Modelpublic

mlx-community/SoulX-Singer-bf16

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

1likes93downloads

Model Card

SoulX-Singer-bf16

Apple MLX safetensors checkpoint for `Soul-AILab/SoulX-Singer`, including both SoulX-Singer SVS and SoulX-Singer-SVC weights.

Collection: https://huggingface.co/collections/mlx-community/soulx-singer-mlx-6a1c0525d0911ea400102840

This is not a pure end-to-end MLX audio runtime yet. The weights are converted to an MLX-friendly safetensors layout, while full audio generation currently uses the official PyTorch model structure through `ailuntx/SoulX-Singer-MLX`.

TL;DR


Precision	`bf16`
Disk	2.6G
Components	`svs/` and `svc/`
Runtime / bridge	`ailuntx/SoulX-Singer-MLX`
Official model	`Soul-AILab/SoulX-Singer`
Official code	`Soul-AILab/SoulX-Singer`

Quick Start

bash

git clone https://github.com/ailuntx/SoulX-Singer-MLX.git
cd SoulX-Singer-MLX
conda create -n soulxsinger-mlx -y python=3.10
conda activate soulxsinger-mlx
python -m pip install -U pip
python -m pip install -r requirements.txt mlx safetensors huggingface_hub hf_transfer

HF_HUB_ENABLE_HF_TRANSFER=1 hf download mlx-community/SoulX-Singer-bf16 --local-dir ./models/SoulX-Singer-bf16
hf download openai/whisper-base --local-dir pretrained_models/openai__whisper-base

Run a short SVS bridge test:

bash

PYTORCH_ENABLE_MPS_FALLBACK=1 \
SOULX_WHISPER_MODEL=pretrained_models/openai__whisper-base \
python scripts/inference_mlx_bridge.py \
  --model ./models/SoulX-Singer-bf16 \
  --component svs \
  --device mps \
  --prompt_wav_path example/audio/zh_prompt.mp3 \
  --prompt_metadata_path example/audio/zh_prompt.json \
  --target_metadata_path example/audio/zh_target.json \
  --control melody \
  --n_steps 1 \
  --cfg 1 \
  --save_dir outputs_mlx_bridge/svs

Run an SVC bridge test:

bash

PYTORCH_ENABLE_MPS_FALLBACK=1 \
SOULX_WHISPER_MODEL=pretrained_models/openai__whisper-base \
python scripts/inference_mlx_bridge.py \
  --model ./models/SoulX-Singer-bf16 \
  --component svc \
  --device mps \
  --prompt_wav_path example/audio/zh_prompt.mp3 \
  --target_wav_path example/audio/music.mp3 \
  --prompt_f0_path example/audio/zh_prompt_f0.npy \
  --target_f0_path example/audio/music_f0.npy \
  --n_steps 1 \
  --cfg 1 \
  --save_dir outputs_mlx_bridge/svc

Variants

Variant	Disk	Notes
`SoulX-Singer-4bit`	774M	smallest checkpoint; MLX affine quantized
`SoulX-Singer-8bit`	1.4G	smaller checkpoint; MLX affine quantized
`SoulX-Singer-bf16`	2.6G	recommended high-quality baseline
`SoulX-Singer-fp32`	5.6G	full-precision conversion baseline

Layout

text

SoulX-Singer-bf16/
|-- config.json
|-- config.yaml
|-- mlx_manifest.json
|-- svs/
|   |-- model.safetensors.index.json
|   `-- model-00001-of-000xx.safetensors
`-- svc/
    |-- model.safetensors.index.json
    `-- model-00001-of-000xx.safetensors

This repo has 11 SVS shard(s) and 11 SVC shard(s).

Validation

Local Apple Silicon validation:

Test	Result
Official PyTorch/MPS SVS, `n_steps=1`	generated 6.71s WAV, 24kHz mono
`SoulX-Singer-bf16` component load	`svs/` and `svc/` safetensors are indexed and loadable
bf16 PyTorch bridge SVS, `n_steps=1`	generated 6.71s WAV, 24kHz mono, RMS about `0.029`

For quantized checkpoints, the bridge loader dequantizes MLX affine tensors into the official PyTorch module shapes for compatibility testing. Native all-MLX inference is planned as a later runtime step.

License

The converted weights follow the upstream SoulX-Singer Apache-2.0 release.

Citation

bibtex

@misc{soulx-singer-mlx,
  title  = {SoulX-Singer-MLX: Apple MLX safetensors port of SoulX-Singer},
  author = {ailuntx},
  year   = {2026},
  url    = {https://github.com/ailuntx/SoulX-Singer-MLX},
}

@misc{soulxsinger,
  title={SoulX-Singer: Towards High-Quality Zero-Shot Singing Voice Synthesis},
  author={Jiale Qian and Hao Meng and Tian Zheng and Pengcheng Zhu and Haopeng Lin and Yuhang Dai and Hanke Xie and Wenxiao Cao and Ruixuan Shang and Jun Wu and Hongmei Liu and Hanlin Wen and Jian Zhao and Zhonglin Jiang and Yong Chen and Shunshun Yin and Ming Tao and Jianguo Wei and Lei Xie and Xinsheng Wang},
  year={2026},
  eprint={2602.07803},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2602.07803},
}