機械学習による表面筋電信号を用いた日本語単音の認識

抄録

The method and results of recognizing single Japanese sounds using surface electromyography (sEMG) signals generated from muscles around the mouth are presented. We determined six features of the waveforms of four muscles (a total of 24 indexes) and recognized 45 single Japanese sounds. We used machine learning with a neural network to improve sound recognition. The neural network has a 24 node input layer, a 100 node intermediate layer, and a 45 node output layer. Each index of a sound was entered into the input layer; the probabilities of the sound were output to the output layer. They were compared, and the output (sound) with the highest probability was determined to be a recognition sound. We used the cross-entropy loss as the loss function and gradient descent as the machine learning method. Machine learning, which built a neural network, has dramatically increased recognition; it stands at approximately 94%.