This document discusses an experiment on silent speech recognition using electromyography (EMG) and electrode arrays. The experiment tested four setups using either 16 or 35 EMG channels attached to the face. Results showed that recognition performance was better when using more training sentences, with optimal context widths for feature extraction differing based on the number of channels and training sentences. Principal component analysis preprocessing led to more consistent relative word error rate improvements between 10-18% across setups.