Unsupervised Neural Transcription of Percussive Audio

Transcription of percussive audio without human-labeled data is still a challenging area of research for music information retrieval. This work presents a deep learning solution that learns to transcribe drums autonomously without requiring human-annotated datasets. The strategy uses a neural transcription model alongside a fixed synthesizer module that, collectively, iteratively improve the drum transcription by maximizing the accuracy of reconstructed audio—without any human-annotated datasets. The experimental results indicate that the unsupervised system provides performance that is on par with fully supervised models, with added scalability. These results indicate that self-supervised learning has the potential to actually improve the accuracy of transcription for drums, opening the door for its wider application to automatic music analysis and generative sound modeling.