任意話者間の音声特徴パラメータの変換

概要

論文の詳細を見る
Speech contains individual, emotional and linguistic information. We can recognize this information unconsciously, but speech recognition system by computer can't obtain this information. The differences in individuals have been a problem for many years in the field of speech recognition. In this paper, the PARCOR coefficients of each speaker's vowel are analyzed by principal component analysis, and the overall characteristic values and eigenvectors are obtained. Suppose these eigenvectors indicate characteristics of a vowel, and this difference indicates the individuality between speakers. This assumption means that eigenvectors of speaker A correspond respectively to those of speaker B according to descending order of eigenvalues. Based on this assumption, the overall characteristic values of speaker A are mapped to those of speaker B. Effectiveness of this method is shown by experiments.
東海大学の論文
2001-03-30