Tavus releases realistic avatar model

Software startup Tavus announced a new video rendering system named Phoenix 4.5, designed to generate responsive digital avatars for live calls. According to a post by researcher Minh Anh Nguyễn, the system creates complete upper-body movement including shoulder shifts and posture adjustments in real time, rather than splicing an animated face onto static body footage. The company claims the software translates spoken audio into realistic video within 134 milliseconds.

This technical leap matters because interactive video bots have historically felt rigid and uncanny. Older models restricted movement to the lower facial region, which created visible seams and awkward body stillness. By making digital humans react faster and move their entire upper torso while speaking, the technology aims to make interactions with virtual assistants, healthcare guides, and online tutors feel far more natural.

It is still unclear how effectively the system performs during laggy internet connections or when managing overlapping speech. Fast rendering solves part of the latency equation, but true conversational flow depends on context and emotional pacing that simple speed cannot guarantee. As these virtual personas become smoother and easier to generate from a single picture, society will also need clearer safeguards to prevent fraudsters from spoofing real individuals in live video calls.