I would assume it's related to limitations in the early wake word detection hardware and software which require pretraining of the weights, which were usually h

I would assume it's related to limitations in the early wake word detection hardware and software which require pretraining of the weights, which were usually hard coded into the low power edge system. Nothing would stop a modern system from training a new wakeword in a cloud system, if enough samples were available.