Research & Papers

MEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification

A new AI model can read your mind by decoding MEG brain signals.

Deep Dive

Researchers have developed MEGConformer, a Conformer-based AI model that decodes speech and phonemes directly from non-invasive MEG brain scans. The system achieved a top score of 88.9% accuracy for speech detection and 65.8% for phoneme classification on the LibriBrain 2025 benchmark, winning the official Phoneme Classification Standard track. The model uses a compact architecture adapted to raw 306-channel MEG signals with specialized data augmentation and normalization techniques.

Why It Matters

This is a major leap toward scalable, non-invasive brain-computer interfaces for communication and assistive technology.