Handling multiple rooms not possible?

[supple cliff]

Hey there,

I'm using Assist in fully local mode, see my config attached, using Wyoming satellites.

I was trying to use a voice command to turn on the lights in two rooms, but Assist understands the two rooms as a single one: "Schalte das Licht im Wohnzimmer und Esszimmer an." ("turn on lights in living room and dining room")
I see Whisper correctly understands it: INFO:wyoming_faster_whisper.handler: Schalte das Licht im Wohnzimmer und Esszimmer an.

Is that currently not possible, or am I missing something?

Thanks and greetings,

Andy!

[vast anvil]

Yes, only one area by default. That's enough for most people.
But it's quite possible to implement this yourself. A simple option is to create automations (custom sentences) with predefined phrases, a more complex option is to work with multiple slots.

[steady tapir]

Using an llm can also accomplish this, but that of course requires some heavy hardware for a good experience unless you use cloud

[supple cliff]

Yeah the automation path I still need to look at. I was hoping for an interface one day where you can just add custom intents/sentences, instead of creating an automation...
Working with multiple slots sounds like modifying the built-in intents? so have a copy of those as custom_component: https://github.com/OHF-Voice/intents/tree/69b193ed423d1fdc286da5603a2b3845f9b04399/sentences/de
or how would I go is that?

[supple cliff]

What hardware are we talking about? I'm running it all on an Intel Core i3 12100 4x 3.30GHz, so not just an RPi. And I also have Frigate running on it with a few cameras and we're still sitting on maybe 20% CPU power.
Maybe I can also make use of the Google Coral TPU I have and use for Frigate?

Happy for some links where I can start reading ❤️ Thx!

[steady tapir]

For llm you'd need a pretty strong gpu. I've seen people use a 4060ti and get fair performance. A TPU unfortunately can't run a LLM, it doesn't have the memory and compute capabilities 🫤

[supple cliff]

ah sad. yah that's out of scope for now sadly. Thx for the replies though!

[vast anvil]

It can be said that the repository you linked to contains phrases for the trigger, while the logic itself is located in the system component. You will need to create your own version of automation. Start by reading the "custom sentences" documentation . Your trigger will look something like this: "Schalte das Licht im {area1} und {area2} an".

[supple cliff]

Actually, that's exactly what I was looking for: https://www.home-assistant.io/integrations/conversation/