content-understanding

This paper evaluates full-duplex speech models' ability to decide when to speak, finding that models like Moshi and PersonaPlex primarily respond to being addressed or silence rather than content-driven triggers such as false claims or hazards, identifying a gap in content understanding.