The variation of prosody with text type

It is widely understood that unnatural prosody is one of the key problems facing text-to-speech synthesis. There is much interest in using recordings of human speech to automatically train prosody prediction models, which in turn can be used to predict highly natural prosody. However, most work in this field to date has ignored the fact that human speakers adopt different speaking styles depending on the type of text being read. For example, weather reports have quite different prosody to children´s stories. Nevertheless most text-to-speech systems are only capable of speaking in one `neutral´ speaking style. To remedy this, we are constructing databases (for three languages-Dutch (Belgium), English (US) and French (France) of human speech, which contain recordings based on a variety of text types. This paper describes the design methodology for these databases and the results of some pilot experiments carried out to verify the validity of the prosodic markup scheme, and to choose three text types with prosodically diverse characteristics