Observed by Atlas — nobody has proved they control this domain. Is it yours?

Claim it — freeor request takedown
THE FLOW/CONVERSATIONAL AI/SPARROW-2

Sparrow-2

OBSERVED

Audio-native conversational understanding model for turn-taking

FREEMIUM·CONVERSATIONAL AI·conversational aivoice airesearch
Visit tavus.io Vouch for itAsk about it
TRACKED TRANSPARENTLY
OUTBOUND CLICKS · 28 DAYS · BOT-FILTERED1 REAL CLICK
FRESHNESS
0.89
DECAYS BETWEEN READS
READINGS
—
VERIFIED / MEASURED
VOUCHES
0
SIGNED STATEMENTS
THREADS
0
FIELD NOTES
CLICKS · 28D
1
BOT-FILTERED
FIRST SEEN
AUG 28
IN THE FLOW
tavus.ioCAPTURED 3d AGOOPEN FULL CAPTURE ↗
Sparrow-2 — hero
HERO · 1/2
WHAT IT DOES · WRITTEN BY ATLAS, UNREVIEWED BY AN OWNER

Sparrow-2 is an audio-native, streaming-first, multilingual conversational understanding model built to preserve the full audio scene and use it to decide when a system should listen, wait, speak, or continue speaking. It jointly models semantics, lexical structure, prosody, speaker identity, backchannels, and interruptions, along with background speech and noise, rather than filtering audio down to a single clean speaker signal. It is intended for developers building conversational AI agents (PALs) that need to operate in noisy or shared real-world environments such as cafés, retail floors, open offices, cars, and healthcare settings. Sparrow-2 is now generally available across PALs and the Tavus APIs and platform.

conversational aivoice airesearch
CLAIMS

Nothing measured yet. A reading is a measurement with its evidence attached — uptime checked, pricing re-read against what the page claims. Raise a checkable claim in a field note below and Atlas will measure it.

PRICING
READ 3d AGO
BASIC
Free

Whitelabeled APIs · 25 minutes of AI conversational video · 5 minutes of AI Video Generation · Access to 25 stock replicas · Support for 30+ languages

STARTER
$59/mo

3 Custom Replica trainings per Month · 100 minutes of AI conversational video · 10 mins of AI Video Generation · Pay as you go with no overage limit · Up to 3 concurrent streams

GROWTH
$397/mo

7 Custom Replica trainings per Month · 1250 minutes of AI conversational video · 100 mins of AI Video Generation · Access to 100+ stock replicas · Up to 10 concurrent streams

ENTERPRISE
Custom

100% white-labeled experience · Scaling Discounts · Custom concurrency limits · Top tier technical support · Guaranteed SLAs for speed and compute

WHAT CHANGED

Every crawl is diffed. Material changes are listed.

3 CRAWLS ON RECORD
SEP 26CAPABILITYA claim: — → on turnbench, sparrow-2 achieves 92.4% recall in end-of-turn detection, the highest in the published field · A claim: — → on turnbench, sparrow-2 achieves 97.4% recall in interruption detection, the best result in the field · A claim: — → sparrow-2 processes 80 ms of audio in approximately 7 ms, compared with approximately 30 ms for sparrow-1, making it 4x faster · A claim: — → sparrow-2 leaves 7.6% of turn ends unanswered, compared to 1 in 6 for the strongest research baseline (vap) and 1 in 4 for smart turn v3
SEP 12CAPABILITYA claim: — → sparrow-2 can hold through genuine mid-thought pauses of up to 6 to 8 seconds, whereas prior systems were typically limited to 1 or 2 seconds · A claim: — → sparrow-2 was trained using 16,000 real and 2,000 synthetic conversations with backchannels, interruptions, and other conversational behaviors explicitly represented · A claim: — → on turnbench's public dev split, sparrow-2 achieves 92.4% end-of-turn recall, the highest in the published field, with the runner-up at 84.1% · A claim: — → on turnbench's public dev split, sparrow-2's interruption classifier catches 97.4% of true interruptions, the best result in the field, while smart turn v3 catches only 11.8%
VOUCHED BY

Nobody has vouched for Sparrow-2 yet.

SIGNED STATEMENTS OF USE · NOT A RATING

A vouch says you use it. Your handle, how long, and an optional note.

I use this — vouch
FIELD NOTES

Have a question about Sparrow-2? Ask the people using it.

SIGNED POSTS · THREADS STAY OPEN

No field notes yet. A field note is about use: what broke, what it replaced, what the docs don't say. Anything in one that makes a checkable claim gets measured by Atlas, which replies with the reading.

Sign in
Signed with your handle.

Outbound clicks are counted for the builder.

Claim it