AudioForge

Remove background noise and polish your voice for calls, streams, and recordings. Everything is processed on your Windows PC.

AudioForge sits between your microphone and the apps you talk in. It cuts background noise, evens out your level, tames harsh "s" sounds and sudden peaks, and lets you shape your tone. A guided setup tests its suggestions on your own recording before you apply them. No account and no cloud service: your audio never leaves your PC.

Download for Windows · Set up in five steps · What it does · No sound?

Important

To use AudioForge in Discord, OBS, Zoom, or Teams, you also need a virtual audio cable. AudioForge doesn't install one. It is tested with the free VB-CABLE; other virtual cables should work but haven't been tested. You don't need a cable to monitor yourself on headphones.

What It Does

Your goal	What AudioForge gives you
Remove background noise	RNNoise or DeepFilterNet noise suppression, a speech-aware noise gate, and hum cleanup.
Sound consistent	A compressor with speech-aware automatic makeup gain, plus peak protection that stops clipping.
Shape your voice	A ten-band EQ you can drag or type into, Auto-EQ, and guided Auto Voice Setup.
Tame harshness	A de-esser for sharp "s" and "sh" sounds and persistent harshness.
Stay in control	Presets, undo and redo, a global mute shortcut, level meters that flag clipping, and health advice that says what to change.

More screenshots: processing cards and Auto Voice Setup

Processing cards

Auto Voice Setup

Screenshots use example devices and settings.

Download

Current version: v1.18.2 (release notes). This is the source-checkout version; the latest published package may be older. Open the latest release and download one of these files. Not sure? Choose the installer.

• Installer (recommended): AudioForge-v<version>-win64.msi, where <version> is the release number. Run it, then open AudioForge from the Start menu. It installs for your Windows user only.
• Portable folder: AudioForge-v<version>-win64-ultra.7z. Extract the whole archive, then run AudioForge.exe. If Windows can't open .7z files, use 7-Zip. "Ultra" is the compression level, not a different edition.

Both include everything the app needs; you don't need Python or Rust. The other files on the release page (source code, validation evidence, and checksums) are for verification and redistribution.

Verify your download: the executable, the DLLs and the MSI aren't code-signed, so Windows SmartScreen may show "Windows protected your PC" the first time you run them. Download only from this repository's Releases page, then check the file two ways:

• Run Get-FileHash -Algorithm SHA256 <file> in PowerShell and compare the result with the matching line in AudioForge-v<version>-SHA256SUMS.txt.
• With the GitHub CLI, run gh release verify-asset v<version> <file> --repo FueledByRedBull/audio-forge. It checks the file against the release's signed attestation.

If both match, choose More info > Run anyway.

Set Up

1. Install a virtual cable (once). Install VB-CABLE and restart if its installer asks you to. Skip this step if you only want to listen on headphones.
2. Choose devices in AudioForge. Set Input to your microphone and Output to CABLE Input (VB-Audio Virtual Cable). To hear yourself first, choose your headphones as Output instead.
3. Point your app at the cable. In Discord, OBS, Zoom, or Teams, set the microphone (input device) to CABLE Output (VB-Audio Virtual Cable).
4. Start and check. Press Start and speak. The AudioForge meters and your app's input meter should both move, and a short test recording in your app should contain your processed voice.
5. Tune and save. Noise suppression starts with RNNoise, the low-latency default. Run Auto Voice Setup for a guided starting point, or adjust the EQ and dynamics yourself. Setup starts with a three-second check of your microphone level and says how much to raise or lower it when a change helps. Press Ctrl+S to save a preset when it sounds right.

Opening AudioForge normally leaves processing stopped. Stop keeps the window open. While audio runs, closing the window, Ctrl+Q and Quit AudioForge in the tray ask Stop Audio and Quit? You can quit, keep running in the tray (offered only while the tray icon shows) or cancel. Don't ask again skips the question, and Ask before quitting while audio is running under Settings > Tray and background brings it back. With close-to-tray turned on in the same menu, closing the window hides it and audio keeps running; Ctrl+Q and Quit AudioForge then quit without asking. Signing out of Windows or shutting it down never asks.

Login startup is opt-in and off by default. Packaged copies offer Settings > Tray and background > Configure login shortcut for this copy. When launched by that shortcut, AudioForge stays in the tray and waits up to 60 seconds for your exact saved input and output devices before starting, preserving output mute. Missing devices leave processing stopped. Windows Settings or Task Manager can disable the shortcut; AudioForge reports it as configured without overriding that choice. Before moving a portable copy, remove its shortcut with the same switch and configure it again from the new location. See the startup lifecycle and testing limits.

No sound?

• Check that processing is started and Mute is off.
• Check the pair: AudioForge's Output is CABLE Input, and your app's microphone is CABLE Output. The two names are easy to swap.
• In Windows Settings, microphone privacy must allow desktop apps to use your microphone.
• Watch the meters to see where the signal stops. If AudioForge's input meter stays still, check the microphone. If AudioForge's output moves but your app's meter doesn't, check the cable selection in your app. While the output is muted, its meter is labelled MUTE, and any level it shows is from before the mute.
• Still stuck? Open an issue and attach the file from Settings > Help > Export Diagnostics. The export excludes audio and real device names; check it before you share it.

Everyday Controls

• Processing mode: Normal runs the full chain. Bypass skips voice effects, adaptive input cleanup and the lookahead limiter, and keeps the fixed 80 Hz high-pass and, while the limiter is on, the final true-peak limiter and ceiling. Raw also skips the input filter, for troubleshooting.
• Compare Recording: after calibration, hear the same passage as original, current, and proposed processing before you apply anything. The Review page also draws the proposed EQ over the current one.
• EQ graph: hover to read the frequency and gain of the total curve. Double-click a band handle, or press Delete on the selected band, to reset that band as one undo step. In Bypass and Raw the graph says "EQ not running".
• Shortcuts: Ctrl+S saves your preset, Ctrl+Shift+S saves a copy, and Ctrl+Z / Ctrl+Shift+Z undo and redo. Ctrl+Alt+M mutes from anywhere once you enable it under Settings > Tray and background.
• Unsaved changes: quitting or switching presets offers Save, Discard, or Cancel.

How Auto-EQ and Auto Voice Setup make their suggestions

• Two EQ stages. Calibration writes a correction stage. Your own edits go in a separate tone stage, so changing gains, templates, filter types, or slopes keeps the calibration. Older presets keep their combined EQ as the tone stage. The graph draws each band as a thin curve, the correction stage as a dashed line and the fixed input filter (the DC blocker and an 80 Hz high-pass, which Gentle and Strong cleanup keep) as a dotted line. The thick curve is what the signal really gets: the input filter plus the EQ, or the input filter alone while the EQ is off. Adaptive cleanup also follows the signal beyond that: it raises the high-pass to 100 or 120 Hz when it detects rumble and notches hum, and neither is drawn.
• Your voice stays yours. Speech alone can't reveal a microphone's exact frequency response, so automatic targets keep the recorded voice and apply bounded tonal preferences. Natural / No Added Tone stays neutral. Warm / Full Voice adds low-mid body with restrained upper presence. Subtle aims for small tone changes, never more than 2 dB from flat; Full strength aims for the tone preset's whole curve, up to 5 dB.
• A level check comes first. Voice Setup and Auto-EQ listen for three seconds before any take and say whether the microphone is too hot, too quiet or silent. For a hot or quiet voice they give the level they heard and, where a level change helps, the change to make. A good level continues with one click. A hot or quiet level can be checked again or continued anyway; silence can only be checked again. One stray full-scale sample doesn't reject a take; clipping needs a flat top of about 125 microseconds or more, or more than 20 separate hits.
• Results in plain words. Each reason a recording was not used says what was withheld and what to do. If a hot microphone leaves no full-size correction within the headroom rule, Auto-EQ offers the largest correction that is no worse than your current EQ, or says by about how much to lower the microphone level. When the correction's own boost is most of the excess, that amount counts only your recording's part, and the advice says so.
• Suggestions have to earn a change. Auto Voice Setup tries gate and noise suppression options before calibrating compression. It tests speech followed by noise without resetting the processing, so pause noise is checked too. It keeps your current settings unless a candidate passes safety and speech-preservation checks and improves a separate part of the recording. DeepFilterNet settings stay within DeepFilterNet, because automatic switches back to RNNoise failed clean-speech checks. CPU timing is only a pass/fail limit; it never ranks sound quality.
• Compression and loudness. EQ is fitted first. Compression calibration then tunes the compressor with your requested automatic makeup gain. If peaks at that loudness leave too little headroom, Voice Setup lowers the target by whole decibels until the final check passes and says so in the review. With the limiter on, automatic makeup and a voice that is not too hot, the check lets the limiter take up to 3 dB off the loudest peaks. If no allowed loudness keeps headroom, the result stays advisory; for a microphone that is too hot, lower its input gain by the amount shown and run it again.
• Verified before it sticks. A suggestion stays temporary until a second passage checks it through the full chain. Adjustments reuse a valid take, and your previous sound comes back if verification can't succeed. Recorded checks can't reproduce live device dropouts, so confirm the result in your destination app.

Full feature reference

Sound processing:

• Noise suppression with RNNoise or DeepFilterNet (both included).
• Noise gate with Threshold Only, VAD Assisted, and VAD Only modes, smooth speech-confidence gain control, and an automatic threshold that tracks the room's noise floor in the VAD modes.
• Ten-band EQ with bell, notch, shelf, and pass filters, 12–48 dB/octave pass slopes, click-free bypass, drag or keyboard editing on the graph, a readout under the pointer, and per-band reset.
• Auto-EQ that uses speech detection, rejects unreliable parts of a capture, and declines to correct when a safe correction isn't supported.
• Dynamic-EQ de-esser, compressor with speech-aware automatic makeup gain, and lookahead limiter.
• True-peak protection with 320 samples of lookahead (6.67 ms at 48 kHz), in addition to other processing and device delays. Turning the limiter off keeps that delay, so toggling it never shifts your timing. Turning it back on while audio is above the ceiling applies the ceiling immediately.
• Mono alignment for stereo microphones and adaptive 50/60 Hz hum removal.
• Latency calibration for each input/output pair. It measures your route and improves the reported estimate; it can't reduce physical delay.
• Undo and redo for manual edits, presets, Auto-EQ, and Auto Voice Setup.

Reliability and diagnostics:

• Input/output peak lines that catch the highest sample between updates (output: before the limiter) and show a CLIP tag once six samples at full scale arrive within three seconds. One touch of full scale only warns that peaks are too high. The output tag stays dark while the limiter is on, because the output is held at its ceiling, and lights when the limiter is off. Bypass and Raw monitor skip the lookahead limiter, but the final true-peak limiter and the ceiling still apply there.
• Health indicators that give the number and the action, for example how many decibels to lower the microphone level, plus technical counters behind Details.
• Automatic stream restart after device errors.
• Device refresh keeps your selection when the device is still available. Streams prefer 48 kHz when the device supports it.
• Settings > Help > Export Diagnostics writes a size-limited support file. It includes configuration, health fields and the last setup result, replaces device names with report-specific codes, and excludes audio, environment variables, secrets, and arbitrary paths.

Compatibility

AudioForge is built for Windows 10 (version 1809 or later) and Windows 11, 64-bit. Linux and macOS aren't supported.

Hardware testing so far covers Windows 11 at 48 kHz with USB microphones and VB-CABLE, including 30-minute runs and model switching, on earlier release candidates. The final release binaries haven't repeated that full hardware run. Windows 10, analog inputs, 44.1 kHz, device unplug and replug, and sleep recovery aren't hardware-tested yet. Each release's evidence archive records its automated validation; see the release workflow. Measured sound-processing decisions are indexed in evaluation/README.md.

How It Works

Audio passes through these stages in order:

1. Input cleanup: DC removal and one high-pass filter
2. Noise gate
3. Noise suppression
4. De-esser
5. Ten-band EQ
6. Compressor
7. Limiter and true-peak protection

The noise gate runs before suppression: with the default threshold gate, moving it after suppression clipped the starts of phrases in testing. See Processing mode above for Bypass and Raw. The latency shown in the app includes engine and noise-suppression delay, plus your measured route when you calibrate it. End-to-end latency still depends on your devices, driver, buffer sizes, and routing.

Help and Contributing

• Report a bug or request a feature: open an issue with your app version, what you expected, and steps to reproduce.
• Report a security issue: follow SECURITY.md.
• Contribute: follow CONTRIBUTING.md.
• See what changed: read the changelog.

Quick Start From Source

These steps are for building AudioForge, not for running the download. The engine is written in Rust and the interface in PySide6. You need Windows 10 (1809+) or 11 x64, PowerShell 7+, Git, CPython 3.13.15 x64, uv, rustup with the toolchain selected by rust-toolchain.toml, MSVC C++ Build Tools, GitHub CLI (gh), and 7-Zip. Install these prerequisites separately; dev.ps1 does not install system tools or download Python.

git clone https://github.com/FueledByRedBull/audio-forge.git
cd audio-forge

.\dev.ps1 bootstrap -PythonPath (py -3.13 -c "import sys; print(sys.executable)")
.\dev.ps1 doctor
.\dev.ps1 run

Without the Python Launcher (py), pass your existing Python executable's path to -PythonPath. For an existing .venv313 checkout, add -VenvPath .venv313 to each command, and add -DryRun to print a command's steps without running them.

The fallback release is pinned once in release-assets.json; it supplies runtime assets independently of the application version. Hydration verifies each asset against the manifest.

RNNoise is the default suppression backend. dev.ps1 run verifies the bundled assets and enables DeepFilterNet for that process, then restores the shell's environment. Direct source launches can set AUDIOFORGE_ENABLE_DEEPFILTER=1 after hydrating assets. Packaged builds enable verified bundled DeepFilter assets automatically. External DLL/model paths require an explicit opt-in; see runtime configuration.

CONTRIBUTING.md explains what bootstrap and doctor do and covers development checks and runtime configuration, RELEASING.md covers packaging and publication, and evaluation/README.md records the measured DSP decisions.

Roadmap

Planned work and explicit holds live in versioned GitHub milestones and issues labeled roadmap.

License

AudioForge's original source is MIT-licensed; see LICENSE. This source checkout uses PySide6 and LGPLv3 Qt modules. Complete component terms and library replacement instructions are available under Settings > Help > Licenses, in third-party notices, and in the corresponding-source instructions. Previously published portable and MSI packages containing PyQt6 retain their GPLv3 distribution terms; this source migration does not relicense them.

Acknowledgments

• RNNoise by Jean-Marc Valin
• DeepFilterNet by Hendrik Schroter and contributors
• Silero VAD contributors