vosk for Home Assistant (app)

About

Vosk is an app that provides offline speech-to-text to Home Assistant using the Wyoming protocol. Once it is running, the Wyoming integration discovers it automatically, so setup is finished from that integration. Speech models are downloaded automatically from HuggingFace, and larger models can be placed in a shared folder for better recognition. It has three modes: open-ended, which uses the raw transcript; corrected, which matches results against sentence templates; and limited, which only allows sentences from those templates. Templates are written per language in YAML and support optional words, alternatives, lists and rules, plus an in and out form that lets a spoken phrase send a different transcript. It needs Home Assistant OS, runs on amd64, aarch64 and armv7, and requires no special permissions.

Automatic summary of the docs and configuration of rhasspy/hassio-addons. Something wrong, or are you the maintainer and want your own text here? Get in touch

Setting it up

Add its app repository under Settings, Apps, then install and start it.

Permissions

Needs no special permissions.

Configuration

Correct sentencesStrictness when correcting sentences, where 0 is the most strict and larger values get less strict. This is only used when a YAML file exists for the model's language at /share/vosk/sentences/<LANGUAGE>.yaml.
Limit sentencesWhen enabled, only sentences from the file /share/vosk/sentences/<LANGUAGE>.yaml can be spoken.
Allow unknownWhen enabled with limit_sentences, sentences that are not part of the templates (or "no correct" patterns) will be returned as empty strings.
Preload languagePreloads the speech-to-text model for the selected language. Other models are loaded as requested.
Debug loggingEnable debug logging. Useful for seeing each wake word detection in the logs.