In this paper, the authors propose a training system for unvoiced vowel recognition. In the unvoiced vowel recognition system, three pairs of electrodes were placed on the facial muscles and Surface Electromyogram (SEMG) signals were recorded. The vowel is recognized by the binary pattern which is calculated from the activity values of the measured SEMG signals using thresholds. However, the recognition accuracy was low for some speakers who were not able to adapt themselves to the stable movement of the facial muscles for unvoiced speech. To address this issue, the activity values measured from multiple speakers were classified using Self-Organizing Map (SOM) algorithm and the classification results determined the standard pattern of the activity values for each vowel. The proposed training system visualizes the similarity between the measured activity pattern and the standard activity pattern by displaying each activity pattern's position on the self-organized map. Experimental results of 16 speakers showed that the learning rate of the proposed training system was superior to our previous system in which only the recognition result was provided by the character and the sound.
