found that this algorithm surpasses the accuracy of
the previous methods. However, the resulting models are not perfect, as their estimations still contain mistakes.
In this work, we expose the most common sources of errors in the estimations, aiming to help the development of even more accurate models. This was done in three steps, as described in the following.


https://www.qmul.ac.uk/dmrn/dmrn18/