This repository contains all three model versions (v1, v2, v3), converted directly from the original PyTorch checkpoints to safetensors format for use with MLX on Apple Silicon. No fine-tuning or quantization was applied — the weights are numerically identical to the originals.

Models

Each version is stored in its own subfolder:

Version	Subfolder	Weights	Paper
DeepFilterNet v1	`v1/`	~7.2 MB (float32)	arXiv:2110.05588
DeepFilterNet v2	`v2/`	~8.9 MB (float32)	arXiv:2205.05474
DeepFilterNet v3	`v3/`	~8.3 MB (float32)	arXiv:2305.08227

Model Details

All versions share the same audio parameters:

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

Version	Embedding hidden dim
v1	512
v2	256
v3	256

Files

convert_deepfilternet.py   # PyTorch → MLX conversion script
v1/
  config.json              # v1 architecture configuration
  model.safetensors        # v1 weights
v2/
  config.json              # v2 architecture configuration
  model.safetensors        # v2 weights
v3/
  config.json              # v3 architecture configuration
  model.safetensors        # v3 weights

Usage

Python (mlx-audio)

python

from mlx_audio.sts.models.deepfilternet import DeepFilterNetModel

# Load v3 (default)
model = DeepFilterNetModel.from_pretrained("mlx-community/DeepFilterNet-mlx")

# Load a specific version
model = DeepFilterNetModel.from_pretrained("mlx-community/DeepFilterNet-mlx", subfolder="v1")

# Enhance a file
enhanced = model.enhance("noisy.wav")

Swift (mlx-audio-swift)

swift

import MLXAudioSTS

let model = try await DeepFilterNetModel.fromPretrained("mlx-community/DeepFilterNet-mlx", subfolder: "v3")
let enhanced = try model.enhance(audioArray)

Converting from PyTorch

To re-create these weights from the original DeepFilterNet checkpoints:

bash

# Clone the original repo to get the pretrained checkpoints
git clone https://github.com/Rikorose/DeepFilterNet

# Convert each version
python convert_deepfilternet.py --input DeepFilterNet/DeepFilterNet  --output v1 --name DeepFilterNet
python convert_deepfilternet.py --input DeepFilterNet/DeepFilterNet2 --output v2 --name DeepFilterNet2
python convert_deepfilternet.py --input DeepFilterNet/DeepFilterNet3 --output v3 --name DeepFilterNet3

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

Requires torch and mlx to be installed.

Origin

Original model: DeepFilterNet by Hendrik Schroeter
License: MIT (same as the original)
Conversion: PyTorch → safetensors via convert_deepfilternet.py

Citations

bibtex

@inproceedings{schroeter2022deepfilternet,
  title={{DeepFilterNet}: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering},
  author={Schr{\"o}ter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
  booktitle={ICASSP 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2022},
  organization={IEEE}
}

@inproceedings{schroeter2022deepfilternet2,
  title={{DeepFilterNet2}: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio},
  author={Schr{\"o}ter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
  booktitle={17th International Workshop on Acoustic Signal Enhancement (IWAENC 2022)},
  year={2022},
}

@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}
}