Abstract
A minimum generation error (MGE) criterion has been proposed for model training in hidden Markov model (HMM)-based speech synthesis to minimize the error between generated and original static parameter sequences of speech. However, dynamic properties of speech parameters are ignored in the generation error definition. In this study, we incorporate these dynamic properties into MGE training by introducing the error component of dynamic features (i.e., delta and delta-delta parameters) into the generation error function. We propose two methods for setting the weight associated with the additional error component. In the fixed weighting approach, this weight is kept constant over the course of speech. In the adaptive weighting approach, it is adjusted according to the degree of dynamicity of speech segments. An objective evaluation shows that the newly derived MGE criterion with the adaptive weighting method results in comparable performance for the static feature and better performance for the delta feature compared with the baseline MGE criterion. Subjective listening tests exhibit a small but statistically significant improvement in the quality of speech synthesized by the proposed technique. The newly derived criterion improves the capability of HMMs in capturing dynamic properties of speech without increasing the computational complexity of the training process compared with the baseline criterion.
