[!NOTE]
AnyCaptions is free and still in development. It listens to the audio you are already hearing — games, videos, calls, any app — and shows a floating caption overlay on top of everything. Speech recognition and translation run entirely on your machine: no audio or text ever leaves your computer, and the app works offline once installed.
Features
- Live transcription of system audio: captured via WASAPI loopback — no microphone or virtual cable needed, just pick the output device you already use.
- Real-time translation into 37 languages: from Portuguese to Japanese, with automatic detection of the spoken language.
- Fast English ↔ Portuguese path: runs on the CPU (~90 ms per sentence), with the full 37-language model on the GPU.
- Voice dictation: speak and your words are typed into whatever application you are using.
- Transparent, click-through overlay: when locked, clicks pass straight through the caption; recover it any time from the tray or the global shortcut (
Ctrl+Alt+L, configurable).
- In-app model download: the large translation model is a one-click download with integrity verification; everything else ships in the installer.
- Customizable captions: size, colors, width and position.
- UI in 6 languages: Portuguese (Brazil), English, Spanish, French, Russian and Simplified Chinese.
- Private by design: no history, no telemetry, nothing sent to the internet.
Shortcuts
| Shortcut | Action |
|---|
Ctrl+Alt+L | Lock or recover the caption overlay (default, can be changed) |
Ctrl+Space | Start or stop voice dictation (default, can be changed to F6–F24 or Ctrl/Alt + key) |
Tech stack
| Layer | Technologies |
|---|
| Desktop | Tauri 2 · Rust · WebView2 |
| Interface | Vanilla TypeScript · Vite |
| Audio capture | WASAPI loopback |
| Local inference | NeMo-Speech.cpp · CTranslate2 · GPU (Vulkan) with CPU fallback |
Website (site/) | Astro · ASP.NET Core · PostgreSQL |
| Tests | Vitest · Rust tests |
| Code quality | TypeScript strict mode · Clippy · rustfmt |
Models
All models run locally:
Inference runs on NeMo-Speech.cpp (Apache-2.0) for recognition and Riva, and on CTranslate2 (MIT) for OPUS-MT.
Licenses: Nemotron under OpenMDW-1.1; Riva Translate — Licensed by NVIDIA Corporation under the NVIDIA Open Model License; OPUS-MT under CC-BY-4.0 (credits: Helsinki-NLP/OPUS-MT, Jörg Tiedemann). Selector flags from flag-icons (MIT).
Getting started
The easiest way to use AnyCaptions is the Windows installer from anycaptions.app.
To build from source on Windows 11, install Rust 1.90+ (MSVC, x86_64-pc-windows-msvc), the Visual Studio C++ build tools with the Windows SDK, the WebView2 Runtime, and Node.js 24.15 or newer. Then run:
git clone https://github.com/EduardoViniciusJ/AnyCaptions.git
cd AnyCaptions
npm.cmd --prefix frontend ci
npm.cmd --prefix frontend run tauri -- dev
Models and the inference libraries are not required to compile — only to transcribe and translate (see backend/models/README.md).
From the project root:
cargo test --manifest-path backend/Cargo.toml
cargo clippy --manifest-path backend/Cargo.toml --all-targets -- -D warnings
npm.cmd --prefix frontend run test -- --run
npm.cmd --prefix frontend run typecheck
Your data
Audio and captions stay in memory and are never written to disk — there is no history. Settings are stored locally in %APPDATA%\com.anycaptions.app, and models live in the models/ folder next to the executable. Nothing is sent to the internet; the only network access is the optional in-app model download.
