Piet Mertens

Automatic pitch contour stylization using a model of tonal perception

Abstract A new quantitative model of tonal perception for continuous speech is described. The paper illustrates its ability for automatic stylization of pitch contours, with applications to prosodic analysis and speech synthesis in mind, and evaluates it in a perception experiment. After a discussion of the psycho-acoustics of tonal perception, and an overview of existing tonal perception models and systems for automatic analysis of intonation, the model and its computer implementation are described in detail. It includes parameter extraction, segmentation into syllables, perceptual integration of short term pitch change, tonal segment computation, and pitch contour stylization. This is followed by a perception experiment in which subjects are asked to distinguish original signals from resynthesized signals with automatically stylized pitch contours. The aim of this experiment is to show the usefulness of the model as a basis for intonation representation, and to study the influence of the model parameters. It is shown that the stylization obtained with the model is an economic representation of intonation which can be useful for speech synthesis and prosodic analysis.

Parsing dependency grammar using ALE

This paper describes a technique for parsing dependency grammars using a bottom-up chart parser originally designed for phrase-structure grammars, using typed feature structures as the only data structure. Each lexical item is represented as a tree where nodes indicate lexical elements (the anchor, its dependents and governor) and edges (branches) indicate dependency relations between these elements. Nodes may carry additional features, including one for node saturation. Trees combine into derived trees provided that node and edge features unify. The ALE system is used to implement an active chart parser where a chart edge represents a tree, and two adjacent edges are combined into a more saturated tree.