A duplex speech-to-speech model changes the premise: The intelligence layer consumes audio and produces audio directly. The model can attend to what was said and how it was said—content and delivery ...