5th International Conference on Spoken Language Processing
Speech generation systems can benefit from the prediction of abstract prosodic labels from text input. Earlier methods of prosodic label prediction have relied on hand-written rules or on statistical methods such as decision trees. Statistical methods have the advantage of being automatically trainable and are portable to new domains. This research presents a new method for automatically training an abstract prosodic label predictor, transformational rule-based learning. This method is automatically trainable. Results will be presented for pitch accent location and phrase boundary prediction.
Bibliographic reference. Fordyce, Cameron S. / Ostendorf, Mari (1998): "Prosody prediction for speech synthesis using transformational rule-based learning", In ICSLP-1998, paper 0682.