Abstract
Electromyography (EMG) is a method for capturing silent speech by measuring the electrical activity of facial muscles using EMG sensors. However, distinguishing between different phonemes solely from the EMG signal becomes challenging when they are pronounced in close proximity or have similar manners of articulation. The objective of this research is to develop a framework for correcting speech units based on EMG signals. This framework leverages a unit-language model trained on normal speech units to improve the accuracy of silent speech synthesis. The proposed unit correction framework demonstrates its efficacy in enhancing the intelligibility of synthesized speech derived from EMG signals. The developed framework holds the potential for application in various domains, including assistive communication devices and speech rehabilitation therapies.