Interspeech'2005 - Eurospeech
Determining the position of breaks in a sentence is a key task for a text-to-speech (TTS) system. We describe some methods for phrase break prediction in which the whole sentence is considered, in contrast to most previous work which has focused on using local features. Three approaches are described: by analogy, where the breaks from the best-matching sentence in our training data is used for the unseen sentence; by phrase modelling, in which we build stochastic models of phrases to segment unseen sentences; and finally, using features derived from a syntactic parse tree. Our best result, obtained on the MARSEC corpus and using a combination of parse derived features and a local feature, gave an F score of 81.6%, which we believe to be the highest published on this dataset.
Bibliographic reference. Read, Ian / Cox, Stephen (2005): "Stochastic and syntactic techniques for predicting phrase breaks", In INTERSPEECH-2005, 3233-3236.