Talk the way you talk. Get the sentence you meant, typed at your cursor in almost any app, and not a word leaves your computer.
Hold a key (fn (🌐) on a Mac, Right Ctrl on Linux and Windows), speak, and let go. Live Transcribe turns what you said into clean, punctuated text and types it where you are working: a message, an email, a document, a terminal. The "um"s are gone. "Monday, no wait, Tuesday" comes out as "Tuesday". On the Mac, add your names and jargon once, and they are spelled your way.
Speech-to-text and a small language model run on your computer, with MLX on an Apple silicon Mac and with OpenVINO and sherpa-onnx on Linux and Windows. There is no account, no cloud and no telemetry, and once the models are downloaded it works offline. On the Mac, a live transcript window shows your words as you speak and tidies each line in place.
Free and open source (MIT) · Mac, Linux and Windows · 31 languages · Download or build from source
| Get | Notes | |
|---|---|---|
| Mac | LiveTranscribe.dmg, signed with Developer ID and notarized. It keeps itself up to date. |
Apple silicon, macOS 14 or later. Everything below. |
| Linux | a .deb, an .rpm or an .AppImage |
x86-64, glibc 2.35 or later, on a Wayland desktop that allows typing: COSMIC works; GNOME, KDE Plasma and X11 don't yet. Hold-to-talk needs a udev rule to read the keyboard: the deb and rpm install it, and the AppImage needs it added by hand. |
| Windows | live-transcribe_X.Y.Z_x64-setup.exe, not signed yet, so SmartScreen warns (More info, then Run anyway) |
Windows 10 or 11, x86-64. |
Linux and Windows have dictation, the tray and Settings; the live transcript, dictation history, Undo AI Edit, snippets, vocabulary and per-app settings are Mac-only for now. The install commands are in each release's notes; first start and every setting are in Using Live Transcribe.
A reply dictated into Slack on the 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.)
Real outputs from the dictation eval, at the default Medium cleanup level:
| You say | Live Transcribe types |
|---|---|
| Um, I think we should, uh, push the launch by a week | I think we should push the launch by a week. |
| Let's meet on Monday, no wait, Tuesday | Let's meet on Tuesday. |
| The budget is fifty thousand, I mean sixty thousand | The budget is sixty thousand. |
| We're flying into Boston, scratch that, into New York | We are flying into New York. |
| Tell Daniel, sorry, tell Maria the draft is ready | Tell Maria the draft is ready. |
| So, uh, what time does the, um, the train leave | So what time does the train leave? |
| I'll be about ten minutes late, the train is running slow today | I'll be about 10 minutes late. The train is running slow today. |
| see you soon smiley face emoji | See you soon! 🙂 |
| email me at john dot smith at example dot com | Email me at john.smith@example.com. |
And in anything that takes several lines, such as a document, an email or a chat message, Deep lays out lists you never announced, and leaves a series inside a sentence alone:
| You say | Live Transcribe types |
|---|---|
| Hey Kofi, two things about your invoice: the total looks too high and the tax number is missing. Can you fix it? | Hey Kofi, Two things about your invoice: - The total looks too high - The tax number is missing Can you fix it? |
| For the community garden we need: compost, raised beds, a water barrel, gloves, and volunteers for Saturday. | For the community garden we need: - Compost - Raised beds - A water barrel - Gloves - Volunteers for Saturday |
| Here's how to renew the parking permit: first log in to the council portal, second upload your proof of address, third pay the fee, and finally print the receipt. | Here's how to renew the parking permit: 1. Log in to the council portal. 2. Upload your proof of address. 3. Pay the fee. 4. Print the receipt. |
| three things went wrong last sprint we underestimated the migration work the test environment was down for two days and nobody owned the release notes | Three things went wrong last sprint: - We underestimated the migration work - The test environment was down for two days - Nobody owned the release notes |
| three changes i would suggest to the pricing page: shorten the headline, show the annual plan first and add a comparison table. happy to talk it through | Three changes I would suggest to the pricing page: - Shorten the headline - Show the annual plan first - Add a comparison table Happy to talk it through. |
| Last Sunday we cleaned the creek bank, planted thirty saplings and ate lunch under the old fig tree. | Last Sunday we cleaned the creek bank, planted thirty saplings and ate lunch under the old fig tree. |
| Please send the order to the Leeds warehouse. No, sorry, the Bristol warehouse. We close for stocktake on the thirtieth. | Please send the order to the Bristol warehouse. We close for stocktake on the thirtieth. |
| Languages like Python, sorry, not Python, Rust are a good fit for this kind of service. | Languages like Rust are a good fit for this kind of service. |
These are cases from the set Deep is tested on
(Training/eval/layout.jsonl, described in
Training), which it is never trained on; the left
column is the text as the speech model writes it, and the right column is Deep's own answer,
unedited. Deep got 216 of the 229 cases in
that set right, about 0.36 s at p50 on an M4 Pro. At Medium, the default, a list you announce
("first… second…", "number one… two…", "bullet point…") and a letter's greeting and sign-off are
laid out by rules alone.
12 of 12 self-corrections resolved. 10 of 10 filler clips cleaned. 65 of 65 clips laid out as meant. 641 ms at p95. In the 65-clip dictation eval at Medium, dictating into a multi-line field, speech-to-text plus cleanup of sentence-length dictations took 280 ms at p50 and 641 ms at p95, well inside the 1.2 s target.
Measured on an M4 Pro Mac with synthetic speech: the left column is the script a macOS text-to-speech voice read aloud. Stopping the recorder and inserting the text are not included in those times. Three of the 65 clips fell back to the uncleaned transcript, with fillers still removed: a plain sentence, a sentence the speech model broke at a hesitation, and a spoken "comma" it heard as "common" (details). The Linux and Windows app cleans up the same way, on the same model; it runs on the CPU there, about four times slower than on a Mac.
- Say it naturally, get what you meant. At the default Medium level, fillers disappear, spoken self-corrections are resolved, spoken lists and letters are laid out and numbers are written in digits ("twenty one chairs" → "21 chairs"); say emoji, punctuation and line breaks at any level. Every edit is checked against what you said, and on the Mac ⌃⌥Z puts your own words back.
- Dictate from any app. Hold the key, or double-tap it for hands-free. A small circle by the mouse pointer shows what is happening, and Esc cancels. On the Mac the key is fn (🌐) or a shortcut of your own; on Linux and Windows, Right Ctrl or one other key.
- Choose how much it edits. None, Light, Medium, High, or Deep, which also follows a correction back into an earlier sentence ("…is tomorrow. No, sorry, the after tomorrow." → "…is the day after tomorrow."), fixes grammar and lays out emails and lists.
- Make it yours (Mac). Snippets insert saved text when you say their phrase, and vocabulary spells your names and jargon your way.
- Dependable, app after app. On the Mac, text goes in through Accessibility and is read back, or is pasted with your clipboard put back, so terminals, browsers and Electron apps work too. Linux types through the Wayland input method, or pastes; Windows types as keystrokes. Line breaks go only where they belong, and nothing goes into a password field it can recognise.
- Private by design. Speech-to-text and cleanup run on your computer, with no account, no cloud and no telemetry. The Mac's dictation history stays on this Mac, and you can turn it off; Linux and Windows keep none.
- A live transcript, too (Mac). Watch your words appear as you speak, each line cleaned in place and every session saved.
- Easy to start, easy to trust. Guided setup and VoiceOver support on the Mac, more than 1,000 tests, a bench for your own recordings, a tool that retrains the adapters on your Mac, and your own models if you prefer.
Live Transcribe recognises 31 languages, and the default speech model works out which one you are speaking. Sinhala comes out in Sinhala script with English words in English letters, as people type it: "meeting එක cancel කරන්න".
| Arabic | Cantonese | Chinese (Mandarin) | Czech |
| Danish | Dutch | English | Filipino |
| Finnish | French | German | Greek |
| Hindi | Hungarian | Indonesian | Italian |
| Japanese | Korean | Macedonian | Malay |
| Persian | Polish | Portuguese | Romanian |
| Russian | Sinhala | Spanish | Swedish |
| Thai | Turkish | Vietnamese |
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. The default speech model is Qwen3-ASR 0.6B, fine-tuned for Sinhala by Nerdstorm; the Models tab in Settings offers others.
| Guide | What's in it |
|---|---|
| Using Live Transcribe | Installing, first start, dictating and every Settings tab, on the Mac, Linux and Windows, plus the Mac's Dictation History and live transcript |
| Cleanup and spoken commands | The five cleanup levels, what OutputGuard and Deep's check reject, and the phrases that become emoji, punctuation, line breaks and addresses |
| Snippets, vocabulary and apps | Saved text, your names and jargon, and how each app gets its text and line breaks |
| Privacy | What is kept where, what goes over the network, and how to remove it all |
| Known limitations | What doesn't work well yet, and what isn't built |
| Development | Building and signing, the tests and their audio clips, the bench and the dictation eval, training the adapters, licence notices and icons |
| Linux and Windows app | Building, packaging and changing the Rust app |
| Releasing | The signing credentials, and writing the changelog, building, notarizing, publishing and offering a release as an update |
| Architecture | For contributors: how dictation and cleanup work step by step, with diagrams, where each step's code is on the Mac and on Linux and Windows, and where to make a change |
| Design notes | Why it is built the way it is |
| Training the adapters | The self-correction and Deep adapters' datasets, training and evaluation |
Version 1.0, free and open source under the MIT License.
- Download it from GitHub Releases, or build it from source (below). What changed in each release is in the changelog.
- The Mac app was tested on an M4 Pro Mac with macOS 27 and Xcode 27. It targets macOS 14 or later but has not been run on older systems, nor on 8 GB Macs. With the default models it uses about 3 GB of memory.
- The Linux app has been run on COSMIC. GNOME, KDE Plasma and X11 desktops are not supported yet. The Windows app is built and tested in CI, which also installs it and starts it. On an Intel desktop CPU on Linux, the default speech model peaked at 5.3 GB of memory over a long test run, and the cleanup model at 2.5 GB.
- Tested with English and Sinhala speech. It recognises 29 other languages too; see Languages.
- Issues and pull requests are welcome; see Reporting a problem.
- What doesn't work well yet is in Known limitations.
Open an issue, in whatever form suits you. These help, when you have them:
- your computer and its system (the Mac and macOS version; the Linux distribution and desktop;
the Windows version), and the app's version (About Live Transcribe in the Mac's menu bar,
livetranscribe --versionon Linux and Windows, from a terminal); - the app you were dictating into;
- what you said, what was typed, and what you expected;
- for a Mac crash, the report macOS offers to send, or the one in Console › Crash Reports; on
Linux and Windows, the terminal output when it is started with
LIVETRANSCRIBE_LOG=infoset (on Windows, runlivetranscribe.exefrom a terminal in its install folder).
On the Mac you need an Apple silicon Mac and Xcode 26.4 or later with its Metal Toolchain
component (make doctor checks what is missing); from the repository root:
make runbuilds the app in Release and opens it. make on its own lists every target: the tests, the bench
and the eval, releases, and upkeep such as the licence notices and the icons. Building for the
first time, signing your builds so the permissions survive a rebuild, and disk space: see
Development. On first launch, Set Up Dictation walks you through
microphone access, Accessibility and the fn key while the models (about 2 GB) download; see
First launch.
Linux and Windows are a Cargo workspace in linux-windows/: run packaging/fetch-sherpa-onnx.sh
once, then cargo test (Linux also needs Tauri's build packages). Building and packaging it:
linux-windows/README.md.
Audio and transcripts never leave your computer, and there is no telemetry. The app goes online only to download models (from Hugging Face; on Linux and Windows also from GitHub) and, in a downloaded Mac release, to check GitHub for a new version about once a day if you allow it. On the Mac, dictation history is on by default: every completed dictation (what you said, what was typed, the app and timings) is kept unencrypted on this Mac, never synced, until you turn history off, limit how long it is kept or clear it in Settings › History. Linux and Windows keep nothing about what you say. On Linux, hold-to-talk needs a udev rule (installed by the deb and rpm) that lets any program you run read the keyboard, as any X11 program always could. Where each file lives, and how to remove everything: docs/privacy.md.
- Two apps, one behaviour. The Mac app is Swift, with MLX. The Linux and Windows app is Rust (a Tauri tray app), with OpenVINO and sherpa-onnx. The Mac app is the reference: its tests write fixtures that the Rust tests must reproduce, so the same text rules and the same cleanup checks run on all three systems.
- Vertical slices. Each feature's code lives together: on the Mac, one folder per slice of
Packages/LiveTranscribeKit, and inlinux-windows/one crate per slice with the same name. Slices depend on protocols, which the unit tests replace with fakes, andApp/AppComposition.swiftconstructs every concrete implementation. - Menu bar and tray apps. Dictation has to be available in every app, so Live Transcribe lives in the menu bar or the tray. On the Mac it becomes a regular app with a Dock icon only while one of its windows (transcript, history, Settings, setup) is open.
App/ the Mac app: menu, windows, composition root, app icon
LiveTranscribe.xcodeproj the Mac app's project (hardened runtime, no App Sandbox)
Packages/LiveTranscribeKit/ the Mac app's feature code, as vertical slices
linux-windows/ the Linux and Windows app, a Cargo workspace
Fixtures/ golden and cleanup fixtures the Mac's tests write and the Rust tests replay
Config/ signing settings
scripts/ releases, the changelog, licence notices, test audio and icons
docs/ the guides indexed above
design/ the app icon, drawn as SVG
site/ the website, published to GitHub Pages
CHANGELOG.md what changed in each release, written by make changelog
Makefile building, tests, the bench, releases and upkeep (make lists them)
Each step, with diagrams, and where to change what: Architecture. Why it is built this way: Design notes.
The app downloads its models. They are not part of this repository and not covered by its licence. The cleanup adapters, two 10 MB LoRA adapters for Qwen3-1.7B (one that resolves self-corrections at Medium and High, and Deep's), are part of this repository.
| Role | Model used | Original model | Licence |
|---|---|---|---|
| Speech-to-text (default) | Mac: Nerdstorm/Qwen3-ASR-0.6B-Sinhala-8bit. Linux, Windows: Nerdstorm/Qwen3-ASR-0.6B-Sinhala-OpenVINO | Qwen3-ASR-0.6B by the Qwen team, Alibaba Cloud, fine-tuned for Sinhala by Nerdstorm on OpenSLR 52 (Google, CC BY-SA 4.0) | CC-BY-SA-4.0 |
| Cleanup | Mac: mlx-community/Qwen3-1.7B-4bit. Linux, Windows: Nerdstorm/Qwen3-1.7B-MLX-4bit-OpenVINO, the same weights converted | Qwen3-1.7B by the Qwen team, Alibaba Cloud | Apache-2.0 |
| Self-correction adapter | bundled (Sources/Cleanup/Adapter) |
trained on synthetic data in this repository (Training/) |
MIT |
| Deep adapter | bundled (Sources/Cleanup/DeepAdapter) |
trained on synthetic data in this repository (Training/) |
MIT |
| Voice activity detection (Mac live transcript) | mlx-community/silero-vad | Silero VAD by the Silero team | MIT |
Settings › Models offers other speech-to-text models to download and switch to, without a
restart, each credited there with its licence. The Mac offers seven: the default, Qwen3-ASR 0.6B and
1.7B, Parakeet TDT v2 (English) and v3 (European languages), Whisper large-v3-turbo and Cohere
Transcribe. Linux and Windows offer four: the default, Parakeet v2 and v3 and Cohere Transcribe; the
last three are builds that sherpa-onnx (k2-fsa, Apache-2.0)
quantised to 8 bits for ONNX. Cohere Transcribe can't tell which language it hears, so its row also
chooses which of its 14 it writes (English at first); the others find the language themselves. They
are listed in
speech-models.json,
each pinned to a tested commit or, for sherpa-onnx archives, a SHA-256, and both apps read that
list. Another model there takes, on the Mac, any Hugging Face repository that mlx-audio-swift
can load or a folder on your Mac with the files such a repository has, and on Linux and Windows a
model converted for OpenVINO in the models folder.
The cleanup and voice activity models can be changed on the Mac in Settings › Advanced.
The Mac app is built with mlx-swift,
mlx-swift-lm,
mlx-audio-swift,
swift-huggingface,
swift-transformers and
Sparkle. Every Swift package dependency, including
indirect ones, is MIT or Apache-2.0 licensed. Some bundle third-party code under other
permissive licences (MLX includes the BSD-licensed PocketFFT, for example), so a redistributed
build must carry those notices too. The app does: About Live Transcribe in the menu bar shows
every package's licence, collected from Package.resolved by
scripts/generate-acknowledgements.sh. The Linux and Windows app is built with
Tauri and other Rust crates (see linux-windows/Cargo.lock), OpenVINO
(Intel), sherpa-onnx (k2-fsa, Apache-2.0) and ONNX Runtime (Microsoft, MIT); its packages carry
OpenVINO's, sherpa-onnx's and ONNX Runtime's licences.
MIT © 2026 Nerdstorm. The licence covers this repository's code only. The models (see Models and credits) and the software the apps are built with have their own licences.