Key points are not available for this paper at this time.
In the past, text-to-speech systems, while effective in functional contexts, were not considered viable substitutes for professional human narrators in audiobook production due to their limited expressive variability. Literary audiobooks, in particular, rely on systematic modulation of pitch, tone, rhythm, and tempo to convey narrative atmosphere, emotional nuance, and character differentiation. Recent advances in artificial intelligence (AI), however, have enabled synthetic voices to approximate these expressive qualities and to perform differentiated narrative roles. As AI-generated voices are increasingly being deployed in commercial audiobook productions, it remains an open empirical question whether they meet listeners' expectations for literary narration. This study reports the results of an experiment evaluation in which self-identified audiobook listeners assessed literary audiobook excerpts performed either by AI-generated voices or by professional human voice actors. Overall, differences in listener evaluations between AI-generated and human voices were minute. Expectations regarding whether a narrator was human or artificial did not moderate these evaluations. Perceived narrator gender played only a marginal role, while the presence or absence of direct speech did not interact with voice type. Listener age and gender showed no systematic effects. Instead, affinity for technology interaction emerged as the strongest (but still weak) predictor of differential evaluations between AI-generated and human voices. Taken together, the modest size of all observed effects suggests that from the listener's perspective, AI-generated voices currently function as a viable substitute for human audiobook narration.
Schwabe et al. (Fri,) studied this question.