Surface Electromyography-Based Speech Detection Amid False Triggers for Artificial Voice Systems in Laryngectomy Patients

Abstract

Laryngectomy, a surgical intervention for laryngeal cancer, effectively treats the condition but results in the loss of natural speech. Voice restoration post-laryngectomy typically involves manual control, limiting patients’ ability to multitask while speaking. Surface electromyography (sEMG) offers a hands-free alternative for controlling artificial voice systems. However, challenges arise from daily, orofacial activities like chewing or coughing, activating the same muscles used for sEMG control, potentially causing false triggers. To address this, we perform a detailed analysis of facial and neck muscles during speech and non-speech activities to identify potential false triggers for sEMG-controlled artificial voice systems. We propose a five-step algorithm to prepare noisy sEMG data for analysis and to detect accurate speech onset and termination times within the muscle activity. A two-stage classification approach is suggested to effectively distinguish speech from non-speech activities. The classifier in the first stage detects the presence of any activity versus non-activity with an F1-score of 95.8%, while the classifier in the second stage recognizes speech among other activities with an F1-score of 96.3%. This research marks a significant advancement in differentiating speech from other daily activities, thereby minimizing false triggers in sEMG-controlled artificial voice systems.