Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997

        

Prosodic Phrasing Without Syntax via Part-Of-Speech Amalgamation

Arman Maghbouleh

Stanford University, Department of Linguistics, Stanford, California, USA Apple Computer, Inc., Cupertino, California, USA

This paper describes a set of regression models for predicting prosodic phrasing, ToBI break indices, from text. Two notable features of the models are the use of a variable-length window of information, and the use of easily extractable features, such as automatically determined part-of-speech tags. The models are especially useful for text-to-speech systems which often do not have access to computationally expensive information such as syntactic parses or derived information structure. This paper reports results of training and testing on different portions of the WBUR Radio News Corpus. Results include 87% correct prediction of the presence or absence of a phrase break, which corresponds to 78% correct detection and 8% false detection of the presence of a phrase break.

Full Paper

Bibliographic reference.  Maghbouleh, Arman (1997): "Prosodic phrasing without syntax via part-of-speech amalgamation", In INT-1997, 215-218.