DeepFilterNet
github.com
- Free plan
- Open source
- Windows
- macOS
- Linux
- API
- Self-hosted
What DeepFilterNet is
DeepFilterNet is a free, open-source framework for speech enhancement on full-band 48 kHz audio. Its command-line tool suppresses noise in WAV files and currently supports only 48 kHz sampling. A LADSPA plugin handles real-time noise suppression; with PipeWire filter chains, it can provide a virtual microphone or audio output for VoIP applications such as Zoom or Discord. The plugin's minimum latency is 20 ms from STFT processing, with extra delay possible from its LADSPA host. The Rust command-line tool uses DeepFilterNet2 by default, while the Python interface documents DeepFilterNet and DeepFilterNet2. An optional post-filter may slightly over-attenuate very noisy sections. The framework supports Linux, macOS and Windows, but training is tested only under Linux. The Python package installs as deepfilternet, with an optional train extra for data loading. Training uses HDF5 datasets and a configuration file for train, validation and test sets. Precompiled binaries and a Python wheel are available. The code is licensed under either MIT or Apache License 2.0, at the user's choice.
Who it suits
DeepFilterNet suits developers or audio users who need noise suppression for 48 kHz WAV files or a real-time LADSPA workflow. Linux users can also use its documented training tools.
Good to know, and what to watch
- Command-line WAV noise suppression
- LADSPA plugin supports real-time noise suppression
- PipeWire filter chains can create a virtual microphone
- Available on Linux, macOS and Windows
- Choose between MIT and Apache License 2.0
- Command-line tool currently supports only 48 kHz
- Training is tested only under Linux
- Plugin adds at least 20 ms processing latency
- Post-filter may over-attenuate very noisy sections
Plans
Every plan and priceDeepFilterNet
Free and open source
Open source
The verdict
DeepFilterNet offers both file-based and real-time noise suppression, with a Python interface and training path as well. Its 48 kHz requirement and plugin latency are important constraints to consider.
What the maker publishes
- Audio coverage
- The framework targets 48 kHz full-band speech audio. source
- CLI limitation
- The precompiled deep-filter command currently accepts only WAV files sampled at 48 kHz. source
- Distribution
- The project provides precompiled binaries through its GitHub releases and a Python wheel through PyPI. source
- Embedded performance
- The DeepFilterNet2 paper reports a real-time factor of 0.04 on a notebook Core-i5 CPU and describes embedded real-time use. source
- Framework platforms
- The framework supports Linux, macOS, and Windows; training is only tested under Linux. source
- Hugging Face demos
- Hendrik Schröter's Hugging Face profile lists two DeepFilterNet Spaces, including one marked Runtime error. source
- Latency
- The LADSPA plugin has a minimum latency of 20 ms from STFT processing, plus latency from the LADSPA host. source
- Latest release
- GitHub lists version 0.5.6 as the latest release and says it fixes issues 421 and 293. source
- Licence
- The repository is free and dual-licensed so users may choose MIT or Apache License 2.0. source
- License
- The repository code is free and open source under either the MIT License or Apache License 2.0, at the user's option. source
- Limit: ASR use
- A Hacker News commenter cautioned that DeepFilterNet would not improve an automatic-speech-recognition model such as Whisper. source
- Limit: Colab installation
- A 2025 GitHub issue reports metadata-generation failure while installing deepfilternet 0.5.6 in Google Colab. source
- Limit: dependency setup
- A Hacker News user reported frustrating Poetry setup involving keyring behavior over SSH. source
- Limit: package installation
- A GitHub issue reports that installing deepfilternet 0.5.3 failed because the required deepfilterlib 0.5.3 distribution was unavailable. source
- Models and processing
- The Rust command-line tool defaults to DeepFilterNet2, while the Python interface also documents use of DeepFilterNet and DeepFilterNet2 models. source
- Noise suppression
- The command-line tool suppresses noise in WAV audio files, and currently supports only 48 kHz sampling rate. source
- Noise suppression
- Its command-line tool suppresses noise in WAV recordings. source
- Package activity
- SkillFed lists the package as dormant, with its latest release dated 2023-08-31. source
- Package downloads
- SkillFed reports 78,622 monthly downloads for deepfilternet in its displayed 30-day window. source
- PipeWire integration
- The LADSPA plugin can be used with PipeWire filter chains to create a virtual microphone or audio output that suppresses noise, including for VoIP apps such as Zoom or Discord. source
- PipeWire integration
- The LADSPA plugin can be used in PipeWire filter chains to create a virtual noise-suppressed microphone. source
- Post-filter
- An optional post-filter slightly over-attenuates very noisy sections. source
- Postfilter
- An optional postfilter is documented as slightly over-attenuating very noisy sections. source
- Praise: CPU efficiency
- A Hacker News commenter reported adequate inference speed on plain CPU using an ONNX-based deployment. source
Checked against the maker's pages on 15 Oct 2026.