Using voice to wake up mirror from sleep

[A Former User]

Hi (copain), it’s a PIR sensor job and not an hotword job

PIR sensor is very low cost (5€) and add it to your RPI

After, A2D (Assistant2Display) can manage PIR sensor with screen sub-module for real integration with other sub-modules

[mimic974]

I would add a PIR sensor later but where the Mirror is i’d prefer to call it to show me informations , so it is not possible at all ?

[mumblebaj]

@mimic974 Not sure if this would be helpful in anyway, I have MMM-Networkscanner running on my Mirror and it wakes my mirror when it detects my phone on my network. So when I leave it turns the screen off and wakes up when I am back on the network.

[artieikon]

Thanks for the link to the MMM-assistant, I was about to start working on something similar myself.

From what I remember (truth be told I havent looked at the google assistant in a while) the google assistant does not support customized keywords ie. “Mirror” only things like “Ok Google”.
Anything that does a custom keyword usually uses snowboy (as MMM-assistant does), Picovoice, DeepVoice, or some kind of a custom tensorflow derivative. I have not looked into any of these other solutions because I’m not too much of a fan of having to say “Ok Google” (not to mention the whole “big brother” things that come with it).

Also one thing to note is that some of these solutions (especially a custom tensorflow and potentially snowboy) wont have the accuracy of a google assistant since they wont have the raw voice data that google has.
Sounds like it will come down to: how much legwork do you want to do to create a solution that fits your needs, how accepting are you of a potential “always listening and sending data” solution, and how important accuracy is.

Personally I’m probably going to try and figure out how to do a custom hotword that doesn’t leave my local network, and have that be a trigger for Google’s assistant. No big bro, and only have to worry about training a model for a single trigger word

[A Former User]

@artieikon: sure, I try PicoVoice but i Think it’s too complex for newbies user
I have planed to test tensorflow… maybe it’s the solution

in all case snowboy project will turn off soon… so, I search a solution

[artieikon]

There is also MMM-voice that you can check out. Module hasn’t been updated for 3 years, but the forum post has still been getting some action with people sounding pleased with it.

Took a quick look at what it uses. Something called Sphinx SPHINX.
Never head of it before, but it seems to be developed by Carnegie Mellon so it should be pretty legit. Definitely much less used and known than tensor or any of the others, but being made by CMU means it should be pretty good stuff (although maybe not currently or ever “production ready”)

[sdetweil]

@artieikon sphinx is not near as good as snowboy for hotword detection.

I am sure some team will take over snowboy to keep it going

[A Former User]

that’s why i don’t use it

I prefer my own node-snowboy npm library actually and I don’t find better

[artieikon]

Ahh well thank you for saving me the trouble of looking into it. Figured it wouldnt be as good just based on the adoption