武田 一哉
音素的特徴の動的性質を用いた調音結合の正規化
Abstract:This report proposes an approach for normalizing coarticulatory variations
in connected speech. The proposed scheme expands acoustic realization of
speech into distinctive features and on the basis of dynamic property of
each feature, normalizes them independently. In this report, the
effectiveness of the scheme is discussed focusing on an phonetic feature
“voiced”. First, for the extraction of the“voiced” feature from the speech,
a static measurement is designed by optimizing the parameters of
preliminary measurements based on acoustic-phonetic knowledges.
Second, the extracted feature's dynamics are analyzed based on the
discrimination performance of the static measurement. The experiment
suggests that the dynamics of the measurement can be a good cue for
detecting voice-onset and offset timing. Finally, a measurement which
combines static and dynamic measurements shows 12% better accuracy
than the static measurement. Through these experiments, the effectiveness
of the proposed scheme in normalizing the coarticulatory variation is
confirmed.