Offline dictation into whichever window has focus
Murmure is a desktop dictation app. Press a shortcut, speak, and the text is typed into whichever window has focus: a document, a terminal, the prompt box of an AI assistant. Recognition runs offline on the CPU, with NVIDIA's Parakeet TDT 0.6B v3 model. The app is built on Tauri, in Rust and TypeScript.
The local speech-to-text comparison sets this kind of dictation against tools that transcribe meetings or audio files, and looks at what each one needs from your machine.
CPU-only Parakeet model covering 25 European languages
- Parakeet instead of Whisper: the project's site says Parakeet matches whisper-large's accuracy with 0.6 billion parameters against 1.5 billion. No GPU is needed.
- 25 European languages: the list covers French, German, Spanish and Polish, and also Maltese, Russian and Ukrainian.
- Windows, macOS and Linux: MSI, setup.exe or WinGet on Windows 10 and later, DMG builds for Apple Silicon and Intel Macs, and .deb, .rpm, pacman or AppImage packages on Linux.
- Custom dictionary: entries nudge the decoder toward the names and jargon you add. Find-and-replace rules, regex included, run after transcription.
- Prompt Mode: optionally, an LLM translates or rewrites the dictation before it is inserted, through Ollama or any OpenAI-compatible server.
- Local API: once enabled,
POST http://127.0.0.1:4800/api/transcribetakes a WAV file and returns the text.
Language can't be forced; one speaker at a time
You can't pick the language. Parakeet detects it from the audio, and no setting forces one. With a weak microphone or background noise, the output falls back to English, which the docs call the most commonly reported issue. French, English, German and Swedish are recognized best; Greek and some smaller languages trail behind.
One speaker at a time. Murmure doesn't tell speakers apart, and its FAQ describes it as a single-speaker dictation tool. For meeting notes, Meetily records and transcribes the call; for recordings processed on a server, Scriberr transcribes audio files. On a Mac, BetterWispr handles both dictation and meeting notes in one app.
Offline, except for update checks. According to its privacy policy, Murmure contacts GitHub on every launch to look for a new version, and the app has no switch to turn that off. The check carries no audio or text. Pointing Prompt Mode at a remote server sends your transcriptions there, and Murmure warns about it during setup.
Light on hardware, heavy to download. The site recommends 2 GB of free RAM and 1 GB of disk. The v1.11.3 installers weigh from 600 MB (Windows setup.exe) to 865 MB (.rpm). Without a GPU, the docs warn, Prompt Mode's LLM is very slow.
On Wayland, shortcuts live in your desktop settings. Wayland has been supported since v1.9.0, but Murmure registers no global shortcut there: you bind a desktop shortcut to murmure --transcription, and only toggle mode works, with no push-to-talk. The .deb needs GLIBC 2.38, as shipped with Ubuntu 24.04. The FAQ rules out mobile and web versions because those platforms block clipboard and keyboard access, and there is no Flatpak.
Free, and essentially one developer's project. Murmure has no paid tier ("No business model," says the site) and takes donations through Tipeee. The code is AGPL-3.0-or-later, so derivative works must stay open source. The repository dates from October 2025, and v1.11.3 shipped on August 31, 2026, after a minor release every one to two months through 2026. One developer authored 554 of the 603 commits on the main branch.