This year, we took part in the LibriBrain Competition 2025, an international challenge presented at NeurIPS 2025 aimed at advancing the decoding of language from non-invasive brain signals using the large-scale LibriBrain dataset. The competition seeks to foster progress in brain–computer interfaces, with the long-term goal of restoring communication abilities in individuals with speech impairments and enabling novel forms of human–machine interaction based on neural data.
Our system, neural2speech, achieved first place in the Phoneme Classification Standard Track. This track focuses on the prediction of phoneme classes directly from MEG (magnetoencephalography) recordings, under a constrained setting where only the official training data can be used, making robustness and generalization key challenges.
In our paper, “MEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification”, we describe the technical approach behind our solution. We adapt a Conformer architecture, originally designed for automatic speech recognition, to operate directly on raw MEG signals from 306 sensors, effectively capturing both temporal dependencies and spectral characteristics of neural activity. Our method incorporates instance-level normalization to reduce distribution shifts across data splits, a dynamic chunk-averaging data loader to improve phoneme classification performance, and class-balancing strategies based on inverse square-root frequency weighting to address class imbalance.
These design choices result in a robust and competitive system, allowing neural2speech to stand out among the submitted solutions and demonstrating a meaningful step forward at the intersection of neural signal processing and speech technology.