Exploring Dimensionality Reduction Techniques in Word Classification using Surface EMG Signals

Surface Electromyography (sEMG) is a revolutionary technology in the field of unvoiced speech recognition that captures and analyzes electric signals produced by facial muscles during speech. Despite the significance of providing an interface for recognizing and synthesizing spoken words, research in this area does not investigate and compare the impact of dimensionality reduction techniques to reduce complexity. This study aims to develop a frame-based word classification system incorporating dimensionality reduction techniques using surface EMG speech signals. To attain this objective, several ML models including Decision tree, KNN, Random Forest, XGB, and Bidirectional LSTM are developed as classifiers integrating the dimensionality reduction algorithms Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA). The study found that the Random Forest classifier along with LDA performed remarkably in the experiment yielding an accuracy of 96.28%. Again, the XGB classifier showed an impressive classification accuracy of 93.45% with PCA. However, both KNN and DT performed poorly across both dimensionality reduction techniques, and the neural network (BLSTM) showed around 85% accuracy with both LDA and PCA. This research provides important insights into the application and effectiveness of dimensionality reduction techniques in improving the performance of different machine learning models for sEMG-based unvoiced speech recognition.