CSL-EMG_Array: An Open Access Corpus for EMG-to-Speech Conversion

We present a new open access corpus for the training and evaluation of EMG-to-Speech conversion systems based on array electromyographic recordings. The corpus is recorded with a recording paradigm closely mirroring realistic EMG-to-Speech usage scenarios, and includes evaluation data recorded from both audible as well as silent speech. The corpus consists of 9.5 hours of data, split into 12 sessions recorded from 8 speakers. Based on this corpus, we present initial benchmark results with a realistic online EMG-to-Speech conversion use case, both for the audible and silent speech subsets. We also present a method for drastically improving EMG-to-Speech system stability and performance in the presence of time-related artifacts.