Prosody generation with a neural network

Abstract

The use of neural networks in speech synthesis has been especially successful in the domain of prosody generation. The approach presented here differs from others in a) the transformation from a simple input to an output vector consisting of different parameters and b) the use of subcorpora that allow specialized networks. The network operates in a prominence-based synthesis system, where prominence is the most important parameter and is, consequently, the input parameter for the network. The output is not yet evaluated formally but the synthetic speech sounds natural and lively.