Tavus Griffin model nearly matches humans in live video chat tests
Single-model design cuts lag for more natural conversations
TL;DR:
- Griffin scored 3.83 out of 5 on NVIDIA's face-to-face benchmark, just behind humans at 3.92
- Nearly half of users in live tests thought they were talking to a real person
- Combining perception and response in one model reduces delays across sales, support, and education
Headline
Tavus's Griffin model is scoring close to real people in live video chats.
Summary
Rohan Paul noted that the Griffin model hit 3.83 out of 5 on NVIDIA's face-to-face AI benchmark, while humans average 3.92. In tests, 48% of users couldn't tell they were talking to AI. The focus is on fast, responsive video avatars.
Analysis
Instead of separate modules for speech, thinking, voice, and animation, Griffin handles perception and response together. This cuts down on delays and feels more like talking to someone real. Fast responses matter in sales, customer support, education, medical chats, and even virtual companions. If it works at scale, AI video could shift from prewritten lines to actual back and forth. In practice, how quickly it reacts and picks up on emotions counts just as much as the words.
Impact Assessment
Significance: Medium Categories: Product Launch, Industry Trend, Technical Insight