Comparison of feature evaluation criteria for speech recognition based on electromyography

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyography (EMG) signals offer a non-invasive method for human-computer interaction.
• Accurate classification of EMG signals is crucial for applications like speech recognition.
• Feature selection and classifier choice significantly impact EMG-based classification performance.

Purpose of the Study:

• To compare 14 feature evaluation criteria and 4 classifiers for isolated Thai word classification using EMG signals.
• To identify an optimal combination of feature selection criteria and classifiers for Thai word recognition.
• To evaluate classification performance in both audible and silent speech modes.

Main Methods:

• EMG signals were recorded from 10 subjects speaking 11 Thai number words (audible and silent).
• Signals were preprocessed, and 22 widely used EMG features were extracted.
• Features were evaluated using 14 criteria (independent and dependent), with the top nine selected for classification.
• Four classifiers were employed with 10-fold cross-validation to assess performance.

Main Results:

• The best average accuracies were achieved using dependent criteria (DC) with Fisher's least square linear discriminant (D_FLDA) and a linear Bayes normal classifier (LBN).
• Audible mode classification accuracy reached 93.25%.
• Silent mode classification accuracy reached 80.12%.

Conclusions:

• The combination of D_FLDA for feature selection and LBN for classification demonstrates high efficacy for Thai word recognition from EMG signals.
• This approach is effective for both audible and silent speech modes, highlighting its robustness.
• The findings provide a near-optimal strategy for EMG-based Thai speech classification.