AI company Tavus has introduced a new model, which has been designed to see, hear, speak and react like humans during live video conversations. The company has named it Griffin. According to Tavus, it is an End-to-End Human Interaction Model (HIM), which is designed for real-time Face-to-Face video conversations.
Griffin has been designed differently from traditional AI chatbots. Usually in AI system, first the voice is converted into text, then the Large Language Model prepares the answer and after that a separate system creates voice or video. Griffin handles all of these processes in a single full-duplex video-to-video system.
How will Griffin work in live video calls?
According to Tavus, Griffin can simultaneously see and hear the user, respond and even use facial expressions. For example, the AI can nod while the user speaks and incorporate something seen in the video into its response.
The company also conducted a small test of Griffin-Lite. In this, 26 out of 54 people i.e. about 48 percent of the participants felt that after one minute of conversation, they were talking to a human being. In the previous system test, only one person out of 41, or about 2.4 percent of the participants, was able to believe this.
However, it was a small and short test. The participants were told in advance that they were going to talk to another person. Therefore, it is not yet clear how human-like Griffin will seem in long-term conversation.
How was the performance in Benchmark?
Tavus also mentioned NVIDIA’s VideoFDB results. In the Generation Track related to Conversational Behavior, Griffin-Lite scored 3.83 out of 5, while Human Reference Recordings had a score of 3.92. In comparison, Gemini 2.5 + Anam’s score was 2.80.
At the same time, in Perception Track, which understands conversation signals, Griffin-Lite’s score was 3.73, while humans’ score was 4.20.
Can be used in AI Tutor and Companionship
Such an AI model can be used as a virtual tutor in studies. The student can seek help by showing his problem in a video call and the AI can respond based on that visual information. However, larger and longer tests will be required to understand its true usefulness.
Similarly, it can also be used to provide conversation and company to elderly people. But here the question is also important whether the company of AI reduces the real contact between people.
The risk of AI Scam may also increase
AI that responds like humans can also make fraud more credible. Such AI can be used to ask for money or sensitive information by impersonating another person.
Currently, Tavus is making Griffin-Lite available as preview features only to select trusted testers. The company is working on features related to Safety and AI Disclosure before wide release. Its public launch date has not been announced yet.
