Local voice dictation for macOS. Press a key, speak, and get clean text right where your cursor is — 100% on your Mac.
Murmure is a tiny menu-bar app. You hold a key, talk, and it transcribes your speech locally (no cloud, no account), runs it through a local AI to clean up filler words and punctuation, and pastes the result at your cursor — in any app, including the browser.
It's the privacy-respecting, free, open-source take on tools like Wispr Flow or Typeless.
- Truly private — speech recognition (whisper.cpp or Parakeet) and AI cleanup (LM Studio / Ollama) run entirely on your Mac. Cloud models are strictly opt-in.
- Fast transcription — Whisper
large-v3-turboby default, or Parakeet v3 (NVIDIA, on the Neural Engine) — about 10× faster, slightly less accurate. - Clean text, not a transcript — an LLM removes "um", "uh", stutters and repetitions, fixes punctuation and casing, while keeping your words and style. French numbers are written as digits ("vingt-trois" → 23).
- Your engines, your order — list the AI models you want (Ollama on CPU, MLX on GPU, LM Studio, or Ollama Cloud) and drag them into priority order: the first one that answers does the cleanup, the others are fallbacks. When another app keeps the GPU busy, MLX models can step aside automatically.
- Reads the conversation you're answering — optionally, Murmure reads the text shown in the window (e.g. a Claude chat) so the cleanup restores mis-heard names, files and terms, and stays on topic. That screen text never goes to a cloud model (cloud engines get the prompt without it).
- Pastes anywhere — native fields, Electron apps, and web inputs (Chrome included).
- Live feedback — a compact card shows a real-time audio spectrum while you speak, then the phase (transcription, touch-up) with a timer and, optionally, the model at work. Need the text now? A lone tap on Shift skips the AI touch-up and pastes the raw transcript.
- Live preview — your words appear in real time above the overlay as you talk (Apple's on-device speech), then vanish; the final pasted text still comes from the transcription model.
- Learns your vocabulary — correct a transcript and Murmure remembers the term, then reuses it to transcribe better next time.
- History — dictations grouped by the field they went into, with "Copy all"; for dictations into Claude, the app icon and a clean digest of the conversation on screen.
- Editable shortcuts — rebind every command from the menu, with conflict warnings.
- Parallel dictations — fire several at once; they're delivered in the order you started them, so your text never gets scrambled. The overlay follows the screen of the field you're typing in.
- Optional audio replay — enable “Keep audio for 20 minutes” in Settings to replay or re-transcribe recent dictations from History. Local WAV copies expire automatically (cleanup resumes on launch if the app was closed), with a 50 MB cap that removes the oldest first. Disabled by default; disabling or clearing history deletes saved audio. Reanalysis uses the current transcription model and updates the text directly in History without a popup.
- Mutes system audio while recording (like Typeless), and restores it after — and, if you enable them, your Google Cast TVs and speakers too (via
catt). - English & French interface, switchable from the menu.
Installs the dependencies, downloads the speech model, builds and signs the app, and launches it. Everything works out of the box:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Hugo291/Murmure/main/install.sh)"The installer pulls in Homebrew, whisper-cpp, LM Studio and ollama if you don't already have them.
Grab the latest Murmure.dmg, open it, and drag Murmure into Applications.
Murmure is free and self-signed (not notarized by Apple), so the first launch needs one extra step:
- Double-click Murmure → macOS says it can't verify the developer. Click Done.
- Open System Settings › Privacy & Security, scroll down, and click Open Anyway next to the Murmure message — then confirm.
(Equivalent in Terminal:
xattr -dr com.apple.quarantine /Applications/Murmure.app.)
The DMG ships the app only — the speech and AI engines aren't bundled. Install them once with Homebrew:
brew install whisper-cpp ollama # speech-to-text + local AI cleanupThen, from Murmure's menu, download a transcription model and an AI model in one click each.
Requirements: macOS 14 (Sonoma) or later.
Murmure lives in the menu bar (the 🎙 mic, top-right). On first run, grant three permissions in System Settings › Privacy & Security:
| Permission | Why |
|---|---|
| Microphone | to hear you |
| Input Monitoring | to detect the Fn key globally |
| Speech Recognition | optional — for the real-time live preview (on-device) |
| Accessibility | to paste text at the cursor |
Tip: in System Settings › Keyboard, set "Press 🌐/Fn key to" → "Do Nothing" so Fn is free for dictation.
| Shortcut | Action |
|---|---|
| Fn | Start / stop dictation (tap to start, tap again to transcribe & paste) |
| Esc | Cancel the current recording or processing |
| ⇧ Shift (tap alone) | Skip the AI touch-up and paste the raw transcript |
| ⌘L | Paste the last dictation again |
| ⌘R | Rewrite the selected text (strip filler & repetition, keep the meaning) |
All shortcuts are editable: menu › Settings › Commands › Edit shortcuts…
Press Fn, talk, press Fn again. The text lands at your cursor. That's it.
Fn ─► record (AVAudioEngine) ─► whisper.cpp / Parakeet ─► AI engines, in your order ─► paste at cursor
live spectrum overlay on-device transcription filler/punctuation cleanup
The speech models live in ~/Library/Application Support/Murmure/. AI cleanup walks your
engine list top to bottom and uses the first engine that answers, so cleanup keeps
working when a server is off.
Tune all of this from Settings:
- AI touch-up — raw text, light cleanup (default), or full rewrite.
- Touch-up engines — add models by family: Ollama · CPU, Ollama Cloud, Ollama · MLX and LM Studio (every downloaded language model, loaded or not). Drag rows to reorder, right-click to move or remove; each row shows whether its server is off or the model is missing. Test engines times each one. No model yet? Download fetches Gemma in one click, for Ollama or LM Studio.
- Read the screen — use the visible conversation to fix mis-heard words (local engines only).
- GPU busy: MLX engines last — when another app holds the GPU, CPU engines go first.
- Transcription model — Parakeet v3 or any whisper size (tiny → large-v3), downloaded from Hugging Face in one click with a live progress bar.
- Show the model under the spectrum — transcription model while recording/transcribing, cleanup engine during touch-up.
- Dictation language — French, English, or auto.
- Language — interface language (English by default, French available).
Models such as kimi-k3:cloud, glm-5.3:cloud or gpt-oss:120b-cloud run on ollama.com through
your local Ollama server (sign in with ollama signin). Add one from Ollama Cloud in the engine
list only if you accept that the dictated text is sent to ollama.com. The screen text is never
sent: a cloud engine keeps its place in the list but gets the prompt without it.
With local engines, Murmure makes zero network requests for its core function. Audio, transcripts and your learned vocabulary stay in ~/Library/Application Support/Murmure/ on your Mac. The only exception is an Ollama Cloud engine you add yourself: it receives the dictated text (never the screen text). The real-time live preview uses Apple's speech recognition with requiresOnDeviceRecognition = true, so it stays on-device too. The only downloads are the one-time model files (from Hugging Face / Ollama) during install.
git clone https://github.com/Hugo291/Murmure.git
cd Murmure
./install.sh # full setup (deps + model + build + launch)
# or, if deps are already in place:
./build.sh # just build & sign Murmure.app into ~/Applications
swift build # plain debug buildPure Swift / AppKit, built with SwiftPM — no Xcode project required.
- Swift 6 · AppKit · SwiftUI (settings windows)
- whisper.cpp and FluidAudio (Parakeet, Core ML) for on-device transcription
- Ollama / LM Studio for LLM cleanup
- Accelerate / vDSP for the live spectrum and voice-activity gate
CGEventTapfor global shortcuts, the Accessibility API for cursor-position pasting
Issues and pull requests are welcome. Murmure is intentionally small and focused — please open an issue to discuss larger changes first.
MIT © 2026 Hugo. Use it, fork it, ship it.