Decoding articulatory and phonetic components of naturalistic continuous speech from the distributed language network

Area of Science:

• Neuroscience
• Speech Science
• Biomedical Engineering

Background:

• Speech production involves a distributed brain network.
• Current speech-brain-computer interfaces (speech-BCIs) often focus on limited superficial brain regions.
• Stereo-electroencephalography (sEEG) offers a less invasive method to access distributed brain activity.

Purpose of the Study:

• To investigate the potential of sEEG for decoding speech components from widespread cortical sites.
• To identify neural correlates of articulatory and phonetic components in continuous speech production.
• To explore novel neural targets for advanced speech-BCI development.

Main Methods:

• Utilized stereo-electroencephalography (sEEG) to record neural activity from eight participants during continuous speech production (reading aloud).
• Trained linear classifiers on broadband gamma activity to decode articulatory components (place/manner of articulation) and phonemes.
• Employed nested five-fold cross-validation to evaluate decoding performance.

Main Results:

• Achieved classification accuracies for place of articulation (18.7%), manner of articulation (26.5%), and phonemes (4.81%).
• Highest accuracies reached 26.3% (place), 35.7% (manner), and 9.88% (phonemes).
• Effective decoding electrodes were distributed across multiple brain regions, including sensorimotor, frontal, temporal, and fusiform cortices, indicating widespread neural representations.

Conclusions:

• Speech components are represented across a distributed cortical network, not localized to a single area.
• sEEG enables decoding of continuous speech components using minimally invasive techniques.
• Findings provide crucial insights into language neurobiology and identify promising neural targets for future speech-BCIs.