Voiced-Unvoiced Classifications of Japanese speech are discussed. Voiced-unvoiced classification of short frames using the criteiria of weighted distance is proposed by Atal [1]. We adopted these methods to the Japanese speech, and reached the conclusion that the bias terms of determinants of the weights are significant for Japanese speech. We propose several mothods with regard to this. We tried several modifications of the Atal's method, and got considerably better results than direct application to Japanese speech data. In these methods it is assumed that the distributions of these parameters have normal distributions, where the Bays-like likelihood ratio tests are efficient. However, when the distributions of the parameters in the unvoiced class are checked precisely, these distributions apparently have two peaks. This breaks the assumption for the Bays-like likelihood ratio tests. The relation between the methods of classifiction and the fact that unvoiced parameters are distributed in two peaks is discussed.