Why
Why I built it
Drum practice requires clean stems, a log of what I actually played, and a click that breathes with natural tempo drift—raw model beat predictions alone are far from a usable practice tool.
What I built
What I built
- music-sep CLI: Demucs (htdemucs_ft, CUDA) stem separation and ADTOF drum transcription.
- DrumLab: 20–55 ms edits on a 96-bar chart; localhost-only with Host/Origin checks on every write and raw hits saved as an audit trail.
- sticky-click: Beat This! downbeats smoothed by a tempo-continuity tracker that ignores fill spikes, giving a click track that follows live playing.
Role & AI
How I split the work with AI
I test every change myself behind the e-drum kit, listening for groove alignment before iterating on stems, notation, and practice flow with coding agents.
Trade-offs
What testing made me change
- Pretrained models handle separation and transcription; the real work is shaping their output into charts you can actually practice with.
- Drum fills trigger false downbeat spikes, so the adaptive click enforces bar-level continuity—always verified by listening through full songs.
Outcome
Outcome
Powers my personal stem, notation, and MIDI workflow, with the adaptive click kept as a separate listening-tested experiment.