Study finds real-time voice AI hears words but misses vocal delivery
arXiv·medium signal
arXiv 2606.26083 evaluates four leading production real-time voice systems (including OpenAI's realtime stack) and shows they transcribe lexical content well but largely ignore paralinguistic delivery — tone, emphasis, emotion. For builders shipping voice agents, this is a concrete reliability gap: the agent 'hears but does not listen,' so downstream reasoning loses cues a human would act on. It quantifies a failure mode that marketing demos rarely expose.