EEG-Translator: A Cross-Modality Framework for Subject-Specific EEG and Voice Reconstruction from Imagined Speech.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Non-invasive brain-computer interfaces (BCIs) show promise for communication in individuals with speech impairments.
• Decoding speech-related electroencephalography (EEG) signals is key, but cross-domain reconstruction is needed for better performance.

Purpose of the Study:

• To propose a cross-modal EEG translation framework to reconstruct overt speech EEG from imagined speech EEG.
• To enhance subject-specific speech synthesis and improve BCI system performance.

Main Methods:

• Integrated a diffusion-based model with Generative Adversarial Network (GAN) training for cross-domain EEG reconstruction.
• Incorporated spectrogram-based features and a fusion of spatial and spectral losses.
• Trained the model on both EEG signals and spectrogram features to preserve time-frequency properties.

Main Results:

• Reconstructed EEG improved class decoding accuracy by 6.2% compared to original imagined EEG.
• The framework successfully preserved EEG class information and time-frequency domain properties.
• Category-wise analysis provided linguistic insights into decoding performance variations.

Conclusions:

• Diffusion-driven EEG translation shows significant potential for advancing speech BCIs.
• Integrating deep learning with linguistic insights is crucial for improved neural signal reconstruction and interpretation.
• This approach offers a pathway to more effective communication tools for individuals with speech disabilities.