Evaluation of a neural network segmental duration model for Portuguese

This paper presents a segmental duration model, that, as far as the authors know, is the first published for European Portuguese, with objective and subjective evaluations. The model is aimed at TTS applications and is based on an ANN, trained with a resilient back-propagation algorithm. Using a substantial amount of training data and a carefully selected set of input factors, the standard deviation of the error of segmental duration estimations reaches 19 ms and the correlation coefficient goes above 0.9. Several models have been published for other languages with objective and subjective good performances. The methodology of construction of the model, the importance of the used factors and the neural network will be presented, together with the evaluation of the model, allowing a comparison with other models for other languages.

VIVO