DETECTAgents & automation
JEV End-of-Turn Detector for Voice Agents
FP4 Brain@thefp4brain
0:09This project evaluates JEV as an end-of-turn detector for voice agents, comparing its performance against another open turn detector on labeled and streaming ASR datasets.
tested jev (@typesafeai) as an end-of-turn detector for voice agents
one noul question per 500ms of speech: "has the speaker finished their thought?"
turns out it holds surprisingly well
ran it on 3,500 labeled turns against LiveKit's open turn detector, same transcripts:
human-labeled set: 95% vs 86%
streaming ASR set: 90% vs 90%, but jev cuts people off 2.6x less
it's nor perfect though, because in terms of latency you get about 340ms over the network vs 3ms for the local alternative
jev is very plug and play, so it's useful to quickly validate ideas that depend on classification subproblems
and hey, you can always train an embedding model for your specific scenario if you want it better/faster/local