Multi-stream HMM for EMG-based speech recognition

A technique for improving the recognition accuracy of EMG-based speech recognition by applying existing speech recognition technologies is proposed. The authors have proposed an EMG-based speech recognition system that requires only mouth movements, voice need not be generated. A multi-stream HMM (hidden Markov model) and feature extraction technique are applied to EMG-based speech recognition. 3 channel facial EMG signals are collected from ten subjects when uttering 10 Japanese isolated digits. One channel corresponds to one stream. By examining various features, we found that the delta component of the static parameter leads to higher accuracy. Compared to equal stream weighting, the individual optimization of stream weights increased recognition accuracy by 4.0% which corresponds to a 12.8% reduction in error rate. This result shows that multistream HMM is effective for the classification of EMG.