Acoustical Correlates of Prominence: a Design for Research

In this project, all known acoustical correlates for prominence at the sentence level, such as F0, intensity, duration and spectral quality will be investigated. The speech material will be selected from the Dutch Polyphone corpus as recorded by SPEX and KPN Research Leidschendam. The sentences from this corpus are suitable to investigate systematically (with high speaker variability) the acoustical correlates of prominent words. 5000 Dutch speakers each read aloud five phonetically rich sentences, which were recorded over the telephone. The prominent words are detected through a perception experiment, in which naive listeners have to mark emphasized spoken words (this is an operational definition of prominence). Then the acoustical differences between the prominent and non-prominent words will be investigated. The aim of this research project is not only to describe the acoustical correlates but also to develop a predictive model. This model, most probably in the form of an artificial neural network, will have to predict prominence using all relevant acoustical information. Furthermore, some perception experiments with manipulated speech are planned, in order to investigate the effect of F0, intensity, duration and spectral quality on the perception of prominence.