Segmental duration control with asymmetric causal retro-causal neural networks

The generation of pleasant prosody parameters is very important for speech synthesis. A Prosody generation unit can be seen as a dynamical system. In this paper sophisticated time-delay recurrent neural network (NN) topologies are presented which can be used for the modeling of dynamical systems. Within the prosody prediction task left and right context information is known to influence the prediction of prosody control parameters. This can be modeled by causal-retro-causal information flows [1]. Since information being available during training is partially unavailable during application, there is a structural switching from training to application. This structural change of the information flow is handled by two asymmetric architectures. These proposed new architectures allow the integration of further a priori knowledge. By this we are able to improve the performance of our duration control unit within our text-to-speech (TTS) system Papageno.