DeepFilterNet Documentation

To create a virtual microphone that suppresses noise for applications like Zoom or Discord, use a Pipewire filter-chain source.

1. Configure your filter-chain using the template provided in filter-chain-configs/deepfilter-mono-source.conf.
2. Run the Pipewire filter-chain module with your configuration file.

To debug the setup, you can increase the log level using the RUST_LOG environment variable.

Once datasets are prepared in HDF5 format and a configuration file is created, run the training script.

Usage: python df/train.py <DATA_CONFIG_FILE> <DATA_DIR> <BASE_DIR>

• <DATA_CONFIG_FILE>: Path to your dataset.cfg.
• <DATA_DIR>: Directory containing the HDF5 datasets.
• <BASE_DIR>: Directory where logs, checkpoints, and config will be stored.

python df/train.py path/to/dataset.cfg path/to/data_dir/ path/to/base_dir/

For developers working within the repository, follow these steps to build and install the necessary components using maturin and poetry.

1. Install build dependencies:

   pip install maturin poetry

2. Install Python dependencies via Poetry:
   • Option A (Global/Environment): poetry -C DeepFilterNet install -E train -E eval
   • Option B (In-repo development): poetry -C DeepFilterNet install -E train -E eval --no-root (requires setting PYTHONPATH=$PWD/DeepFilterNet).
3. Build and install libdf (required for enhance.py):

   maturin develop --release -m pyDF/Cargo.toml

4. (Optional) Build `libdfdata` for training functionality:
   ```bash
maturin develop --release -m pyDF-data/Cargo.toml

Note: If HDF5 errors occur, try building with the static feature: maturin develop --release --features hdf5-static -m pyDF-data/Cargo.toml

If you are using Ubuntu, you must install the following system dependencies to build and run the project:

• build-essential
• cmake
• libfontconfig1-dev
• libasound2-dev

sudo apt -y install build-essential cmake libfontconfig1-dev libasound2-dev

To ensure accurate reproduction of metrics, it is recommended to use the Python implementation installed via pip.

pip install deepfilternet

Before training, you must convert your audio files into HDF5 format using prepare_data.py.

Usage: python df/scripts/prepare_data.py --sr <SR> <TYPE> <AUDIO_FILES_TXT> <HDF5_DB>

• --sr: Sampling rate (e.g., 48000).
• <TYPE>: One of speech, noise, or rir.
• <AUDIO_FILES_TXT>: Text file containing paths to audio files.
• <HDF5_DB>: Output HDF5 filename.

pip install h5py librosa soundfile
python df/scripts/prepare_data.py --sr 48000 speech training_set.txt TRAIN_SET_SPEECH.hdf5

You can install DeepFilterNet as a Python package using pip. It is recommended to install torch and torchaudio first.

To install the standard package:

pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
pip install deepfilternet

To install with data loading functionality for training (Linux only):

pip install deepfilternet[train]

pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
pip install deepfilternet
# Or install DeepFilterNet including data loading functionality for training (Linux only)
pip install deepfilternet[train]

You can install the LADSPA plugin using one of two methods:

1. Release Build: Download the pre-compiled libdeep_filter_ladspa from the DeepFilterNet releases page.
2. Manual Build: Build the plugin from source using cargo.

Note: The plugin uses a model without lookahead, resulting in a minimum latency of 20 ms (STFT processing), plus any additional latency introduced by your LADSPA host (e.g., Pipewire).

To run the interactive demo on Linux, use the df-demo package via cargo with the ui feature enabled.

cargo +nightly run -p df-demo --features ui --bin df-demo --release

To create a virtual output that suppresses noise from applications (like browsers or video players), use a Pipewire filter-chain sink. The output will contain speech while suppressing most background noise.

1. Configure your filter-chain using the template provided in filter-chain-configs/deepfilter-stereo-sink.conf.
2. Refer to the Pipewire wiki for detailed information on Pipewire filter chains.

DeepFilterNet requires the Rust toolchain. You can install it using the official rustup script.

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh