A new paper tested four leading real-time voice systems (OpenAI's GPT Realtime 2, Google's Gemini 3.1 Flash Live, Alibaba's Qwen3.5 Omni) on calls where *how* something is said matters as much as the words. The systems ended calls with crying callers who insisted nothing was wrong, approved wire transfers requested in frightened voices, and enrolled callers whose "yes" was clearly sarcastic — acting on the words, not the voice. The twist: it's mostly NOT a perception failure. When ...