The ESCA Workshop on Speech Synthesis
September 25-28, 1990
This paper describes work done on the intonation module and the syntax-intonation interface of the Edinburgh University CSTR Text-to-Speech system, a linguistically sophisticated speech output system. It is clear from evaluation studies of our system's intonation that the single biggest problem with the output of the intonation accent placement rules is their failure to take account of pragmatic deaccenting, either of anaphora or of semantically redundant lexical items. As has been frequently pointed out in work on synthetic intonation, the information which governs such deaccenting (pragmatics, semantics, discourse structure, etc.) is not currently available to automatic systems: it is therefore necessary to develop heuristic approaches which will minimise the occurrence of such errors in the short (and probably medium) term. Three classes of such heuristics are currently being implemented in our system, and these are described: it is expected that further heuristics will be revealed by future investigations.
Some of these heuristics are dependent upon the particular semantic domain (in a system giving railway timetable information, items such as "train" and "platform" might be redundant), and some are domain-independent (deictic modifiers should always have the same effect, as should contrastive stress). It is crucial that as much lexical and syntactic informa- tion as possible is available for their development, as incorrect treatment of deaccenting phenomena will seriously degrade the acceptability of synthetic speech.
Bibliographic reference. Monaghan, Alex I. C. (1990): "Treating anaphora in the CSTR text-to-speech system.", In SSW1-1990, 113-116.