Vocalm

Calm your calls. Free, on-device, real-time AI noise cancellation and meeting recorder for macOS and Windows — a Krisp-style virtual microphone with no cloud, no subscription, and no audio ever leaving your machine.

your mic ──► Vocalm (AI denoise) ──► virtual device ──► Zoom / Teams / Meet / Discord
their audio ◄── real speakers ◄── Vocalm (AI denoise) ◄── virtual device ◄── meeting app

Both directions are cleaned: your voice going out, and the noise coming in from everyone else's microphones.

Features

• Two-way noise cancellation — DeepFilterNet3 (Krisp-class quality, ~40 ms) or RNNoise (~10 ms, ultra-light), switchable live. Adjustable suppression strength.
• Meeting recording — one click; when incoming cleanup is on, recordings are stereo: you on the left channel, everyone else on the right (diarization-ready).
• On-device transcription — Whisper (whisper.cpp, Metal-accelerated), timestamped transcripts, auto language detection. Model (~148 MB) downloads once.
• Meetings library — rename, tag participants, reveal in Finder, view transcripts.
• Simple by default, tweakable when wanted — one master toggle for non-technical users; engines, routing, and DSP stats live under Advanced.

Setup

macOS — one installer, exactly like Krisp

Run scripts/build_installer.sh and open the resulting dist/Vocalm-<version>-Installer.dmg. The pkg inside installs Vocalm.app plus its own virtual audio devices — "Vocalm Microphone" and "Vocalm Speaker" — and restarts CoreAudio (asks for your password once). No BlackHole, no extra downloads.

Then:

• In your meeting app: microphone → Vocalm Microphone, speaker → Vocalm Speaker
• In Vocalm: pick your real mic and your real speakers. Done.

The devices are rebranded builds of the GPL-3 BlackHole driver (see drivers/NOTICE); Vocalm also auto-detects plain BlackHole or VB-CABLE if you already use those.

Windows — install from source with one shell command

No .exe installers needed — everything builds from source:

git clone https://github.com/edincampara/vocalm.git
cd vocalm
powershell -ExecutionPolicy Bypass -File .\install.ps1

The script installs prerequisites via winget (Rust, CMake, VS Build Tools), builds Vocalm, creates a Start Menu shortcut, and launches it. The Whisper model auto-downloads on first launch; the noise-removal models are embedded in the binary.

For the virtual microphone, install the free VB-CABLE once: winget install --id VB-Audio.VBCable — then in Teams/Meet set microphone = CABLE Output. (Windows can't have branded "Vocalm" devices yet: virtual audio there means a kernel driver that must be Microsoft-signed, which requires an EV certificate — money. The app itself is fully free and works today with VB-CABLE.)

Updating (any platform)

Updates ship through GitHub — from the repo folder:

# Windows
powershell -ExecutionPolicy Bypass -File .\update.ps1

# macOS
git pull --ff-only && scripts/build_installer.sh   # rebuilds app + installer DMG

Settings, recordings, and models live outside the repo, so updating never touches them. Every push to main is build-verified on Windows and macOS by GitHub Actions (including a DeepFilterNet3/RNNoise smoke test); tagging v* publishes prebuilt binaries as a GitHub Release.

Building

Rust 1.70+ and CMake (brew install rustup cmake, or rustup.rs + CMake on Windows).

cargo build --release          # → target/release/vocalm
scripts/package_mac.sh         # → dist/Vocalm.app + dist/Vocalm-<version>.dmg

The DMG is ad-hoc signed (no paid Apple developer account): first launch is right-click → Open. Notarization would need an Apple Developer ID ($99/yr).

Windows builds use the same code (WASAPI via cpal): cargo build --release with the MSVC toolchain.

Verifying quality offline

cargo run --release --bin denoise-wav -- noisy.wav clean.wav df       # DeepFilterNet3
cargo run --release --bin denoise-wav -- noisy.wav clean.wav rnnoise  # RNNoise
vocalm --transcribe <meeting-folder>                                  # headless transcription

How it works

1. cpal captures audio (CoreAudio / WASAPI), downmixed to mono.
2. Audio is resampled to 48 kHz (rubato) and chopped into 10 ms hops.
3. Each hop runs through the neural engine (DeepFilterNet3 via tract ONNX inference, or RNNoise via nnnoiseless) on a dedicated DSP thread — lock-free ring buffers on both sides, so audio callbacks never block. Measured: 0.37 ms per 10 ms frame.
4. The clean signal goes to a virtual loopback device (outgoing) or your real speakers (incoming). A second identical pipeline handles the reverse direction.

Same architecture as Krisp (virtual device + on-device DNN); Vocalm uses open models and standard free loopback drivers instead of a proprietary model + signed driver.

Roadmap

• Speaker diarization ("who said what") via an on-device model (e.g. sherpa-onnx); stereo recordings already separate you from everyone else.
• Google / Outlook calendar integration (auto-name and auto-record meetings).
• Menu-bar / tray mode, start at login; notarized installers bundling the drivers.