This would be immediately obvious in a cursory analysis of performance. On-device transcription is not only computationally infeasible, it would also require model capabilities far beyond what is currently SOTA.

Google had (and has afaik) significant challenges implementing multiple wake-word detection for precisely this reason.

Transcribing a couple of words accurately on-device without a major performance penalty (so that it can be running in the background always) is just _barely_ coming out now.