Phonemes recognition of Japanese vowels on a power spectrum using neural network with backpropagation
A. Tsubota, Akira Taguchi, Hideo Mitsui, M. Sone · 2002
Summary form only given, as follows. An attempt was made to recognize the phonemes 'a, i, u, e, o' (known not only as Japanese vowels, but also as international vowels) by means of a neural network (NN) consisting of three layers with backpropagation. The normal power spectrum which is analyzed by the FFT is used as input data. The spectrum of high frequency elements is very small compared with that of low frequency elements. In the low frequency region, the characters of some vowels are similar. Thus, the average of the recognition rates is not so good. In the present work, an attempt was made to improve the recognition rates. For example, the authors used input data that weighted only high frequency elements of the power spectrum, in order to evaluate the character of the high frequency as well as that of the low frequency region. As a result, the average of recognition rate was improved from 84% to 89%. The rate of recognition of 'e' was improved from 56% to 86%.>