#2025.3.1 BUG: writing makes satellite talk

1 messages · Page 1 of 1 (latest)

snow mist
#

I just updated to 2025.3.1 and now, for some reason, when I chat with the Assistant it plays my question through the wyoming satellite I'm using.
I've not change anything in my configuration but suddenly it's doing this weird thing. Has anyone noticed something like this?

slim bolt
#

what does your assist pipeline look like? are you using an LLM?

snow mist
# slim bolt what does your assist pipeline look like? are you using an LLM?

I'm using Piper as STT, Llama 3.2 3B exposed by Ollama, Faster-whisper for TTS. Only when the request is handed to the LLM it plays back through the satellite and only if the chat screen is in focus. So, from the browser, if I type something and move to another tab it doesn't play what the LLM received as the question

#

I don't see errors anywhere, everything is working as expected. I suspect the issue comes from enabling the announce features for all types of satellites. It wasn't available before the update but now it is

#

I tried debugging following the pipeline but I don't know how it was before this update. But I'm seeing the pipeline calling HassBroadcast during the intent-progress phase. In this phase the message broadcasted is what the LLM understood from my written command and is broadcasting it to my satellite

#

This could be the reason why I'm seeing this behaviour. Even though I don't see the pipeline calling Piper as it does when using the voice

slim bolt
snow mist
#

This issue also cropped up today after I updated. It wasn't happning before. That post was first written on February 8th

slim bolt
#

could be unrelated but its a simalar sort of issue. so might be at least semi related. it was mentioned in the other thrread that dropping to llama 3.1 solved. its perhaps worth giving that a shot?

snow mist
#

I don't have the hardware to run Llama3.1 and I don't see how that would make any difference. The LLM is not involved in the talking back functionality

#

I'm trying to understand, I don't mean to be an asshole

slim bolt
#

yeah i get you. its a strange bug. but it does seem that 3.2 might be sending additional responce commands to other devices for some reason

#

based upon the other thread

#

if you switch to a basic assist pipeline that has no llm attached does it still do it?

snow mist
snow mist
slim bolt
#

did you upgrade from 2025.3 to 2025.3.1 or did you have an older version?

snow mist
#

From 2025.3 to 2025.3.1

slim bolt
#

its super strange. i do thing its a different presentation of the same bug from the other thread. but we dont know whats causing it. for some reason the llm is not just sending a normal responce. it is also trigger an action to TTS to another device

#

i dont think its specific to 2025.3.1 but upgrading it does seem to be what triggered it for you

#

have you tried rolling back the update to see if it starts working normally again?

snow mist
#

Should I try? I think I'm gonna break something if I do

slim bolt
#

i think its worth a shot, can take a full backup beforehand just in case

snow mist
#

I'm sorry, I don't feel like doing it. I have a full backup but I need to keep my house online at least for now

slim bolt
#

fair enough

#

always best to try these kinda things when you have time to fix it fi you need to

snow mist
#

I did a reboot before closing for today: now it broadcasts part of the answer through the satellite

#

I think there's something seriously broken in this release

#

Like proper f*ed. I've had very bad luck with updates. First z2m that broke color temperature for some of my bulbs, now this...

slim bolt
#

yeah i saw zigbee2mqtt had some issues recently. but that got all cleaned up

snow mist
#

Not gonna get into that in this topic to avoid OT. I can only say that I'm starting to see issues right now and maybe it's not so much due to z2m but HA

slim bolt
snow mist
#

Yeah, I need to. I know it's a big project and I can accept some weirdness with voice. But issues with temperature control is really not

slim bolt
#

i think the z2m thing was 2 big projects (z2m and ha) releasing updates very close to each other and was an issue that only showed up when both dropped

snow mist
#

I'll restore the backup tomorrow, even though the announce feature is what I was waiting for. But if that means everything else starting to break, no thanks

#

Thank you for your patience, I really appreciated it

slim bolt
#

no worries. its a strange issue and will take some time to narrow down most likely. if its a more widespread issue we will see some more reports over the weekend

#

good luck

snow mist
#

If you have any way to escalate this I'd like that and be available for all the tests needed

slim bolt
#

ill mention it to others but your best bet for getting eyes on it is by opening an issue on github

snow mist
#

Will do!

snow mist
#

@slim bolt I opened an issue and one other user reported having the same issue. Another user in another thread here noticed Assist spamming HassBroadcast after the announce feature was introduced. I noticed the same and think is the root cause of this issue

slim bolt
#

yeah its a issue thats cropping up. it does seem to only happen with some llm's

snow mist
#

It's not related to the LLM used. I reverted back to Qwen and it did it anyway

#

Hope I didn't bother you with the tag

slim bolt
#

you mentioned when you dont have an llm linked to the pipeline it doesnt do it.
its a really strange issue on how things interact. some are causing it and some are not and some seem to randomly start causing it

#

give it some time for the issue to get some activity and specially when its not the weekend 🙂 may get some updates