Voice assistant bad experiences?

[harsh igloo]

Anyone having issues with the official voice assistant devices they sell? I’m having a horrible experience with them compared to Alexa or Google home. The device doesn’t listen half the time, takes forever to do the task. Sometimes it speaks and sometimes it doesn’t. I’m saying the wakework really clearly and it doesn’t activate. I also can’t say like “set bedroom lights to 50%” it simply does not know the command

[stone stream]

There's reason it's called "Preview Edition".
Don't expect same experience as from huge vendors. HA is tinkerer paradise, not ready-made solution. A lot depends on you.

[spare finch]

To add to this, using an LLM in the backend, exactly like Apple, Amazon and Google do, is completely possible with those devices!

The cost for this is having a machine, ideally with an Nvidia GPU running locally.

There's also a reason the 'Smart' assistants are becoming visibly dumber over the years. Google and others be cutting functions to save cash.

I personally had 16 Google minis at my 'google peak'.
And felt a sharp decline over the years with the quality of its answers and command following. They also kept cutting functions from it.

I now have not one, but 3 VPEs and 6 ESP32S3BOX3 (That name is a mouthful). And could not be happier. But I do run a local server for this.

An alternative is to use the VPEs with a cloud service such as GPT/Claude. This is also very easy and should not cost too much depending on your usage.

I do suggest running this locally tho, as it gives you an entire control over your stack.

From which model is used, to how the personality should be, and even the voice.
(Bonus, you can clone any voice if that's your jam)

[harsh igloo]

I mean even if it’s a preview edition I still expect it to be better than its current state for the price tag.

It’s practically unusable with 50/50 accuracy

[harsh igloo]

Yeah I might hook up a LLM to these, any recommendations? QWEN?

[spare finch]

Qwen 3 4B was my main driver for a while.

Right now I run GPT-0SS20B. But if that's too big, the Qwen3.5 Family just got released!

If it's a Paid LLM that you mean. I'm unfortunately not knowledgable on those as far as what the best for this would be :/

[stone stream]

Get used to the fact that in case of Alexa/Google you're getting device dirt cheap, because you're main product. They're often sold with loss.

[spare finch]

100%! Those are HEAVELY subsidize in order to gain market share.

And eventually... they'll pull the plug and start charging for the service. Once they deem that this department is going from 'fiscal growth' to 'fiscal optimization'

[stone stream]

BTW if you want similar experience cheaper but more DIY, you might grab Respeaker Lite. Or, if you're fancy, Respeaker XVF3800 - it promises to deliver good beam forming.

[harsh igloo]

I’m well aware but I don’t really give Alexa anything worth it anyway

Thanks for the suggestions