Run an Intel GPU-accelerated Wyoming protocol speech-to-text service for your Home Asssistant in Docker.
Utilizes whisper.cpp and Wyoming Whisper API client.
Tested on an Intel Arc A380.
- Intel UHD Graphics for 11th generation Intel processors or newer
- Intel Iris Xe graphics
- Intel Arc graphics
- Intel Server GPU
- Intel Data Center GPU Flex Series
- Intel Data Center GPU Max Series
- Open the
docker-compose.yamland change the necessary values like user and path for all services. - Ensure the
group_addmatches yourrendergroup id. - Install by running
docker compose up -d.
Add this as a service for the Wyoming Integration. When setting it up, enter your host IP (127.0.0.1 if you run the container on the same host and HA is in a container too) and port 7891.
The initial request will be relatively slow, but subsequent ones will be significantly faster.
To improve accuracy, especially of difficult, uncommon words, you may use the initial prompt option.
In the .env file, simply change the value of
- PROMPT=""
It can set to words that are common in your commands, words that whisper.cpp is having difficulty understanding, or some brief instructions. It seems to work best when you just give it all the areas, devices and actions that control them, without any particular structure, like so:
- PROMPT="turn off turn on open close stop play run set timer cancel temperature degrees weather tv lights lamp curtains roomba thermostat AC bedroom kitchen entryway corridor living room gym"
By default, the container is set to use large-v2, but you are free to set MODEL environment variable in the .env file to any supported model.