A system for left-to-right intonation specification from text

This paper describes some computational strategies employed in the implementation of the model outlined in papers by Ladd and Ladd & Monaghan in this volume. There are two main categories of problern involved in extracting intonation from text: the first derives from the (currently) very limited nature of higher-level information deducible from text; the secend consists of prpblems of interpretation which would exist even if perfect high-level analyses of text were available. Both these categories can be resolved with reasonable success in our Left-to-Right process model by using various levels of representation and employing computational techniques such as default spefification and recursion. Factcrs affecting intonation at a high level include semantic, syntactic and pragmatic considerations, most of which are not explicit in text. Our model uses a small number of abstract PITCH ACCENT types in conjunction with limited syntactic and pragmatic information and a number of default clauses to generate a wide range of intonation contours. 1 LEVELS OF REPRESENTATION Any adequate model requires explicit representation of all relevant levels, and considerable thought was given to determining precisely which levels were relevant to intonation. The desire to avoid speakerspecific representations as much as possible, and the adoption of intonational tunes (see Ladd, this volume), led to the choice of three descriptive levels: abstract phonological, abstract phonetic and concrete phonetic. 1.1 Abstract Phonological Representation. This level defines a fairly abstract intonational contour by specifying accent location and degree (major or minor). Also represented at this stage are register steps and phrasal boundaries (see Ladd, ibid.). The current system specifies default locations for all these elements automatically, and other items (such as non-default boundaries or extra register steps) can be entered by hand. 1.2 Abstract Phonetic Representation. The phonological representation is mapped onto this level using the tune chosen for the utterance: each accent degree phoneme is specified as the appropriate accent type from the tune, to give a number of abstract targets generally three for major and two for minor accents. Boundaries are treated in a similar manner, but register steps are not interpreted at this level. *Centre for Speech Technology Research, University of Edinburgh.