[draugadrotten]

I am switching out the Alexas in the house to the home assistant voice devices, which can leverage a local LLM without any cloud whatsoever

https://www.home-assistant.io/voice-pe/

The process is still complicated enough to be "enthusiast" (aka nerd) territory but it is getting better with every release. It will still be here in 10 years, nobody can take it away from us.

[dwayne_dibley]

I think the audience here might be ok with '"enthusiast" (aka nerd) territory' hardware.

[elif]

" Why did you pick these default wake words and not something like “computer” or “okay assist”? A wake word should be uncommon in everyday conversations at home or in media, such as music or TV, to minimize the risk of the device activating unintentionally. “Nabu”, “Jarvis”, and “Mycroft” ... "

They hardcoded the wake words in hardware...

Why not just use LLM common sense to say "does this really sound like a purposeful activation?"

Or put a GPU in there, or export the call to your PC like they require for text to speech?

For being a DIY thing, they made it inexplicably hard to D

[greycol]

They actually have thought it out well. The short of it is you can set up a device to constantly stream to home assistant so that you can use any wake word. That comes with the draw back of more power use and more importantly higher cpu use for each microphone you add. It's still possible (https://github.com/dscripka/openWakeWord)

The alternative for a dedicated low power device is to have trained a model to run on smaller micro controllers so that it can run locally on low powered devices (https://github.com/kahrendt/microWakeWord) this is what they have chosen for their dedicated devices.

This choice also comes with much higher default privacy. Which is great as home assistant offers cloud integration and the fact that they put privacy first in this area makes it much easier to trust they do the right thing in other areas.

A brief run down on the subject https://www.home-assistant.io/voice_control/about_wake_word/

[arcrwlock]

Because then you'd be running a full speech to text model all the time, and an LLM any time any speech is detected.

[stogot]

How many hours did you spend on this? I want to do it, but I don’t have much time.

[timwis]

For what it's worth, I was worried about that as well, but I found it to be _fun_, and I actually look forward to playing with it, even before bed. Weird, I know, but I'm enjoying it and didn't think I would.