Unstable DAW

an AI-native DAW built in a weekend at Music Tech Hackathon Montreal (August 22-23, 2026) for the Stability AI track with a team of three. the idea: hum a tune into your laptop, get it back as notes you can edit, then ask for a band to play along.

the front half is pitch-to-MIDI. you record or upload a vocal, it gets analyzed for tempo and key (18/18 correct on both across our test fixtures), pitch-tracked, and turned into MIDI clips that land in a proper studio view: multiple tracks on a timeline, a piano roll, snap to grid, split, duplicate, undo/redo, mute/solo.

the back half is generation. you type something like "bass, drums and piano, bossa nova" and an LLM turns that into an arrangement. the trick for keeping your melody intact is a guide track: the MIDI gets rendered to audio first, and Stable Audio transforms that guide instead of generating from nothing, so the tune survives while the instrumentation changes. the output is time-aligned back onto the grid and can be split into stems with Demucs.

you can also map MIDI to sampled real instruments or describe a made-up timbre in text and have it generated. everything exports: WAV stems, MIDI, the original recordings, and the metadata for each generation.