NoiseGate

Real-time microphone noise cancellation for Windows. Pure-Rust inference, WASAPI low-latency capture/render, system-tray UI. Pipes cleaned audio into VB-Cable so any app (Zoom, Teams, Discord, OBS, browser calls) sees a noise-free mic.

Status: pre-alpha scaffold. Builds on Windows 10/11 with the MSVC toolchain. Tested on (none yet — you're the first).

Why

Existing noise-cancellation tools either cost money (Krisp), require an RTX GPU (NVIDIA Broadcast), or are Linux-only (NoiseTorch). NoiseGate is a free, open-source, lightweight alternative for Windows that runs on any CPU, with a swappable model so you can opt into a state-of-the-art network when you want one.

Stack

• Audio I/O: WASAPI shared low-latency, event-driven, MMCSS Pro Audio scheduling.
• Routing: VB-Cable (free virtual audio cable, vb-audio.com).
• DSP model: RNNoise by default (via the nnnoiseless crate — pure-Rust port). Optional ONNX path for newer models like DeepFilterNet3 from Hugging Face.
• UI: tray-icon + winit event loop.
• Lang: Rust 2021, single static binary, ~10 MB stripped (RNNoise default), ~25 MB with ONNX runtime.

Architecture

physical mic ─► WASAPI capture ─► ring A ─► DSP thread (denoiser) ─► ring B ─► WASAPI render ─► VB-Cable Input
                                                                                                        │
                                                               other apps choose "CABLE Output" as mic ◄┘

Three dedicated MMCSS-priority threads. Lock-free SPSC ring buffers (8 frames ≈ 80 ms headroom). 480-sample (10 ms) frames end-to-end — the native frame size for both RNNoise and DeepFilterNet, so no reblocking inside the DSP path.

Noise suppression model

Two supported backends, both real-time:

Backend	Default?	Quality	Install size	How
RNNoise (via nnnoiseless)	✅	★★★ Good. Excellent for stationary noise (fans, hum). Older classical-RNN architecture.	+0 MB (model embedded)	Just cargo build.
ONNX (DeepFilterNet3 etc.)	opt-in (--features onnx)	★★★★★ State of the art.	+15 MB (onnxruntime.dll) + ~12 MB model file you supply	Build with --features onnx, point config at an ONNX file. See below.

For most users, RNNoise is what shipping software did until ~2022 and is still good enough for clean voice calls. Step up to ONNX/DFN3 when you have non-stationary noise (kids, traffic, music) and the extra CPU is worth it (typically 3-7%).

Install

Option A — Download a prebuilt .exe (easiest, once releases are cut)

Grab the latest binary from the Releases page, then jump to Set up audio routing below.

No releases yet? Build from source (Option B). A GitHub Actions workflow will start producing prebuilt binaries on every tagged release.

Option B — Build from source

You need three things on your Windows 10/11 machine:

#	Component	Download	Notes
1	Rust toolchain	rustup-init.exe — or visit rustup.rs	Pick the default (stable, x86_64-pc-windows-msvc).
2	MSVC C++ build tools	Build Tools for Visual Studio 2022	In the installer, check "Desktop development with C++" (gives you link.exe + the Windows SDK).
3	VB-Cable virtual audio driver	vb-audio.com/Cable (direct ZIP)	Extract, run VBCABLE_Setup_x64.exe as Administrator, reboot.
4	Git (probably already installed)	git-scm.com/download/win	Used to clone the repo.

Then in PowerShell:

git clone https://github.com/Yashsomalkar/noisegate.git
cd noisegate
cargo build --release
.\target\release\noisegate.exe

First build pulls ~300 MB of crates and takes 5-10 minutes. Subsequent builds are seconds.

A tray icon appears in the system tray (bottom-right corner of your taskbar). Right-click it for the menu: Enable, Open log folder, Quit.

Set up audio routing

Once NoiseGate is running, point your communication apps at VB-Cable instead of your real microphone:

1. Open Windows Sound Settings → Sound → Input.
2. Verify CABLE Output (VB-Audio Virtual Cable) is in the device list. If not, the VB-Cable install didn't finish — re-run step 3 above.
3. In Zoom / Teams / Discord / OBS / browser settings, choose CABLE Output as the microphone.

NoiseGate captures from your real mic, denoises, and writes the cleaned signal into CABLE Input. Apps that listen to CABLE Output then receive your noise-free voice.

Sanity test: open the Windows Voice Recorder app, set its mic to CABLE Output, record 10 seconds with a fan / typing in the background. Toggle NoiseGate's tray Enable off and re-record. The difference should be obvious.

Picking a specific microphone

By default NoiseGate captures from your system default mic. If that's a Bluetooth headset, Windows will switch the headset into HFP/Hands-Free mode as soon as we open the mic — that's a Windows-wide behavior, not a NoiseGate bug, and it sounds awful (16 kHz mono, glitchy). Pick your USB or built-in mic instead:

# See what's available:
.\noisegate.exe --list-devices

# Run with a specific mic (substring match on the friendly name):
.\noisegate.exe --mic "USB"
.\noisegate.exe --mic "Yeti"
.\noisegate.exe --mic "Realtek"

For a permanent choice, copy the device id from --list-devices into %APPDATA%\NoiseGate\config.toml:

input_device_id = "{0.0.1.00000000}.{...your-id...}"

Configuration

%APPDATA%\NoiseGate\config.toml — created on first run:

input_device_id = ""        # empty = default mic
output_device_id = ""        # empty = auto-detect VB-Cable
enabled = true
attenuation_db = 100.0       # 6.0 = subtle, 100.0 = max
auto_start = false

Logs at %APPDATA%\NoiseGate\logs\noisegate.log. Tune verbosity with RUST_LOG=noisegate=debug.

Cargo features

Flag	Default	What it does
rnnoise	✅ on	RNNoise backend via nnnoiseless. Pure-Rust, model embedded, no extra runtime deps.
onnx	off	Adds ONNX Runtime as a dependency so you can load any raw-audio noise-suppression ONNX model (e.g. DFN3). Needs onnxruntime.dll on PATH.

Build with the ONNX backend in addition to RNNoise:

cargo build --release -p noisegate -F onnx

To get DeepFilterNet3 quality:

1. Build with -F onnx.
2. Download the DFN3 ONNX export from Hugging Face: https://huggingface.co/Rikorose/DeepFilterNet3.
3. Drop onnxruntime.dll next to noisegate.exe (download from https://github.com/microsoft/onnxruntime/releases — pick the win-x64 zip).
4. Point model_path in config.toml at the ONNX file.

License

Code: dual MIT / Apache-2.0 — your choice.

The bundled RNNoise model (via nnnoiseless) is BSD-licensed — fine for any use including commercial. If you opt into the ONNX backend with DeepFilterNet3 weights, those are research / non-commercial; for commercial use, retrain on your own data or pick a different ONNX model.

Cost

$0 for personal use. Every component is free. No driver-signing certs required (we don't ship a driver — VB-Cable does).

Not included (yet)

• macOS / Linux backends (the audio-io crate is Windows-only; the rest is portable).
• Far-end denoising (cleaning the audio you hear from the call — only your mic is cleaned).
• Acoustic echo cancellation (combine with an AEC frontend if needed).
• Auto-update.

Credits

Built on top of excellent open-source work:

• RNNoise by Jean-Marc Valin / Xiph.Org — the recurrent-network noise-suppression model that powers the default backend.
• nnnoiseless by jneem — pure-Rust port of RNNoise; this is the actual crate doing the work.
• DeepFilterNet by Hendrik Schröter et al. — the modern speech-enhancement model that the optional ONNX backend can run.
• ONNX Runtime + the ort Rust bindings — power the optional onnx backend.
• VB-Cable by VB-Audio — the free virtual audio driver every Windows audio-routing app depends on.
• windows crate by Microsoft — official Win32 bindings for Rust (WASAPI, MMCSS, COM).
• tray-icon + winit — system-tray UI and event loop.
• ringbuf — the lock-free SPSC buffers that connect the audio threads.
• Inspired by NoiseTorch (Linux-only equivalent that uses RNNoise).

Issues & discussion

• Bug reports / feature requests: open an issue.
• Questions, ideas, "is X possible?" — Discussions.