A voice assistant that listens without phoning home

What it does

Rhasspy is a fully offline voice assistant toolkit built for home automation. You define commands in a lightweight grammar template, and it converts recognized speech into structured JSON events that trigger actions in Home Assistant, Node-RED, or OpenHAB. It supports a wide range of languages using open-source backends like Kaldi, Pocketsphinx, and DeepSpeech.

The interesting bit

Rather than shipping your voice to a cloud NLP service, Rhasspy expects you to describe exactly what sentences it will hear. This grammar-first approach keeps inference local and accuracy high. Under the hood, version 2.5 is a collection of independent services that coordinate over MQTT via a superset of the Hermes protocol, so you can replace the speech recognizer or intent handler without disturbing the rest of the stack.

Key highlights

• Runs fully offline with no required cloud connection.
• Plugs into Home Assistant, Node-RED, Jeedom, and OpenHAB via MQTT, HTTP, and WebSockets.
• Handles uncommon words and custom pronunciations with automated phonetic assistance.
• Exposes every web-interface feature through a comprehensive HTTP API.
• Supports over a dozen languages, from English and German to Mandarin and Hindi.

Caveats

• The README explicitly targets “savvy amateurs or advanced users,” so expect a learning curve.
• Open-ended transcription is possible, but the system is optimized for rigid, pre-defined grammars.
• The project is split across many sub-repositories, which may complicate troubleshooting.

Verdict

Rhasspy is for home-automation tinkerers who value privacy and want local voice control. If you are looking for a conversational cloud assistant that works out of the box, the authors suggest Mycroft instead.