Abstract: In the advancing field of rehabilitation technology and human-machine interfaces, surface electromyography (sEMG) has emerged as a critical non-invasive method for interpreting human intentions, particularly in developing advanced prosthetics and silent speech recognition systems. However, despite its potential, challenges such as noise interference and the necessity for precise electrode placement have constrained its accuracy. This paper explores the application of advanced deep learning (DL) models, including Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Deep Neural Networks (DNN), and Convolutional Neural Networks (CNNs) in both 1-dimensional (1D) and 2-dimensional (2D) formats to improve the interpretability of sEMG signals for silent speech recognition. The proposed setup utilized a multithreading queuing (MTQ) based novel three-channel low-cost sEMG data acquisition system for English vowel recognition. The two channels are responsible for collecting and extracting data, while the third channel helps visualize data in real-time. It involves data acquisition using disposable electrodes across key facial muscles, followed by employing a range of DL models to process and classify the sEMG signals. Our findings suggest that advanced DL models, particularly the CNN-2D model, outperformed other state-of-the-art methods by achieving 90% accuracy in vowel recognition, showcasing the potential of deploying low-cost hardware with new predictive paradigms in sEMG analysis.

Keywords: Surface Electromyography (sEMG), Silent Speech Recognition, Deep Learning Models, Rehabilitation Technology

Cite:
Rajendra Kachhwaha, Rajesh Bhadada,"Decoding the Unspoken: Deep Learning-Based Recognition of Silently Spoken English Vowels Using sEMG Signals", IARJSET International Advanced Research Journal in Science, Engineering and Technology, vol. 11, no. 1, 2024, Crossref https://doi.org/10.17148/IARJSET.2024.11115.

| DOI: 10.17148/IARJSET.2024.11115