Programa del congreso

Automatic Speech Recognition (ASR) systems’ performance often decreases in noisy environments, in situations where the message should remain private, or when used by individuals affected by speech production impairments. The aim of this study is to assess the potential of SSI as an alternative to provide a more robust, accessible, and secure kind of communication. Specifically, it evaluates and compares the performance of four different types of machine learning models (Naive Bayes, Support Vector Machines, Random Forest, and K-nearest neighbours) to recognise 10 English words pronounced silently. To do this, surface electromyography (sEMG) signals produced by 13 subjects when articulating the words were recorded using a 5-channel bipolar electrode arrange. Although the average accuracy across all models was relatively similar, Naive Bayes models exhibited the most consistent performance across different subjects, which makes them a solid candidate for Silent Speech systems. The Naive Bayes models achieved an accuracy of 79.49% for the classification of five numbers, 81.80% for five aviation-related commands, and 67.56% when classifying a dataset of digits and commands together. These results confirm the potential of machine learning models for developing Silent Speech systems based on EMG signals that may achieve robust and accurate silent communication.