A SPEECH SYNTHESIZER USING FACIAL EMG SIGNALS

Abstract

This paper proposes a novel phoneme classification method using facial electromyography (EMG) signals. This method makes use of differential EMG signals between muscles for phoneme classification, which enables a speech synthesizer to be constructed using fewer electrodes. The EMG signal is derived as a differential between monopolar electrodes attached to two different muscles, unlike conventional methods in which the EMG signal is derived as a differential between bipolar electrodes attached to the same muscle. Frequency-based feature patterns are then extracted using a filter bank, and the phonemes are classified using a probabilistic neural network, called a reduced-dimensional log-linearized Gaussian mixture network (RD-LLGMN). Since RD-LLGMN merges feature extraction and pattern classification processes into a single network structure, a lower-dimensional feature set that is consistent with classification purposes can be extracted; consequently, classification performance can be improved. Experimental results indicate that the proposed method with a fewer number of electrodes can achieve a considerably high classification accuracy.