A New Frontier for
Conversational Voice
Shaan Fulton, Dylan Westland, and the Dreamweaver team • March 2026
Abstract
Conversational AI has always forced a tradeoff. Sequential pipelines flatten voice into text, wait for a frontier LLM, then synthesize sound—trading quality for speed at every step. End-to-end speech models sound natural but lack frontier intelligence, reliability, and tool use. Vox 1.2 sidesteps this entirely. A predictive architecture generates and refines its response while you speak—frontier reasoning, emotional voice analysis, and high-quality synthesis across 40+ languages, with no compromise.
The Predictive Duplex Architecture
A continuous, parallel processing pipeline that reasons while you speak, eliminating the latency of sequential turns.
As the user speaks, audio streams in.
Simultaneously, we speculatively predict, compute, and synthesize the response.
The response is held in memory, waiting for the user to finish.
The moment speech ends, audio plays. Standard pipelines are still thinking.
Trajectory Prediction
A lightweight language model continuously predicts the user's semantic trajectory.
Conversational
Dynamics
Predictive Backchanneling
The model injects natural acknowledgments—"yeah", "right", contextual laughter—while the caller is still speaking. Duplex audio means these don't interrupt the caller's stream.
Acoustic Empathy
Parallel audio analysis feeds caller sentiment, speaking rate, and emotional state to the LLM in real time. A frantic caller gets a calm, measured response—not just the right words, but the right cadence.
Expressive TTS
Character voice and vocabulary are trained on the same dataset, not decoupled into separate acoustic and language models. Every line carries the character's true vocal identity. 40+ languages out of the box.
We're a team of founders from UC Berkeley rolling out deliberately. If you want to pilot Vox in your business, we'd love to talk.
Shaan & Dylan