What is Tavus Griffin: The Human Interaction Model that passed the Turing Test
Follow Us
All AIs that you have seen on video calls share one common trait. The moment when you finish your sentence, there is always the pause. The pause during which the machine records what you said, processes it and produces speech and animation. According to Tavus, the new Griffin model puts an end to the pause. They claim that nearly half of the people considered it human.
SurveyAlso read: Best coffee machines for your home under Rs 15,000
What is Tavus Griffin
Griffin is a HIM (Human Interaction Model) in Tavus’ words. It is an alluring term for a very straightforward concept. Typically, most live avatars operate through the relay process. One technology converts spoken words into text, another generates a response, and another gives voice to the response. Another one makes animated facial expressions. Each process involves delays and loss of information – your intonation, for example, or camera footage.
The Griffin system is one example of a video-to-video system which listens and observes as it speaks. Subsecond by subsecond, it determines if it needs to speak, nod, use “mm-hm” or be silent. This is how Griffin interrupts you and gets interrupted. Think about it – most technologies still wait for their turn like the teller at the bank.
The 48% claim
What really stood out was the 48% figure. In a study by Tavus using an independent research platform, 54 out of the total number of people tested believed they had communicated with a human after watching a one-minute video call. The previous stack of Tavus scored 1 out of 41, translating to 2.4%. They were told that they would be matched with someone else.
Also read: Gemini 4 Argon: Google is giving defenders a Cyber AI with no guardrails
Looks like passing the Turing test, but a narrow win indeed. Sixty seconds. Fifty-four people. Study conducted and performed by the company selling the technology, and no external replication of the findings. Original Turing test involved text only, while video Turing test is something that has been defined by Tavus. Some became wary within 20 seconds. Double the length of the conversation to an hour, and we do not know what happens.
How it works
Two engines operate simultaneously. One conversational model consumes audio and video data and produces control signals for speech, emotions, expressions and gestures. Another generation engine translates those control signals into voice and face. The former one can clone voices based on approximately 10 seconds of audio data. The latter one creates 720p video data in 320 millisecond increments based on just one picture reference, even the shadows of the chair.
Speed is the whole point. According to Tavus, Griffin achieves 0.43 second delay between audio input and face reaction time, that is twice as fast as the next fastest known technique. On NVIDIA’s VideoFDB dataset, evaluated by NVIDIA, Griffin got 3.83 out of 5 scores on generation task comparing to human score 3.92. On perception task, Griffin got 3.73 out of 5 comparing to human 4.20, that is the highest among 15 models.
Can you use it
Not yet. At present, there is Griffin-Lite, which is a sneak peek into the research for selected users. The reasons why the company Tavus is waiting before its full release are safety and transparency. According to the company, the same characteristics that make Griffin very natural are the reasons why it makes the users think that Griffin is not an AI. Presently, the customers of Tavus develop upon their Phoenix, Raven and Sparrow products.
Also read: What is SynthID Bio: Google’s watermark for AI designed proteins
