DeepFilterNet

Overview

DeepFilterNet is a speech-enhancement framework for cleaning noise from full-band 48 kHz audio. It suits developers and technically comfortable users who want to process recordings or build noise suppression into a live audio chain. Its open-source code and multiple integration routes are useful, but the 48 kHz limit and plugin latency make it a poor fit for workflows that need broader format support or minimal delay.

Key features

File processing and model choice

The command-line tool suppresses noise in WAV files and defaults to DeepFilterNet2. The Python interface documents both DeepFilterNet and DeepFilterNet2, offering a route for users building their own workflows. The command-line tool currently accepts only 48 kHz audio, so recordings at other sampling rates need another step before processing.

Real-time audio

A LADSPA plugin enables live noise suppression. Used in PipeWire filter chains, it can create a virtual microphone or audio output for apps such as Zoom and Discord. That makes DeepFilterNet adaptable to desktop audio routing, but it is not a turnkey call app: users need a compatible host and must configure the chain. The plugin has a minimum 20 ms latency from STFT processing, with additional delay from the LADSPA host, a meaningful drawback for latency-sensitive audio.

Installation and training

Precompiled binaries are distributed through GitHub releases, and the Python package installs with pip as deepfilternet. An optional train extra adds data-loading functionality for training on Linux. Training expects HDF5 datasets and a configuration file covering train, validation, and test sets. The framework supports Linux, macOS, and Windows, but training is tested only under Linux; users planning to train on another platform should not assume the same workflow.

An optional post-filter slightly over-attenuates very noisy sections. That may help suppress difficult noise, but it can also reduce speech in those passages.

Pricing

DeepFilterNet costs nothing. The code is open source under either the MIT License or Apache License 2.0, at the user's option. There is no free trial, paid tier, seat count, or usage quota to weigh against the free plan; the trade-off is that users work with a framework and its audio-routing setup rather than a packaged call interface.

Platforms

The framework supports Linux, macOS, and Windows. Training is tested only under Linux. The real-time route depends on the LADSPA plugin and, for PipeWire filter chains, PipeWire; the command-line WAV workflow is limited to 48 kHz.

Who it's for

DeepFilterNet is a good choice for developers, Linux audio users, and technically confident people who want free noise suppression in scripts, applications, or a PipeWire setup. It is less suitable for users who want a polished, guided interface, need to process audio outside 48 kHz, or cannot tolerate the plugin's minimum latency and host overhead.

Pros and cons

Pros

• Open source with no price: MIT or Apache-2.0 licensing gives users a free framework they can integrate into their own workflows.
• File and live-audio options: Command-line processing and a LADSPA plugin cover both WAV files and real-time routing.
• Multiple integration routes: Precompiled binaries and a Python package support different development and deployment approaches.

Cons

• 48 kHz ceiling: The command-line tool supports only this sampling rate, constraining file-processing workflows.
• Live setup and delay: PipeWire filter chains require configuration, and the plugin adds at least 20 ms plus host latency.
• Training platform caveat: Training is tested only under Linux, despite broader framework support.

Alternatives

NoiseTorch-ng is a free Linux option if you want a noise-suppression alternative centered on PulseAudio or PipeWire. For a free, open-source choice specifically on macOS, consider NoNoise Mac. Windows users with a supported RTX PC or workstation can consider NVIDIA Broadcast, which offers free AI audio and video effects; NVIDIA ShadowPlay is another free Windows option, but requires Windows 10 or 11, GeForce driver 551.52 or later, and a supported GPU.

Kleervox may suit users who prefer a virtual microphone for calls and want a free tier with neural noise cancellation and acoustic echo cancellation, capped at one hour of cleaning per day. DeskBroadcast is a macOS alternative with a free plan that is fully functional forever but carries a video watermark; its Pro plan costs $79, paid once for three Macs. Noise Suppression for Voice is another free option for Linux, macOS, and Windows.

SoliCall Noise Firewall is worth considering for buyers who want a paid option spanning desktop, web, and self-hosted platforms; contact SoliCall for pricing.

Verdict

Choose DeepFilterNet if you want a free, open-source framework for 48 kHz noise suppression and are prepared to manage command-line processing or configure a live audio chain. Its flexible integration paths are the main reason to choose it. Look elsewhere if you need a consumer-ready interface, other sampling rates, or lower latency in live audio.