timbrica-drum-separator

Timbrica drum-part separator (kick / snare / toms) — ONNX fp16

Three spectral U-Nets that split a drum-only track into kick, snare and toms (hi-hat + cymbals = input − sum of the three). Trained by Timbrica for the in-browser tool https://timbrica.com/ru/drum-separator (intl: https://tembrica.com/en/drum-separator). This repo is the CDN mirror for non-Russian traffic; Russian traffic is served same-origin.

Files

Snare «r2» (2026-09-19)

The first snare was cut short at 8 of 22 epochs and had only heard the kits of one soundfont. drums-snare-r2-fp16.onnx is trained on the same grooves played through those kits and a real acoustic kit (Naked Drums). Measured paired, in the shipping configuration (three masks, α-Wiener 2.0, rest = mix − sum), against the file it replaces:

The old file stays in the repo: pages cached before the switch still ask for it.

I/O (identical for all models)

• input mag — float32 [1, 2, 2048, 512]: magnitude STFT of the stereo drum mix, n_fft = 4096, hop = 1024, 2048 frequency bins (DC..Nyquist-1), 512 frames per chunk (≈ 11.9 s at 44.1 kHz).
• output mask — float32 [1, 2, 2048, 512]: soft mask in [0, 1], multiplied with the input STFT (phase of the mixture kept), inverse STFT → the part.
• fp16 = float16 weights and compute, with a Cast at the input and at the output; runs on onnxruntime-web webgpu and wasm (graphOptimizationLevel: 'basic'). Masks match the fp32 reference to 0.0002–0.002 mean abs; in the audio that is |ΔSI-SDR| ≤ 0.004 dB.

Training data & provenance

• Groove MIDI Dataset (Gillick, Roberts, Engel, Eck, Bamman — Google Magenta, 2019), CC BY 4.0 — 1 150 human drum performances as MIDI.
• Rendered to audio with FluidSynth and 8 General-MIDI kits from FluidR3_GM.sf2 (Frank Wen, MIT); 9 200 mixed tracks with isolated kick / snare / toms targets; 22 epochs on one RTX 4060.
• Snare «r2» additionally: the same grooves played through Naked Drums (Wilkinson Audio, CC BY 4.0, https://creativecommons.org/licenses/by/4.0/) — close microphones only, so that the mixture stays the exact sum of its parts.
• Architecture: the mask U-Net described in «Toward Deep Drum Source Separation» (Mezza, Giampiccolo, Bernardini, Sarti — Politecnico di Milano, 2024). The LarsNet authors' checkpoints (CC BY-NC) were not used for training or shipping — only for a local A/B listening test.

None of the training audio is redistributed here — only the trained weights.

License

Weights © Timbrica, 2026 — CC BY-NC-SA 4.0 for third parties (https://creativecommons.org/licenses/by-nc-sa/4.0/). Commercial licensing: [email protected].