Real-time noise suppression that won't melt your CPU

What it does DeepFilterNet strips noise from full-band audio (48kHz) using deep filtering — a learned approach to speech enhancement. It ships as a pre-compiled Rust binary (deep-filter), a Python package, a LADSPA plugin for PipeWire, and a full training framework. The goal is perceptually clean output without the computational bloat that typically accompanies deep-learning audio pipelines.

The interesting bit The project is deliberately split: Rust handles the real-time STFT/ISTFT loop and data loading (libDF, pyDF), while PyTorch manages the neural network. This lets you run inference on a Raspberry Pi-class device or drop the Python entirely and use the standalone binary. The LADSPA plugin turns it into a virtual noise-suppression microphone on Linux — handy for video calls without buying a Shure SM7B.

Key highlights

• Three published model iterations (DeepFilterNet through v3) with arXiv papers and audio samples for each
• Pre-compiled deep-filter binary needs no Python; Python API available for GPU-backed batch processing
• LADSPA plugin integrates with PipeWire for real-time mic noise reduction
• Training pipeline supports custom datasets in HDF5 format with configurable over/under-sampling
• Cross-platform framework (Linux/macOS/Windows), though training is Linux-only

Caveats

• Standalone binary only accepts 48kHz WAV files; other formats or sample rates need the Python path
• Training setup is involved: HDF5 dataset creation, config files, and Rust build dependencies
• Demo UI requires nightly Rust toolchain and Linux

Verdict Worth a look if you need deployable, low-latency speech enhancement — especially for embedded or real-time streaming use. Skip if you just want a one-click desktop app; the project is a framework with sharp edges, not a consumer tool.