hn.today

Show HN: I made an audio-to-MIDI model 60x more efficient

kobi.music5 points1 comments
Screenshot of Show HN: I made an audio-to-MIDI model 60x more efficient

Dettato is an open-source audio-to-MIDI transcription model with 20 million parameters, presented as substantially more accurate and far smaller than other open systems. On the reported benchmark, it scores 60.8 composite F1, ahead of Mirelo A2M v1.1 Pro at 56.1, MuScriptor 1.3B at 47.4, and YourMT3+ at 45.7. Dettato also leads on onset, frame, offset, and drum measures, though Mirelo scores slightly higher on the RWC Classical set. Across nine datasets, Dettato performs especially well on Bach10 and synthetic Slakh2100, while results vary by repertoire. The creator cautions that benchmark scores may not predict performance for a particular use case.

The project’s central argument is that transcription quality depends more on model architecture than simply adding training data. Recasting audio-to-MIDI as a local pitch-lane detection problem, rather than a broad sequence-modeling task, enabled a compact system to perform well. Training used synthetic recordings and Creative Commons solo instruments mixed through the URMP dataset, with timbres varied using roughly 20 soundfonts; no copyrighted recordings were used. The model and its outputs are available for commercial use. Dettato has important limits: it can struggle to separate instruments, and its output is beat-aligned performance MIDI rather than playable notation. The creator frames copyright-conscious training as a spur to innovation and a way to build capable music tools that support, rather than replace, musicians.

Read on kobi.music1 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.