Tavus's real-time turn-taking model — it reads the whole audio scene every 10ms to decide when a conversational AI should listen, wait, or speak. It reasons over noise rather than cancelling it, and it ships inside Tavus's platform.
Last verified · 2026-10-23 · by Moe Ameen
Published 2026-09-09 · Updated
Short answer
Sparrow-2 is Tavus's real-time turn-taking model, announced on August 27, 2026. It decides when a conversational AI should listen, wait, or speak by reading tone, pauses, speaker identity, backchannels, interruptions, and noise every 10 milliseconds. On TurnBench's dev split, Tavus reports 92.4% end-of-turn recall and 97.4% interruption recall. It powers Tavus avatars and makes no content itself.
Key takeaways
Sparrow-2 is Tavus's real-time conversational-understanding model, announced on August 27, 2026. Its job is turn-taking: deciding, moment to moment in a live conversation, whether the person talking to an AI has finished a turn, is just pausing to think, or is interrupting. It sits inside Tavus's Conversational Video Interface (CVI) and PAL agents, alongside the company's Phoenix (video rendering) and Raven (emotional intelligence) models, and it powers Tavus's real-time avatars.
Its defining idea is to treat the whole incoming audio stream as evidence. Instead of the traditional voice-activity-detection-plus-timeout approach, Sparrow-2 models tone, pauses, speaker identity, backchannels ("mm-hm"), interruptions, and ambient noise together — updating its read of the conversation at a native 10ms frame rate, up from Sparrow-1's 40ms. Tavus frames this as a fresh take on the "cocktail party problem": understand the noisy room rather than scrub it. Notably, its handling of noise is understanding, not cancellation.
Tavus positions Sparrow-2 as state-of-the-art on TurnBench, an independent, triple-annotated benchmark with a public leaderboard. On the dev split it reports the highest end-of-turn recall in the field (92.4%) and the highest interruption recall (97.4%). TurnBench is third-party; treat the specific figures as Tavus-reported until independently reproduced.
Two boundaries matter for anyone doing content work. First, it ships inside the Tavus stack, reachable through the Tavus API, the no-code PAL Maker, and CVI; Tavus's announcement describes no open-source release. Second, it is infrastructure for live agents: it writes no script, renders no video, edits no audio, and publishes nowhere. It decides when the AI should talk. That is the whole product.
Sparrow-2 does not make content — it makes conversations feel human. But the moment you run a Sparrow-2-powered agent live, you generate something a content engine can use: recordings. A recorded demo of your avatar handling a tricky sales question, a customer conversation you have consent to share, a screen capture of the agent working — that raw conversational footage is the exact input Kompozy is built to turn into publishable content.
Point Kompozy at the recording and it produces Clipped Shorts: the strongest moments, reframed to 9:16, 1:1, and 16:9 and captioned automatically, so one demo call becomes a week of platform-native shorts instead of a file that sits in a drive.
And where you have no footage to clip, Kompozy generates net-new avatar content the live model never touches. Persona Shorts render a talking-head avatar reading a script you wrote — a product explainer, a launch announcement, a feature walkthrough — while Persona HeyGen handles longer, multi-scene pieces.
Everything stays on one voice via the Persona Brief, and autopilot schedules the batch across eight social platforms plus blog and email. Clean division of labor: Sparrow-2 runs the live conversation, Kompozy turns the recordings and your scripts into finished, scheduled content. (Kompozy's avatar video runs on HeyGen, not Tavus, so this is a repurposing-and-generation layer around your agent, not a plug-in.)
An AI avatar video from the Kompozy workspace, made on HeyGen rather than Tavus. It shows the format and has nothing to do with Sparrow-2.
Best for:
Not ideal for:
Sparrow-2 is Tavus's real-time turn-taking (conversational-understanding) model, announced August 27, 2026. It reads the whole audio scene — tone, pauses, speaker identity, backchannels, interruptions, and noise — roughly every 10ms to decide when a conversational AI should listen, wait, or speak. It powers Tavus's real-time avatars via the Tavus API, PAL Maker, and CVI.
Neither, going by the Tavus announcement. Tavus offers it through its own platform and describes no open-source release, and it does not cancel noise. It reasons over background noise to time conversational turns.
Tavus reports it ranks #1 on the dev split of TurnBench, an independent benchmark, with 92.4% end-of-turn recall and 97.4% interruption recall. Treat the specific numbers as vendor-reported until reproduced independently.
No. It is infrastructure for live conversational agents and generates no media. To turn agent recordings into shorts, or to make net-new avatar video, and publish it, pair it with a content engine like Kompozy, which clips, captions, generates persona video, and distributes across the eight social platforms plus blog and email.
VAD plus a fixed silence timeout only detects speech and guesses the turn is over after a set pause. Sparrow-2 predicts the actual conversational state — done, thinking, or interrupting — from the whole scene every 10ms, so it holds through thinking pauses and catches interruptions far more reliably.
We checked every figure and vendor claim on this page against the sources listed below on 2026-10-23, including the announcement date, frame rates and TurnBench recall numbers in Tavus's announcement; we removed a failure-rate figure and self-hosting claims the announcement does not state. Kompozy plan and product details come from kompozy.io and kompozy.io/pricing.
Sources