Persona-2.0-Live

Today we're releasing persona-2.0-live, our most capable real-time avatar model yet. It sets a new standard for realism, expressiveness, and control, and it is the only model in the world with precise, word-level control over expression. Write [concerned] halfway through a sentence and the face changes on the very next word.

It starts at $0.06/min, scales to thousands of concurrent sessions, and is live on every plan today. See for yourself.

A big upgrade, without the tradeoffs

persona-2.0-live is a step change from persona-1.5-live. Emotion now changes on the word instead of about once a second, and every frame carries four times the pixels.

Big upgrades usually come with tradeoffs. We worked hard to avoid them here.

• Just as fast. The model still turns audio into video in about 180 ms.
• Just as scalable. It easily scales to thousands of concurrent sessions.
• Same price. 6 credits per minute, with extra minutes from $0.06.
• Any face. Alongside the stock library, you can create a custom avatar of the same quality from a single image, on Enterprise plans.

Direct the face, word by word

Expressive voice models already take direction in the script. Drop [sad] into a reply and the voice changes from that word on. With persona-2.0-live the face follows the same tag at the same moment, so one line of script directs the whole performance.

[happy] Good news, the refund went through this morning,
[concerned] but the replacement is still on backorder until March.

There are eleven emotions to direct with: neutral, happy, sad, angry, surprised, concerned, confused, disgusted, excited, scared, and skeptical. Each switch lands on its word, to within a single 40 ms video frame.

That precision is where the delight comes from. A tutor lights up the instant you get the answer right. A buyer in a sales role-play turns skeptical halfway through your pitch.

Hooking it up takes minutes:

• ElevenLabs agents. With an expressive ElevenLabs voice model, the agent writes the tags in its reply and our embed components pass them to the avatar. There is no code to write; the system prompt tells the agent which tags to use.
• OpenAI Realtime. The model calls a set_emotion function that our components declare for it.
• Your own stack. The SDK pins an emotion to a position in the audio you send.

// The reply opens happy.
client.sendAudio(firstChunk, { emotion: 'happy' });

// TTS word timestamps put "but the replacement" 2.1 s into this chunk.
client.sendAudio(nextChunk, { cues: [{ offsetMs: 2100, emotion: 'concerned' }] });

With ElevenLabs, we do exactly this for you. It sends word-level timings along with the audio, and we use them to sync each tag to its word.

Measured against the field

You can feel the difference in the first few seconds of a call. You should be able to measure it too.

So we ran EmoteBench, a blinded human study: 250 participants, 19,500 judgments, five real-time avatar APIs. Participants rated whether each avatar expressed the right emotion, looked natural, and changed when prompted. persona-2.0-live outperformed the closest competitor in every category: 2.2× on expression, 1.9× on naturalness, and 4× on emotion control.

Clearer emotions, clearer intent. Shown a muted clip of an avatar expressing an emotion, participants picked the intended emotion 81% of the time for persona-2.0-live. Anam, Tavus, and LemonSlice landed between 32% and 37%, where guessing scores 14%.

The usual worry with an expressive avatar is that it looks overdone. Participants saw the opposite: more expressive, and more natural. With the sound on, they rated persona-2.0-live's avatar very or completely natural on 61% of clips, against 32% for the runner-up, and said it showed the intended emotion clearly on 79%.

Precise emotional changes, too. 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 persona-2.0-live 53% of the time, against 13% for the runner-up, Tavus. Prompted to hold instead, persona-2.0-live's avatar kept its expression 86% of the time, while the other avatars drifted off theirs on 28% to 41% of clips.

You be the judge. Here is one of the study's screens; each letter is the same line on a different system.

The benchmark page breaks the results down by emotion and by avatar, and the full report covers the method.

Try it

Talk to it on our home page, create an avatar and embed it in a website in minutes, or drive it through our API with your own agent stack.

Much more to come

Keyframe started with two brothers, a garage, and a handful of 4090s. Our ambition was never just realistic avatars. We believe face-to-face AI can make personalized learning, care, and other high-touch experiences accessible to anyone, anywhere.

This is the first of several major model releases. Much more to come.