Watch two people talk and you will notice that the listener is never still. They nod at the right moments, raise an eyebrow at a surprise, smile when the other person smiles. Those signals tell the speaker to keep going, slow down or explain more. Take them away and a conversation feels cold, even when the words are perfect.
Most digital avatars and video agents get the speaking part right and the listening part wrong. They move their lips, then freeze while you talk.
What Meta built
In 2025, Meta's FAIR research group released Seamless Interaction, a family of models that generate facial and body motion for both people in a two-person conversation. The key design choice is that the model takes audio from both sides of the conversation. That lets it produce listening behavior, such as head nods and smiles, and not only gestures while speaking.
The models can also react to what the other person does on video, for example mirroring a smile or matching gaze. Their output can be rendered as 2D video or as 3D avatars using Meta's Codec Avatar technology, with controls for how expressive the avatar should be.
The data behind it
To train these models, Meta also released the Seamless Interaction Dataset: more than 4,000 hours of face-to-face interaction from more than 4,000 participants, who consented to being recorded. It is one of the largest datasets of its kind, and it makes a simple point: you cannot teach a model to listen without recordings of people listening.
Why it matters for Social AI
Two-person conversation is the simplest case. Real interactions involve groups, where attention shifts between several people and listeners react to each other as well as to the speaker. They happen in meetings, classrooms and dinners, not only in a recording studio.
As avatars, tutors and video agents move into those settings, they will need data that captures group listening: who looks at whom, how a room reacts to a joke, how a quiet participant signals they want to speak.
Crosstalk sources consented recordings with a separate video feed for every participant, labeled for gaze, expression and turn-taking, across two-person and group settings. Explore the catalog or read the paper.
Sources: Seamless Interaction paper (arXiv), Meta FAIR research page, Meta AI blog.