Which ElevenLabs voice is best for audiobooks?
Thirty-two voices from the ElevenLabs Voice Library, chosen to span gender, age, accent, recording tier and popularity, read the same 30-second nonfiction passage. 125 paid listeners heard them in pairs and picked the one they would rather listen to for an audiobook. Listen before you read.
Listen first
Which voice would you rather listen to for an audiobook?
Two of the 32 voices read the same passage. Hear both, choose one, then see which voices they were, where each finished in the study, and how the listeners voted when the two met. The clips are the study's own stimuli, exactly as the listeners heard them: ElevenLabs' direct output, nothing normalised.
All 32 voices, ranked, with their clips
Voice names are the creators' own Voice Library titles. Each voice met eight others in the study (§4.4), so most random pairs here were never judged directly; for those the reveal shows the two ranks only. One pair in the study is about nine judgments, so a 7–2 is a hint, not a verdict; the ranking in Table 1 pools all 1,122.
Question Among voices in the ElevenLabs Voice Library, which do ordinary listeners prefer for an audiobook, and which properties of a voice (gender, age, accent, recording tier, popularity, level) predict that preference?
Study 32 English voices were drawn from the 5,401 library voices tagged for narration or education, stratified so that gender, age, accent, category (professional studio recording vs instant clone) and usage are each spread rather than left at the library's skew. Every voice read one 73-word first-person nonfiction passage with identical eleven_v4 settings. Each voice met eight others in a connected design of 128 pairs; 125 paid US listeners made 1,122 forced-choice judgments, no listener hearing any voice more than twice.
Findings Marie - Audiobook Narrator and Measured finished first, winning 66.2% of its pairings (Wilson 95% CI 55 – 76) and ranking first in 36% of listener-resampled replicates. Verity (64%) and Nigel (61%) follow; the top eight cannot be ordered among themselves. Alder - Warm, Raspy and Comforting was last at 24%, below even in every replicate. Across properties: studio-recorded professional voices were not preferred to instant clones (professional won 46% of 649 mixed pairs, CI 42 – 50); library popularity did not predict preference; voices tagged "old" were disfavoured; gender made no difference; and the louder clip of a pair won 53.6% of the time (CI 51 – 56), with level explaining a measurable part of the voice-level order.
Scope One neutral nonfiction passage of about 30 seconds, one render setting, lay listeners hearing each voice at most twice. The result is listener preference between short readings in a fixed context, not long-form narration, fiction, or dialogue.
Key findings
- Marie - Audiobook Narrator and Measured is the best-supported first place: 47 wins of 71 (66.2%), first in 36% and top-five in 81% of 3,000 replicates. Verity and Nigel are next; the top eight overlap and are not ordered.
- The bottom four are clearly below even: Clarice (39%), Shelton (36%), Angela (30%), Alder (24%). Alder lost 53 of 70 pairings.
- Professional voices did not beat instant clones. ElevenLabs' "professional" tier is a studio recording with an actor agreement; "instant" clones are user uploads. In the 649 pairs that crossed tiers the professional voice won 46.1% (CI 42 – 50); the model coefficient is -0.05 log-odds (CI -0.21 to +0.12). The tier is a provenance label, not a quality grade.
- Popularity predicts nothing. Usage ranged from under a million to nearly a billion characters a year; the usage coefficient is +0.005 per tenfold (CI -0.05 to +0.07). Pairs of a very-high-usage voice against a low-usage one split 53–50. The two famous audiobook voices in the set, Adam and Bill Oxley, finished 18th and 15th.
- Level matters. ElevenLabs ships these voices at levels 15 dB apart (-32.6 to -17.3 dB RMS). The louder clip won 53.6% of pairs (CI 51 – 56), +0.039 log-odds per dB (CI +0.010 to +0.069); the quietest voice, Alder, finished last, though the next quietest, Piper Monroe, finished 16th. Level and strength correlate at r = 0.34. Nothing was normalised, by design (§4.3): the level is part of what you get when you choose the voice.
- Voices tagged "old" were disfavoured (-0.19 log-odds against middle-aged, CI -0.42 to +0.01; mean win rate 46% against 53% for "young"). Gender made no difference (49% female wins in mixed pairs, CI 45 – 53), and neither did side of presentation (50.0% left).
- Accent effects are small and uncertain at two voices per non-American accent; the American coefficient is +0.14 (CI -0.11 to +0.40). Both Irish voices finished in the bottom five, which is two voices, not an accent result.
1What we tested
The ElevenLabs Voice Library holds more than 21,000 shared voices, 5,401 of them English voices tagged for narration or education. It offers no quality signal: no rating, no score, no listener data. What it does offer are creator-supplied tags (gender, age, accent, a descriptive word, a use case), a "professional" badge for studio recordings made under an actor agreement, and usage counts. Anyone choosing a narrator from the library is choosing on those, or on a preview.
This study asks two things of that catalogue. First, which of a broad sample of its voices do listeners actually prefer for an audiobook. Second, and the reason the sample is stratified rather than random, whether any of the catalogue's own signals predict that preference: does the professional badge mean a better voice, does popularity, does the age or gender tag, does accent. A random draw would reproduce the library's skew (69% male, 63% middle-aged, 47% American) and leave the smaller cells too thin to say anything; the 32 voices were instead drawn to spread every property (§4.1).
The passage is one 73-word first-person nonfiction paragraph (§4.2): plain prose with no humour, grief or jargon, so that no voice is rewarded for acting. Every voice read it with identical settings. The design, the passage and the analysis were written down before any judgment was collected.
2Results
2.1 Ranking
Table 1 ranks the 32 voices by win rate over all 1,122 judgments; Bradley-Terry strength, which adjusts for the opponents each voice happened to face, gives nearly the same order and is shown alongside. Marie - Audiobook Narrator and Measured won 66.2% of its pairings and is first in 36% of replicates; no other voice is first in more than 22%. The eight voices at the top (Marie, Verity, Nigel, Kendis, Michael Moody, Heather, Cody, Lydia) have overlapping intervals and the study cannot order them. The middle of the table is a crowd near even. The bottom four separate cleanly: each is below even in at least 95% of replicates.
Side of presentation did not matter: the left-hand clip was chosen on 50.0% of judgments (CI 47 – 53). Each voice was on the left in three to five of its eight pairs by design (§4.4).
Figure 1 (top of page) shows every voice's win rate with its interval; each voice appeared in about 70 pairings (eight pairs, about nine judgments each).
Table 1. All 32 voices ranked by win rate. Tags are the creators' Voice Library tags. "Professional" is ElevenLabs' studio-recorded tier; "Instant" is an instant clone. Strength is Bradley-Terry log-strength with a listener-resampled interval (3,000 replicates); P(top 5) and P(bottom 5) are the share of replicates in which the voice's win rate finishes in those places. Level is the RMS of the shipped clip in dB FS.
| # | Voice | Tags | Tier | W–L | Win rate | 95% CI | Strength | 95% CI | P(top 5) | P(bottom 5) | dB |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Marie - Audiobook Narrator and Measured | female · middle-aged · American | Professional | 47–24 | 66.2% | 55 – 76 | +0.84 | 35 – 142 | 81% | 0% | -22.0 |
| 2 | Verity - Calm & Soft | female · young · British | Professional | 45–25 | 64.3% | 53 – 74 | +0.57 | 7 – 115 | 73% | 0% | -22.7 |
| 3 | Nigel - Classic, Polished and Smooth | male · old · British | Instant | 43–27 | 61.4% | 50 – 72 | +0.36 | -11 – 87 | 53% | 0% | -20.5 |
| 4 | Kendis - Cool and Calculated | male · middle-aged · African American | Instant | 43–28 | 60.6% | 49 – 71 | +0.57 | 5 – 116 | 48% | 0% | -23.8 |
| 5 | Michael Moody - Grandfatherly warm and deep | male · old · American | Professional | 42–28 | 60.0% | 48 – 71 | +0.39 | -12 – 94 | 42% | 0% | -19.6 |
| 6 | Heather - Warm, Raspy American Female | female · middle-aged · American | Professional | 43–29 | 59.7% | 48 – 70 | +0.44 | -2 – 95 | 43% | 0% | -23.9 |
| 7 | Cody - Energetic and Upbeat Educator | male · young · Canadian | Instant | 39–30 | 56.5% | 45 – 68 | +0.08 | -43 – 59 | 25% | 1% | -21.0 |
| 8 | Lydia - Engaging and Enticing | female · young · American | Instant | 40–31 | 56.3% | 45 – 67 | +0.42 | -10 – 95 | 22% | 0% | -25.2 |
| 9 | Brian | male · young · US Southern | Professional | 39–31 | 55.7% | 44 – 67 | +0.18 | -34 – 71 | 19% | 1% | -21.7 |
| 10 | Adam - Classic Scottish Storyteller | male · old · Scottish | Instant | 40–32 | 55.6% | 44 – 66 | +0.18 | -26 – 63 | 13% | 0% | -21.9 |
| 11 | Michael C. Vincent - Suspenseful Storyteller | male · middle-aged · American | Instant | 38–31 | 55.1% | 43 – 66 | +0.45 | -4 – 101 | 20% | 1% | -20.4 |
| 12 | Valf - Gentle Meditation & Mindfulness | male · young · American | Instant | 37–32 | 53.6% | 42 – 65 | +0.11 | -38 – 60 | 10% | 1% | -24.2 |
| 13 | Mira Gold - Dystopian and Methodical | female · young · American | Instant | 38–33 | 53.5% | 42 – 65 | +0.18 | -29 – 70 | 11% | 1% | -23.1 |
| 14 | Ethan - Warm & Trusting | male · young · British | Professional | 36–33 | 52.2% | 41 – 64 | +0.01 | -56 – 55 | 9% | 2% | -19.2 |
| 15 | Bill Oxley - Mature American Male | male · middle-aged · American | Instant | 37–34 | 52.1% | 41 – 63 | +0.15 | -29 – 63 | 7% | 1% | -23.5 |
| 16 | Piper Monroe - Engaging Podcast Host | female · young · Scottish | Professional | 35–35 | 50.0% | 39 – 61 | +0.16 | -29 – 61 | 3% | 2% | -32.6 |
| 17 | Maribeth - A Southern Sweetheart | female · middle-aged · US Southern | Instant | 36–36 | 50.0% | 39 – 61 | -0.03 | -59 – 49 | 4% | 5% | -17.3 |
| 18 | Adam - Steady, Rich and Round | male · old · British | Professional | 35–36 | 49.3% | 38 – 61 | +0.03 | -52 – 59 | 5% | 7% | -22.0 |
| 19 | Alistair - Clear, Neutral and Informative | male · middle-aged · British | Instant | 35–36 | 49.3% | 38 – 61 | -0.27 | -75 – 19 | 2% | 4% | -19.7 |
| 20 | Christy | female · young · American | Instant | 34–35 | 49.3% | 38 – 61 | -0.08 | -63 – 46 | 3% | 7% | -19.8 |
| 21 | Kim T. - Calm Guidance | female · old · African American | Professional | 33–36 | 47.8% | 36 – 59 | -0.04 | -56 – 49 | 2% | 6% | -21.4 |
| 22 | Harriet - British Actress | female · old · British | Instant | 33–36 | 47.8% | 36 – 59 | -0.25 | -73 – 20 | 1% | 6% | -25.7 |
| 23 | Sammy - Australian Young Girl | female · young · Australian | Instant | 33–38 | 46.5% | 35 – 58 | -0.02 | -50 – 46 | 2% | 12% | -20.3 |
| 24 | Blue - Low, Calm with Grit | female · old · Canadian | Professional | 32–38 | 45.7% | 35 – 57 | -0.16 | -72 – 35 | 1% | 14% | -27.2 |
| 25 | WarrenT - Business, Upbeat | male · middle-aged · Australian | Professional | 32–38 | 45.7% | 35 – 57 | -0.20 | -68 – 28 | 0% | 15% | -24.4 |
| 26 | James - Natural and Clear Narrator | male · young · American | Professional | 31–38 | 44.9% | 34 – 57 | -0.09 | -53 – 42 | 0% | 13% | -21.8 |
| 27 | Clare - Soft and Warm | female · middle-aged · Irish | Professional | 28–40 | 41.2% | 30 – 53 | -0.52 | -108 – -2 | 0% | 37% | -24.3 |
| 28 | Ann - Expressive, Clear and Textured | female · middle-aged · British | Professional | 28–41 | 40.6% | 30 – 52 | -0.37 | -91 – 15 | 0% | 40% | -22.1 |
| 29 | Clarice - Expressive and Suspenseful | female · middle-aged · British | Instant | 27–43 | 38.6% | 28 – 50 | -0.66 | -119 – -17 | 0% | 56% | -26.0 |
| 30 | Shelton | male · old · American | Professional | 25–45 | 35.7% | 26 – 47 | -0.66 | -123 – -13 | 0% | 71% | -23.9 |
| 31 | Angela | female · old · American | Professional | 21–50 | 29.6% | 20 – 41 | -0.62 | -128 – -9 | 0% | 96% | -22.2 |
| 32 | Alder - Warm, Raspy and Comforting | male · old · Irish | Professional | 17–53 | 24.3% | 16 – 35 | -1.18 | -191 – -65 | 0% | 99% | -32.5 |
2.2 What predicts preference
Each judgment was modelled as a logistic choice on the difference between the two voices' properties: gender, age, accent, tier, log usage and clip level, with an intercept for side. Table 2 gives the coefficients with listener-resampled intervals. Two things are clearly non-zero: level, and nothing else at conventional confidence; "old" is the next strongest and sits just inside zero.
Table 2. Property model: log-odds that the voice with more of the property is chosen, per unit difference. P(>0) is the share of 3,000 listener-resampled fits in which the coefficient is positive.
| Property | Log-odds | 95% CI | P(>0) |
|---|---|---|---|
| Female (vs male) | -0.027 | -0.20 to +0.16 | 38% |
| Young (vs middle-aged) | +0.070 | -0.12 to +0.27 | 77% |
| Old (vs middle-aged) | -0.194 | -0.42 to +0.01 | 3% |
| American accent (vs other) | +0.139 | -0.11 to +0.40 | 87% |
| British accent (vs other) | +0.035 | -0.27 to +0.34 | 60% |
| Professional (vs instant clone) | -0.048 | -0.21 to +0.12 | 29% |
| Library usage, per ×10 characters/year | +0.005 | -0.05 to +0.07 | 59% |
| Level, per dB RMS | +0.039 | +0.01 to +0.07 | 100% |
| Left-hand side (intercept) | -0.009 | -0.17 to +0.15 | 46% |
The professional tier is the result that matters most for anyone picking a voice. ElevenLabs recorded these voices in studios under contract, and 78% of the narration pool carries the badge; the natural assumption is that it marks quality. In this sample it does not. Professional voices averaged a 47.8% win rate against 52.4% for instant clones, and in the 649 pairings that crossed tiers the professional voice won 46.1%. Of the top eight, 4 are professional; of the bottom eight, 7 are. The badge tells you how the voice was made, not how it sounds.
Usage is the other null. The library's most-used voices are the ones that were early and well placed, and the study finds no relationship between how many characters a voice renders a year and whether listeners prefer it. Within this sample the mid-usage band did worst (40%), the high band best (55%), with low and very-high in between; that is noise around a flat line, not a trend.
Table 3. Mean win rate and mean strength of the voices in each group. Groups of two voices (the non-American accents) describe those two voices, not the accent.
| Group | Voices | Mean win rate | Mean strength |
|---|---|---|---|
| Female | 16 | 49.2% | -0.01 |
| Male | 16 | 50.7% | +0.01 |
| Young | 11 | 53.0% | +0.14 |
| Middle-aged | 11 | 50.8% | +0.04 |
| Old | 10 | 45.7% | -0.19 |
| Professional (studio, PVC) | 17 | 47.8% | -0.07 |
| Instant clone | 15 | 52.4% | +0.08 |
| American | 12 | 51.3% | +0.13 |
| British | 8 | 50.4% | -0.07 |
| African American | 2 | 54.2% | +0.27 |
| Scottish | 2 | 52.8% | +0.17 |
| US Southern | 2 | 52.9% | +0.08 |
| Canadian | 2 | 51.1% | -0.04 |
| Australian | 2 | 46.1% | -0.11 |
| Irish | 2 | 32.7% | -0.85 |
| Usage low | 8 | 52.1% | +0.06 |
| Usage mid | 8 | 39.8% | -0.43 |
| Usage high | 10 | 54.5% | +0.20 |
| Usage vhigh | 6 | 53.1% | +0.14 |
2.3 Level
The clips were not normalised (§4.3): the study measures the voice as ElevenLabs delivers it, and ElevenLabs delivers these 32 voices across a 15 dB range. That decision has a measurable price. The louder clip of a pair won 53.6% (CI 51 – 56), the level coefficient is +0.039 log-odds per dB, and level correlates with fitted strength at r = 0.34. The two quietest voices, Alder (-32.5 dB) and Piper Monroe (-32.6 dB), finished 32nd and 16th; the loudest, Maribeth (-17.3 dB), finished 17th. So level is a real advantage but not the ranking: the property coefficients in Table 2 are estimated with level in the model, and the quietest half and loudest half of the voices split about evenly on gender and tier (§4.3), so level adds noise to the voice order rather than bias to the property conclusions.
3Interpretation and limitations
The defensible findings are the shape of the table and the two nulls. Marie is the best-supported choice for this kind of passage; the bottom four should be avoided for it; and neither the professional badge nor popularity is a reason to pick a voice. The old-age tag is a probable disadvantage for first-person nonfiction. Gender is not a factor, which bears on the common practice of matching narrator gender to author gender: nothing here supports it.
What this study does and does not show
It shows which of 32 library voices lay listeners preferred reading one neutral nonfiction paragraph, and that the library's own quality-looking signals (tier, usage) do not predict that preference.
It does not rank the library: 32 of 5,401 voices were sampled, chosen to spread properties, and a voice absent from the sample is simply unmeasured. It does not test fiction, dialogue, humour or long-form narration, where pacing and expressiveness would matter and a "measured" voice might not win. It measures preference at ElevenLabs' default render settings and shipped level; a user who normalises audio removes the level effect reported here.
Specific limitations:
- Sample raised after a first look. The design specified three judgments per item (384). After the first 382 were inspected the target was raised to nine per item (1,152) to tighten the voice-level intervals. This is a data-dependent continuation; the property conclusions were the same at 382 and the final ranking is reported on all 1,122. Two listeners whose sign-in locations were outside the United States were excluded (28 judgments).
- One passage, one setting. Every voice read the same 73 words at
stability 0.5, similarity 0.75, style 0. A different text or setting could reorder the middle of the table; the design bought precise property estimates at the cost of generality across texts. - Level not normalised. Chosen deliberately and reported in §2.3; readers who would normalise should discount the quietest voices' positions accordingly.
- Tags are creator-supplied. Age, accent and descriptive words are what the uploader typed. "Old" means the creator called it old.
- Two voices per non-American accent. Enough to keep accent in the model, not enough to say anything about an accent.
- Forced choice, no tie. A listener with no preference contributed a coin flip, which widens intervals but does not bias them.
- Listeners. 125 raters contributed a median of 7 judgments (maximum 20); no rater heard a voice more than twice. 52% of judgments came from mobile devices, so many clips were heard on phone speakers or earbuds. Median time on a pair was 65 s against two 26-second clips.
4Methods
4.1 Pool and sample
The ElevenLabs shared Voice Library was listed on 3 November 2026: 21,143 voices, 12,132 English, 5,401 English with the narrative_story or informative_educational use case, 5,211 of those verified for eleven_v4 with a preview. Voices above one billion characters a year were excluded as over-exposed. From the remainder 32 were drawn by stratified selection to balance gender (16/16), age (11 young, 11 middle-aged, 10 old), accent (12 American, 8 British, two each Australian, Canadian, Irish, African American, US Southern, Scottish), tier (17 professional, 15 instant) and usage band (8 under 1M characters/year, 8 at 1–10M, 10 at 10–100M, 6 at 100M–1B). Two well-known audiobook voices (Adam, Bill Oxley) were included on purpose as the test of whether fame tracks preference.
4.2 Passage
One paragraph, 73 words, written for the study: first-person general nonfiction in plain prose, about 30 seconds at a narration pace, with nothing in it to act. One passage rather than several because, with 32 arms, the budget had to buy property main effects; a genre interaction would need several times the data.
4.3 Rendering and level
Each voice rendered the passage once through the ElevenLabs API with eleven_v4, 44.1 kHz PCM, stability 0.5, similarity boost 0.75, style 0, speaker boost on; identical for all 32, so that the only difference between clips is the voice. Clips were shipped as WAV and not loudness-normalised: whatever ElevenLabs returns for a voice is what a user of that voice gets. Measured RMS spans -32.6 to -17.3 dB FS. Before data collection level was checked against the properties under test; the quieter and louder halves split roughly 7/4 on gender and on tier, close to random, so level is a covariate in the analysis rather than a confound of the property estimates. No clip clips (worst case two saturated samples in 1.2 million).
4.4 Design
A complete round-robin on 32 voices is 496 pairs and spends most of its budget separating adjacent voices. Instead a circulant graph with offsets 1, 3, 5 and 9 pairs voice i with i+d (mod 32) for each offset: 128 pairs, every voice in exactly eight, every voice reachable from every other (Bradley-Terry needs a connected design), and offsets chosen so that each property is crossed in at least 58% of pairs. Side was balanced by local search so that each voice is on the left in three to five of its eight pairs.
4.5 Listeners and task
Recruitment, payment, blinding and assignment follow the Woodstitch platform methodology. Listeners are paid US adults who hear two clips and answer "Which voice would you rather listen to for an audiobook?" with no tie option. Each session was capped at 16 judgments and no listener heard any one voice more than twice, so no single listener can decide a voice's position. Each item was judged by up to nine different listeners.
4.6 Analysis
Primary statistic: each voice's win rate over its pairings with a Wilson interval, which sets the ranking; rank probabilities come from the voice's win-rate rank in each of 3,000 listener-resampled replicates. Bradley-Terry strength (MM algorithm, weak prior) with a percentile interval from the same replicates is reported alongside as the opponent-adjusted view. Property model: logistic regression of the choice on the difference in each property between the two voices (female, young, old, American, British, professional, log10 usage, dB RMS) with an intercept for side, intervals from the same listener bootstrap. Group means in Table 3 are descriptive. Rater 7 (the author's test account) is excluded from every count.
5Data and reproducibility
The per-judgment dataset is published at study004_judgments.csv (1,122 rows: item, the two voices and their library names, side, choice, time spent, anonymous listener index). All 32 clips are on this page as the listeners heard them. The voice list with library identifiers and tags, the design, the passage and the analysis code are available on request.
6Funding and competing interests
This study was funded by SquireWorks Audiobooks, an audiobook production company that narrates books with ElevenLabs voices and needed a defensible way to choose them; its interest is in the answer, not in any particular voice winning. Woodstitch, which ran the study, sells listening studies of this kind. Neither has any relationship with ElevenLabs beyond being an API customer, nor with any of the voice creators. ElevenLabs was not consulted and did not see the results before publication.
7References
- Bradley RA, Terry ME. Rank analysis of incomplete block designs: I. The method of paired comparisons. Biometrika 1952;39(3/4):324–345.
- Hunter DR. MM algorithms for generalized Bradley-Terry models. Ann. Statist. 2004;32(1):384–406.
- Efron B, Tibshirani RJ. An Introduction to the Bootstrap. Chapman & Hall, 1993.
- Wilson EB. Probable inference, the law of succession, and statistical inference. J. Am. Stat. Assoc. 1927;22(158):209–212.
- ElevenLabs Voice Library. https://elevenlabs.io/voice-library. Listed 3 November 2026.
8Version history
- 1.0 — 11 November 2026. First publication, 1,122 judgments.