Speech decoding from stereo-electroencephalography (sEEG) signals using advanced deep learning methods

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-computer interfaces (BCIs) offer a way to restore communication by decoding brain signals.
• While invasive methods like micro-electrode arrays and electrocorticography are common for speech BCIs, stereo-electroencephalography (sEEG) is less explored.
• Restoring communication for individuals with speech impairments is a significant challenge in neuroscience and medicine.

Purpose of the Study:

• To investigate the efficacy of stereo-electroencephalography (sEEG) for decoding spoken words.
• To compare the performance of deep learning models against traditional methods for speech decoding using sEEG data.
• To explore the potential of sEEG-based BCIs for speech restoration.

Main Methods:

• Utilized recently released sEEG data from epileptic participants speaking Dutch words.
• Implemented and compared three decoding methods: linear regression, a recurrent neural network (RNN)-based sequence-to-sequence model, and a transformer model.
• Applied advanced deep learning techniques to decode speech waveforms directly from sEEG signals.

Main Results:

• Both RNN and transformer models significantly outperformed the linear regression method in speech decoding.
• No significant performance difference was observed between the RNN and transformer models.
• Speech decoding was achievable using only a subset of the sEEG electrodes, indicating the importance of electrode location.

Conclusions:

• Decoding speech from sEEG signals is feasible and effective using deep learning.
• The precise location of sEEG electrodes is a critical factor influencing speech decoding performance.
• sEEG presents a viable, yet underutilized, modality for developing advanced speech-restoring brain-computer interfaces.