Prosody phrase boundary prediction with ensemble learning

It´s necessary for TTS systems to predict the boundaries of natural prosodic phrases for synthetic speech. However, it´s lengthy and costly to annotate enough prosodic phrases manually for the model training. In this paper, we propose a practical method to extract and predict prosodic phrases with raw speech corpora even without manual prosodic phrase annotations. We extract prosodic phrase automatically from four Mandarin speech corpora that are different in narrators or recording text scripts. To alleviate the effects of such differences, we apply two ensemble learning algorithms, AdaBoost and MultiBoosting with C4.5 as the base classifiers, to predict the phrase boundaries derived from different corpora. Experiments show the two algorithms are efficient in reducing classification error, especially when training data is combined with multiple speech corpora. The best performance is achieved by MultiBoosting, in which the F-measure increases from 44.1% to 54.1%, compared with C4.5.