Geometry of orofacial neuromuscular signals: speech articulation decoding using surface electromyography

Area of Science:

• Biomedical Engineering
• Neuroscience
• Signal Processing

Background:

• Electromyogram (EMG) signals offer a potential pathway for speech restoration in individuals with speech impairments.
• Current EMG-based speech neuroprostheses face challenges in data collection and computational resources.

Purpose of the Study:

• To present data and methods for decoding speech articulations using surface electromyogram (EMG) signals.
• To develop efficient neural networks for EMG-based speech prostheses.

Main Methods:

• Collecting EMG signals from facial, jaw, and neck muscles during speech articulation.
• Employing EMG-to-speech translation techniques.
• Utilizing the manifold of symmetric positive definite matrices as an embedding space for EMG data.

Main Results:

• The manifold of symmetric positive definite matrices provides a natural embedding space for EMG signals.
• An algebraic interpretation of manifold-valued EMG data was achieved using linear transformations.
• Distribution shifts in EMG signals across individuals were analyzed and quantified.

Conclusions:

• The proposed approach demonstrates potential for data- and parameter-efficient neural networks.
• This method is significant for developing practical EMG-based speech restoration systems.
• The findings address key challenges in EMG signal processing for neuroprosthetics.