Contents
What it does
- Live, both ways. Type in either box and the other translates as you go. New words show grey until the engine settles.
- From any app. Select text anywhere and press Ctrl+Alt+T (Mac: just tap Right ⌘); the translation appears in a bubble by your cursor, with a match score. Tab to its language and type to switch ("spa" → Spanish). Click your text in the bubble to see all of it, and drag the bubble by its top bar to put it anywhere: once moved, it stays put like a sticky note until you close it.
- Snip & translate. Ctrl+Alt+S (Mac: ⌃⌥S), draw a box around text on screen (images, subtitles, apps that block copying), read it in your language. On Mac, it's read by Apple's built-in text recognizer (the Live Text engine); on Windows, the snip is cleaned up first (enlarged, dark themes flipped, contrast boosted). Either way, screen text comes through word for word. The image stays in memory and is never saved.
- Dictation. Tap or hold Right Ctrl (Mac: Right ⌥ Option) and talk. Transcribed on your machine; optionally typed into the app you're using, as said or translated.
- Read aloud. Natural male and female voices in about 75 languages. Each word lights up as it is spoken; slow (0.75×), normal and fast (1.25×) speeds and a volume slider, which take effect mid-sentence. Press Esc in any app to stop it.
- Conversation mode (beta). Two people who speak different languages take turns through one computer; each phrase is translated and read aloud to the other. See how it works and its limits.
- Engines. Google Translate (free), or Claude, ChatGPT or Gemini with your own key. The AI engines add a tone setting (natural, casual, polite, formal) and a "who it's for" note so pronouns come out right (settings → translation; dimmed while Google Translate is selected).
- Back-translation with a match score. See your translation turned back into your language, with a score for how much of your meaning survived (see below). Theme colors by default, or colorblind-friendly colors with symbols (settings → appearance).
- Detect language, pinned languages, pronunciation for non-Latin scripts, 30-item history, always on top (fades while you work elsewhere), 37 themes (star your favorites to keep them on top), search in settings, self-updates.
- In your language. Every menu, button and message in any of the 133 languages Vox2 translates (settings → app language; it follows your computer's language to start). Right-to-left languages flip the layout. The translations are machine-made and checked by translating them back; Vietnamese has been read in full by a person. .
Shortcuts
| Windows | Mac | Action |
|---|
| Ctrl+Shift+Space | ⌃⌥V | show / hide Vox2, ready to type |
| Ctrl+Alt+T | tap Right ⌘ | translate the selected text in any app (a quick tap on its own, so Right ⌘ + a key still works as usual) |
| Ctrl+Alt+S | ⌃⌥S | snip an area of the screen and translate it |
| Right Ctrl | Right ⌥ Option | dictate: tap to start (stops when you go quiet, or tap again), or hold to talk |
| Esc | Esc | stop reading aloud or dictating, from any app (does nothing otherwise) |
| Ctrl+L | ⌘L | listen to the translation, again to stop (while Vox2 or the bubble is in front) |
| Ctrl+Shift+C | ⇧⌘C | copy the translation (in the bubble: ⌘C / Ctrl+C) |
| Ctrl+Shift+S | ⇧⌘S | swap the two languages (while Vox2 is in front) |
| Tab | Tab | go round the languages and text boxes; on a language, just type its name ("spa") |
| Ctrl+P | ⌘P | pin on top (while Vox2 is in front) |
| Ctrl+Shift+F | ⇧⌘F | fit the window to its text; press again for a wider reading width, again for a compact window (handy for long text), again to fit (while Vox2 is in front) |
| Ctrl+H | ⌘Y | open / close history (while Vox2 is in front) |
| Ctrl+, | ⌘, | open / close settings (while Vox2 is in front) |
All configurable in settings → shortcuts. Hover a button in the bottom bar to see its shortcut.
Conversation mode (beta)
For talking with someone who speaks another language, with one computer between you.
How it works today
- Put your language in one box and theirs in the other (say English on top, Vietnamese below), and turn on conversation mode with the speech-bubble button in the bottom bar. Both mic buttons light up.
- You tap the mic on your side and speak. When you pause (or tap the mic again), Vox2 writes the translation in the other box and reads it aloud in their language.
- They tap the mic on their side and answer. The translation is read aloud to you.
- Repeat. Starting to talk stops any reading still in progress, and both sides stay on screen as a written record.
Tips: use the speakers, not headphones, and put the computer between you. For a natural pace, set settings → dictation → "after a tap, stop when quiet for" to 2 s. Turning conversation mode on also turns on dictation, which it needs.
Limitations (why it's in beta)
- Every turn starts with a tap. Each person has to tap their mic before speaking. That's fine at a desk, but awkward face to face, when you'd rather look at each other than at the screen.
- One computer, one microphone. It works best in a quiet room, with each person close enough to the mic.
Where it's heading: hands-free turns. After Vox2 reads a translation aloud, it starts listening on the other side by itself, walkie-talkie style, so a conversation flows without touching the computer after the first tap. Tracked in #11.
Dictation models
Pick one in settings → dictation → voice model. The Whisper models are OpenAI's open speech-recognition models, run entirely on your computer: free, private, and they work offline once downloaded. All of them understand 99 languages. Bigger models are more accurate, especially with accents, background noise and mixed languages, and take more memory and download time.
| Model | Size | Good for |
|---|
| Whisper Large v3 Turbo · most accurate | 1.5 GB (about 760 MB on graphics cards without half-precision support) | The best accuracy Vox2 offers on your own computer. A faster version of Whisper's largest model (released 2024): near-flagship accuracy at a fraction of the work. Wants a computer with a capable graphics card (any Apple Silicon Mac, or a recent Windows PC); without one it's slow. On an M1 Pro it transcribed a 6-second sentence in about 3.5 s, about as fast as Small. |
| Whisper Small · default | 590 MB | A good balance of speed and accuracy on almost any computer. |
| Whisper Base · lighter | 210 MB | Quicker to download and lighter on memory; fine for clear speech in a quiet room. |
| Whisper Tiny · lightest | 120 MB | For older or low-memory computers. Noticeably rougher; mostly useful when the others are too slow. |
| OpenAI · online | none | Sends your audio to OpenAI after each pause (models gpt-4o-mini-transcribe by default, gpt-4o-transcribe, or whisper-1). Uses your own OpenAI key (the same one as the ChatGPT translation engine) and is billed to it. Needs internet; not private in the way the local models are. |
Each local model downloads once, the first time you dictate with it, and settings marks it "downloaded" after that. "Keep the voice model ready" decides whether it stays loaded in memory between dictations.
Why not Parakeet? Some dictation apps recommend NVIDIA's Parakeet v3, which is very fast. It only covers English and 24 European languages (no Vietnamese, Chinese, Japanese, Korean, Arabic…), and isn't available in a form that runs inside Vox2's web view, so it's not offered.
Install
Windows
- Download
Vox2-setup.exe from the latest release and run it. It installs per user, no admin needed.
- The installer isn't code-signed, so Windows shows "unknown publisher". Choose More info → Run anyway.
- Vox2 updates itself from GitHub releases (signed updates). Switch to "ask me" or off in settings → window & updates.
Windows 10/11, 64-bit.
macOS (beta)
- Download
Vox2-mac.dmg, open it and drag Vox2 into Applications.
- Open Vox2. It isn't notarized by Apple yet, so the first time macOS says it can't verify the developer. Go to System Settings → Privacy & Security, scroll down and click Open Anyway. You only do this once.
- Vox2 opens a permissions screen listing what macOS needs you to allow. Click allow on each row and switch Vox2 on in the System Settings page that opens; the rows tick off as you go, and restart now appears where macOS needs one. You can come back to it any time from settings → permissions.
- Vox2 updates itself from GitHub releases, like on Windows, and keeps its permissions across updates.
Apple Silicon Macs (M1 and newer). Tested on macOS 26.
macOS notes
Permissions. macOS asks for each of these separately; the permissions screen shows them all in one place.
| Permission | Used for |
|---|
| Input Monitoring | hearing your shortcuts while you're in other apps |
| Accessibility | typing what you dictate into other apps, copying the selected text to translate it, keeping Vox2's shortcuts from also reaching the app in front |
| Screen Recording | snip & translate (only the area you box is read, on your Mac) |
| Microphone | dictation (only shown when dictation is on) |
Why permissions survive updates. Mac builds are signed with Vox2's own certificate (free, self-made; not Apple-issued). macOS remembers permissions per signature, so every update counts as the same app. Unsigned builds would lose their permissions on every update.
Shortcuts look Mac-native. Settings shows them the way the menu bar does (⌃⌥T). The shortcuts that work from any app are ⌃⌥ + a letter (V show/hide, S snip), one left-hand press that macOS and most apps leave free, and translating a selection is a quick tap of Right ⌘ on its own. Vox2 keeps those keys from also reaching the app in front (this needs Accessibility, which Vox2 already asks for). If another app uses the same combo (e.g. a window manager like Rectangle), change either one in settings → shortcuts. The ones that only work while Vox2 is in front use ⌘ (⌘, settings, ⌘Y history, ⌘P pin, ⇧⌘F fit).
If a shortcut does nothing. Check settings → permissions. If a row stays "needed" even though Vox2 looks switched on in System Settings, an old entry is in the way: select Vox2 there, remove it with −, add it again with +, then restart Vox2. To clear every Vox2 entry at once, run this in Terminal and allow them again:
tccutil reset All com.chris.translator
Back-translation on Mac. Google's back-translation service often refuses requests from Vox2 on Mac ("too many requests"), so on Mac the ↩ line and its score come from the same Google service as the main translation. Pronunciation (pinyin, romaji…) still comes from the back-translation service; if Google refuses at that moment, only that line is left out.
Not yet on Mac. Intel Macs, and Apple notarization (which would remove the "Open Anyway" step).
Back-translation and the meaning check
The goal: after every translation, you know whether your meaning came through, and when Vox2 can't tell, it says so. The Meaning Check page has the north star, a diagram of where we are and what's next, and every test.
With back-translation on (settings → translation), Vox2 translates the result back into your language and shows it under the translation as a faint ↩ line, tagged with a score such as 92% match. Hover the tag for a short key.
What it measures. How much of your meaning survived the round trip: your text → the translation → back into your language. Good translations often come back reworded ("Good morning" → "Good day"). Counting shared words would mark those as failures, so the score compares meaning, not exact words, using a sentence-embedding model.
Tiers
| Score | Tier | Meaning | Color |
|---|
| 85–100 | high | meaning kept | theme accent |
| 65–84 | moderate | check the details | theme text |
| below 65 | low | likely off: something dropped, added, reversed or mistranslated | theme warning |
can't check (no score) means part of your text came through untranslated (gibberish, a garbled snip, text already in the other language): the ↩ line just repeats it, so a score would mean nothing. The untranslated words are underlined.
With colorblind-friendly colors on (settings → appearance), the tiers use the Okabe–Ito colorblind-safe palette instead, plus a symbol, so the tier never depends on color alone: blue ✓ (high), amber ! (moderate), vermillion ✕ (low), in a darker shade on light themes.
Targeted checks. On top of the score, Vox2 compares your text with the ↩ line for meaning flips: a "not" that appeared or disappeared, he/she swapped, an opposite word, a changed day or month, part missing, a changed number, a changed unit (days → weeks, mg → g, °F → °C), or an "unless" that became "if". Any of these caps the score at 60, the word behind it is underlined in the ↩ line, and the hover card says what changed and which of your words didn't come back. (English for now.)
How well it works. Every change is measured on translations into 11 languages with planted meaning errors, and the results (current numbers, what's strong, what's weak, and the full history) are on the Meaning Check page. In short: it's strongest on dropped parts, changed numbers and a "not" added or removed, weakest on swapped pronouns and opposite words, and says "can't check" instead of guessing when text wasn't translated.
More: how the score is computed, exactly, all test reports, and how to re-run the test.
How it works
| Part | Tech |
|---|
| Shell | Tauri 2: a ~5 MB Rust binary hosting the UI in the system web view (WebView2 on Windows, WebKit on Mac), no bundled browser |
| UI | plain HTML/CSS/JS in resources/, no framework or build step |
| Global shortcuts, typing into other apps, selection grab | Rust: keyboard hook (rdev on Windows; on Mac a listen-only Core Graphics event tap in hotkey.rs), enigo simulated input, arboard clipboard |
| Mac permissions | permissions.rs: checks and prompts for Input Monitoring, Accessibility, Screen Recording and Microphone |
| App language | Every word through tr() (resources/i18n.js), looked up by its English wording in resources/i18n/<code>.json. Machine-translated, checked by translating back and scoring with the meaning model, hand-fixed where flagged (details) |
| Speech to text | Whisper (tiny, base, small or large-v3-turbo; see Dictation models) running locally via transformers.js + ONNX Runtime, WebGPU when available. Kept loaded after use for 2, 10 or 30 min depending on the computer's memory, and released early if memory runs low |
| Back-translation score | paraphrase-multilingual-MiniLM-L12-v2 sentence embeddings via transformers.js, local; released from memory when idle |
| Screen text | xcap capture, read locally: Apple's Vision text recognizer on Mac (ocr.rs), Tesseract.js on Windows and as the Mac fallback |
| Voices | Microsoft neural voices (Edge Read Aloud protocol, ), OpenAI, or system voices |
Privacy
- No account, no server, no analytics.
- Text goes only to the engine you choose. Voice, screen snips and the back-translation score are processed locally (models download once from Hugging Face).
- Settings and history are stored in your user profile. API keys are kept in the system credential store (Windows Credential Manager; the macOS Keychain on Mac), never in the settings file. Keys saved by versions before 0.4.6 are moved there automatically on first launch.
- The free Google Translate and Microsoft voice services are unofficial endpoints and may break; the API engines are the official route.
Roadmap
The principle: a light core, with bigger features as optional add-ons that load only when you use them and unload after (the way dictation and the meaning check already work). Ordered by value for the effort, highest first: what makes Vox2 more trustworthy and useful day to day comes before bigger projects and paid certificates. Each item has an issue for discussion, all under the roadmap label. Ideas and requests are welcome in Issues. Tags: value · effort · platforms.
- Meaning check: sharpen the targeted checks. Shipped through 0.4.20: Phases 1, 1.1, 1.2 and 1.2b. Next: confirm on a second fresh set (run 6), word lists beyond English. Numbers and status: the Meaning Check page. High · small · both. (#28)
- Meaning check: show what changed. Started: the words behind a flag are underlined in the ↩ line, and the hover card lists your words that didn't come back. Next: highlight added content too. High · small · both. (#27)
- Meaning tests: more engines and an independent judge. Every test so far uses Google for everything, one engine grading itself. Re-run with DeepL, Microsoft Translator, Apple's on-device translator and an AI engine (forward and back), score the same pairs with a research judge (COMET-Kiwi or an AI judge), and add native-speaker spot checks (Vietnamese first). Decides #29 with numbers. High · small · testing. (#52)
- Meaning check: translate back with a different engine, so one engine can't agree with its own mistake (Google's own translations score higher than professional ones today). Candidates: DeepL, Microsoft Translator, Apple's on-device translator on Mac (free, private), an AI engine; built in once #52 shows which helps. High · small · both. (#29)
- Explain this (AI-assisted). Select a phrase and ask what it really means: slang, idioms, how formal or rude it is, cultural context, how a native speaker would say it. High · small · both. (#21)
- Personal glossary (AI-assisted). Names and terms that always come out your way: family names and nicknames, work terms, preferred words. Medium-high · small · both. (#23)
- More translation engines: DeepL and Microsoft Translator, with your own key (both have free tiers). ()
Done: Vox2 in your language (0.4.25): the whole app in 133 languages, see App languages. macOS version (beta), see macOS notes. Meaning check says "can't check" instead of 100% when text came through untranslated (#38).
Build
cargo install tauri-cli --version "^2" --locked
cargo tauri build # installer in src-tauri/target/release/bundle/nsis/
Needs Rust and, on Windows, the Visual Studio C++ build tools (on macOS, Xcode command-line tools). Release builds that publish updates also need the updater signing key in TAURI_SIGNING_PRIVATE_KEY.
Mac builds come from GitHub Actions (.github/workflows/build.yml): run the build workflow by hand for a test .dmg. Releases (Windows + Mac together) come from the release workflow (Actions → release → Run workflow, type the new version), which also bumps the version and dates the changelog, so they can be made from any computer; publishing a release made on a PC also makes build.yml attach Vox2-mac.dmg and the Mac update to it. Signing uses repo secrets: APPLE_CERTIFICATE, APPLE_CERTIFICATE_PASSWORD and APPLE_SIGNING_IDENTITY (Vox2's own certificate), plus TAURI_SIGNING_PRIVATE_KEY for updates.
Layout: resources/ is the UI (app.js wires everything; engines.js, dictation.js, tts.js, ocr.js, langpicker.js are the pieces). src-tauri/src/ is the native side (main.rs, hotkey.rs, overlay.rs, permissions.rs, tts.rs).
License
GPL-3.0. Color themes are from Monkeytype (GPL-3.0). Turtle and rabbit icons are from Lucide (ISC). Fonts: Roboto Mono and Lexend Deca (SIL Open Font License).
