摘要 |
PURPOSE:To prevent a decrease in consonant rough classification rate in noise environment by using two specific pieces of similarity information together. CONSTITUTION:When a consonant group standard pattern is generated, the values of a low-band power dip and a high-band power dip are extracted by following vowels by using many consonant data segmented from words to find a distribution of power dips, which is used to generate a standard pattern consisting of a mean value, variance and covariance. Then, the similarity obtained by a consonant similarity calculation part 12 and the similarity obtained by a consonant similarity calculation part 11 are used to decide a consonant group by a rough classification decision part B. Thus, the similarity as to a consonant group by following vowels obtained from the values of the low-band power and high-band power and the similarity obtained by the matching with the consonant standard pattern are used together to prevent the decrease in the rough classification rate due to noises. |