That’s STT, curious about wake word. I am running microWakeWord so if you are running openWakeWord then maybe there is something different / an issue.
Shoot my bad. Microwakeword on tje device with hey jarvis.
Have you experienced the Disconnecting bug with these devices? I think @msp1974 had a fix in for this to make it more resilient but im still seeing a ton of these disconnects. the lucys_room is an echo show 8 and i’ve explicitly chosen 2.4ghz channel as i’ve read that its more stable. master_bedroom is a fb portal which is connected to 5ghz so i dont think its related to my network.
Logger: custom_components.vaca.assist_satellite
Source: custom_components/vaca/assist_satellite.py:112
Integration: View Assist Companion App ([documentation](https://github.com/msp1974/ViewAssist_Companion_App), [issues](https://github.com/msp1974/ViewAssist_Companion_App/issues))
First occurred: July 19, 2026 at 9:00:13 PM (123230 occurrences)
Last logged: 7:18:59 AM
* Satellite vaca_2b53c04e0 has been disconnected. Reconnecting in 10 second(s)
* Satellite master_bedroom_parents_assistant has been disconnected. Reconnecting in 10 second(s)
* Satellite lucys_room_lucys_assistant has been disconnected. Reconnecting in 10 second(s)
I only connect to my 2.4 GHz IoT network but no I haven’t really had any issues with them disconnecting frequently
Just a follow up here. I debugged a ton and it seemed like the wyoming connection kept getting disconnected/reconnected with different ports and that was causing the pipelines to fail. What (sadly) fixed it was a HA restart and now i have a stable connection from the echo show 8 (lucys_room) and will test out the same from the master_bedroom portal device. I am guessing it’ll work just fine.
Hi All,
I’m looking for some best-practice advice on getting a locally-hosted Conversation Agent working but using Home Assistant Cloud for the Speech-to-Text and Text-to-Speech part (the ability to remote-in to my Home Assistant from anywhere is worth the monthly fee, which is why I’ve gone this route).
I have followed this thread as best as I can, but the sheer amount of changes and development that people have been up to (especially Nick and Nathan) over just several months is just mind-boggling. 
The object of the exercise is to be able to rip-out my Alexa boxes entirely and have a local LLM/AI do all the things that I currently use Alexa for, although this is mainly limited to turning lights on and off, setting timers, and getting answers to small bits of trivia (like “How old is Reese Witherspoon?”).
My current setup is as follows;
- Latest Home Assistant on latest Home Assistant OS, running on a Hyper-V instance on my Home Server.
- The Home Server has 64Gb memory and has an Nvidia GeForce GTX 1060 (6Gb memory) GPU, which I know is underpowered for this kind of work but it’s the best I can do for the foreseeable future.
- Ollama is also running on this Home Server to provide the actual LLM bit, and is currently using the latest “Qwen3 4B Instruct Q4_K_M” model as this seems to fit in the 6Gb of GPU memory nicely.
- 1x ‘Home Assistant Voice PE’ box as the satellite/endpoint.
- 17 Entities ‘exposed’ out of the 300+ present in HA.
I know different users have been playing with different Models (like Qwen vs Gemma) and getting interesting performance results, so I would also like some advice on what Model I should focus my current efforts on, given my limited GPU. I’m looking to minimise the gap between the end of the question and the beginning of the spoken answer.
Best wishes to all,
AlanG
As far as using Nabu Casa STT and TTS goes, there really is no difference, that is just a UI config. Regarding your 1060, I don’t think you’ll have great result but worth trying as a first step. You will definitely want to use llama.cpp, not ollama. IMO Ollama wastes too many resources and hides too many config knobs you will be leaving too much on the table with that limited amount of space.
I’d suggest starting with Gemma4 E4B QAT unsloth/gemma-4-E4B-it-qat-GGUF · Hugging Face
Agreed on the model.
CTX size is what I’m concerned about. Think what… 4K CTX after all is said and done?
Could do small summary jobs though. With speech to phrase and local processes first. May not be that bad.
Hi Nick,
Thanks for the advice. I’ve been looking at the ‘config.ini’ and ‘prompt’ sections of your Github for help, and whereas I’ve been able to put together a good Prompt, it is difficult to tell what settings in your config.ini are set for your specific GPU.
Could you please post a bare-bones config.ini suitable for my 6Gb GPU that I can tinker with?
At the moment I am using LM Studio as my back-end as this is based on Llama.cpp and the config.ini parameters can be directly translated into LM Studio settings. I know that it’s slower than ‘raw llama.cpp’ so I am just using it to just for initial development.
In other news, I’ve picked up an 8" Facebook Portal which I have semi-jailbroken and plan to use as a Home Assistant satellite/endpoint. More details to follow…
Nothing in it is really GPU specific. Most things are use case specific with the main things you’d want to change being:
- remove the ubatch and batch settings as they increase VRAM usage
- Probably can fit only 1 slot and less context
- You might need to use q8_0 kv cache
Hi Nicolas,
Firstly a big thank you for all the work you’ve put in to this - it was extremely helpful.
I’ve managed to tweak the config to get Qwen3-ASR and Gemma4-12B+MTP working extremely fast on my 5060Ti 16GB - speech is processed in under 1s and Gemma answers in under 1.6s for almost any query.
The only problem I have is that I cannot get the OpenAI STT to send speech to llama.cpp without waiting out the 5 seconds.
I’m using a FutureProofHomes Satellite and have tried various combinations of vad_speech_threshold and mic_gain but cannot get it working.
Any ideas?
If you’re using my STT integration then enable debug logging and it will save the audio clips and the logs for each session. That will show what is going on.