NUROSIS LABS
How thought becomes text
The instrument, the model, and the evidence — everything behind the applications, in the open.
I am thirsty, please bring water.
FIG. 01LIVE DECODING VIEWER· ILLUSTRATIVE TRACE
GLASS-TO-VOICE
189 MS END-TO-END · INSIDE CONVERSATIONAL TOLERANCE
STAGE WIDTHS DRAWN TO SCALE · 1 PX ≈ 2 MS
THE MODEL
A foundation model for the speaking brain
Every 8 ms window is encoded, attended, and voiced — end to end inside conversational latency.
CAPABILITIES
Frontier scale, human grade
- 01Open vocabulary
- Decodes arbitrary sentences, not a menu of commands. If you can think it, Relay can say it.
- 02Session-robust
- Stable across days and headset placements, without per-session retraining rituals.
- 03Person-adaptive
- A short enrollment pass tunes the model to a new speaker; shared priors do the rest.
- 04Low latency
- A streaming architecture keeps end-to-end delay inside conversational tolerance.
- 05Hardware-agnostic
- Works across non-invasive recording hardware instead of locking you to one device.
- 06Private by architecture
- Decoding runs on local hardware where possible; raw neural data never has to leave the room.
THE BET
Scale dissolves the invasiveness gap.
One model, many brains, every device — until whose head it came from stops mattering.
RESEARCH
Why speech, why now
Invasive implants have shown that the brain's speech code can be read — but surgery will never scale to the millions who need it. Meanwhile non-invasive recording has quietly crossed a threshold: dense, high-fidelity signals, collectable anywhere.
Language models learned to write by reading everything ever written. A model of the brain has to be raised the same way: Nurosis trains a single large network across many brains, sessions, and devices until the representation of speech stops depending on whose head it came from. That is the bet.