I've been scouting the internet for days and after a lot of digging, I decided to resort to my last hope and ask: I'm talking about the "hey Google", "Alexa" and so on voice activations. All results I could find suggested recurrent neutral networks but I have some serious doubts about that for several reasons:
Running an rnn requires a lot of resources, both memory and CPU. None of which are available in both Alexa and Google home devices-they are really underpowered and have hardly any resources available to spare.
"Hey Google" is also available on pixel phones and while they have plenty of system resources, to my mind this would seriously impact battery life.
So my question is if anyone knows what might be going on here - we are looking at incredible accuracy on hardware that's significantly less potent than a first gen raspberry pi?