Artificial neural network approach to the modelling of prosody in the speech synthesizer of the Czech language

In this contribution we try to describe the optimal choice of phonetic and phonologic parameters, which are necessary for prosody modelling. The rule-based approach [5] or the artificial neural networks (ANN) can be used for prosody control. According to our experience ANNs are able to achieve better results. If the prosody of the speech synthesizer is controlled by an artificial neural network (ANN), an optimisation of the ANN topology is one of the most important problems. The application of a supervised neural network has been used for prosody parameters determination in the process of prosody modelling. The pruning of neural networks based on the GUHA method [10] or the utilization of the synaptic weights sensitivities can be suitable tools for the minimization of the number of input parameters, and for the reduction of the neural network structure redundancy. The automatic system, designed for the preprocessing of written text, training the ANN by the speech of suitable speaker and prosody modelling are the main goals of our research. The ANN dedicated for prosody control is able to model prosodic parameters in a quality, which may be comparable with natural speech. The specific attributes of national languages must be taken into account. From this point of view the Czech, similarly as the other Slavonic languages, is more difficult than English or German.