Understand what's on screen. Keep enjoying what's underneath.
Translate the dialogue in a game, subtitles in a video, a menu, or text in an image. Press a shortcut, select a region, and read the translation in a quiet, customizable overlay.
Release readiness · Acceptance matrix · Non-invasive testing · Published releases · Getting started · Upgrade guide
DeskTranslate 2.0.0 brings a rebuilt Windows experience, managed local OCR, continuous dialogue handling, contextual translation, title profiles and an editable subtitle overlay. Download the Windows release. Binaries are currently unsigned; see the Windows warning below.
- Choose once or stay live. Snip a region for a single result, or watch a dialogue box for changes. Static frames skip OCR; obsolete requests cannot overwrite newer text.
- Follow your application. Bind a normalized subtitle region to a chosen window's client area. It follows movement and resizing, waits while covered or minimized, and reconnects after a restart only if you enable it. This uses visible-screen capture; keep the region unobstructed.
- Recognition runs locally. Managed, verified RapidOCR/ONNX models replace executable-path setup. Optional Tesseract remains available for existing installations. Japanese, Korean, Chinese, English and Spanish fixtures are included in the test corpus.
- Choose how to translate. Start with keyless quick translation, connect a conventional API, or use contextual AI locally or in the cloud. OCR and translation are independent choices.
- Context for dialogue. AI providers receive bounded recent lines, relevant approved terms and optional story notes. Save these with a title profile. Glossaries belong to a language pair; quick translation applies exact whole-term matches locally. Natural, literal, subtitle and game styles help preserve tone and terminology.
- An overlay that fits your content. Editable subtitle presets, typography, colors, gradients, outlines, shadows, line limits, source/translation layout, position locking and click-through. Enter edit mode with
Ctrl+Alt+E. - A setup for every story. Save, load, update, duplicate, rename and exchange profiles for games and shows. Profiles include language, OCR, provider/model, glossary, context, shortcuts and overlay settings. Keys stay in the OS vault; imported profiles require a new local capture selection.
- A first translation you can see. Guided setup installs the right recognition pack and processes an authored sample with actual OCR and your chosen provider. Back, retry and Set up later remain available.
- Built for everyday use. Searchable languages/models, global shortcuts, tray controls, light/dark and Windows high-contrast support, previewed diagnostics, and explicit stable/beta update checks. Runtime failures pause translation without opening the main window over your game.
- Privacy you can understand. Screenshots remain in memory. Cloud services receive recognized text and configured context, never screenshots. API keys use the OS credential vault. DeskTranslate has no app analytics and opts out of ONNX Runtime telemetry before inference. Local servers have their own logging policies. An optional bounded session transcript stays in memory until you explicitly copy or export it. Privacy and dependency scope.
Save a setup for each game, show or task, then load it when you're ready. The examples above are synthetic; keys stay in your local OS vault. Read the privacy policy or report a problem.
Download Windows x64, primarily Windows 11. The installer's configured minimum is Windows 10 build19041; that setting does not establish current Microsoft support or successful runtime/hardware qualification. Download DeskTranslate-2.0.0-Setup-x64.exe from the official release page. Installation is per user and needs no administrator account. Normal users need no Python, pip, Visual Studio, Tesseract, or manual DLL installation. See the acceptance matrix.
For portable use, extract the entire ZIP into a writable folder and launch DeskTranslate.exe. Keep _internal and portable-mode beside it. Portable settings, logs, and downloaded models live in DeskTranslateData beside the executable; installed builds use %LOCALAPPDATA%\DeskTranslate. Credentials remain in your Windows account's vault. Move the entire folder to retain portable data.
Windows binaries are currently unsigned. No legitimate publisher certificate is available, and Windows/SmartScreen may warn. Download only from this repository. Published SHA-256 checksums and hosted build attestations verify integrity/provenance; they do not establish an Authenticode publisher identity. Runtime/build SBOMs, a source-matched provenance statement, matching library sources and third-party notices accompany the release. Trusted signing remains a future release improvement.
V1 did not persist language preferences. V2 starts with new settings and does not remove an unrelated legacy installation or Tesseract. Read upgrade and data retention before replacing an existing setup.
- Launch DeskTranslate and follow Your first translation. Choose your reading and translation languages; choose Japanese or Korean explicitly when reading those scripts.
- Choose Google · online · no API key for simple setup, or use your configured provider. Configure translation… opens Providers; Return to first translation setup resumes your language choices and requires a new sample test. An existing local-provider choice stays selected when you reopen setup. Google uses a public web endpoint whose availability is not guaranteed.
- Install the selected language pack and press Translate sample. Downloads show progress, support cancellation, and must pass SHA-256 verification. Setup is complete only after an actual successful sample or screen translation. You can reopen it in Preferences.
- Press Ctrl+Alt+T, drag tightly around the text, and press Enter. Escape cancels. Resize a selection using its corners.
- Read the overlay. For continuing dialogue, use Ctrl+Alt+L and select the dialogue area.
For a moving game or video window, choose Choose application, then Region inside application and select its subtitle area. Keep the area visible. A covered, minimized, missing or ambiguous target waits safely; it never falls back to another application's pixels. Fixed regions and entire-monitor one-shot capture remain available.
Use a tight text region for changing video backgrounds. DeskTranslate keeps the overlay outside that region to prevent it from translating itself. If a region leaves no space for an overlay, use a smaller region or one-shot mode; the result can appear in the main window.
The status at the bottom of every page tells you whether capture is running, paused or stopped. Pause becomes Resume; a failed request offers Retry translation. If a language pack is missing, review its size and choose Install pack and start live or Install pack and translate once. Downloads start only after that click. Stop, cancellation or a settings change cancels automatic continuation.
| Action | Default shortcut |
|---|---|
| Select and translate once | Ctrl+Alt+T |
| Select and start live | Ctrl+Alt+L |
| Pause / resume | Ctrl+Alt+P |
| Stop | Ctrl+Alt+S |
| Reselect region | Ctrl+Alt+R |
| Show / hide overlay | Ctrl+Alt+O |
| Copy latest translation | Ctrl+Alt+C |
| Edit overlay | Ctrl+Alt+E |
Shortcuts are configurable and checked for conflicts. First-run setup adds Shift (then Win if needed) when another app owns a default shortcut; the welcome screen shows your assigned shortcut. Closing exits by default; keeping the app in the tray is an explicit preference.
Launching DeskTranslate again restores its existing window. Separate portable folders remain independent. Quit hides the app while it finishes closing; setup cannot reopen or start new work during shutdown.
Choose Midnight mint, Anime subtitles, Cinematic, Paper & ink, or High contrast in Appearance, then customize. Font family, size, weight, italic, text/source colors, separate text/background opacity, gradient direction, outline, shadow, spacing, padding, corner radius and alignment are editable. Position it at the top or bottom center, drag/resize it, or reset the layout.
Click-through lets a game receive mouse input. Edit mode temporarily restores interaction and pauses an active session; finish editing and resume when ready. A reduced-motion option disables the brief appearance fade. Long translations shrink within a bounded range and clip at the configured line limit; Copy retains the complete text. The main window also contains the full result.
Translation targets retain v1's 133-language catalog. OCR sources depend on the selected engine: managed source packs are listed separately from optional Tesseract languages. Provider language availability varies; a listed language does not promise universal OCR or translation accuracy.
API providers may charge usage fees. Trial credits or free quotas are not promised. In Providers, select a service, enter a key if required, and Save and test connection. Choose a discovered model and Apply translation settings. Official provider links are available in the app. Model names are not permanently hardcoded.
Advanced endpoint/API controls are hidden until requested. Known embedding, image, audio and realtime-only models are excluded from text-model lists; unknown compatible models remain selectable with unverified support. Native OpenAI requests disable response storage and streaming; compatible servers retain Chat Completions. Provider setup and recovery explains privacy, costs and common errors. Glossaries support previewed JSON/CSV exchange; imported terms remain a draft until Apply.
Find local AI servers probes Ollama and LM Studio on localhost. DeskTranslate does not download large LLMs or silently fall back from local to cloud. Start/load the server yourself, test it, choose a model, and apply. Translation quality and speed depend on that model and your hardware.
Core algorithms, concurrent stale-result rejection, settings recovery, ten provider request/response contracts, model integrity, UI flows and privacy boundaries have automated tests. Routine and pre-deployment checks run headlessly. Native desktop probes are separate, bounded manual checks for affected boundaries that cannot be verified below the UI; older reports qualify only their recorded scope/source. The installer and portable app are built with PyInstaller and Inno Setup.
The validation report distinguishes measured evidence from unqualified scenarios. The corpus now contains47 authored fixtures across19 managed source languages, including outlined, shadowed, translucent, low-contrast, vertical, mixed-layout and ruby text. The regional Japanese check reads all eight authored vertical layouts and the pixel-font line exactly, with18/20 positive cases exact overall; two punctuation/name cases and ambiguous-background false text remain. The eight layouts also match across all three image treatments, including ruby after contrast/scaling. Warm regional calls on this host ranged from0.98–1.64seconds, with a1.20-second median across20 cases; rapid subtitles may outpace it. The first cold call took15.05seconds, including worker initialization. These small authored tests do not establish accuracy or throughput for every game or video.
- Regions stay within one monitor. Negative origins and mixed-scale coordinate transforms are tested; physical mixed-DPI, hot-plug and exclusive-fullscreen coverage still need a hardware matrix. Use windowed/borderless mode if capture is blank. Protected content cannot be promised.
- Auto recognition uses a Chinese/Latin family and is not universal. Choose the source script explicitly. Vietnamese accents, Chinese punctuation/Traditional characters and Japanese punctuation/name formatting retain fixture errors. Japanese columns use regional recognition with upright captions and conservative ruby handling; enable Vertical columns in Recognition. Paced authored moving/VN/subtitle/ruby/column workflows pass; this does not establish accuracy for every real game or film. Other managed languages retain experimental vertical ordering. A verified model download guarantees integrity, not recognition accuracy; reinstalling the same pack will not fix these errors.
- All adapters have contract/failure tests. Actual Ollama and LM Studio lifecycle tests have also run, but the tiny test model failed plain-text translation-quality checks. Paid-account APIs and a real LibreTranslate server remain unqualified. Choose and evaluate a model for your content; no universal accuracy or compatibility claim is made.
- The current implementation uses local CPU OCR and visible-window following. Cloud OCR, vision uploads, production GPU inference, universal automatic script routing, multiple simultaneous regions, speech and persistent history are not shipped. DirectML and Windows Graphics Capture remain documented investigations.
- macOS/Linux source portability is unqualified. The separate legacy Mac project is not the DeskTranslate 2 release.
Use the hash-pinned CPython 3.12.15 Windows x64 runtime prepared by tools/install_python.ps1 for release builds. Application source accepts Python 3.12–3.14; other interpreters need their own release qualification.
./tools/install_python.ps1
.tools/python-3.12.15/python/python.exe -m venv .venv
.\.venv\Scripts\python -m pip install --require-hashes -r requirements-bootstrap.lock
.\.venv\Scripts\python -m pip install --require-hashes --no-build-isolation -r requirements-windows.lock
.\.venv\Scripts\python -m pip install --no-deps --no-build-isolation -e .
.\.venv\Scripts\python -m desktranslate.\.venv\Scripts\python -m pytest -q --basetemp=.test-tmp
.\.venv\Scripts\python -m ruff check src tests tools
.\.venv\Scripts\python -m ruff format --check src tests tools
.\.venv\Scripts\python -m mypy src
.\.venv\Scripts\python -m desktranslate --smoke-test
.\.venv\Scripts\python -m pip_audit --require-hashes --disable-pip -r requirements-windows.lock
.\.venv\Scripts\python tools/library_sources.py
.\.venv\Scripts\python tools/build.py --installer --qualificationInstall Inno Setup 6 and put ISCC on PATH or set ISCC_PATH to its verified compiler. Build outputs are in dist/: portable ZIP, per-user installer, runtime/build SBOMs, provenance and checksums. Version metadata comes from desktranslate.__version__; a mismatched tag fails the build. Signing and release qualification, CONTRIBUTING, architecture, troubleshooting, and the original audit explain the workflow. See SECURITY for private vulnerability reporting.
DeskTranslate 2 is licensed under MIT. Third-party libraries and downloaded OCR models retain their own licenses. Qt/PySide6 license notices and replaceable shared libraries are included in binary distributions; see third-party attribution.