Had this explained once, I might miss a detail, but it’s like this:

The only way not to drain your battery is to program in selective key words.

“But then its always listening” yes, but also, no.

Imagine someone speaking into a microphone, and seeing their voice bounce around on a oscilloscope.

This compresses the audio a LOT, and makes it very difficult to discern the differences between words.

But if you were trained to notice the pattern for a specific word, like “Siri”, then you could ignore all the other shapes, conserving your battery.