https://kyutai.org/blog/2026-07-10-muscriptor/ says
We're releasing MuScriptor, the best open model for multi-instrument transcription to date, created in collaboration with Mirelo. Give it a recording: pop, classical, metal, jazz, whatever, and it transcribes the notes played by every instrument into MIDI, without needing to know in advance what instruments are present.
Woah, that's much more interesting, as it seems to handle more than just the lead instrument/piano, as this Show HN does. Thanks for sharing that, very useful!
I wonder what the intent is behind releasing the weights under a CC-BY-NC license. Is it about restricting any commercial use of the weights, including for inference?
Wow, this works way better than I expected. Fed it "Shimmer and Out" by Lotus, the intro of which is just a pretty simple guitar riff based around triad shapes - but it's the kind of thing I've seen Chordify just mangle. This produced a transcription that sounds 1:1 to my ear.
Could totally see myself using this as part of the tools I use to learn music - and I could also see someone eating Chordify's lunch by hooking up a model like this and doing note-level theory analysis in addition to whatever spectral analysis model they're using.
ah, thanks mate. this isn't even their most powerful model - i'd recommend using the Mirelo API to access their most powerful variant or self-host on a GPU.
you're not wrong. the main problem is the public instance of muscriptor (https://muscriptor.kyutai.org/) is poorly handled in terms of backend infrastructure, and i often run into errors and rate-limiting. this is just a more niche version of it on my own GPU so the public can use it more.
https://kyutai.org/blog/2026-07-10-muscriptor/ says We're releasing MuScriptor, the best open model for multi-instrument transcription to date, created in collaboration with Mirelo. Give it a recording: pop, classical, metal, jazz, whatever, and it transcribes the notes played by every instrument into MIDI, without needing to know in advance what instruments are present.
Could totally see myself using this as part of the tools I use to learn music - and I could also see someone eating Chordify's lunch by hooking up a model like this and doing note-level theory analysis in addition to whatever spectral analysis model they're using.