EMG to speech mapping using scalogram-spectrogram features for silent speech recognition

Silent speech Recognition systems (SSRs), which rely upon non-vocal signals like electromyography (EMG) from facial muscle tissues, offer a promising answer by permitting speech production without sound. This study seeks to enhance communication for those individuals by using a machine learning model, getting to know to correctly translate EMG signals from silent speech into audible, intelligible speech, thereby bridging the distance between non-verbal and verbal exchange. SSR improves the quality of life for those with speech impairments by providing necessary communication choices. This project investigates a whole new SSR method, which involves using surface electromyography signals (sEMG) and audio. A machine learning model was trained to recognize keywords via silent speech by an addition of a dataset of audio and sEMG signals. The scalograms and spectrograms are plotted for audio and EMG signal respectively and then it is mapped using the EfficientNet B0. It is then converted into textual content and then audible speech with the help of a speech synthesis system with an accuracy of 76%. This innovative technology opens new avenues in terms of silent communication, especially for customers who can talk without vocalizing, thereby improving accessibility and conversation for those who need it.