Some relationships between preference and intelligibility among synthetic voices

The relationship between preference and intelligibility among synthetic voices was examined. Preferences among the speech produced by three text-to-speech systems were assessed using an A/B paired comparison procedure, and the intelligibility of the synthetic voices was measured using a transcription task. The text-to-speech systems tested were DEC talk, Prose 2000, and Infovox. Phoneme specific sentences served as stimuli and enabled a preliminary examination of the relationship between intelligibility and preference as a function of specific phoneme categories. Results indicated a close relationship between mean intelligibility and preference for a voice. In some cases, preference was found to vary with the intelligibility of individual classes of phonemes, accounting for some of the variability associated with preference decisions. These results will be discussed in terms of previous research involving preference and intelligibility. Work supported by NIH and USAF.