How Far Are We From the Realization of Non-Invasive Speech BCIs? A Deep Learning Scaling Law-Guided Clinical Analysis
Zihan Zhang, Yu Bao, Xiao Ding, Tianyi Jiang, Xia Liang, Juntong Du, Kai Xiong, Bing Qin, Ting Liu · 2025
Early brain-computer interfaces (BCIs) based on motor imagery have achieved notable success. Since 2023, pioneering work on high-performance BCIs - speech BCIs - has garnered widespread attention. These BCIs enable real-time translation of spontaneous language-related neural activity into natural language, offering the potential to restore speech in patients with conditions such as Amyotrophic Lateral Sclerosis, who have lost their ability to communicate. However, these speech BCIs necessitate craniotomy, making them unsuitable for patients who are unwilling or unable to undergo neurosurgical procedures. This limitation underscores a pressing, unresolved question in the field: how far are we from achieving non-invasive speech BCIs? To address this, we adopted interdisciplinary approaches and clinically collected the first dataset for training non-invasive speech BCI. To draw cautious conclusions, we performed a large number of analyses on the experimental data from multiple perspectives and derived the ‘non-invasive speech BCI scaling law’, which suggests that approximately 15 years’ worth of data from a single participant is required to realize non-invasive speech BCIs-posing a substantial barrier to clinical application. We therefore explore key strategies to overcome this challenge, for reducing data requirements and advancing the development of next-generation BCIs.