Decomposition of multichannel electromyographic signals for a silent speech interface

Silent speech interfaces (SSI) provide ways capture silently spoken human speech. In electromyography (EMG) based SSIs, surface electrodes attached to the user’s face record electric signals, which are used to infer speech. It is an important goal to increase wearing comfort and decrease system setup time by reducing the number of devices attached to the user’s face. By using multi-electrode arrays, the activity of facial muscles in different positions can be recorded using one single device. Because of the increased data dimensionality and the similar signals recorded in neighboring electrodes, source separation methods can be used to locate the activity of different muscles and to detect noise. We conduct experiments to evaluate the performance of different source separation setups in preprocessing for direct EMG-to-Speech conversion. Objective intelligibility measures are used to evaluate the quality of synthesized speech. Independent Component Analysis (ICA) performs well for source separation on EMG signals and can be used to improve the quality of the preprocessed data. We propose a novel heuristic for automatic selection of promising sources identified by ICA. The heuristic is based on the correlation of spatial spectra of the weight distributions of the single independent components. Reprojecting the selected independent signal sources to their original dimensions can improve the intelligibility of the synthesized speech. We also show that while Stationary Subspace Analysis (SSA) is able to remove white noise from EMG data, it is not suitable for unsupervised artifact removal in EMG arrays.