Sparrow-2
Sparrow-2 is the demo page for Tavus Lab's real-time conversation turn-taking model: when talking with PAL, you can see who has the floor, who is responding, and who just interjected; the official website states that it understands pauses, interruptions, talkative speakers, and background noise, making it suitable for observing the rhythm of AI conversations.
Updated: 2026-10-27
What is Sparrow-2?
The official website demo page of Sparrow-2 describes the theme as turn-taking you can watch happen: users converse with a PAL, and the page synchronously displays who holds the floor, who is responding, and who just interjected. The official website states that the model understands the entire audio scene, including pauses, interruptions, backchannel responses, multiple speakers, and background noise; Tavus' documentation places it in the CVI pipeline, collaborating with perception, speech, and rendering layers.
Who is it for?
- Conversational engineering teams evaluating real-time voice or video conversation products.
- Dialogue experience researchers who need to observe turn-taking, interruption and backchanneling behaviors.
- Developers who want to understand the role of the Conversational Flow layer in Tavus CVI.
- An agent product manager who needs to explain to their team "why the AI interrupts or stays silent".
Key features
5Dialogue Turn-Taking Visualization
The official website describes it as a live demo, where you can see the judgment process of turn ownership, backchannel responses and interruptions during the demo, instead of just seeing the final reply.
Full-Scene Understanding
The official website states that it not only judges turn-taking but also understands the entire audio scene, including pauses, interruptions, backchannel responses, multiple speakers, and background noise.
Real-time Link Component
Tavus' documentation places Sparrow in the CVI pipeline, responsible for turn-taking and interruptibility, and works together with the perception, speech, and rendering layers.
Behavioral Objectives Stated on the Research Page
The official website states that the model decides when to listen, wait, speak, and respond, aiming to reduce missed turns, misjudged interruptions, and awkward conversation errors.
Platform and Research Access
The official demo page provides <b0>TRY IN PLATFORM</b0> and <b1>READ RESEARCH</b1>; the Tavus official website states that PALs can be built through APIs, PAL Maker, and enterprise solutions.
How to use Sparrow-2
- First, treat it as a demo and evaluation page, observe the judgments during pauses, backchannels and interruptions using real multi-person dialogue samples, and then discuss product integration.
- Before integration, read Tavus's research page and documentation to confirm the boundaries of responsibilities between Sparrow and other layers such as Raven and Phoenix, and avoid treating the turn-taking model as a complete system.
- To test background noise and multiple speakers, record the on-site conditions and compare the page judgments observed with human auditory perception.
- The official website states that it ranks first on its own TurnBench leaderboard.
- such rankings are only for reference and should be retested according to your own conversation distribution.
Traffic & popularity
Data month Sep 2026
Sparrow-2 receives about 1.9K visits per month.
Monthly visits
3 monthsTraffic sources
- Search34.8%
- Direct32.4%
- Referrals13.5%
- Social5.9%
- Display ads5.5%
- Other8.0%
Top countries
- United States100.0%
Estimated from third-party traffic data.
