Feature Selection for Improved Phone Duration Modeling of Greek Emotional Speech

In the present work we address the problem of phone duration modeling for the needs of emotional speech synthesis. Specifically, relying on ten well known machine learning techniques, we investigate the practical usefulness of two feature selection techniques, namely the Relief and the Correlation-based Feature Selection (CFS) algorithms, for improving the accuracy of phone duration modeling. The feature selection is performed over a large set of phonetic, morphologic and syntactic features. In the experiments, we employed phone duration models, based on decision trees, linear regression, lazy-learning algorithms and meta-learning algorithms, trained on a Modern Greek speech database of emotional speech, which consists of five categories of emotional speech: anger, fear, joy, neutral, sadness. The experimental results demonstrated that feature selection significantly improves the accuracy of phone duration modeling regardless of the type of machine learning algorithm used for phone duration modeling.