I noticed an interesting detail

Before this USC study came out, I had just hit a wall with MAI Realtime for a few days. It handles mixed Chinese-English speech and multi-turn interruptions more smoothly than mainstream voice modes on the market, but it starts getting confused when dealing with emotional content or implied meanings. In my testing, if you ask it "Are you sure?", text input recognizes it as a challenge, but voice input might interpret it as confirmation. The study's conclusion matches my experience perfectly: AI is far better at reading comprehension than listening comprehension, especially regarding paralinguistic information like tone, pauses, and stress.