HTDemucs ft (vocals specialist) - MLX
The vocals model of Meta's fine-tuned bag htdemucs_ft (Demucs v4: 4 HTDemucs models, each fine-tuned for one stem; bag weights [0, 0, 0, 1] for vocals), as a single-model MLX checkpoint for Apple Silicon. Only this one model is kept, so it runs 4x faster than the full bag. It is the same model as htdemucs_ft_vocals.onnx in kickvn/htdemucs-onnx.
Used by KICK Studio on macOS through a Swift bridge built on kylehowells/demucs-mlx-swift and mlx-swift.
Files
demucs-mlx.zip (~95 MB, sha256 c882cfa9518a6433ef6dac424c9cd5116e54055aff4a451e1d545abf26f75e51) contains a folder demucs-mlx/:
htdemucs.safetensors- fp32 weights (~168 MB)htdemucs_config.json- demucs-mlx-swift bag config with one model (num_models: 1)
The folder loads with demucs-mlx-swift as model htdemucs (pass it as the model directory). Use only the vocals stem (index 3); accompaniment = mix - vocals.
How it was made
Taken from the htdemucs_ft MLX conversion in iky1e/demucs-mlx: member 3 (vocals) is copied out unchanged, with its tensor prefix renamed model_3. to model_0. (script tools/aligner-candidates/extract_demucs_mlx_member.py, Python standard library only). No retraining, no quantization.
Checks (Apple M4)
- Vocals of a 221.6 s song: about 19-25 s, versus 39-40 s for the ONNX model on ONNX Runtime CPU.
- Against the ONNX model on the same song: correlation 0.9992, SNR 27.8 dB.
License and credits
MIT. Original model by Alexandre Defossez / Meta AI (facebookresearch/demucs). MLX conversion format by kylehowells/demucs-mlx-swift (MIT).
Citation
@inproceedings{rouard2022hybrid,
title={Hybrid Transformers for Music Source Separation},
author={Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
booktitle={ICASSP 23},
year={2023}
}