Jenila audio models
Collection
ONNX versions of open-source audio models. • 3 items • Updated
HTDemucs (Core ML, Apple Neural Engine)
Core ML (FP16, ML Program) version of the network of HTDemucs (Meta AI), the four-stem music source separation model (drums, bass, other, vocals). Weights come from the MLX checkpoint of demucs-mlx.
The model has one fixed input shape, a 7.8 s stereo segment at 44.1 kHz
(343,980 samples), and does not contain the STFT / iSTFT: the caller
computes the standardised spectrogram and rebuilds the stems. It is meant to
run on the Apple Neural Engine (cpuAndNeuralEngine).
| Input | Shape |
|---|---|
x, standardised complex-as-channels STFT (n_fft 4096, hop 1024) |
1 x 4 x 2048 x 336 |
xt, standardised waveform |
1 x 2 x 343980 |
| Output | Shape |
|---|---|
spec, per-source STFT, still standardised |
1 x 4 x 4 x 2048 x 336 |
wave, per-source time branch, still standardised |
1 x 4 x 2 x 343980 |
The three files make up one .mlpackage: htdemucs_core_manifest.json is its
Manifest.json, htdemucs_core.mlmodel is Data/com.apple.CoreML/model.mlmodel
and htdemucs_core_weight.bin is Data/com.apple.CoreML/weights/weight.bin.
| File | SHA-256 |
|---|---|
htdemucs_core_weight.bin |
ba066ced62ece0684a7f2712379ba214c1bf4e72700bbd9d0be5f3af84bf71ef |
htdemucs_core.mlmodel |
5cb09cd63734a8577025f1cd804e519b90968e778fc5b8dd5fee31fdbb509d66 |
htdemucs_core_manifest.json |
253d660963dbccad579967b72e25e4e86ceac69550dd80211cf2608403a8e5d4 |
Against the MLX stems, the output differs by 41-50 dB (stem SNR). GELU and GLU are rebuilt on the Neural Engine's tanh, and the first time decoder's transposed convolution is computed as four interleaved convolutions.
License: MIT, see LICENSE.
- Downloads last month
- 8
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support