Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text (arxiv.org)

In a recent study, researchers found that significant advancements in decoding words from non-invasive brain recordings can occur without relying on actual brain data. By examining the methods outlined in d'Ascoli et al. (2025), the study revealed that a neural network could predict speech effectively due to implicit cues in the timing of word intervals rather than direct brain signals. This shortcut facilitated a comparable accuracy of 22.0% with synthetic signals devoid of brain information, just slightly below the 22.3% accuracy from real recordings. To enhance the fidelity of the brain-to-text interpretation, the authors implemented a simple but crucial adjustment: processing each window of speech independently instead of collectively. This modification allowed the neural network to better focus on the nuances of word-specific information derived from brain data, leading to improved performance. The new approach, termed SimpleB2T, achieved a word error rate of 36.6%, nearing results from invasive methods, and significantly boosted the effectiveness of existing strategies for aggregating predictions and leveraging pre-trained language models. This research highlights the importance of avoiding shortcuts in AI and offers a more robust framework for advancing non-invasive speech decoding technologies.