Griffin-Lite
#10 v Hlasoví agenti v reálném časeTavus · lite · od 1. Oktober 2026 (Research Preview) · 2× · naposledy 05. 10. 2026
Griffin-Lite is a research preview of a "Human Interaction Model" (HIM) for real-time video/voice conversation, unveiled by Tavus on October 1, 2026. The model processes audio and video in a unified video-to-video system, detects pauses, gestures and facial expressions, and responds in real time with synthesized speech, expression and gesture. In a company-run study, 48% of testers (26 of 54) believed their conversation partner was human after a one-minute video call; on NVIDIA's VideoFDB benchmark, Griffin-Lite achieved top scores among tested AI systems. Access is currently restricted to select, vetted testers; an official price and a public release date for broad availability have not yet been published.
Vlastnosti
| Real-Time Streaming | Generates 720p video in real time in 320 ms chunks from a single reference image, full-duplex video-to-video architecture |
| Latency | Audio-to-video latency averages 0.43 seconds on H100 GPUs, about half that of the fastest published streaming diffusion method |
| Platform | Not yet integrated into the Tavus platform; access only via request form for vetted testers |
| Price | Not published – Griffin-Lite is only available as a free research preview for select testers, no public pricing |
| Release Date | October 1, 2026 (research preview, limited to select testers) |
| Voice Cloning | Speech model can clone a voice from about 10 seconds of reference audio; audio packets as small as 10 ms |