Discussion
I’ve been having a lot of fun making and using voice-powered interactions—all completely local, all built in Home Assistant’s automations directly.
It’s amazing how far you can get with “““AI””” that is actually just some clever scripting that never touches an LLM.
This was likely how earlier versions of Google Assistant worked, back when Google Assistant was actually helpful and really good as an assistant around the home. Now it’s using the Gemini LLM and is next to useless.
@cassidy What do you use to interpret the audio?
@worik I’ve tried whisper and Home Assistant’s cloud processing, the latter of which is admittedly not local but has been more reliable with my current hardware. I might beef up my Home Assistant server at some point for better local STT, or I just need to spend some time configuring it better.
Those are both LLM models, I thought.
Whisper is Open AI and Homeassistant is all in on AI, are they not?
@worik Whisper was developed by OpenAI but it’s an open source STT engine. My understanding is that it’s a more traditional ML transformer rather than anything to do with LLMs. I imagine Nabu Casa is using the same or a similar engine.
Something Google got right with Google Assistant on home devices was that you could say “Send feedback” and then it would actually send a feedback report to the Google Home/Google Assistant teams. As we migrate to Home Assistant Voice Previews, I realized I didn’t have a way for my family to replicate this, but to send the feedback to me.
Now I have a nice little feedback list! And it’s not super robotic sounding, either! Triggered by voice along with a timestamp for future debugging.
Sentence triggers are a super power. For real. And while building this, I saw someone in the Home Assistant forums use Jinja templating with a random filter to give Home Assistant multiple possible responses, which I loved… and maybe took too far by composing a whole reply from a bunch of pieces.
It's so satisfying, though, and really does make it feel natural and “smart” even though it’s predetermined. And it’s *fast.*
Install bonfire.cafe
Get the full app experience
![Home Assistant automation showing "sentence" triggers, including several variations to capture utterances:
- (send|share|submit|leave|log|record|save) [some] feedback
- I [have|have got|'ve got|got] [some] feedback
- I (want|need|would like|wanna) to (send|share|leave|give) [some] feedback
- (give|add) [some] feedback [to the list]
- feedback
- [I have] [a piece of|some] feedback [for you|to share]](https://cdn.masto.host/mastodonblaedefamily/media_attachments/files/116/648/165/620/306/274/original/a6709682229ecb55.png)
