Consonant confusions and perceptual spaces for natural and synthetic speech

Abstract

Earlier research in our laboratory has demonstrated that synthetic speech is less intelligible and more capacity demanding than natural speech. These differences appear to be related to processes responsible for encoding the input signal into a segmental phonemic representation. There are several hypotheses that could account for the greater difficulty involved in synthetic speech perception. One hypothesis is that synthetic speech is structurally equivalent to natural speech degraded by noise. An alternative hypothesis is that the acoustic‐phonetic structure of synthetic speech is impoverished in comparison to natural speech in that a minimal set of acoustic cues are used to implement phonetic segments. The two hypotheses lead to different predictions about the nature of synthetic consonant confusions in relation to confusions of natural speech degraded by noise. To resolve this issue, we carried out multidimensional scaling analyses of confusion matrices for synthetic consonants produced by DEC‐Talk, Prodse‐2020, and the Votrax Type‐N‐Talk, and for natural speech at several S/N ratios. The results of the analyses tend to support the second hypothesis: The properties of the perceptual spaces obtained for the synthetic consonants differed considerably from those obtained for the natural consonants. [Work supported by AFOSR.]