Speech prosody control using weighted neural network ensembles

Ensembles of artificial neural networks (ANNs) show improved generalization capabilities that outperform those of single networks. However, for aggregation to be effective, the individual networks must be as accurate and diverse as possible. This paper presents a new statistical model for prosody control that combines weighted ensembles of ANNs with feature relevance determination. This approach allows the individual networks to be accurate and diverse. The weighted neural network ensemble model was applied for both, phone duration modeling and fundamental frequency modeling. A comparison with state-of-the-art prosody models based on classification and regression trees (CART), multivariate adaptive regression splines (MARS), or ANN, shows a 12% improvement compared to the best duration model and a 24% improvement compared to the best F₀ model. The neural network ensemble model also outperforms another, recently presented ensemble model based on gradient tree boosting.