Sparrow-2

What is Sparrow-2?

Sparrow-2 is a real-time turn-taking model from Tavus that decides when a conversational AI should speak, listen, or wait. Instead of filtering out background audio, it reads the whole acoustic scene, including interruptions, pauses, and short acknowledgments. Developers use it inside Tavus conversational video agents through the API and platform.

Top Features:

• Turn endings: catches 92.4 percent of real turn endings on the TurnBench benchmark.
• Interruptions: spots 97.4 percent of interruptions so the AI stops talking quickly.
• Scene awareness: treats background noise as useful context rather than cancelling it.

Use Cases:

• Video agents: give Tavus AI humans more natural timing in live customer conversations.
• Noisy settings: keep voice assistants responsive in cafes, retail floors, cars, and clinics.
• Support calls: tell a quick mhm apart from a caller who wants to speak.

Who Can Use Sparrow-2?

• Developers: add better turn detection to conversational products through a single setting.
• Product teams: build sales, support, or training agents that feel less robotic.
• Researchers: study turn-taking behaviour through a live demo showing real-time detections.

Pricing

• Free (free): 25 minutes of conversational video, stock replicas, and full API access.
• Starter ($59 per month): 100 minutes, three custom AI humans, and three concurrent conversations.
• Growth ($397 per month): 1,250 minutes, seven custom AI humans, and ten concurrent streams.
• Enterprise (contact sales): volume discounts, white labelling, custom concurrency, and dedicated SLA support.

Pros and Cons

Pros:

• Benchmark lead: tops TurnBench on both turn endings and interruptions at once.
• Fast reactions: processes audio in 10 millisecond frames for quicker, human-like reply timing.
• Easy switch: developers turn it on by changing a single configuration value.

Cons:

• Platform lock-in: it runs inside the Tavus stack, not as a standalone model.
• Bundled pricing: costs come bundled into Tavus conversation minutes and plan limits.
• Developer focus: using it well requires API work and some conversational design knowledge.

FAQs:

1) What does a turn-taking model do?

It decides when an AI should talk, keep listening, or yield.

2) How can developers try it?

Try the live browser demo, or enable it through the Tavus platform.

3) Is it the default model?

Tavus set it to become the default model in September 2026.

4) Does it work in noisy places?

Yes, it reads the whole audio scene and suits cafes, cars, and offices.

5) What did it replace?

It succeeds Sparrow-1 and processes audio four times faster than before.