Home Assistant is open source home automation that puts local control and privacy first. Powered by a worldwide community of tinkerers and DIY enthusiasts. Perfect to run on a Raspberry Pi or a local server. Available for free at home-assistant.io
You can split it in a variety of ways, but I think the Green will be fine if you're using Whisper and Piper.
If you want to run something with an OpenAI compatible API on another machine, that will work as well, and I believe they are expanding other API interfaces. Basically, if you have something capable of responding in the way HA wants for it's intents, you can do that. Your network roundtrip may not make it faster to respond though.
Sure, HA talks to its voice integrations using the Wyoming Protocol, see here for a list of available services: https://github.com/rhasspy/wyoming
You can, for example, follow the link to piper, run the docker container on your home server and then add a new "Wyoming Protocol" device in HA (you will be asked for IP/Port). You can then select your new STT service in the settings for Voice Assistant.
To answer your question, yes.
However, I’ve found that unless you’re going to host your own AI engine with a GPU, you’re stuck with OpenAI or Google Gemini. Luckily I’ve been able to use both with just the free tier.
Whisper and Piper run locally although Piper is a little slow on the green for my tastes. 4 second responses work, but are also frustrating when you’re not sure if it heard you or not. You’re also stuck with just the built in intents (phrases) so you need to be specific in how you say stuff.
Ultimately I set up a cloud based and a local based voice workflow and then I setup a ping probe to a IPs. If I detect the internet is up, it uses the cloud workflow. If the internet is down, it swaps to local to keep the house voice controlled.
Thank you! Honestly, all I'm trying to do is have it play some music in rooms, control lights and create timers during cooking. I was hoping whisper would be enough for that and I can get the 4s response time down with running it on a ryzen chip than the green.
I’ve also been using Speech-to-Phrase, which has a limited number of supported phrases BUT works much faster than the 4s, so it’s a good tailback when the primary methods aren’t working.