Accent label prediction by time delay neural networks using gating clusters

In this paper a new neural network (NN) architecture for data driven prediction of accent labels—perceptual accents and pitch accents—for speech synthesis is presented. Within the proposed NN architecture, gating clusters are applied in a time delay (TD) framework. Gating clusters enable the dynamic adaptation of a network structure depending on the actual input to the NN. In the proposed TD framework, gating clusters are used to adapt the network structure such that only available input feature vectors from the actual context window are treated. The proposed NN architecture has been successfully applied for accent label prediction within our text-to-speech (TTS) system. Prediction accuracy for our German corpus was determined at 86.1%. On an english corpus the achieved accuracy was 84.5%. This result is superior to results achieved on the same corpus with an approach based on classification and regression tree (CART) techniques [1]. The results were achieved with a simpler feature set than that used in [1].