An Optimized EMG Encoder to minimize soft speech loss for speech to EMG conversions

초록

Electromyography (EMG) to speech conversions is a standard problem to facilitate speech impaired individuals for communication via EMG (EMG to Speech). However, dataset acquisition is a cumbersome process and highly dependent on acquisition configuration. The availability of EMG signals can be made by tackling the inverse problem (Speech to EMG). In this paper, we propose an optimized EMG encoder which enhanced EMG feature extraction and in turn leads to improvements in soft speech units’ representations. To validate the efficacy of our proposed enhanced EMG encoder, we utilized state-of-the-art speech to EMG generative adversarial network (STE-GANs). We witnessed a significant improvements in synthesized EMG signals after utilizing proposed EMG encoder which improves soft speech losses by producing enhanced speech units during training of STE-GANs. The extensive results are presented on public dataset.