Setu Labs · Research
What we are working on.
Our research follows one belief: the interface between people and machines is a learning problem. The sensor sets the ceiling. What decides the rest is how fast a model can come to know one person.
New release · v0.1.0
SETU Neural EEG
Trained models for the next human interface. Our EEG model suite and the software behind it, released in the open.
All research
Model releasev0.1.0 · EEGSETU Neural EEG: trained models for the next human interface.
Task-trained neural networks, fine-tuned EEG encoders and command classifiers, with open-source training, adaptation and evaluation code.
919 inference checkpoints · 5.86 GB of trained artifacts · Open source
Silent speech, from muscle to text.
SETU-EMG decodes deliberately mouthed speech from eight channels of facial muscle activity, with no audio at inference time. 16.16% word error rate on the benchmark test partition.
Technical summary. Internal results, not yet peer reviewed.
Signal and position, together.
SetuMontageNet combines measured EEG with electrode locations to study motor-imagery decoding. Missing channels stay masked; a separate policy turns decoder outputs into computer actions.
Architecture overview. Performance evaluation and live-hardware studies will be reported separately.