A bilingual speech neuroprosthesis driven by cortical articulatory representations shared between languages.

Area of Science:

• Neuroscience
• Computational Linguistics
• Biomedical Engineering

Background:

• Previous speech decoding research focused on single languages.
• The neural basis of bilingual speech production across languages remained unclear, especially post-paralysis.

Purpose of the Study:

• To investigate shared versus unique neural activity for speech production in bilingual individuals.
• To develop a brain-computer interface capable of decoding multiple languages from neural signals without language specification.

Main Methods:

• Utilized electrocorticography (ECoG) to record brain activity from the speech-motor cortex.
• Employed deep-learning and statistical natural-language models for English and Spanish decoding.
• Developed a generalized syllable classifier based on shared articulatory representations.

Main Results:

• Successfully decoded sentences in both English and Spanish from a single bilingual participant with paralysis.
• Demonstrated that decoding models leverage shared articulatory representations across languages.
• Showcased the efficacy of transfer learning to improve decoding performance between languages.

Conclusions:

• Bilingual speech production relies on shared cortical articulatory representations that are robust to paralysis.
• A single, language-agnostic decoder can be developed for multiple languages.
• This approach offers a pathway for restoring communication in paralyzed individuals who speak multiple languages.