The models listening for the initial hook of "okay google" or "hey cortana" etc are stored and executed locally on the device-- usually in a low power manner.

The models listening for the initial hook of "okay google" or "hey cortana" etc are stored and executed locally on the device-- usually in a low power manner. This comes with a set of constraints: Certain activation phrases are more catchy than others, with hard consonant sounds and enough length for the alg to grab on to. Allowing users to pick any name opens this up to a high false positive rate, where the device would 'respond' to arbitrary inputs rather than specifically its name when called. Training models on new trigger words and then deploying them back down to the device is in its infancy, and voice detection has a long way to go (another example of this is the difficulty around responding to a specific person's voice, with its signature sounds etc). I could see this happening in the future as the tech improves.