Original model: DeepFilterNet3 by Hendrik Schröter
Paper: DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement
License: MIT (same as the original)
Conversion: PyTorch → safetensors via the included convert_deepfilternet.py script

No fine-tuning or quantisation was applied — the weights are numerically identical to the original checkpoint.

Files

File	Description
`config.json`	Model architecture configuration
`model.safetensors`	Pre-converted weights (8.3 MB, float32)
`convert_deepfilternet.py`	Conversion script (PyTorch → MLX safetensors)

Model Details

Parameter	Value
Sample rate	48 kHz
FFT size	960
Hop size	480
ERB bands	32
DF bins	96
DF order	5
Parameters	~2M

Usage

Swift (mlx-audio-swift)

swift

import MLXAudioSTS

let model = try await DeepFilterNetModel.fromPretrained("iky1e/DeepFilterNet3-MLX")
let enhanced = try model.enhance(audioArray)

Python (mlx-audio)

python

from mlx_audio.sts.models.deepfilternet import DeepFilterNetModel

model = DeepFilterNetModel.from_pretrained(version=3, model_dir="path/to/local/dir")
enhanced = model.enhance("noisy.wav")

Converting from PyTorch

To re-create this conversion from the original DeepFilterNet checkpoint:

bash

python convert_deepfilternet.py \
  --input /path/to/DeepFilterNet3 \
  --output ./DeepFilterNet3-MLX \
  --name DeepFilterNet3

The input directory should contain a config.ini and a checkpoints/ folder from the original repo.

Citation

bibtex

@inproceedings{schroeter2023deepfilternet3,
  title={DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement},
  author={Schr{\"o}ter, Hendrik and Rosenkranz, Tobias and Escalante-B., Alberto N. and Maier, Andreas},
  booktitle={INTERSPEECH},
  year={2023}
}