I have a fresh HA install with full local assistant and Ollama integration to make it more human like.
I have one HA Voice device and one Satellite1, both are attached to the same HA assistant, but linked to different rooms, with a goal to have one device in each room eventually.
For now both are sitting on my desk and HA Voice has it's mic muted to prevent both activating on wake word, so I'm only expecting Satellite1 to wake up and respond.
However, when I talk to Satellite1, sometimes the response comes from HA Voice and then repeats again from Satellite1, but it's not predictable, seems random.
Is there a mechanism to ensure response is sent to the same voice device when multiple devices are configured?
#Multiple assist devices issues
topaz steeple
heady zealot
This is clearly an bug; you need to figure out which system component is responsible for it.
Each satellite uses a unique dialogue stream with a unique identifier, and they do not exchange data.
Are your devices updated to the latest firmware version? As far as I remember, there was a discussion of similar issues on the community website or here.
wary scroll
I have 5 Koalas (Respeaker Lite based satellite, with VPE-like firmware), and Satellite1. No problems at all with any of them (except of rare occasional reboots..)
trail shoal
I had this problem with my VPEs. It was being caused by llama3. Somehow on its own it was deciding to broadcast to other VPEs. I switched to qwen2.5 and haven't had a problem since.
fierce jungle
I have a fresh HA install with full local assistant and Ollama integration to ma...
so, some models have this issue and it randomly crops up. sometimes instead of responding they think "I am a voice assistant i must speak therefore i must call a TTS tool" so they call the tool separately from a normal response and it hits other devices.
the general solution is to add something to the prompt to tell it to not do it. "respond only in plain text, do not call any TTS tools" or something like that. a little bit of prompt engineering might be needed to get it to work with your setup.
topaz steeple
when trying different Ollama models and prompts, how do you verbose log and debug it? I tried verbose debug in ollama server, but it's still not showing content of API calls from HA, so I assume it must be done on HA side, but I don't know where and how to access it. I was able to log Whisper to see how poorly it decodes my speech, but that's a different problem for the future.