use case · wild
"Query by humming" used to need a server. Here it's a ~2 MB model on your device. Hum, sing, or whistle a tune; CREPE traces the pitch contour frame-by-frame, and we snap that wobbly line to a sequence of musical notes — the melody you could search a catalogue with, or drop into notation. Try the sample melody, or record your own.
Pitch contour (voiced frames, note grid):
Transcribed note sequence:
A hum isn't clean — it slides into notes and wavers. We take CREPE's per-frame f0, keep the voiced frames (confidence ≥ 0.5), round each to the nearest equal-tempered note, then merge runs of the same note (bridging tiny gaps) and drop anything shorter than your "shortest note". What's left is a note-event sequence — the same representation a query-by-humming search or a notation exporter would use. All on-device, in a Web Worker.