It’s less CPU intensive to monitor for a single trigger phrase (“OK Google”) after which the more general, CPU-intensive speech to text is performed (either on

It’s less CPU intensive to monitor for a single trigger phrase (“OK Google”) after which the more general, CPU-intensive speech to text is performed (either on device or in the cloud). The devices aren’t translating all speech to text all of the time and then checking it against the phrase “OK Google”. They’re looking for the trigger phrase.