Author here. Super excited about our new model as it solves a huge challenge that Voice AI agents run into today: timing.
We tried every end-of-turn model on the market, and none of them really worked well in production. False negatives mean waiting an awkwardly long time to respond (very frustrating), and false positives mean the agent cutting people off (even worse).
I'm proud of how our model stacks up! On LiveKit's public end-of-turn benchmark, Tango catches 96% of turn endings, responds in about 800ms on average, and barges in 5% of the time. It's also the fastest of the 11 systems tested once you require at least 90% detection.