Selection of important input parameters for a text-to-speech synthesis by neural networks

Speech signal synthesis in real time, with an unlimited vocabulary is very complicated for all languages. These synthesizers usually work in the frequency domain and the fundamental frequency contours F₀ must be determined for all phonemes or diphones by conventional equipment based on the linguistic rules. Our effort is to minimize the difference between the synthetic speech of the synthesizer, which is usually more monotonous and natural speech of people. Because of this, a special building block is included into the synthesizer for prosody control. The multilayer artificial neural network (ANN) is used for prosody control in our case. In this part of the synthesizer the fundamental frequency is “a little bit” modified in such a way so that speech can sound as natural as possible. The number of input training parameters for ANN training must be generally kept as small as possible because of the optimal generalization ability of the network. An original method for the determination of the most important input parameters for the training of ANN for prosody control is described in the paper. The method is based on the data mining from the database of the training patterns by the GUHA method