Timbrica drum-part separator (kick / snare / toms) β ONNX fp16
Three spectral U-Nets that split a drum-only track into kick, snare and toms (hi-hat + cymbals = input β sum of the three). Trained by Timbrica on 2026-08-15 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
| file | bytes | sha256 |
|---|---|---|
drums-kick-fp16.onnx |
19681453 | f3dd5b5e3b23dafaa41d401c90595bb8ea7f121424c3b90885d751e3e96cb0c7 |
drums-snare-fp16.onnx |
19681453 | f80c524d7df71f2cfa8442c7935b581cdff28ad32157336391dd6724b58d5336 |
drums-toms-fp16.onnx |
19681453 | 35bc5fe22f162cb3d8aa9b2b17f960006d5bc9b98b6452a944e7338f0afd360e |
I/O (identical for the three models)
- input
inputβ 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 (β 11.9 s at 44.1 kHz). - output β float32
[1, 2, 2048, 512]: soft mask in[0, 1], multiplied with the input STFT (phase of the mixture kept), inverse STFT β the part. - fp16 = weights-only float16 (Castβfp32 per initializer, compute graph fp32);
runs on onnxruntime-web
webgpuandwasm(graphOptimizationLevel: 'basic'). Masks match the fp32 reference to 0.0002β0.002 mean abs.
Training data & provenance
- Groove MIDI Dataset (Gillick, Roberts, Engel, Eck, Bamman β Google Magenta, 2019), CC BY 4.0 β 1 150 human drum performances as MIDI.
- Rendered to audio with FluidSynth and 8 General-MIDI kits from FluidR3_GM.sf2 (MIT); 9 200 mixed tracks with isolated kick / snare / toms targets; 22 epochs on one RTX 4060.
- Architecture follows LarsNet (Mezza, Giampiccolo, Bernardini, Sarti β Politecnico di Milano; code MIT). The LarsNet authors' checkpoints (CC BY-NC) were not used for training or shipping β only for a local A/B listening test.
License
Weights Β© Timbrica, 2026 β CC BY-NC-SA 4.0 for third parties (https://creativecommons.org/licenses/by-nc-sa/4.0/). Commercial licensing: support@timbrica.com.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support