Tavus Releases Face-to-Face Interaction Model Griffin, Claiming It as the First to Pass the Real-Time Video Turing Test

Nearly Half Thought It Was Human After a One-Minute Chat: Tavus Introduces Face-to-Face Interaction Model Griffin

Tavus has introduced Griffin, a face-to-face AI interaction system, calling it the first model to pass a real-time video Turing test. In official evaluations, 54 participants engaged in a one-minute, one-on-one video call, and 26 (48%) believed they were speaking with a real human. Under identical test conditions, only 1 out of 41 participants (2.4%) mistook Tavus's previous-generation cascade system for a human. This technology aims to solve the awkward interactions common with digital avatars and voice assistants: you pause for two seconds mid-sentence to find the right word, and the AI assumes you are done and cuts in; or you share intensely exciting news, yet the on-screen avatar remains expressionless until, two or three seconds later, it mechanically blurts out, "That's great."

Many turn-taking dialogue architectures rely on a Cascade System: they first convert your speech to text, generate a response, and then synthesize speech and avatar video. The process functions like a relay race:

1. Leg 1 (Automatic Speech Recognition, ASR): Converts speech to text; the system determines you have finished speaking before proceeding to response generation;
2. Leg 2 (Large Language Model, LLM): Generates a text response based on the transcript;
3. Leg 3 (Text-to-Speech, TTS): Synthesizes the LLM's text output into audio;
4. Leg 4 (Avatar Drive): Drives facial animations and lip-sync based on the synthesized audio.

Every handoff risks adding latency, and downstream stages often lose raw context. Converting speech strictly to text discards hesitations in vocal tone; leaving camera feeds out of the reasoning model prevents gaze and gestures from informing decisions. Furthermore, avatars driven strictly by pre-generated audio cannot nod or offer backchannel acknowledgments in real time while you are still speaking.

Tavus terms Griffin a Human Interaction Model (HIM). It continuously processes incoming audio and video while simultaneously generating responses—a bidirectional simultaneous communication mode known as full-duplex. While you speak, it continues listening and watching; while it vocalizes, it continues processing your movements and interruptions. Its dialogue engine re-evaluates at sub-second intervals: should it wait, nod, backchannel, or begin answering right now?

How Does It Listen, Watch, and Respond Simultaneously?

Enabling a machine to converse as fluently as a human requires coordinating conversational judgment, turn-taking timing, and audio-video streaming generation cadence in lockstep. Griffin operates through two core engines working in tandem: the Continuous Conversational Modeling Engine and the Streaming Audio-Visual Generation Engine.

Continuous Conversational Modeling: Continuously Evaluating "Mini-turns"

Turn-based systems typically wait until you finish speaking before formulating a reply.

In Griffin, the dialogue engine never stops running throughout the entire conversation. It decomposes interaction into ultra-short "mini-turns." Within every single mini-turn, it executes several actions concurrently:

• Real-Time Listening and Gaze: The input streams remain permanently open, regardless of whether the user is talking or the avatar itself is speaking;

• Turn-Taking Management: If you pause mid-sentence and shift your gaze upward, the system uses visual context to deduce that you are "still thinking," proactively remaining silent and waiting rather than cutting in upon detecting silence;

• Backchanneling: During prolonged user speech, the system does not need to wait for a turn transition; it can insert an "uh-huh" or a subtle nod within a mini-turn, signaling active, attentive listening;

• Barge-in Handling: If the user interrupts, the system instantly re-evaluates whether to persist, pause, or yield the floor entirely.

Audio Generated in Small Chunks: Playing Immediately While Computing Ahead

Griffin's speech system utilizes continuous latent representations: sound is initially encoded into a compact vector of continuous values, which a decoder subsequently reconstructs into a playable waveform. Its primary stages are:

• Tavec Continuous Codec: A convolutional autoencoder that maps 48 kHz high-fidelity audio into continuous latent vectors—processing 100 frames per second, with each frame containing just 40 continuous values, completely eliminating discrete codebook lookups. One minute of raw audio comprises up to 2.88 million sample points, but Tavec compresses this into 6,000 latent frames, drastically shortening sequence length.

• Autoregressive Diffusion Transformer (VDiT): Capable of voice cloning with as little as a 10-second reference vocal prefix. Rather than waiting for an entire sentence of text to finish generating, it generates audio incrementally as control signals arrive from the dialogue engine.

• Audio Packets as Small as 10 Milliseconds: Paired with a fully causal decoder, generated audio packets can be as compact as 10 milliseconds. The very first packet is pushed to speakers and begins playing before subsequent speech has even been computed.

Generating the Entire Avatar and Scene from a Single Reference Image

Tavus's previous video models relied heavily on 3D priors—pre-introduced three-dimensional structural knowledge—which, according to the team, constrained the expression of intricate hand gestures, expansive body movements, and dynamic backgrounds.

Griffin shifts to diffusion-based streaming visual generation: from a single reference image, it generates full-pixel 720p video in real time, encompassing the subject, hand gestures, clothing folds, chair movement, and dynamic background shadows.

The core engineering breakthrough lies in a three-stage distillation pipeline that converts a multi-step bidirectional diffusion model into a streaming-capable generator:

1. Distribution Matching Distillation (DMD): Distills a large teacher model requiring dozens of inference steps into a high-speed student model needing very few steps, though it initially can only generate entire video segments in a single shot;
2. Teacher Forcing Autoregressivization: Restructures the model into an autoregressive architecture conditioned on historical frames, enabling sequential generation of individual latent chunks;
3. Self-Forcing Anti-Drift Training: Trains the model to roll out long sequences conditioned on its own previously generated frames, downsampling earlier history frames to lower resolutions to mitigate long-horizon error accumulation and bound computational overhead.

Ultimately, each latent block requires only 3 diffusion steps. At 25 fps, Griffin streams 720p video in blocks of 8 frames (corresponding to 320 milliseconds) in real time, reacting dynamically to gaze, expression, and gesture control signals sent by the dialogue engine.

The "Simon Says" game requires players to mimic an action only after hearing the "Simon says" prefix, and to refuse if the prefix is omitted. In the demo video, the system must simultaneously parse spoken commands, monitor visual gestures, evaluate whether an action is warranted, and generate animations such as raising hands, touching its face, or touching its hair. Understanding rules, interpreting visual actions, and responding within a live interaction are unified here, demonstrating how visual input directly shapes dialogue and how full-scene generation enables rich physical expression.

Evaluation Metrics: Understanding the Context and Responding Appropriately

NVIDIA's VideoFDB benchmark is designed specifically to evaluate full-duplex audio-visual dialogue: evaluating whether models comprehend their conversational partner via vision and audio, and whether their own voice, expressions, and gestures coordinate appropriately.

The benchmark compiles 237 segments from real two-person video calls, spanning 11 critical nonverbal dynamics recognized in interpersonal communication studies (including gaze shifts during pauses, emotional expressions, gestures, interruptions, and immediate laughter responses). An LLM judge evaluates the "Perception" and "Generation" tracks independently on a scale of 0 to 5.

The two tracks are scored separately and cannot be treated as a composite ranking of general model intelligence.

The official announcement also summarized its progress by stating it is "about 37% higher than the next-best system" and "closed the gap to human performance by more than 12x." If calculated from the published overall generation scores, 3.83 is approximately 36.8% higher than 2.80; the gap to the human reference of 3.92 was narrowed from 1.12 points for the next-best system down to 0.09 points for Griffin—the former being roughly 12.4 times larger than the latter. What can be independently verified are these evaluation scores and score deltas; this does not imply that response speed increased by 37% or that overall capability expanded twelvefold.

NVIDIA's benchmark reveals two distinct comparison tracks:

1. Generation Track (evaluates naturalness and emotional alignment in the AI's own speech and physical behavior):

   • Traditional cascade pipelines utilizing speech-driven avatars, such as "Gemini 2.5 + Anam," scored only 2.80;
   • Griffin-Lite achieved 3.83 points, sitting just 0.09 points shy of real human reference footage (3.92 points) and significantly outpacing peers.
   • On the TOR-Alignment metric (takeover-rate alignment), Griffin-Lite attained 62.8%, compared to 31%–44% for cascade avatars.

2. Perception Track (evaluates the AI's ability to understand user audio-visual context and respond appropriately):

   • Across 15 evaluated configurations, Griffin-Lite ranked first with 3.73 points;
   • It surpassed the audio-only MiniCPM-o 4.5 (3.44 points), Google Gemini 2.5 Flash Native (3.17 points), and audio-visual OpenAI gpt-realtime (2.75 points).

Regarding latency and temporal dynamics, NVIDIA's benchmark reports that Griffin has a median response latency of 1892 ms for generation and 2232 ms for perception, with turn-taking alignment rates of 62.8% and 73.8%, respectively. For comparison, the human baseline is 900 ms (78%) and 1400 ms (90%). Notably, MiniCPM-o 4.5 achieved an audio-visual perception response latency of 720 ms, demonstrating that obtaining the highest total benchmark score does not equate to having the fastest response time. Furthermore, the 0.43-second average reported by Tavus on H100 hardware reflects strictly the audio-to-video generator latency; 10 ms represents the duration of sound contained in a single audio packet, and 320 ms represents the playback duration of one video chunk. Both are output block granularities and must not be conflated with end-to-end conversational turnaround latency.

The Camera Is On, But the Feed Doesn't Always Inform the Conversation

In side-by-side evaluations of mainstream models, the NVIDIA team highlighted two systematic failure modes prevalent among dialogue models that claim "visual capabilities":

• Failure Mode 1: Captioning Collapse Upon receiving video frames, the model instinctively treats the input as an "image captioning" task. For instance, across up to 87% of multimodal evaluation segments, Mini-Omni2 fails to converse naturally, blurting out descriptions such as: "I see you are wearing dark clothes...", "You appear to be sitting on a chair...", reducing conversation to descriptive captioning.

• Failure Mode 2: Visual-Stream Ignorance Other models ingest video frame tokens in the background, yet their generated responses—both in timing and textual content—remain virtually indistinguishable from text produced in audio-only mode (as observed in gpt-realtime-mini). Visual inputs rarely alter what is said or when it is said.

In pure visual rendering, Tavus benchmarked Griffin-Lite's video generator against 4 mainstream streaming diffusion models (5 models in total) for audio-driven video synthesis:

• On H100 hardware, the average latency from audio input to corresponding visual change on-screen was 0.43 seconds, approximately half the latency of the fastest competing baseline.
• It ranked 1st across visual quality metrics DOVER (perceptual video quality), FID (frame distribution fidelity), and THEval (comprehensive talking-avatar benchmark).
• In lip-sync confidence (LSE-C), Griffin-Lite ranked 2nd with a score of 7.27. Tavus explained that LSE-C rewards exaggerated mouth movement beyond natural levels; in daily communication, excessive lip articulation degrades realism. Meanwhile, THEval—a holistic benchmark that accounts for naturalness—placed Griffin at the top.

The promotion celebration demo clearly illustrates this coordination: while the user is still speaking, the avatar smiles, raises its hands in celebration, and its tone shifts to excitement. Because the continuous dialogue engine feeds vocal tone, facial expressions, and gesture controls simultaneously to the generation pipeline, celebration emerges across sound and vision in a coordinated response. In face-to-face interaction, when and how one responds profoundly shapes whether the conversational partner feels genuinely heard.

How Does It Coordinate in Everyday Interactions?

Beyond static benchmarks, what specific capabilities does Griffin display in dynamic interactions? Demonstrations released by Tavus showcase several concrete scenarios:

Scenario 1: Emotional Resonance and Attentive Listening Feedback

In an interpersonal advice demo, a user expresses uncertainty about how to politely decline someone after a date. As the user talks, the avatar maintains focused eye contact, nods gently, and lets out a soft laugh when the user makes a self-deprecating remark. It then offers advice that is direct yet tactful. The entire interaction maintains a coherent social rhythm, eliminating the need for the user to pause after every sentence to wait for a system response.

Scenario 2: Improvising Dialogue with Visual Props

In an interactive storytelling demo, a user proposes playing a story-building game based on random props and shows two items to the camera sequentially:

1. First, a green plush toy: Rather than mechanically reporting "I see a toy," the AI immediately adapts the story's protagonist into a "green furry undersea monster disguised as a regular fish," appointing the little creature as the special agent captain;
2. Next, a 3x3 Rubik's Cube: The AI incorporates the prop effortlessly, describing it as a "high-tech underwater navigator whose colors must be aligned to reveal hidden treasure coordinates."

This interaction demonstrates that when environmental visual elements change, the system internalizes them directly into the conversational context without breaking conversational flow.

Scenario 3: Step-by-Step Hands-On Tutoring

In a Rubik's Cube tutorial demo, the user notes that the first two layers are solved and only a blue bar remains on the top layer. The AI adopts an incremental coaching approach, issuing only three or four explicit clockwise or counter-clockwise rotation steps at a time, followed by a proactive request: "Hold the top face steady so I can take a look once you finish those steps." After verifying alignment, it proceeds to the next formula. The published video includes clips marked at 2x speed, making it unsuitable for calculating raw response latency. What stands out is its ability to parcel instructions into chunks while waiting for the user to complete physical actions and display the outcome.

Scenario 4: Temporal Tracking and Timed Prompts

In a circuit board soldering demo, the user tightens screws and prepares to grab tools, explicitly instructing the AI: "I just plugged in the soldering iron, it takes about 12 seconds to heat up, please remind me when time is up." The AI acknowledges, continues listening, and then speaks up: "12 seconds are up." Later, as the user solders components, the AI counts aloud "One Mississippi, Two Mississippi, Three Mississippi" to help regulate pressing duration.

A One-Minute Chat, 26 Judged It Human: What Does This Test Show?

Among the announcements, Tavus's most prominent claim is designating Griffin as the "first model to pass the real-time video Turing test."

The specific experimental setup and statistical denominators are as follows:

• Methodology: Participants recruited via an independent third-party platform were told they would join a 1-minute video call with "another participant" to discuss "what they are most looking forward to this year." The partner on screen was actually PAL, an avatar driven by Griffin-Lite;

• Statistical Results: Out of 54 participants, 26 (48%) believed after the call that they were conversing with a real human;

• Historical Baseline: Under identical protocols, Tavus's own prior cascade architecture (Phoenix-4.5 + Sparrow-2 + Raven-1) misled only 1 out of 41 participants (2.4%);

• Confidence and Suspicion: Participants who believed their partner was human reported an average confidence rating of 79%; those who detected anomalies reported an average confidence of 81%, with most growing suspicious within the first 20 seconds;

• Dimension Ratings (out of 7): Naturalness scored 5.4, Trustworthiness 5.6, Willingness to talk again 5.8, and Perceived active listening 5.5; "Conversational flow naturalness" scored the lowest among them at 4.9.

Questions probing whether participants suspected the partner was not human were placed at the very end of the survey, after which the AI identity was disclosed to all. Tavus also reported that over half the participants did not contemplate that the partner might be AI during the call, and nearly all individuals in that cohort ultimately judged the counterpart to be human. Among participants who identified the partner as an AI, willingness to converse again remained high at 5.4/7: identifying an AI counterpart did not inherently diminish willingness to engage.

What Does This Claim Actually Signify?

The video Turing test examines whether conversational partners can discern if the entity on screen is human or machine. A video call introduces far more cues into this assessment: whether answers are contextually fitting, whether vocal tone matches facial expressions, whether gestures are natural, and whether nods, laughter, and turn handoffs occur at appropriate moments. In this study, Griffin-Lite was identified as human by 26 out of 54 people—compared to 1 out of 41 for the legacy cascade system—demonstrating a marked increase in perceived human likeness across full face-to-face interaction. Tavus based its claim of being the first model to pass the real-time video Turing test on this outcome.

The calls lasted only one minute, participants were told in advance they would converse with "another participant," and 28 participants still identified the counterpart as AI. Furthermore, the report does not specify a standardized industry pass threshold, meaning these findings cannot be extrapolated to extended conversations, nor do they imply subjective consciousness in machines. What the result demonstrates is concrete: by simultaneously refining voice, facial expressions, and conversational cadence, a short video exchange is now sufficient to make a substantial proportion of people feel they are speaking with a human.

Who Can Use It Now, and What Comes Next?

When technology reaches the point where nearly half of participants in a short-call study mistake it for a human, clear disclosure of AI identity becomes paramount.

Tavus highlighted this double-edged effect in its report: the exact mechanisms that make interaction profoundly natural can also cause users to mistake the AI for a real person. Consequently, Griffin-Lite has not been released via a public commercial API; it is available only as a "Research Preview" to a limited group of trusted testers. Tavus is actively developing safety disclosure features alongside alignment and safety evaluations, planning a broader release once these guardrails are established.

Looking toward future deployment scenarios, Tavus outlined several applications: in educational tutoring, mentors can detect puzzled facial expressions when students struggle and proactively rephrase explanations; in workplace communications, employees can rehearse challenging conversations with an avatar that reacts in real time like a real person; during physical repairs, users can show damaged parts directly to the camera without knowing technical terminology; and it can offer everyday emotional companionship for older adults.

Griffin is not yet integrated into the existing Tavus commercial platform. Currently, platform customers still access the legacy architecture using Phoenix for facial generation, Raven for visual perception, and Sparrow for turn-taking timing; the availability of legacy models does not imply Griffin is open for use. This work remains in an early stage, but its core value lies in reducing the effort users must expend accommodating a system's awkward conversational timing.