What it does

Rhasspy turns voice commands into JSON events that trigger automations in Home Assistant, Hass.io, or Node-RED. You define commands in a profile using a template syntax, and it handles wake-word detection, speech-to-text, intent recognition, and text-to-speech — all without an internet connection. It runs in Docker and exposes a web UI on port 12101.

The interesting bit

The project is essentially a curated integration layer: it bundles existing open-source engines (Pocketsphinx, Kaldi, Porcupine, Fsticuffs, eSpeak, etc.) and maps their uneven language support into a single coherent system. The matrix of which engine supports which language is refreshingly honest — Vietnamese speech-to-text only works with Kaldi, for instance, and several wake-word engines need custom training for non-English use.

Key highlights

• Supports 14 languages, though coverage varies wildly by component
• Wake-word options: Pocketsphinx (broadest), Porcupine (English-only), Snowboy/Precise (require training for most languages)
• Intent recognition is the most uniformly supported layer; Fsticuffs, fuzzywuzzy, and Adapt work offline for all languages
• Text-to-speech defaults to eSpeak for universal offline coverage, with MaryTTS, PicoTTS, and others as alternatives
• Explicitly targets “advanced users” comfortable writing their own Home Assistant automations

Caveats

• This repository contains version 2.4; active development has moved to github.com/rhasspy/rhasspy (version 2.5)
• The author candidly recommends Mycroft if you want something easier and don’t mind cloud processing
• Language support is a patchwork: Mandarin and Hindi have wake-word and STT coverage, but no TTS via MaryTTS; Swedish lacks wake-word support entirely

Verdict

Worth a look if you run Home Assistant, value privacy, and don’t mind trading polish for control. Skip it if you want turnkey setup or natural-sounding voice synthesis — eSpeak’s robot diction is functional, not friendly.