Using Jev as a semantic turn detector in Agora ConvoAI instead of silence VAD
The author plugged Jev into Agora ConvoAI as a semantic turn detector with silence VAD turned off, so every turn decision comes from text at about 275ms and 2,100 tokens per call, handling hesitation and unfinished thoughts.
Jev just solved a problem real-time voice AI has had since the beginning: knowing when a human is done talking. I plugged it into Agora ConvoAI as a semantic turn detector. Agora's silence VAD is off. every turn decision comes from text. "uh-huh" → 0.03. hold. "wait, stop" → 0.86. interrupt. 275ms per call. ~2,100 tokens. It handles the moments silence VADs can't: hesitation, unfinished thoughts, someone in the room talking to their kid instead of the agent. Not a faster silence detector. a model that reads the conversation and asks one question: has this person handed me the floor? http://github.com/AgoraIO-Community/convoai-jev-vad
This post quotes another post: “OpenRouter brings Jev to all LLM calls for routing”