A simplified adversarial architecture for cross-subject silent speech recognition using electromyography.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyography (EMG)-based silent speech recognition faces challenges due to individual variations in speech patterns and physiology.
• Existing adversarial networks offer limited direct contribution to classifier predictions for cross-subject recognition.
• Feature alignment techniques are crucial for addressing domain offsets across speakers.

Purpose of the Study:

• To propose a simplified discrepancy-based adversarial network for improved EMG-based cross-subject silent speech recognition.
• To develop a streamlined end-to-end structure that enhances feature alignment across subjects.
• To overcome limitations of current adversarial approaches in direct categorical prediction.

Main Methods:

• A novel cascaded adaptive rectification network is used for front-end feature extraction from noisy myoelectric signals.
• A Nuclear-norm Wasserstein discrepancy metric is introduced for feature alignment, serving both classification and domain discrimination.
• The network adaptively reshapes feature maps to filter domain-specific information and retain domain-invariant features.

Main Results:

• Achieved an average accuracy of 89.46% on 40 new subjects after training with data from 60 subjects.
• Demonstrated a 10.07% improvement over state-of-the-art models when tested on 10 new subjects with 20 training subjects.
• Outperformed existing methods even with significantly fewer training subjects.

Conclusions:

• The proposed simplified discrepancy-based adversarial network significantly enhances EMG-based cross-subject silent speech recognition.
• The method effectively filters domain-specific noise while preserving critical domain-invariant features.
• This work offers a promising advancement for EMG-based speech interactive applications.