Robust generation of symbolic prosody by a neural classifier based on autoassociators

In this paper a highly robust method to predict symbolic prosody labels for speech synthesis is proposed. This method is based on a two stage approach. In the first stage the characteristics of each symbolic prosody label are captured by autoassociative models, which are trained independently. In the second stage detailed error information obtained from the different autoassociative models is used to train a neural classifier yielding class conditional probabilities. The method has been successfully applied for German and English language. For the latter the exact same data bases as used by Black and Taylor (1997) and Ostendorf and Veilleux (1994) were used to test the method. The results obtained are superior to those reported for the HMM-based and CART-based approaches. Further experiments also demonstrate considerable generalizing ability, yielding high robustness in sparse training material conditions.