This enables systems that learn speech production rather than rely on recorded datasets.
Internally, recent closed-loop work shows repeatable convergence toward target vowel structures from random initial states, then finer stabilisation—with distinct vowel targets from a single adaptive controller. See the homepage for behaviour-level visuals (before/after, F1/F2, convergence, repeatability) and Research progress for narrative.
This is not a shipped “speech API” yet; partner conversations focus on environments where dataset independence and controllability matter.
Speech generated through motor control can be inherently more stable and adaptable in real-world conditions, where noise and variability break conventional pipelines.
Systems that learn to produce speech directly explore alternatives to reliance on large proprietary datasets—changing cost, compliance, and scalability dynamics.
Motor-based systems expose underlying controls, enabling precise manipulation of voice characteristics beyond black-box generation.
A learning framework grounded in physics aligns naturally with robotics and embodied AI, where perception and action are tightly coupled.
This is not a feature layer on top of existing speech systems.
It is a different foundation for how speech is learned and produced.
Early access and partnerships will focus on environments where control, robustness, and data independence matter most.
We are exploring early applications in environments where dataset dependence creates legal, operational, or technical constraints.
Dataset-independent speech generation layer
Designed as a foundational component rather than a black-box model.
Speech systems built without reliance on recorded human voice datasets—relevant for regulated industries, sensitive deployments, and environments where data provenance, licensing, and control are critical.
Example: environments where using scraped or licensed voice data introduces legal or operational risk.
We are beginning to engage with partners exploring these constraints.
Enquire about early access →