Phoneme classification for speech synthesiser using differential EMG signals between muscles.

This paper proposes the use of differential electromyography (EMG) signals between muscles for phoneme classification, with which a Japanese speech synthesiser system can be constructed using fewer electrodes. In distinction from traditional methods using differential EMG signals between bipolar electrodes on the same muscle, an EMG signal is derived as differential between monopolar signals on two different muscles in the proposed method. Then, frequency-based feature patterns are extracted with filter banks, and classification of phonemes is realized by using a probabilistic neural network, which combines feature reduction and pattern classification processes in a single network structure. Experimental results show that the proposed method can achieve considerably high classification performance with fewer electrodes.