TransTCNet: Transformer-Based Temporal-Contextual Network for Low-Latency Typing Interfaces on Edge Devices.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) offers potential for silent, hands-free typing by interpreting muscle activity.
• Character-level sEMG typing is challenging due to subtle temporal variations and overlapping muscle dynamics.
• Temporal features are crucial for accurate typing recognition, accounting for user-specific variations in keypresses.

Purpose of the Study:

• To propose TransTCNet, a novel deep neural network architecture for high-accuracy character-level sEMG typing.
• To evaluate the performance and generalizability of TransTCNet on a public sEMG typing dataset.
• To assess the suitability of TransTCNet for real-time applications and edge devices.

Main Methods:

• Developed a two-stage deep neural network, TransTCNet, incorporating causal convolutional layers and a transformer component.
• Utilized a publicly available 26-class sEMG dataset from 19 individuals for training and validation.
• Evaluated model performance using accuracy, Area Under the Curve (AUC), and prediction confidence metrics.

Main Results:

• TransTCNet achieved a validation accuracy of 96.53%, outperforming baseline models.
• The model demonstrated high generalization across participants with AUC values exceeding 0.994 for all classes.
• Achieved high training accuracy (97.86%) with reliable predictions (>0.9 confidence) suitable for real-time filtering.

Conclusions:

• TransTCNet effectively decodes fine-grained neuromuscular signals for accurate sEMG typing.
• The model's efficiency and low inference cost make it suitable for wearable and edge devices.
• TransTCNet shows promise for real-time applications like adaptive interfaces, VR/AR, prosthetics, and communication systems.