Transfer learning via distributed brain recordings enables reliable speech decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Speech brain-computer interfaces (BCIs) integrate neural recordings and large language models for real-time speech decoding.
• Current BCIs require extensive cortical coverage, limiting scalability due to individual brain variations.

Purpose of the Study:

• To develop scalable transfer learning strategies for neural speech decoding.
• To improve the generalizability of BCIs across individuals with heterogeneous brain organization.

Main Methods:

• Utilized stereo-electroencephalography (sEEG) recordings from a large cohort during a demanding speech motor task.
• Employed a sequence-to-sequence model for decoding variable-length phonemic sequences.
• Developed a cross-subject transfer learning framework to identify shared neural patterns.

Main Results:

• The group-derived decoder significantly outperformed individual models, demonstrating enhanced decoding robustness.
• The transfer learning framework successfully isolated shared latent manifolds across subjects.
• Achieved reliable speech decoding despite variable neural coverage and activation patterns.

Conclusions:

• A pathway toward generalizable neural prostheses for speech and language disorders is highlighted.
• Leveraging large-scale intracranial datasets with distributed sampling and shared task demands is crucial for BCI advancement.
• This approach offers improved scalability and robustness for future speech BCIs.