EmoteBench · Keyframe Labs Research
A new standard for expressive avatars
Avatar benchmarks measure lip sync, latency, and preference. EmoteBench measures natural, emotional expression. When the sentiment is grave we expect to see concern; when it brightens, we expect the avatar to brighten with it.
EmoteBench is a blinded human study of whether real-time avatars do that. 250 participants made 19,500 judgments across five avatar APIs, rating whether each avatar expressed the right emotion, looked natural, and changed when prompted. Keyframe Labs scored highest on every measure, against every competitor.
Expression
Muted clips where participants named the avatar's emotion
95% CI 76–85%
Naturalness
Ratings of very or completely natural
95% CI 54–67%
Control
Prompted emotion switches where participants named both emotions
95% CI 41–64%
Expression
Clearer emotions, clearer intent.
Shown with an avatar expressing an emotion, participants picked which of seven emotions it showed. For Keyframe Labs' avatar, they picked the intended emotion 81% of the time. Anam, Tavus, and LemonSlice landed between 32% and 37%, where guessing scores 14%. No other avatar was more emotionally readable for 239 of the 250 participants.
Share of muted clips where participants chose the intended emotion. About 2,250 ratings per system; the dashed line is chance.
| Emotion | Keyframe Labs | Anam | Tavus | LemonSlice |
|---|---|---|---|---|
| Anger | 97% | 22% | 41% | 67% |
| Happiness | 95% | 49% | 94% | 41% |
| Neutral | 87% | 77% | 30% | 76% |
| Sadness | 84% | 16% | 45% | 8% |
| Fear | 79% | 2% | 12% | 4% |
| Surprise | 69% | 25% | 30% | 15% |
| Disgust | 58% | 49% | 13% | 10% |
Keyframe Labs has the highest share on all seven emotions, Ekman's six basic emotions plus neutral, though Tavus is within a point on happiness. On fear, the three other systems score between 2% and 12%.
Naturalness
More expressive, and more natural.
With the sound on, participants also rated Keyframe Labs' avatar both more natural and clearer in expressing the emotion than every other system. Even the lowest-rated of Keyframe Labs' three avatars outranked every competitor's best.
Looked natural
Showed the emotion clearly
Share of ratings at the top, middle, and bottom of a five-point scale. About 1,300 ratings per system; HeyGen, which has no emotion control, was rated on neutral and happy lines only.
Mean naturalness, 1 to 5, for all fifteen avatars. Keyframe Labs' lowest, Caspian at 3.42, sits above the best from any other provider, Anam's Mia at 3.03.
Control
Precise emotional changes.
Participants watched muted clips in which each avatar was prompted to switch emotion mid-sentence, then named the emotion before and after. They named both correctly for Keyframe Labs' avatar 53% of the time, against 13% for Tavus, 5% for LemonSlice, and 3% for Anam. Prompted to hold instead, Keyframe Labs' avatar kept its expression 86% of the time, while Tavus, LemonSlice, and Anam drifted off theirs on 28% to 41% of clips. For those three, participants could barely tell a prompted hold from a prompted switch.
Solid: share of prompted switches where participants saw the change and named both emotions; chance is 2%. Hatched: share of control clips, prompted to hold, where participants reported a switch anyway.
Head to head
The near-consensus take: Persona-2.0-Live is better across the board.
Nearly every participant saw every system. For each of the 250 participants we compared their average score for Persona-2.0-Live, Keyframe Labs' model, against their average score for each competitor. On recognizing the emotion, 98% of participants put Persona-2.0-Live ahead; on showing it clearly, 93% to 97%; on looking natural, 77% to 87%. Even the closest comparison, naturalness against Anam, went its way for more than three participants in four. Rating against rating on the same line, it still wins 69% to 87% of the time.
Emotion recognized
Emotion shown clearly
Looked natural
One square per participant who saw both systems, 246 to 250 per comparison. Line by line, Keyframe Labs also won 88% to 100% of the lines both systems performed. Every lead's 95% interval excludes zero, both in all sessions and in the half that passed attention checks. HeyGen has no emotion control and was not tested on recognition.
Try it
You be the judge.
Real screens from the study. Pick an answer, then see what the participants chose. Clips play as participants saw them, one system behind each letter. The full report has every breakdown, interval, and example line.
Behind the Expression result above · No sound. Name the emotion the avatar is showing.
Which emotion does the face show?
No sound on this one. Watch the whole clip, then answer.
The face is intended to show surprise
Sound on. Rate the face only, not the voice.
Emotion switching
No sound on this one. Some faces switch emotion part-way through and some don't.
The study
How it was measured
Every system received identical audio for every line, so the only thing that differed between two clips was the avatar. 182 scripted lines across seven emotions, three stock identities per provider, clips captured live from each API and rated blind.
Late October 2026
Clips captured live from each provider's streaming session, on the versions available at the time.
250 participants · 19,500 rated clips
An online research panel. One clip per screen, the system shown as a letter, no line seen twice.
Five real-time avatar APIs
Keyframe Labs persona-2.0-live, Anam cara-4.1-express, Tavus Phoenix-4.5, LemonSlice 2.1, HeyGen LiveAvatar.
Keyframe Labs designed and ran the study
Participants never saw provider names. Every system is reported on every measure it could take.
Code and data: coming soon