Ultrathin crystalline-silicon-based strain gauges with deep learning algorithms for silent speech interfaces

Area of Science:

• Engineering
• Biomedical Engineering
• Artificial Intelligence

Background:

• Wearable silent speech interfaces (SSI) enable communication without vocalization.
• Surface electromyography (sEMG) is a common SSI method but faces challenges with signal quality and scalability.
• Limitations of sEMG include poor signal-to-noise ratio and interelectrode interference.

Purpose of the Study:

• To develop a novel SSI utilizing crystalline-silicon-based strain sensors and a 3D convolutional deep learning algorithm.
• To address the scalability and signal quality issues associated with sEMG-based SSIs.
• To demonstrate a reliable and accurate silent speech recognition system.

Main Methods:

• Developed a novel SSI using crystalline-silicon-based strain sensors with minimized cell dimensions (<0.1 mm²).
• Employed two perpendicularly placed strain gauges to capture biaxial strain information.
• Integrated a 3D convolutional deep learning algorithm for data analysis.
• Attached four strain sensors near the subject's mouth to collect strain data.

Main Results:

• Achieved a high accuracy rate of 87.53% in classifying an unprecedentedly large wordset of 100 words.
• Demonstrated the system's reliability through various analysis methods.
• Compared performance with an sEMG-based SSI, which achieved a significantly lower accuracy rate of 42.60%.

Conclusions:

• The novel SSI utilizing strain sensors and deep learning offers a reliable and highly accurate solution for silent speech recognition.
• This approach overcomes the limitations of traditional sEMG-based SSIs, showing superior performance.
• The developed system holds promise for advancing non-vocal communication technologies.