Information Based Intonation Synthesis

Information Based Intonation Synthesis* Scott Prevost ~4 Mark Steedman Computer and Information Science University of Pennsylvania 200 South 33rd Street Philadelphia, PA 19104-6389 (Internet: prevost©linc, cis. upenn, edu steedman©cis, upenn, edu) ABSTRACT This paper presents a model for generating prosodically ap- propriate synthesized responses to database queries using Combinatory Categorial Grammar (CCG - cf. [22]), a for- malism which easily integrates the notions of syntactic con- stituency, prosodic phrasing and information structure. The model determines accent locations within phrases on the ba- sis of contrastive sets derived from the discourse structure and a domain-independent knowledge base. 1. Introduction Previous work in the area of intonation generation in- cludes an early study by Young and Fallside ([26]), and studies by Terken ([24]), Houghton, Isard and Pearson (cf. [11, 12]), Davis and Hirschberg (cf. [4, 10]), and Zacharski et ah ([27]). The present proposal differs from the earlier studies in the accent assignment rules, and in the representation of information structure and its relation to syntax and semantics. In the CCG frame- work, the information units that are delineated by in- tonation are directly represented, complete with seman- tic interpretations. These interpretations are utilized in making accent placement decisions on the basis of con- trastive properties rather than previous-mention heuris- tics: While such heuristics have proven quite effective in the earlier studies, we believe the model-theoretic ap- proach taken here will eventually lead to the develop- ment of similar heuristics for handling a wider range of examples involving contrastive stress. The remainder of the paper discusses the contrastive stress model, describes the implemented system, and presents results demonstrating the system's ability to generate a variety of intonational possibilities for a given * Preliminary versions of some sections in the present paper were published as [17] and [lS]. We are grateful to the audiences at those meetings, to AT&T Bell Laboratories for allowing us access to the TTS speech synthesizer, to Mark Beutnagel, Julia Hirschberg, and Richard Sproat for patient advice on its use, to Abigail Gertner for advice on TraumAID, and to Janet Pierrehumbert for discus- sions on notation. The usual disclaimers apply. The research was supported in part by NSF grant nos. IRI90-18513, IRI90- 16592, IRI91-17110 and CISE IIP-CDA-88-22719, DARPA grant no. N00014-90-J-1863, ARO grant no. DAAL03-89-C0031, and grant no. R01-LM05217 from the National Library of Medicine. sentence depending on the discourse context. 2. Motivation Meaning-to-speech systems differ from text-to-speech systems in the manner in which semantic and pragmatic information is exploited for assigning intonational fea- tures. Text-to-speech systems for unrestricted text are forced to rely on crude syntactic analyses and word clas- sifications in making judgements about the accentability of words in an utterance, often using the strategy of pre- vious mention whereby a word is de-accented if it (or perhaps its root) has previously occurred in some re- stricted segment of the text (cf. [10], [15]). The text can be divided into such meaningful discourse segments on the basis of cue phrases and paragraph boundaries. Meaning-to-speech systems, on the …

Content truncated to 500 words