Here Cipher

A model catches speech traces in brain noise, but words are still out of reach

CIPHER reports roughly 0.67 to 0.69 WER on the 11-class CVC phoneme task, keeping it far from a practical EEG-to-text system.