voicemate

voicemate

A private voice assistant that runs entirely on your Mac and speaks English and Hungarian. You talk, it listens, thinks, looks things up when needed, remembers what you already researched together, and answers out loud.

  • Local: speech recognition, language model, memory and speech synthesis run on the machine. Only explicit tool calls (web search, page fetch, arXiv, weather) go online, and each one is marked in the UI.

  • Streaming: speech starts while the answer is still being written (about 1-3 s after you stop talking for chat turns).

  • Natural turn-taking: interrupt it any time; it keeps what you heard and treats your words as a correction. Hold (Ctrl+H) lets you pause and think out loud; I’m finished answers everything at once. Turning the microphone off (Ctrl+M) answers what it heard so far.

  • Asks before risky actions: replacing a file pauses the LangGraph agent and asks you.

  • Remembers: facts, past conversations, notes and a research cache; the conversation itself survives page reloads and restarts.

  • One model for everything: choose fast, gemma or qwen; a startup memory check falls back to a model that fits.

  • Transparent: the browser UI shows each pipeline stage live.

voicemate in the browser

A real session: pipeline stages with timings on the left, memory and tool trace on the right.

Quick start

make dev       # create .venv (Python 3.13) and install everything
make models    # pull the Ollama model and download ASR, TTS and embedding weights
make run       # start the UI on http://127.0.0.1:8080

Configuration lives in config/voicemate.toml; the README on GitHub lists the most useful settings.

How it works and how it was built

The API reference below is generated from the docstrings.