A new open-source project called AVTR-1 tries to make AI avatars listen as well as they talk.
Researchers released AVTR-1, an open stack for running real-time, interactive avatar conversations. It centers on a 153-million-parameter motion generator that uses flow matching and looks at both people's audio, not just the avatar's own lines. The team re-engineered the audio encoder to run in streaming mode and tested the full pipeline against two commercial voice agents, measuring exactly where the delays come from. The result runs in real time on both data-center and consumer GPUs, and in benchmarks it beat comparable systems on visual quality and most motion metrics, while staying merely competitive on lip-sync.
The more interesting contribution is not the avatar - it's the yardstick. Most talking-avatar demos just sync a mouth to audio; they don't prove the avatar is reacting to the other person at all. The researchers built a new metric, Reference-Based Directed Granger Gain, that statistically tests whether a listener's movements carry information from the speaker's voice. Run against AVTR-1 and rival dyadic systems, it found real responsiveness - but none in talking-head generators that never see the other participant's audio.
It's a research release, not a shipping product, and open here means a patchwork of component-specific licenses - so read the fine print before building a customer-service avatar on top of it.