Spoken AI personas don't survive long conversations, according to a new benchmark.
Researchers released RoleBreak, an open benchmark built from 310 character based and user centered roles, 6,688 human verified dialogue turns, and 11,743 fine grained evaluation criteria, including 1,856 turns with specific emotional targets. The team tested nine configurations of spoken dialogue systems, spanning full duplex, omni modal, and cascaded ASR-LLM-TTS setups. Across every configuration, even the strongest system hit its first persona failure after just 10.4 turns on average and its first safety failure after 11.6 turns. Systems held onto the right words far better than they held onto the right vocal emotion.
Most voice assistant demos run for a handful of exchanges, which conveniently sidesteps the exact failure window this benchmark exposes. Scaling up the underlying language model delayed persona breaks, but did little for vocal emotion, and the study found a user's tone of voice can shift how the assistant role plays even when the words spoken stay identical.
A natural sounding voice, it turns out, is not the same as a durable one.