Three Steps of Neuron Network Classification for EMG-based Thai Tones Speech Recognition

摘要

In order to overcome the problem existing in original speech recognition (e.g. noise interruption and private data loss), many researchers have investigated to deal with these problems. Electromyography (EMG) from the muscles producing speech was used to replace a voiced signal. Similarly, we aim to develop EMG speech recognition based on Thai language. Tone is the important characteristic of this language. Hence, Thai tone classification is the first work that was explored. This paper proposes the new technique that can classify five Thai tones for EMG-based Thai speech recognition. This method can overcome the limitation of our previous work that we can classify only two tones. EMG was captured from six positions of the strap muscles and facial muscles while a volunteer was uttering 21 Thai isolated words and five tones of each word (total 105 words). The 68 EMG features were calculated, and RES index was used to evaluate clustering capability of each feature. Top five features that have high value of RES index were selected. Neuron Network (NN) was used for tone classification. We found that Modify Mean Absolute Value 2~(nd) type (MMAV2) is the best features. It yielded an accuracy rate of 56.2% for five Thai tones classification. However, it is not enough for our work. In order to improve the accuracy rate, the three steps of NN Classification was proposed. This technique is the series of three networks of NN classifier. Each network will classify different tones, and use distinct features. We obtained an accuracy rate of 80% for five Thai tones classification from this technique.