Speech AI today is built by compressing human recordings.

The limitation of current systems

Modern speech systems depend on massive datasets of recorded human audio. Their capabilities are constrained by what has been collected, licensed, and statistically reproduced.

This creates structural limits: dependence on proprietary data, weak interpretability, and limited grounding in how speech is actually produced.

A different approach

We train neural controllers to produce speech by interacting with a physical acoustic system in real time.

No datasets. No imitation. No predefined articulation paths.

Speech emerges as a learned motor skill.

Evidence of progress

• · Convergent discovery of speech-like structure across independent runs
• · Transition from large-scale search to adaptive motor control
• · Demonstrated multi-vowel production from a single controller
• · Real-time synthesis grounded in physical constraints

Why now

• · Increasing regulatory pressure around data provenance
• · Rising cost and scarcity of high-quality speech datasets
• · Demand for interpretable and controllable AI systems
• · Availability of compute to explore new training paradigms

What this unlocks

If speech can be learned without datasets, the entire speech stack changes.

Initial applications centre on environments where dataset dependence creates legal, operational, or strategic constraints.

• · Systems independent of proprietary voice data
• · Fully controllable speech generation
• · Stronger interpretability and safety narratives
• · New foundations for voice, robotics, and embodied AI

Trajectory

We are expanding from controlled vowel production toward richer sound inventories, robustness, and real-world interfaces—without compromising the core principle of learning through interaction with physics.