DeepFilterNet

Overview

DeepFilterNet is an open-source speech-enhancement framework for processing recordings and building live audio chains. It suits technically comfortable users who want noise suppression they can integrate into their own desktop workflows. Its flexibility and zero price are appealing, but a 48 kHz input ceiling and plugin delay rule it out for some audio setups.

Key features

File processing and models

The command-line tool suppresses noise in WAV files and defaults to the DeepFilterNet2 model. The Python interface documents both DeepFilterNet and DeepFilterNet2, making it the more adaptable route for users building custom processing workflows. The command-line tool currently supports only 48 kHz audio, so recordings at other sample rates need to be converted before processing.

Live noise suppression

A LADSPA plugin enables real-time suppression. With PipeWire filter chains, it can create a virtual microphone or audio output for use with VoIP apps such as Zoom and Discord. That is useful for users willing to configure their audio routing, but this is not a ready-made call interface. The plugin adds at least 20 ms of latency from STFT processing, before the LADSPA host's own delay; that can be a drawback for latency-sensitive workflows.

Installation and training

Precompiled binaries are available through GitHub releases, and the Python package installs with pip as deepfilternet. An optional training extra adds data-loading functionality, but training is only tested under Linux. Training also requires HDF5 datasets and a configuration file defining the train, validation and test sets, so it is aimed at users prepared to manage a machine-learning data pipeline.

Post-filter

An optional post-filter slightly over-attenuates very noisy sections. It may be useful when stronger suppression matters more than preserving every detail in difficult passages, but that trade-off makes it less suitable when natural-sounding output is the priority.

Pricing

DeepFilterNet is free and open source. Its sole plan costs nothing and uses the MIT License or Apache License 2.0, at the user's option. There is no paid tier or free-trial period to weigh against it. The zero price makes it accessible for individual projects, while users must handle their own installation and audio configuration rather than relying on a turnkey call app.

Platforms

The framework supports Linux, macOS and Windows, although training is only tested under Linux. The real-time plugin's PipeWire workflow is specifically relevant to Linux audio routing. Precompiled binaries and a Python wheel provide installation routes, but DeepFilterNet's 48 kHz processing limit applies to the command-line tool.

Who it's for

DeepFilterNet is a good fit for developers, audio enthusiasts and other technically confident users who want to process 48 kHz WAV recordings, integrate suppression into a Python workflow, or configure live noise reduction through PipeWire. It is a weaker choice for people who want a simple call app, need to process audio at other sample rates directly, or cannot tolerate the plugin's baseline delay.

Pros and cons

• Pro: Free, open-source licensing under either MIT or Apache 2.0 lets users choose a license for their project.
• Pro: File processing, a Python interface and a LADSPA plugin support several ways to incorporate noise suppression.
• Pro: PipeWire chains can supply a virtual microphone or output to VoIP apps, including Zoom and Discord.
• Con: The command-line tool handles only 48 kHz WAV audio, limiting direct use with recordings at other sample rates.
• Con: The plugin adds at least 20 ms of processing latency, with more delay from its host.
• Con: Training entails HDF5 datasets and configuration work, and is only tested under Linux.

Alternatives

Kleervox is worth considering for users who want a virtual microphone for calls, neural noise cancellation and acoustic echo cancellation, with a free plan capped at 1 hour of cleaning per day. Noise Blocker is a Windows alternative with a 30-day business trial for up to 20 employees or computers, subject to request and approval. NVIDIA ShadowPlay is another free Windows option, provided the system meets its Windows 10 or 11 and GeForce driver requirements and has a supported GPU.

DeskBroadcast is a macOS alternative with a free version and a one-time Pro plan priced at $79, paid once for three Macs. Linux users seeking a simpler noise-suppression option can consider NoiseTorch-ng, which supports PulseAudio or PipeWire. NoNoise Mac is a free, open-source option for macOS. Krisp AI Meeting Assistant offers a seven-day free trial for readers considering a broader meeting-focused tool. Noise Suppression for Voice is a free option for Linux, macOS and Windows.

Explore more options in AI Noise Cancellation Software and AI Speech Enhancement Software.

Verdict

Choose DeepFilterNet if you want free, open-source speech enhancement that you can incorporate into file processing, Python workflows or a PipeWire audio chain. Its integration options are the main reason to choose it; the 48 kHz command-line limit, configuration demands and plugin latency are the reasons to look elsewhere if you need broad-format processing or a low-delay, ready-to-use call tool.