Tavus's Sparrow-2 models the state of a voice conversation using speech content, speaker identity, tone and surrounding audio. Its technical description says an acknowledgement such as 'mhm' can let the assistant continue speaking, while a genuine interruption tells it to stop. Nearby conversation and unclear audio are classified rather than automatically discarded.
The model updates its conversational-state estimate in 10-millisecond frames. Tavus says it trained on 16,000 real and 2,000 synthetic conversations, with interruptions and acknowledgements represented explicitly. The public demo displays detected conversational events as a user talks to a Tavus assistant.
Tavus reports 92.4% end-of-turn recall and 97.4% interruption recall on TurnBench. These are its preliminary measurements on the benchmark's public development split: 38 conversations covering 7.3 hours of audio. The technical post says results from the private test set were still pending, so the demo's headline ranking should not be read as a verified private-test result.
Tavus lists Sparrow-2 as available through its PAL assistants, APIs and platform. The published results measure turn-taking, not the reliability of an entire voice agent or every noisy-room scenario described in its launch material.