Increasing the Translational Efficacy of Intracortical Speech Brain-Computer Interfaces

Vocal aphasia, a condition characterized by impaired ability to vocalize speech, affects an estimated two million individuals (Ivanova, 2022). A recent study by Willett et al. from the Neuroprosthetics Translational Laboratory (NPTL) has introduced Brain-Computer Interfaces (BCIs) as a novel approach to mitigating the effects of aphasia. In this study, researchers collected intracortical EEG data from a patient and developed a recurrent neural network (RNN) to decode the neural signals into words, boasting a word error rate of just 18.8%.
Building upon this foundation, our paper presents the application of two deep state space machine learning models (dSSMs)—the S5 and Mamba—to improve the accuracy of this task. These model types have gained significant popularity in recent years and have outperformed traditional RNN models in a variety of long time series tasks, motivating their use in our study (Cirone, 2024). After considerable re-engineering efforts to fine-tune our models', our Mamba model achieved a word error rate of 16.7%. This achievement emphasizes the concrete benefits that BCIs could offer to individuals with speech impairments.
Moreover, an essential part of our research was the novel implementation of various machine learning techniques, from the NPTL's Speech-BCI model pipeline, in JAX, an open-source Python library that is particularly well-suited for the intricacies of dSSMs. This endeavor has resulted in the creation of multiple JAX-based functions and classes, which will facilitate the development of other dSSMs in the future. Moving forward, I hope to further enhance the accuracy and utility of our models and investigate the neural representations underlying speech in our models hopefully providing insight into how speech is processed in the brain.