Multi-layer Perceptrons for Subvocal Recognition

A multi-layer perceptron (MLP) system is investigated for recognizing phonemes from EMG data. EMG data was recorded during full vocalization and subvocalization. Motor Unit Action Potentials (MUAPs) were inferred from EMG data and their spectral energies used to train four articulatory feature extraction models based on MLPs. Outputs from the feature extraction networks were then fed forward through a final MLP to predict phonemes. Performance gains against a benchmark with no articulatory feature extraction and only a single phoneme classifying MLP were demonstrated.