Dept. of Science & Engineering
Oregon Health & Science University
The general superpositional approach to intonation posits that the fundamental frequency contour can be quasi-additively decomposed into component curves such as phrase curves, accent curves, and segmental perturbation curves. Currently, these component curves can only be estimated if one assumes a specific superpositional intonation model, such as the Fujisaki model (Fujisaki 1983). The central objective of the presented research therefore is to develop an algorithm for decomposing any arbitrary fundamental frequency contour into its component curves using only the most general assumptions of the superpositional approach. The decomposition algorithm is applied in two ways in this work: (1) to examine the assumptions about the relationship between the shape of the F[subscript 0] contour and the perception of prominence in the superpositional context, and (2) to enable the implementation of the recombinant synthesis method (van Santen et al. 2004:i, van Santen et al. 2005). Recombinant synthesis, also called multi-level unit sequence synthesis, involves searching one corpus for acoustic units while searching another corpus for components of the F[subscript 0] contour, namely, phrase curves, accent curves, and segmental perturbation curves, that are then additively recombined according to the general superpositional approach, and imposed on the acoustic unit sequences using standard pitch modification methods. In this synthesis method, the components of the F[subscript 0] contour are obtained from natural F[subscript 0] contours, and extracting them requires a decomposition algorithm. Besides the two specific applications that are demonstrated in this thesis, the stated objective also has a wider significance. If one can estimate component curves of the F[subscript 0] contours using only the most general assumptions of the superpositional approach, then this would enable accurate determination of (i) the effects of linguistic control factors on the component curves, (ii) their time courses, and (iii) interactions between curve classes. Barring the discovery of fatal flaws in the superpositional concept along the way, the ultimate result of this research would be the development of a - potentially complex -superpositional intonation model that contains a richer collection of curve classes than current models allow, specifies a clear and phonologically well-justified mapping between each curve class and the F[subscript 0] contour, and generalizes well to fit across a wide range of speech materials.
Div. of Biomedical Computer Science
School of Medicine
Mishra, Taniya, "Decomposition of fundamental frequency contours in the general superpositional intonation model" (2008). Scholar Archive. 334.