AnkiBrain

See AnkiBrain on AnkiWeb for more information.

AnkiBrain Voice (Kokoro TTS)

Fully local text-to-speech, available in both Regular (server) and Local mode — the engine never talks to any server after setup. Highlight text on a card and press Speak, hit the speaker icon on any chat reply, or let generated cards embed [sound:] audio automatically.

First use: press any speak control (or Settings → Voice → Install voice engine). AnkiBrain downloads a pinned, checksum-verified runtime (~700 MB download / ~1.6 GB disk): the uv package manager, a standalone CPython 3.11, the Kokoro-82M model + voices, and voice/uv.lock's exact dependency set — all into user_files/voice/. No admin rights, no pyenv, no system Python involved, one progress bar with cancel/retry, and repairs re-use the local cache.

• Languages built in: English (US/GB), Spanish, French, Hindi, Italian, Portuguese, Chinese. Japanese is an optional pack (needs cmake + a C/C++ compiler to build pyopenjtalk unless prebuilt wheels are available in the wheelhouse referenced by voice/runtime-manifest.json).
• Automatic language detection (on by default): each text is spoken with a voice for its detected language (Spanish text → Spanish voice, and so on). The voice selected in Settings is the fallback when the language can't be detected or its pack isn't installed; toggle it in Settings → Voice.
• Requirements: Windows 10+, macOS 12+ on Apple Silicon (PyTorch no longer ships Intel-mac wheels), Linux x86_64/aarch64. Windows boxes with long-path support disabled auto-relocate the engine to %LOCALAPPDATA%\AnkiBrain\voice.
• Air-gapped networks: mirrors work via standard env vars (UV_PYTHON_INSTALL_MIRROR, HF_ENDPOINT, UV_DEFAULT_INDEX/PIP_INDEX_URL) inherited from the Anki process.
• Everything lives under user_files/ (survives add-on updates); delete that folder's voice/ dir to uninstall. The engine spawns lazily and unloads after 15 idle minutes to free RAM.

Local Mode Installation

Local mode runs the ChatAI engine inside a self-contained runtime that AnkiBrain provisions itself: no admin rights, no pyenv, no system Python, no shell scripts and no C++ build tools.

Setup

1. Settings -> Local AI Engine -> Install engine (or the AnkiBrain -> Local AI Engine: Install/Repair... menu item).
2. AnkiBrain downloads a pinned, checksum-verified runtime (~380 MB download / ~1.1 GB disk): the uv package manager, a standalone CPython 3.11, and local_engine/uv.lock's exact dependency set - all into user_files/local_engine/. One progress bar with cancel/retry, and a repair re-uses the local cache and takes seconds.
3. Set your OpenAI API key (Settings -> Local AI Engine -> Set OpenAI API Key...).

• Requirements: Windows 10+, macOS 12+ on Apple Silicon (a locked dependency publishes no Intel-mac wheels), Linux x86_64/aarch64. Windows boxes with long-path support disabled auto-relocate the engine to %LOCALAPPDATA%\AnkiBrain\local_engine.
• Air-gapped networks: mirrors work via the standard env vars (UV_DEFAULT_INDEX / PIP_INDEX_URL, UV_PYTHON_INSTALL_MIRROR) inherited from the Anki process, and a pre-seeded user_files/local_engine/cache/ is reused instead of re-downloaded.
• Repair engine rebuilds the runtime from the lock and leaves your documents alone; Uninstall removes the engine runtime only; Reset documents & data separately deletes the vector store, the imported-document cache, temp files and the saved OpenAI key.
• Everything lives under user_files/local_engine/ and survives add-on updates.
• Status and teardown from a terminal: python3 -m local_engine.state --status (and --uninstall).

Remarks

Linux

Please notice that this addon doesn't work when Anki is installed as a Flatpak. To resolve this, simply install Anki from the official website using the .deb package.