AI avatar model
Tavus CVI
TavusConversational video with perception — the avatar can see the person it is talking to.
Tavus CVI is not one model but a stack: Phoenix renders the face, Raven watches the user through their webcam, and Sparrow decides when to speak. That perception layer is what separates it from a talking-head renderer — it can notice that a candidate is reading from a second screen, or that they stopped mid-sentence to think rather than because they were finished. For screening interviews that behaviour is the whole point, and it is the reason to pay a real premium over a pure lip-sync model.
Sample output
What it actually produces.
One clip is worth more than any benchmark row. Where we have not published a sample yet, the frame below holds the space it will occupy.
Demo clip
No sample of Tavus CVI uploaded yet.
What it animates
Head to toe.
A model that only drives the mouth looks wrong the moment the other person starts talking — nothing on screen moves. Filled means driven, outlined means limited or looped, greyed means static.
- Face
- Full
- Lip-sync
- Full
- Upper torso
- Partial
- Hands
- None
- Full body
- None
Phoenix-3 does full-face reenactment, so brow, eyes and head motion move with the speech rather than only the mouth. Framing is head-and-shoulders: shoulders drift naturally but hands are out of shot.
Source: HumanLike hands-on assessment of Tavus CVI output
Variants
4 ways to run it.
Same family, different quality/cost rung. Picking the wrong one is the most common way teams overpay.
Phoenix-3 (render)
Maxphoenix-3The rendering model. Full-face reenactment rather than mouth-only compositing, so expression and micro-head-motion move with the speech.
Candidate-facing screening calls · High-trust brand video
Raven-0 (perception)
Highraven-0Visual perception. Reads the user’s camera for attention, emotion and environment — this is what makes the conversation feel two-way.
Proctored screening · Engagement signals
Sparrow-0 (turn-taking)
Balancedsparrow-0Decides when to speak. Handles pauses, interruptions and thinking silences without talking over the candidate.
Natural interview pacing
Hummingbird-0 (lip-sync)
Fasthummingbird-0Lip-sync only, applied to existing footage. The cheap path when you do not need conversation.
Localising recorded video · Bulk personalised outreach
Fit
Two audiences, one model.
The same model is a different proposition depending on what you point it at.
Recruitment fit
The most complete option for conversational screening — the perception layer is worth the premium when a real person is on the other end.
Best for
- First-round conversational screening
- Employer-brand videos personalised per candidate
- Proctored assessments where attention signals matter
Strengths
- Reads the candidate’s camera, so the interview reacts rather than just plays
- Turn-taking handles thinking pauses without cutting candidates off
- Replica training lets a real recruiter’s face front the screen at scale
Watch out for
- At roughly $1/min it is 200× the cost of Homo 1 — model your per-interview cost before committing
- Recording a candidate’s webcam raises consent and data-retention obligations; several jurisdictions require explicit notice for automated video assessment
- A trained replica of a real employee needs that employee’s written consent
The numbers
Measured, not asserted.
HumanLike evals
Run on our own harness.
- Wired into HumanLike avatars
- Yes — face_id providerOur eval
Selectable as a video provider on any AI persona; requires a trained Tavus face_id.
Published benchmarks
Vendor and independent figures.
- Lip-sync accuracy
- 90%Estimate
- Utterance-to-utterance latency
- 1000 msEstimate
Published by us as an approximation (~90%).
Published by us as an approximation (~1.0s).
Compare
