Modelling the temporal structure of newsreaders' speech on neural networks for Estonian text-to-speech synthesis

Generation of natural-sounding synthetic speech from a text requires perfect control over the temporal structure of speech flow. The present paper describes an attempt to replace the rule-based durational model, hitherto used in Estonian text-tospeech synthesis, by neural networks (NN). For this aim, fluent speech of radio announcers and newsreaders was analysed and its temporal structure was modelled on neural networks. Analysis of pauses in extended material revealed that if a text is read out with a normal speech rate, it is quite possible to classify the pauses made, so that the results can be used in speech synthesis. For sound durations, certain characteristics of phone context as well as certain syllablelevel features were found to be the relevant input for an NN algorithm. For models of pause durations and positions, however, the prevalent features were variables characterizing text structure (punctuation marks and conjunctions).