Adventists AI Open-Sources DuplexJev, a Speech-to-Decision Model for Voice Agents

Adventists AI today released DuplexJev, an open-weight speech-to-decision model for voice agents.

Instead of transcribing speech and asking language models to write answers, DuplexJev listens to the raw audio and answers the many small decisions voice agents make on every pause: has the caller finished, is this a back channel or a real interruption, how does the caller sound, as multiple-choice questions, all read out in a single forward pass.

"Voice agents rarely fail because they can't answer a question. They fail on the small judgements in between - when to speak, when to stop, whether the caller is getting frustrated," said Jie Jin, CEO of Adventists AI, in a statement. "DuplexJev makes those judgments straight from the voice, in about a tenth of a second, and it is open for anyone to build on."