Selection of Suitable Features for Modeling the Durations of Syllables

Abstract

Acoustic analysis and synthesis experiments have shown that duration and intonation patterns are the two most important prosodic features responsible for the quality of synthesized speech. In this paper a set of features are proposed which will influence the duration patterns of the sequence of the sound units. These features are derived from the results of the duration analysis. Duration analysis provides a rough estimate of features, which affect the duration patterns of the sequence of the sound units. But, the prediction of durations from these features using either linear models or with a fixed rulebase is not accurate. From the analysis it is observed that there exists a gross trend in durations of syllables with respect to syllable position in the phrase, syllable position in the word, word position in the phrase, syllable identity and the context of the syllable (preceding and the following syllables). These features can be further used to predict the durations of the syllables more accurately by exploring various nonlinear models. For analying the durations of sound units, broadcast news data in Telugu is used as the speech corpus. The prediction accuracy of the duration models developed using rulebases and neural networks is evaluated using the objective measures such as percentage of syllables predicted within the specified deviation, average prediction error (µ), standard deviation (σ) and correlation coefficient (γ).