F0パターンの自動推定と目標点の抽出(Session-4 一般(ポスターセッション),第8回音声言語シンポジウム)

概要

論文の詳細を見る
The fundamental frequency (F0) contours of voices manifest the acoustic correlation. of lexical tones in tonal languages (or pitch accents in non-tonal languages), expressive intonation, and such speaker factors as individual vocal ranges. Tonal target points, focusing particularly on tonal F0 peaks and valleys, are used to sparsely specify the prosodic contributions of lexical tones (or pitch accents) and expressive intonation to the F0 contours. This paper presents an approach to automatic extraction of the underlying tonal target points with optimal approximations of the observed F0 contours using a functional F0 contour model. This consists of (1) finding potential target candidates from the observed F0 contours, (2) suppressing the micro-prosodic effects involved in the F0 contours, and (3) selecting appropriate target points from these candidates. This approach is currently evaluated by investigating the errors between the observed and reproduced F0 contours in the spirit of analysis-by-synthesis, conducted on thousands of Chinese speech samples and a few of Japanese and English speech samples. It shows a straightforward way to studying dynamic intonation variations, thanks to its ability to parametric representation of the F0 contours in term of the same way as specifying the prosodic contribution of tone and intonation to the F0 contours.
社団法人電子情報通信学会の論文
2006-12-14