The Chrome Browser can transcribe audio into text, with what I consider good accuracy. It's well out of the realm of a conspiracy theory when it's been demonstrable for a couple decades.
Don't forget energy usage. The phone would need to be on high power mode all the time to run those kinds of algorithms. There's a reason "Hey Siri" has dedicated low-power hardware - it means it can work without burning through the battery.
> it can work without burning through the battery.
It can work by burning through the battery. When you have a browser open or any number of apps, some of them are certainly detecting.