A voice agent needs to decide whether a pause means “I’m finished” or “I’m still thinking.” Veris AI inserted Jev into that decision in one Pipecat agent and reran its simulated caller benchmark.
What changed
- The baseline used 0.8 seconds of silence to end a caller turn. The Jev version classified the caller’s words and the agent’s last line as complete, short pause, or long pause.
- Across the same 100 tasks, run three times per configuration, Veris reports the agent began talking over the caller on 52% of turns without Jev and 11% with it. Those figures are Veris’s measurements, not a run I reproduced.
- Task success barely moved: 50 of 300 calls passed without Jev, versus 53 of 300 with it. Fewer interruptions did not make the agent reliably accomplish more.
- Median reply delay rose from 2.1 to 4.6 seconds. Background television speech still confused transcription and could keep a turn open.
The comparison’s limit
The Pipecat release also changed—1.8 without Jev, 1.12 with it—and Veris chose the hold times. This is evidence about that combined configuration, not a clean causal estimate for Jev or proof it beats other semantic turn detectors. The sensible adoption test measures interruption rate, reply delay, and completed tasks together on your own calls.