翻訳待ち：Verba: Learn a language by talking to an AI never touches the internet

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Uh oh! There was an error while loading. Please reload this page. Notifications You must be signed in to change notification settings Fork 4 Star 4 BranchesTags Open more actions menu Latest commit History 143 Commits 143 Commits Folders and files NameName Last commit message Last commit date .claude .claude .github/workflows .github/workflows .serena .serena .vscode .vscode docs docs public public sample sample scripts scripts src-tauri src-tauri src src website website .gitignore .gitignore CONTRIBUTING.md CONTRIBUTING.md LICENSE LICENSE README.md README.md index.html index.html package-lock.json package-lock.json package.json package.json tsconfig.json tsconfig.json tsconfig.node.json tsconfig.node.json vite.config.ts vite.config.ts Repository files navigation Learn a language by talking to an AI that never gets bored, never judges, and — if you want — never touches the internet. Full conversations with inline corrections, AI-generated graded reading, automatic vocabulary capture, and a voice on both ends. Runs entirely on your machine. ▶︎ Click to play — one session end to end: talk, get corrected, hear the reply. A session is a plan, not a menu. One theme threads today's conversation, reading, role-play, and the words that are due. The coach speaks first and keeps the turn moving. Suggestions are there for the moment you stall — not to answer for you. Voice and dictation, no setup. A 21 MB voice running inside the app: no key, no server, nothing on the network. Why Verba Offline, genuinely. Point it at Ollama and the whole app — chat, reading, speech, dictation — works on a plane. No account, ever. Your provider, your key. Ollama, LM Studio, OpenAI, Anthropic, Gemini, OpenRouter. Swap in Settings; nothing is locked in. It talks back. Speak, get transcribed, get answered out loud. The bundled models download on demand and run in-process. It remembers. Words you stumble on become a spaced-repetition deck, and the deck feeds tomorrow's session. Your data stays put — and it's yours to move. Everything lives in a local SQLite file. There is no server to sign up to, because there is no server. Export it to one file whenever you like, or point Verba at a folder your own sync service watches (iCloud Drive, Google Drive, Syncthing, a USB stick) and pick up on another machine — same words, same history, nothing to set up again. Eight languages, and adding one is a folder. English, Spanish, French, German, Japanese, Turkish, Italian, Portuguese — see CONTRIBUTING.md. Download → Latest release Platform macOS — Apple Silicon (M1–M4) .dmg macOS — Intel .dmg Windows .exe .msi Linux .AppImage .deb · .rpm On a Mac, take aarch64 for Apple Silicon and x64 only for an Intel Mac — the wrong one either refuses to start or crawls under Rosetta. The build is unsigned — here's how to open it anyway Verba has no Apple Developer certificate and no Windows code-signing certificate, and we would rather ship the build than gate it behind either. So both operating systems will refuse to open it on the first try. macOS says "Apple could not verify 'Verba' is free of malware" and offers you Move to Trash. That is not a detection — nothing was scanned and nothing was found. It is what macOS says about any app that hasn't been through Apple's notarisation service. Drag Verba to /Applications, then: xattr -dr com.apple.quarantine /Applications/Verba.app Without a terminal: try to open it once, then System Settings → Privacy & Security → scroll down → Open Anyway. (On macOS 15+ the old right-click → Open trick no longer works.) Windows SmartScreen shows "Windows protected your PC" → More info → Run anyway. Building it yourself sidesteps both, and is the honest option if the above makes you uncomfortable. None of this is unique to Verba — it's the toll every unsigned open-source desktop app pays. First run Pick a language pack and a provider in Settings. Test connection lists your installed Ollama models. Optional — Settings → Speech → Bundled models downloads a voice and a transcriber that run inside the app: Model Size Speed Voice Piper — one voice per language ~21 MB ~7× real time Voice Kokoro — 53 voices, six languages 349 MB ~2× real time Dictation Whisper base / small 208 / 639 MB faster than real time Piper is the default a pack recommends — a fiftieth of the download and several times quicker, which for a conversation turn matters more than the last few percent of naturalness. Hit 🎤 in Talk and say something. Already run your own speech server, or prefer a cloud voice? Any OpenAI-compatible server works (Kokoro-FastAPI, speaches, LocalAI, vLLM), as do ElevenLabs and Deepgram keys. Verba walks bundled → local server → cloud key → OS voice and always tells you which tier is actually speaking. Platform honesty: all four targets are packaged from the same source by CI, but the bundled speech engine has only ever run on macOS arm64. If you are the first to run it on Windows or Linux and no voice comes out, that's a bug worth reporting — not you doing it wrong. Build from source Needs Node and Rust; for offline, Ollama running. npm install npm run tauri dev # develop npm run tauri build # package More What's inside — the design, the speech ladder, the trade-offs we took and what they cost. Contributing — especially language packs. Adding a language is a folder and one import line. License: MIT. Fork it, bundle it, ship it inside your own tool. Topics Resources Readme MIT license Contributing Contributing Activity Custom properties Stars 4 stars Watchers 0 watching Forks 4 forks Report repository