Private dictation for your computer

What you say, and what it types

A reply dictated into Slack on a Mac: hold the shortcut, speak, let go. (Recorded with an earlier version, when the panel was a capsule at the text cursor; it is now a small circle by the mouse pointer.)

Outputs from the dictation test set (synthetic speech) at the default cleanup level.

On the Mac, Dictation History keeps what you said next to what it typed, so you can copy your original wording. It stays on your Mac, and you can turn it off.

What it does

• Works in almost any app: email, chat, documents, browsers, code editors and terminals.
• Removes filler words and fixes punctuation and capitals.
• Follows your corrections: “Monday, no wait, Tuesday” becomes “Tuesday”.
• Lays out spoken lists, letters and line breaks in any field that takes several lines. On the Mac that includes chat apps; on Windows, browser and Electron apps are typed on one line.
• Listens hands-free when you double-tap the key, until you tap it once more.
• Lets you choose how much it edits: None, Standard or Deep.
• Recognises and cleans up a sentence in about 0.3 seconds (280 ms median) on an M4 Pro.
• On the Mac, inserts saved text when you say its phrase, such as “my calendar link”.
• On the Mac, spells your names and jargon correctly once you add them.
• On the Mac, puts back exactly what you said if you press ⌃⌥Z within 30 seconds.
• On the Mac, shows a live transcript as you speak, for longer sessions.

Languages

Live Transcribe recognises 31 languages, and works out which one you’re speaking. Sinhala comes out in Sinhala script with English words in English letters, as people type it: “meeting එක cancel කරන්න”.

Cleanup is written for English. Sinhala skips the cleanup model and is typed as recognised, and the other languages haven’t been tested with cleanup.

Settings on the Mac

Shortcut and cleanup

Hold fn while you speak, or double-tap it to keep listening until you press it again. You can choose another key. Cleanup has three levels: None only applies your spoken commands, snippets and vocabulary, Standard fixes punctuation, capitals and misheard words, removes fillers, follows your corrections and rewords lightly for clarity, and Deep follows a correction back into an earlier sentence, fixes grammar and lays out emails and lists. Linux and Windows have the same levels in a smaller Settings window, with General, Models and Advanced tabs.

Snippets

Save text you use often under a short phrase, such as a link, an address or a template. Say the phrase while dictating and the text goes in exactly as saved.

Vocabulary

Add names and jargon, with the ways speech recognition tends to mishear them, and they’re spelled your way.

Apps

For each app, choose how dictated text gets in, typed through Accessibility or pasted, and whether it may break across lines. Every app takes line breaks out of the box, so lists, letters and “new line” work in chat apps too. Terminals stay on one line, so a line break never runs a command, and so does any field that only takes one line, such as a search box.

Stays on your computer

Speech recognition and cleanup run locally, with Apple’s MLX on a Mac and with OpenVINO and sherpa-onnx on Linux and Windows. There is no account, no cloud service and no telemetry. The app goes online only to download its models (from Hugging Face, and on Linux and Windows for some models from GitHub) and, on the Mac if you let it, to check GitHub for a new version about once a day. Dictation works offline. The Mac keeps your dictation history on the Mac, and you can turn it off; Linux and Windows keep none.

Install

Mac

1. Download the latest release from GitHub and move Live Transcribe to your Applications folder.
2. Open it. Setup walks you through microphone access, Accessibility and the fn key setting, while the speech models (about 2 GB) download.
3. When the menu bar icon shows it’s ready, hold fn in any app and speak.

The app is signed and notarized by Apple, so macOS only asks you to confirm opening an app downloaded from the internet. It keeps itself up to date.

Linux

1. Download the latest release from GitHub: a deb for Ubuntu and Debian, an rpm for Fedora and openSUSE, or an AppImage for anything else, and install it. The release notes give the commands.
2. Start Live Transcribe. It lives in the tray, and downloads its speech and cleanup models (about 2 GB) on first start.
3. When the tray says “Hold Right Ctrl to dictate”, hold Right Ctrl in any app and speak.

Dictation types on Wayland desktops that allow it. COSMIC works; Sway and Hyprland should but haven’t been tried. GNOME, KDE Plasma and X11 desktops are not supported yet: there the app starts and Settings says why. Hold-to-talk reads the keyboard through a udev rule, which the deb and rpm install and the AppImage needs added by hand (the release notes give the commands). The rule lets any program you run read what you type, as any X11 program always could.

Windows

1. Download the latest release from GitHub and run the setup program. It installs for your account, without administrator rights.
2. Start Live Transcribe. It lives in the tray, and downloads its speech and cleanup models (about 2 GB) on first start.
3. When the tray says “Hold Right Ctrl to dictate”, hold Right Ctrl in any app and speak.

The installer isn’t signed yet, so SmartScreen warns about an unrecognised app: choose More info, then Run anyway.

You can also build it from source.

Requirements

• Mac: Apple silicon and macOS 14 or later (tested on macOS 27)
• Linux: x86-64 with glibc 2.35 or later, on a Wayland desktop that allows typing
• Windows: Windows 10 or 11 on x86-64
• About 2 GB of disk space for the models, and on a Mac about 3 GB of memory while running
• Speech in one of the 31 languages