It’s not just google who have AI stuff built into their phones. All recent SoCs I have seen have had NPUs going back the last couple generations. A lot of older or cheap phones won’t have one, but the new devices will.
I don’t see the problem with using the phones normal GPU. This shouldn’t be more insecure than making a call currently is. I am pretty sure android phones don’t have a secure enclave just for making calls as you can give different apps access to calling features, and most calls I make are through third party apps anyway, not via POTS. That being said android is pretty secure anyway provided you don’t give permissions to the wrong app. It’s more secure than your average Linux system, as each app has its own user and is only allowed to access things it has explicit permissions to access. Secure enclaves aren’t all that in my opinion.
And let’s not forget that if the phone can listen to your conversation to detect malicious intent, any country can legally compel Google to provide them with the data by claiming it is part of a law-enforcement investigation.
The point of doing it locally is the audio never gets sent to google directly. That being said they could definitely do some dodgy things by training the ML model to search for words like abortion, drugs, transgender, etc depending on what the laws are in the country the phone is being used in.