A Comparative Study of Acoustic Features for EMG-to-Speech Conversion in Low-Source Dataset

Electromyography (EMG) to audible speech conversions can help individuals with speech disabilities regain their ability to communicate. However, collecting large-scale EMG datasets can be challenging and cumbersome, posing a burden on users. Training a silent speech conversion model that performs well on low-resource datasets remains a challenge. One potential solution is to use acoustic features extracted from more efficient pre-trained models as the target features for prediction. This paper investigates the effects of content features from different pre-trained acoustic models and compare with commonly used MFCCs features in previous studies, exploring their impact in intelligibility, prosody, and conversion similarity.