deepfilternet-slim

deepfilternet-slim distributes the DeepFilterNet C runtime and the standard DeepFilterNet3 ONNX model in platform wheels. It does not depend on PyTorch and does not require a system installation of DeepFilterNet.

The package is intended to be the native runtime layer used by MushroomAgent. Applications receive stable paths from a small typed Python API instead of asking users to configure platform-specific library and model paths.

Installation

pip install deepfilternet-slim

Usage

from deepfilternet_slim import runtime_assets

assets = runtime_assets()
print(assets.library_path)
print(assets.model_path)

The wheel currently contains:

• the DeepFilterNet C API runtime built from the pinned main commit d375b2d8309e0935d165700c91da9de862a99c31 (not the v0.5.6 release tag);
• DeepFilterNet3_onnx.tar.gz, verified with a pinned SHA256 checksum;
• the upstream MIT and Apache-2.0 license texts.

The upstream capi feature normally embeds the default model in the shared library while also accepting an external model path. The build stages a clean copy of libDF and removes that feature edge, so the model is included exactly once in each wheel.

The dedicated Cargo lock updates the compatible tract runtime to 0.21.17. That version avoids selecting Apple AMX kernels on virtualized macOS arm64 runners, where those instructions are unavailable.

Supported wheels

• Linux x86_64
• Linux AArch64
• macOS x86_64
• macOS arm64
• Windows AMD64

Linux wheels are built in the manylinux environment selected by cibuildwheel. The Python module has no CPython ABI dependency, so one py3-none-<platform> wheel serves every supported Python version.

Development

Clone with the pinned upstream source:

git clone --recurse-submodules https://github.com/kiwi-data/deepfilternet-slim.git
cd deepfilternet-slim
python -m build --wheel
python -m pip install --force-reinstall dist/*.whl
pytest -q

Configuration fails if the submodule commit or model checksum differs from the values pinned by this repository.

Releasing

Push a v* tag. GitHub Actions builds and tests every platform wheel, then publishes them to PyPI using Trusted Publishing. Configure a PyPI project named deepfilternet-slim with the GitHub pypi environment before the first release.

Upstream attribution

DeepFilterNet is developed by Hendrik Schröter and contributors and is available from https://github.com/Rikorose/DeepFilterNet. See NOTICE and the license files bundled in each wheel.