I am experimenting with llama3.2:3b model and have similar issues. When I told it to Pause, it paused music, and also turned off all my lamps in the room.
It requires that you specify devices area, which is annoying. I am trying to teach it to use only aliases and ignore areas.
Here is our latest convo:
Update: in Ollama Instructions I added: “Only use devices aliases and ignore devices areas” and reloaded Ollama integration.
Seems working, I keep testing…
If you have a HUM make sure you don’t have a “ground loop” in your setup / try a “ground loop isolator”.
Installed the VPE last night. About the worst thing I can say about it is that it works well as advertised.
Mics are better at 15’ than the ReSpeaker 2-mic hat mics in my Wyoming Satellites. Music playback is also as advertised. The Wyoming Satellite 2-mic hat setup plays back at 44,100 provisioned with PulseAudio; I believe the VPE plays back at 48K - I don’t know anyone that can actually hear the difference, I know I can’t.
If you want better speaker fidelity feed the audio via the 3.5mm line output to a small powered speaker like Cambridge Audio Oontz Angle 3 or something similar. I feed my audio output to a Yamaha AVR input that then outputs zone 2 into a Monoprice whole home amp feeding 6 rooms that have fairly good speakers (Def Tech). I think the audio piece is great. I can rock the house without clipping and pretty good fidelity.
If you are not someone who likes to tinker and build/assemble kits and fuss with provisioning in Debian/Docker/Python, this device is for you.
If you want to experiment and tinker, then not so much.
I will continue to use my Wyoming Satellites, just not in our living room - this is a please my spouse deal, and so far it has not started responding for no reason.
Yeah I have a few aliexpress ones as cheap and effective.
https://vi.aliexpress.com/item/1005005573884974.html
Just so others know just goes inline to the line-out/line-in.
That’s not accurate, as Stuart pointed out. It depends on the amplifier class.
Analogy: You have a 100kW engine in your car. Assume it’s a manual. Assume you rev it high enough to be producing 100kW, but you’re not in gear. You’re going nowhere, but the engine is still burning gas/fuel. That’s an amp turned to max volume, but with no input sound. It’s still amplifying (in case of class A and AB amps).
Who seriously plugs a 200W class A amp (50% efficiency) to a voice assistant?
You guys are joking, either something is wrong with the amp, or with the instrument used to measure power consumption.
By turning down the volume do you mean rotating the ‘volume’ pot left? If so that would result in attenuating the signal and has nothing to do with amplification or power!
Like I said, I have my old amplifiers and speaker combos (and old powered speakers) connected to Chromecast Audio (network music dongles) for music playback as a cheap way to achieve multi-room audio. Again, I do not have that type of setup in all rooms as for most I only have Google Nesa / Google Home voice assistant speakers in that serve dual purpose for voice control and music playback (several paired for stereo).
For multi-room music playback you would generally have them powered on 24/7-365 so they ara always ready to play music on demand without having to manually turn anything on or off (that is why I had to automate that with power plugs).
If you manually turn the volume down then they can not be remotley controlled. That is why I now instead use smart power plugs, but maintaining those automations are a pain.
Long term the goal would be too replace both all Google Nest / Google Home voice assistant and all Chromecast Audio with amplifiers with Home Assistant voice hardware for both voice control and multi-room music playback solution.
Nice setup. I was not criticising it, I was just skeptical about the idle power consumption but it is what it is.
Second part about the volume control, you are right to maintain audible level at the amp, as it an attenuator you are controlling, there is no risk of high power consumption so long as nothing is playing, contrary to what was previously stated in this thread
Its a matter of choice and a question to the current setup of HA where an Opensource Sonos like wireless audio system such as Snapcast can play 24bit 96Khz that is definately in the realms of some audiophile taste.
The raspberry Pi has a whole cottage industry of high end Dac hats, amps, filters and PSU’s.
I say this because the current system lacking wireless audio cuts out a very active Maker and enthusiast market as not all ClassA Amps are £45,000 as that was just shown as an example of the very high end.
https://vi.aliexpress.com/item/1005008307829609.html if someone wants to who are you to say?
Note that in my analogy I didn’t say it would be consuming full power. I implied that it can still be a significant amount though.
This is not accurate for class A and AB amps. I agree that what one is doing is that one is attenuating the signal (which is why fancier amps will actually use negative dB for its volume scale, because that’s actually what it is doing).
However, you need to understand how the transistors and circuitry in class A and AB amps work. Class D is highly efficient though.
In case of a class A amp, your ratio of idle (quiescent) power to active (averaged) power could be close to 1:1 to 1:1.5. For class AB amps it’s more in the range of 1:2 to 1:3. That’s not insignificant.
Nice if they would deliver them internationally.
Thanks for your answer, I do know a thing or two about audio amplification. I agree that class A amps are inefficient but, still, attenuating the signal by means of turning down the volume will have near-zero (apart from noise amplification) effect on power consumption. That’s all.
Apologies to all who read this thread and expect conversation focused on voice assistant, this will be my last post on the amplification matter.
ok, so there you go Home Assistant Voice PE Enclosure remix by Luka Gra | Download free STL model | Printables.com HA Voice PE enclosure bottom step file with mods.
Work in progress for other parts.
I have a strange behaviour with PE. Yesterday i was able to activate a script. Today it does no longer work. The response is that the script does not exist. Any idea to make it work again?
Areas seem to partly work, but ollama in the mix is too unpredictable at the moment, like you say, it’ll some times change devices totally unrelated to what I’m asking for.
I’ll ask it to play/pause the music and it’ll turn off all of the lights instead because it couldn’t do whatever it needed with the music.
Of course this has nothing to do with the PE, but I’ve had to be careful to not expose devices that I don’t want to risk it switching at random times like cutting the power to my UPS or the 3d printer in the middle of a print, or turn on a heater without me knowing.
I hope this is something that can be tweaked in the system prompt of the ollama add-on but I’ve not seen any updates for it and I’m not sure yet where the prompt comes from but I might experiment with building something with a more refined prompt.
I also get loads of false positives where it’ll get a wake words that wasn’t said then turn everything off in the living room (using the S3 box 3); two nights in a row now it’s turned all of the lights and the TV off on us in the living room, then to make matters worse, it yaps on for 5 minutes about the tools it couldn’t find and devices it changed or didn’t, despite the prompt telling it to be “to the point”;
One or twice a day it’ll actually do what I want and then the rest of the time it’s worse than talking my toddler, because at least he’ll just say “no”.
I’m not sure how the folks on YouTube are getting it to actually do useful stuff by saying things like “it’s a bit dark in here” and it’ll turn on lights or open the blinds because it’s nowhere near that in my testing so far.
Anyone tested PE with multiple receivers in the same room? I have a big room (40sqm) with an open kitchen (20mq) and I’d like to have a setup where closest microphone is triggered when I speak and acts according to the room where it is. So when triggering kitchen for audio it will properly group Sonos speakers to kitchen and use those.
My advice would be using assist to let solve HA stuff and only when not resolved then use the LLM as fallback
Ya Assist pipeline itself has a lot to resolve to throw in the opaque and unpredictable nature of LLMs failing in random ways. Hallucinations in the context of home automation can range from very annoying to dangerous.
And the issue it isn’t really deterministic, it’s billions of params of an overfit model so not like you can reliably fix things.
It is likely that the Home Assistant team could develop a RAG implementation to direct the LLM to a specific domain of expected answers with higher degrees of reliability. Or as an end user if you are using open-web-ui and ollama you could do so external to Home Assistant.
I believe OPEN AI paid subscriptions also let you generate custom instances with RAG data to get more specific outcomes.
