An early proof-of-concept Home Assistant add-on that exposes
wyoming-vosk to Assist, so local
speech-to-text can run with Vosk.
The current focus is simple and practical:
- get a German Vosk backend running in Home Assistant Assist
- keep setup lightweight
- prioritize speed over maximum transcription quality
Home Assistant's local Assist setups often use faster-whisper for
speech-to-text. That is usually more accurate for open-ended speech, but on
smaller hardware it can also feel slow.
This add-on explores a different trade-off:
- Vosk is noticeably faster than
faster-whisperin this setup - startup and response feel lighter
- recognition quality in open-ended mode is often worse than Whisper
In short: this project is for people who want to test whether a faster local STT path is "good enough" for their own Home Assistant usage.
This PoC is tuned for a lightweight German test setup:
- language
de - model
vosk-model-small-de-0.15 - mode
open - port
10300
The add-on downloads the configured Vosk model on first start when
download_model: true is enabled.
This project is intentionally early-stage and has a few important limitations:
- It is an early PoC, not a polished production add-on.
- It currently targets one simple use case first: German Assist STT with
vosk-model-small-de-0.15. - Open-ended Vosk recognition is generally less accurate than Whisper.
- Discovery and networking can vary depending on how the add-on is installed;
in some setups, manual Wyoming integration via the host IP and port
10300may be needed. - This repository has not been hardened for all Home Assistant installation styles, hardware classes, or multilingual setups.
- No automatic fallback to Whisper is implemented.
This add-on is built around the following upstream projects:
Vosk itself is licensed under the Apache License 2.0.
Please also note:
- Vosk models may have their own licensing terms
- you should review the model license before distributing or embedding a model
- this add-on currently defaults to
vosk-model-small-de-0.15
What works today:
- running Vosk as a Wyoming speech-to-text backend
- connecting it to Home Assistant Assist
- using a fast local German STT path instead of
faster-whisper
What still needs work:
- cleaner auto-discovery behavior across installations
- more robust docs for all Home Assistant deployment types
- broader model selection and better configuration UX
- optional sentence-correction / limited modes for higher accuracy
See DOCS.md for installation and configuration details.