On Customizing Prosody in Speech Synthesis: Names and Addresses as a Case in Point

This work assesses the contribution of domain-specific prosodic modelling to synthetic speech quality in a name-and-address information service. A prosodic processor analyzes the textual structure of labelled input strings, and inserts markers which specify the intended prosody for the DECtalk text-to-speech synthesizer. These markers impose discourse-level prosodic organization, annotate the information structure, and adapt the speaking rate to listeners in real time. In a quantitative comparison of this domain-specific modelling with the default rules in DECtalk, the domain-specific prosody was found to reduce the transcription error rate from 14.6% to 6.4%, reduce the number of repeats requested by listeners from 2.6 to 1.1, and to sound significantly easier to understand and more natural. This result demonstrates the importance of prosodic modelling in synthesis, and implies an even more important role for prosody in more complicated domains and discourse structures. 2. I N T R O D U C T I O N Text-to-speech synthesis could profitably be used to automate or create many information services, if only it were of better quality. Unfortunately it remains too unnatural and machine-like for all but the simplest and shortest texts. It has been described as sounding monotonous, boring, mechanical, harsh, disdainful, peremptory, fuzzy, muffled, choppy, and unclear. Synthesized isolated words are relatively easy to recognize, but when these are strung together into longer passages of connected speech (phrases or sentences) then it is much more difficult to follow the meaning: the task is unpleasant and the effort is fatiguing [1]. This less-than-ideal quality seems paradoxical, because published evaluations of synthetic speech yield intelligibility scores that are very close to natural speech. For example, Greene, Logan and Pisoni [2] found the best synthetic speech could be transcribed with 96% accuracy; the several studies that have used human speech tokens typically report intelligibility scores of 96% to 99% for natural speech. (For a review see [1]). However, segmental intelligibility does not always predict comprehension. A series of experiments [3] compared two high-end commercially-available text-to-speech systems on application-like material such as news items, medical benefits information, and names and addresses. The result was that the one with the significantly higher segmental intelligibility had the lower comprehension scores. Although there may be several possible reasons for segmental intelligibility failing to predict comprehension, the current work focuses on the single most likely cause: synthesis of prosody. Prosody is the organization imposed onto a string of words when they are uttered as connected speech. It includes pitch, duration, pauses, tempo, rhythm, and every known aspect of articulation. When the prosody is incorrect then at best the speech will be difficult or impossible to understand [4], at worst listeners will be misunderstand it with being aware that they have done so. Arguments for the importance of prosody in language abound in the literature. However, the cited examples of prosodic resolution of ambiguity usually are either anecdotal citations or are illustrated by small sets of carefullyconstructed cited sentences. It is not clear how important prosody is in more normal everyday texts. This brings us to the first question addressed in the current study: how much will prosody contribute to perception of synthetic speech for non-contrived, real-world textual material? 2.1. C u r r e n t A p p r o a c h e s to P r o s o d y in S p e e c h