A cloned familiar voice transfers one personality trait, and never the same one twice

What the paper tested, and why

Johanna Magdalena Kuch, Maud van Lier, Daksitha Senel Withanage Don, Patrick Gebhard, Tanja Schneeberger, and Elisabeth André posted arXiv:2610.18669, "Known-Voice Cloning and the Limits of Personality Transfer for Character Design," on 13 October 2026. A related version of this work appears at CHI EA '26 under the title "What Happens When a Virtual Character Speaks With the Cloned Voice of a Person You Know?" (DOI 10.1145/3772363.3799332).

The question behind both studies: zero-shot voice cloning can now reproduce a recognizable voice from a short recording and attach it to any virtual character, robot, or game avatar. Prior work on familiarity and personalization suggested a familiar voice might carry the real person's perceived personality onto the new target. This paper tests that transfer directly, by having the same virtual character, same visuals, same scripted content, speak in two conditions: a known-voice clone of someone the participant actually knows, versus a cloned unknown voice. Participants first rated the real person from memory, then rated the character under each voice condition. The measure is the absolute distance between the character rating and the remembered real-person rating: a smaller distance means more apparent transfer.

Study 1 cloned a university lecturer's voice, known to the participants from their own course, using OpenVoice, read over about 14 seconds of Oldenburg sentences. 18 participants completed the study; one was excluded for not actually knowing the lecturer, leaving n = 17. Personality was measured with the 21-item BFI-K. This was exploratory, with no correction for multiple comparisons.

Study 2 cloned the voices of two social media content creators, using XTTSv2, over about 43 seconds of trivia statements. Of 208 participants who started, 65 passed preregistered exclusion criteria. Personality was measured with the shorter 10-item BFI-10. This study was preregistered with one-tailed significance tests, also uncorrected for multiple comparisons.

The results that were significant, and the ones that weren't

What actually seems to transfer, versus what doesn't

The paper's own interpretation is that a familiar voice mainly triggers identity activation, recognizing "this sounds like someone I know", rather than full personality reconstruction. Likability moved in the exploratory study but not the preregistered one; the trait that moved was agreeableness in one study and extraversion in the other, with no overlap. The authors explicitly describe the effects as "selective and inconsistent across traits and contexts" and state there is little evidence for a reliable, generalizable personality transfer effect.

Two design differences between the studies could explain part of the inconsistency, and the paper names them as limitations rather than as an explanation that resolves the discrepancy: Study 1 used a single, well-known lecturer's voice over the briefest stimulus (about 14 seconds); Study 2 used two less personally familiar content creators' voices over a longer stimulus (about 43 seconds), with a larger, preregistered, one-tailed sample. The paper does not claim the design difference caused the different trait to move, and neither do we; it is the most obvious candidate explanation available from what the paper reports, not a tested one.

The authors also report, as a stated limitation, that "participants sometimes perceived the cloned voices and virtual characters as artificial," which works against any transfer effect regardless of condition: a voice or face flagged as fake is a weaker vehicle for personality impressions than one perceived as authentic.

What this means for dubbing, character voice, and consent

For localization and character-design teams using voice cloning to attach a real performer's or presenter's voice to a dubbed character or virtual avatar, the direct implication is that a cloned familiar voice does not reliably import that person's perceived personality wholesale. A studio hoping that cloning a beloved narrator's voice onto an unrelated animated character will carry over the narrator's warmth, competence, or authority should not assume that happens by default: in this paper's two tests, it happened for one trait each time, and the trait was not the same trait twice.

The consent framing is where this paper is most direct, and it is squarely inside BLOMEGA's own thesis that AI voice and likeness data needs explicit, rights-cleared consent rather than scraped or assumed permission. The authors write: "we discourage the default use of familiar voices in social robotics, as well as their use without explicit consent," and call for "transparent mechanisms to give consent for voice cloning," including the ability to "keep track of, update, and retract this consent." In their own research, voices were cloned "with permission from the respective voice owners or from materials explicitly authorized for the study," which the paper treats as a baseline requirement, not an optional safeguard. The limited and inconsistent personality transfer this paper measures does not reduce the consent obligation; the authors are explicit that the ethical concern (impersonation, manipulation risk, reputational use of a real person's vocal identity) stands independently of whether personality itself transfers.

Check it yourself

• The paper: arXiv:2610.18669, "Known-Voice Cloning and the Limits of Personality Transfer for Character Design," posted 13 October 2026.
• Full text with both studies' statistics: arxiv.org/html/2610.18669v1.
• The related CHI EA '26 record: DOI 10.1145/3772363.3799332.
• BFI-K (21-item) and BFI-10, the two personality instruments used, Study 1 and Study 2 respectively, for anyone replicating the measure.
• Recompute the agreeableness gap yourself: unknown-voice distance 1.74 minus known-voice distance 1.28 = 0.46, on a scale where this is a rank-biserial effect size of r = -.59, described by the authors' own test statistic (W = 16.0, p = .038).

What would prove this wrong

If a third study, using the same within-subject known-vs-unknown-voice design on a different set of familiar voices before the end of 2027, finds the same trait (agreeableness or extraversion) shifting significantly in two separate samples, that would be evidence of a real, specific, replicable transfer mechanism rather than the selective, study-specific effects reported here. Until then, the fair reading is the authors' own: little evidence for a reliable, generalizable transfer effect, not "no transfer at all" and not "a mechanism we understand."