keyword detection like “Hey Google” is only used to wake up a device from a low power state to perform more powerful listening

That’s more applicable for something like a Google Mini. A phone is powerful enough, especially with the NPU most phones have now, to perform those detecting efficiently without stepping up the CPU state.

Is there some kink of roleplaying AI dev?

Is there some kink on your side in pretending you’re smart? You have no idea who I am or what I know.

Increasing the number of keywords to thousands or more (which you would need to cover the range of possible ad topics) requires more processing power

Again, you’re showing your lack of knowledge here. A model doesn’t use more power if trained on one class or a hundred. The amount of cycles is the same in both instances.

It’s usually smart speakers that have a low powered chip that processes the wake word and fires up a more powerful chip. That doesn’t exist in phones.