We’re sharing some of the challenges we faced building an AI video interface that has realistic conversations with a human, including getting it to under 1 second of latency.

To try it, talk to Hassaan’s digital twin: https://www.hassaanraza.com, or to our "demo twin" Carter: https://www.tavus.io

We built this because until now, we've had to adapt communication to the limits of technology. But what if we could interact naturally with a computer? Conversational video makes it possible – we think it'll eventually be a key human-computer interface.

To make conversational video effective, it has to have really low latency and conversational awareness. A fast-paced conversation between friends has ~250 ms between utterances, but if you’re talking about something more complex or with someone new, there is additional “thinking” time. So, less than 1000 ms latency makes the conversation feel pretty realistic, and that became our target.

Our architecture decisions had to balance 3 things: latency, scale, & cost. Getting all of these was a huge challenge.

The first lesson learned was to make it low-latency, we had to build it from the ground up. We went from a team that cared about seconds to a team that counts every millisecond. We also had to support thousands of conversations happening all at once, without getting destroyed on compute costs.

For example, during early development, each conversation had to run on an individual H100 in order to fit all components and model weights into GPU memory just to run our Phoenix-1 model faster than 30fps. This was unscalable & expensive.

We developed a new model, Phoenix-2, with a number of improvements, including inference speed. We switched from a NeRF based backbone to Gaussian Splatting for a multitude of reasons, one being the requirement that we could generate frames faster than realtime, at 70+ fps on lower-end hardware. We exceeded this and focused on optimizing memory and core usage on GPU to allow for lower-end hardware to run it all. We did other things to save on time and cost like using streaming vs batching, parallelizing processes, etc. But those are stories for another day.

We still had to lower the utterance-to-utterance time to hit our goal of under a second of latency. This meant each component (vision, ASR, LLM, TTS, video generation) had to be hyper-optimized.

The worst offender was the LLM. It didn’t matter how fast the tokens per second (t/s) were, it was the time-to-first token (tfft) that really made the difference. That meant services like Groq were actually too slow – they had high t/s, but slow ttft. Most providers were too slow.

The next worst offender was actually detecting when someone stopped speaking. This is hard. Basic solutions use time after silence to ‘determine’ when someone has stopped talking. But it adds latency. If you tune it to be too short, the AI agent will talk over you. Too long, and it’ll take a while to respond. The model had to be dedicated to accurately detecting end-of-turn based on conversation signals, and speculating on inputs to get a head start.

We went from 3-5 to <1 second (& as fast as 600 ms) with these architectural optimizations while running on lower-end hardware.

All this allowed us to ship with a less than 1 second of latency, which we believe is the fastest out there. We have a bunch of customers, including Delphi, a professional coach and expert cloning platform. They have users that have conversations with digital twins that span from minutes, to one hour, to even four hours (!) - which is mind blowing, even to us.

Thanks for reading! let us know what you think and what you would build. If you want to play around with our APIs after seeing the demo, you can sign up for free from our website https://www.tavus.io.