A Comparison of Deep Learning Methods for Timbre Analysis in Polyphonic Automatic Music Transcription

Automatic music transcription (AMT) is a critical problem in the field of information retrieval (MIR). When AMT faced with deep neural networks, variety timbres different instruments can be an issue that has not been studied depth yet. The goal this work to address by analyzing how timbre affect monophonic first approach based on CREPE network and then improve results performing polyphonic second Deep Salience model performs Constant-Q Transform. method show envelope onsets have high impact shows developed less dependent strength than other state-of-the-art models deal piano sounds such as Google Magenta Onset Frames (OaF). Our for non-piano outperforms model, bass instruments, which F-score 0.9516 versus 0.7102. In our latest experiment we also adding onset detector outperform given work.