CIPHER: Conformer-based Inference of Phonemes from High-density EEG

A new research paper introduces CIPHER (Conformer-based Inference of Phonemes from High-density EEG Representations), a novel AI model developed by researcher Varshith Madishetty that attempts to decode speech directly from brain activity. The system uses a dual-pathway approach combining two types of neural features: event-related potentials (ERP) and broadband data-driven analysis (DDA) coefficients extracted from high-density electroencephalography (EEG) recordings. This represents a significant technical challenge due to the low signal-to-noise ratio and spatial blurring inherent in scalp EEG data.

The model was tested on the OpenNeuro ds006104 dataset containing recordings from 24 participants across two studies with concurrent transcranial magnetic stimulation (TMS). While binary articulatory tasks reached near-ceiling performance, these results proved vulnerable to confounding factors like acoustic onset timing and TMS-target blocking. On the primary 11-class consonant-vowel-consonant (CVC) phoneme task using leave-one-subject-out validation with 16 held-out subjects, performance was substantially lower with word error rates of 0.671 for ERP features and 0.688 for DDA features.

The researchers position CIPHER not as a functional EEG-to-text system but as a benchmark study for comparing neural feature extraction methods in brain-computer interfaces. They deliberately constrain their claims about neural representation to confound-controlled evidence, acknowledging the current limitations in fine-grained speech discrimination from non-invasive brain recordings. The work establishes important methodological standards for future research in this challenging domain of neural decoding.