Listeners blind-ranked 32 ElevenLabs voices; the pro tier lost

A blinded listening test found that the voices ElevenLabs labels "professional" were not preferred over its instant clones for audiobook narration. The research firm Woodstitch had 32 voices from the ElevenLabs Voice Library read the same 73-word nonfiction passage, then played 125 paid US listeners two unnamed clips at a time and asked which they would rather hear narrate an audiobook. Across 1,122 of those forced choices, a professional voice won 46.1% of the 649 matches against an instant clone, according to the report.

The two tiers differ in how a voice is made, and the study suggests that is all the label tells you. A professional voice comes from a studio recording made under an agreement with the actor; an instant clone is built from audio a user uploads. The report calls the tier a provenance label rather than a quality grade. Popularity did no better: how much text a voice reads for ElevenLabs customers each year showed no relationship to whether listeners preferred it, and two well-known audiobook voices included on purpose, Adam and Bill Oxley, finished 18th and 15th.

One factor that did move results was volume. ElevenLabs ships these voices at levels about 15 decibels apart, the clips were not evened out, and the louder clip of a pair won 53.6% of the time. The study is also narrow by its own account: one short nonfiction passage, one set of rendering settings and 32 of the library's 5,401 English narration and education voices, so it says little about fiction, dialogue or hours-long listening. It was funded by SquireWorks Audiobooks, a producer that narrates books with ElevenLabs voices, run by a firm that sells such studies, and has not been peer reviewed; the per-judgment data and all 32 clips are published for anyone to check.