Alex I. C. Monaghan

Symbolic output as the basis for evaluating intonation in text-to-speech systems

Abstract Many text-to-speech (TTS) systems under development in Europe and elsewhere — we discuss in particular the system under development at Edinburgh Universitys Centre for Speech Technology Research (CSTR) — generate intonational properties of synthetic utterances on the basis of an intermediate abstract phonological representation of prosodic features that is quite independent of any acoustic realisation. For evaluating certain aspects of synthetic prosody (notably accent placement and division into domains), this abstract representation is a more appropriate object of evaluation than the final acoustic output of a system, just as word stress and grapheme-to-phoneme conversion are appropriately evaluated in terms of symbol strings rather than acoustic output. By way of illustration we present the results of an evaluation exercise carried out on the sentence-accent assignment rules of the CSTR system, based on just such an abstract representation, which has been useful in improving our rules.

Evaluation of speech synthesis techniques in a comprehension task

Abstract Six types of speech synthesis were evaluated for comprehensibility: standard linear predictive coding analysis/ resynthesis; pitch synchronous analysis/resynthesis; pitch synchronous multi-pulse analysis/resynthesis; and three PSOLA (pitch synchronous overlap-and-add) techniques. The relative comprehensibility of the synthesis types was tested by using the synthesised speech to convey information that subjects needed in order to perform a diagram-based multiple-choice task.

The intonation of textual anomalies in text-to-speech

Abstract This paper examines the intonational characteristics of a number of types of non-word, e.g. numbers, dates, times and other abbreviations, which occur in text and are readily identifiable. These phenomena are often treated as problems for text-to-speech systems, but we take the view that the ease with which they can be identified and the relative predictability of their semantic or pragmatic content makes their intonational behaviour more regular and therefore easier to model than the behaviour of other areas of text. Examples of such phenomena are presented, and heuristics for their treatment by an automatic system are proposed. A formal evaluation of these heuristics is presented, showing a success rate of over 94%. A final discussion outlines the advantages and disadvantages of such a treatment, and suggests lines of future research.