TEMPO-NORMALIZED MEASUREMENT AND TEST SET DEPENDENCY IN OBJECTIVE EVALUATION OF ENGLISH LEARNERS' TIMING CHARACTERISTICS

Shizuka Nakamura1, Hajime Tsubaki1, Yusuke Kondo2, Michiko Nakano2 & Yoshinori Sagisaka1
1GITI, Language and Speech Science Res. Labs, Waseda University, Japan; 2School of Education, Language and Speech Science Res. Labs, Waseda University, Japan

ID 1644
[full paper]

In this paper, we present experimental results on tempo-normalized measurements and sentence sets for objective evaluation of English speech timing by Japanese learners. Phone-independent versus phone-dependent tempo normalizations were compared using raw duration differences between learners and native speakers. To observe the effect of test sentence differences, sentence length was adopted as a criterion. Through experiments, high correlations between subjective judgment and duration differences with normalization showed remarkable advantage of phone-dependent normalization. Large correlation differences between long sentences and short sentences indicated the need of careful choice of test materials. The subjective score estimation by linear regression showed better performance using long sentences and duration differences with phone-dependent normalization than conventional one using all sentences and duration differences without normalization.