Analysis of Efficient Feature for Japanese Vowel Recognition

Recently, speech, especially word recognition using visual information, has attracted significantly interest. In word recognition, the target is not just a word, but also a vowel. However, since vowel frames do not contain many phonemes, vowel recognition rate is less than that of the word. Some research has been done on vowel recognition. This paper analyses features that can effectively recognize five Japanese vowels. The process of our method is as follows. First, the sampled active contour model is applied to detect the lip contour during the input image sequences. The lip size is then normalized. Next, various features, such as the shape, diameter, and the approximate function features, are calculated and compared with the recognition rates using k nearest neighbor method. Two experiments, speaker dependent recognition and speaker independent recognition are carried out with five subjects. We calculated eleven feature sets and found that the feature set including the area and aspect ratio is the most effective. In the speaker dependent recognition, we obtained the recognition rate of 96.6%, and in the speaker independent recognition, we obtained 82.4%