... listens for the keywords ...
To be clear, the listening app is so that the model isn't doing tons of processing for the audio ...
Moldy I would prefer not to use Google much.
The ultra-low-power, wake-word processor in the pixel will be using a google model ("hey google") if it even exists in GOS. I haven't heard of anyone changing it and we cannot as users.
Subverting this ULP processor will be a massive power hog compared to intermittent deep sleep. Software would always be running.
The hardware wake-word might only engage google's assistant so we wouldn't have a choice in that. Moreover, I'm fairly sure the majority of developers wouldn't create several builds to target specific TPUs, so they aren't used--higher power usage, slower inference.
I'd wager all FLOSS AI photo editors are extremely lacking.
https://github.com/T8RIN/ImageToolbox
Looks like this has upscaling, denoise, and background removal.