Murmur

Local, on-device voice-to-text for the desktop.

Hold a key, talk into any app, and get clean, formatted text dropped right at your cursor — transcribed locally with Whisper, polished by AI, and never uploaded.

⬇️ Download · 🌐 Website · 🐛 Issues

---

✨ What it does

Murmur is a tiny menu-bar app that turns speech into text anywhere you can type.

• 🎙️ Hold-to-talk, anywhere — press and hold ⌥ Space in any app, speak, release. The text appears at your cursor.
• 🧠 Transcribed locally with Whisper — runs on your CPU or GPU (Metal on macOS). No audio ever leaves your machine.
• ✨ Polished by AI — a cleanup pass strips filler words, fixes punctuation and capitalization, and formats the result.
• 🔒 Private by design — no account, no sync, no telemetry, no cloud. Download, grant mic access, and talk.
• 🗂️ A real desktop app — searchable history, words-saved stats and a streak, every setting on one screen.
• 🪶 Light & native — a Tauri app (Rust + system webview), ~7 MB installer, with a frosted floating overlay.
• 🌍 Multilingual — English, Spanish, French, German, Japanese, Chinese, with auto-detect.
• 🖥️ Optional remote server — offload transcription to a beefier machine over an OpenAI-compatible endpoint.

🎬 How it works

   you speak              whisper                ai polish              inserted
   ─────────  ───▶  ──────────────────  ───▶  ───────────────  ───▶  ────────────
   hold ⌥ Space      transcribe on-device      filler & grammar      at your cursor
                                                                    ┊
                          ✗ the cloud — never touched ┄┄┄┄┄┄┄┄┄┄┄┄┄┄┘

Audio is captured, resampled to 16 kHz mono, transcribed by Whisper in memory, cleaned up, then written to the focused field via the clipboard + a synthetic paste. Nothing is written to disk or sent anywhere.

🖥️ The desktop app

Home — every dictation, searchable, with stats and your streak.	Voice & model — mic, hotkey, language and the Whisper picker.

🧠 Whisper models

Murmur ships every OpenAI Whisper size — pick the one that fits your machine. Smaller is faster, larger is sharper, and they all run entirely on-device. Models download on first use and you can switch any time.

Model	Download	Speed	Accuracy	Best for
tiny	75 MB	32× realtime	●●○○○	Quick notes
base	142 MB	16× realtime	●●●○○	Everyday chat
small	466 MB	6× realtime	●●●○○	Balanced
medium	1.5 GB	2× realtime	●●●●○	High accuracy
large-v3	2.9 GB	1× realtime	●●●●●	Best quality · default

🔒 Privacy by architecture

Most dictation tools stream your microphone to someone else's servers. Murmur doesn't. Both transcription and the AI cleanup pass run on the one machine you control:

• Processed on-device — audio is transcribed in memory and discarded; never written to disk, never sent anywhere.
• No account, no cloud — no sign-up, no sync, no telemetry, no background phone-home.
• Open source, forever — every line is auditable here. MIT-licensed and free.

⬇️ Install

Grab the latest installer from the Releases page.

macOS

One-liner — downloads the latest universal build, installs it, and clears the Gatekeeper flag:

curl -fsSL https://raw.githubusercontent.com/kurenn/murmur/main/scripts/install-macos.sh | bash

Or install the .dmg manually. Murmur is not notarized by Apple (no paid Developer account), so macOS blocks the unsigned app on first open ("Apple could not verify…" / "will damage your computer"). It's safe — allow it once:

• System Settings → Privacy & Security → click Open Anyway next to the Murmur message, or
• run xattr -dr com.apple.quarantine /Applications/Murmur.app in Terminal.

On macOS 15+ the old right-click → Open no longer bypasses Gatekeeper — use Open Anyway or the command above.

Windows & Linux

OS	File	Notes
Windows	.msi or .exe	Requires WebView2 (preinstalled on Windows 10/11).
Linux	.AppImage or .deb	AppImage: chmod +x Murmur*.AppImage && ./Murmur*.AppImage.

On first launch, grant Microphone access — and on macOS, Accessibility (to paste at your cursor) and Input Monitoring (for the fn-key trigger).

🔨 Build from source

Prerequisites: Rust (stable), Node.js 18+, and the Tauri v2 system dependencies for your OS.

git clone https://github.com/kurenn/murmur.git
cd murmur
npm install            # installs deps + the git pre-commit hook

npm run tauri dev      # run the app with hot-reload
npm run tauri build    # produce a native installer in src-tauri/target/release/bundle

🌐 Remote Whisper server (optional)

Run transcription on another machine (e.g. a desktop with a GPU) and point the app at it from Settings → Remote server. The bundled server exposes an OpenAI-compatible /v1/audio/transcriptions endpoint:

cd server
./install.sh          # creates a venv + installs the Whisper server (use Python 3.12)
./run.sh              # serves on http://0.0.0.0:8000

See server/README.md for details.

🧪 Development & testing

npm test                       # frontend specs (Vitest + Testing Library)
cd src-tauri && cargo test     # backend specs (pure logic: polish, resample, config, …)

A pre-commit hook (installed by npm install) runs tsc + Vitest on frontend changes and cargo test on backend changes; CI (ci.yml) re-runs everything on macOS. Releases are built on native macOS / Windows / Linux runners by release.yml.

🗂️ Project structure

murmur/
├── src/                  # React + TypeScript UI (dashboard, overlay, onboarding, design system)
├── src-tauri/            # Rust backend
│   └── src/
│       ├── lib.rs        # app wiring, tray, hotkey, commands, window setup
│       ├── audio.rs      # microphone capture (cpal)
│       ├── transcribe.rs # Whisper (local) + remote transcription, WAV/resample
│       ├── polish.rs     # AI/heuristic text cleanup
│       ├── inject.rs     # clipboard + synthetic paste (accessibility-gated)
│       └── …             # config, history, state
├── server/               # optional remote Whisper server (Python)
└── docs/                 # the marketing site (GitHub Pages → kurenn.github.io/murmur)

📄 License

MIT © Abraham Kuri Vargas. Free and open source — no trial, no seat count, no paywall.

_{Built with Tauri, Rust and React · transcribed by Whisper, polished by AI, never uploaded.}