thebluescreen/oaf-drums-onnx

Onsets and Frames Drums (ONNX)

Drum transcription with velocity, trained on isolated drum audio (the Expanded Groove MIDI Dataset, 444 hours). A conversion to ONNX of the E-GMD checkpoint of Magenta's Onsets and Frames drums model (Callender, Hawthorne, Engel). Best on drum-only audio such as an e-kit recording, a drum stem or a loop.

• Input waveform: [1, samples] float32, mono, 44.1 kHz, values in [-1, 1], at least 2049 samples. The log-mel spectrogram (librosa, 250 bands, 100 frames per second) is part of the graph.
• Outputs onset_probs and velocity: [1, samples // 441 + 1, 8] at 100 frames per second for the eight trained pitches 36, 38, 48, 42, 51, 53, 49, 75 (kick, snare, toms, closed hi-hat, ride, ride bell and cowbell, crash, clave).
• A hit is an onset probability above 0.5; velocity is int(clip(v, 0, 1) * 127).
• Only the onset and velocity branches are kept. Matches the original TensorFlow graph restored from the same checkpoint to within 1e-3 (measured 4e-5).

Converted by the Silly MIDI Tools scripts in scripts/drums/.

See NOTICE.md for the attribution and the changes made, and LICENSE for the licence text.