Brian Johnson is a staff applied machine-learning engineer at Tavus and founder of aerbits.ai. He develops the real-time systems that help digital humans recognize when to listen, respond, or ignore interruptions.
Johnson studied electrical engineering at the University of Texas at Austin, earned a law degree focused on intellectual property, and managed engineering teams at Zillow and Earnest. His startup aerbits.ai applied drones and computer vision to detecting illegal dumping, giving municipalities another way to identify urban problems.
At Tavus, Johnson works on conversational timing and selective attention for video agents. His contributions include:
- Sparrow-0 and semantic turn detection: Sparrow-0 uses linguistic context to distinguish completed thoughts from ordinary pauses, reducing interruptions without introducing unnecessary delays.
- Sparrow-1 and conversational floor ownership: Sparrow-1 incorporates rhythm, hesitation, intonation, overlapping speech, and individual pacing to determine who should speak and when. Response timing can shift with emotional context: quick exchanges demand speed; sensitive moments benefit from a pause.
- Real-time speaker identification: His work on speaker embeddings and speaker focus helps digital humans distinguish their conversation partner from background voices and other interruptions.
Johnson also advocates integrating Tavus models with established conversational infrastructure. At AI Engineer World’s Fair 2025, he explained how Tavus rendering, turn-taking, and perception can connect with Pipecat, the separate open-source orchestration framework developed by Daily. Building these systems has also sharpened his attention to how silence, hesitation, and attentive listening shape human conversation.
