Which Fish Audio voice is best for audiobooks?
Thirty-two voices from the Fish Audio voice library, chosen to span gender, age, popularity and Fish's own quality check, read the same 30-second nonfiction passage, loudness-matched. 72 paid listeners heard them in pairs and picked the one they would rather listen to for an audiobook. Listen before you read.
Interim snapshot: data collection in progress
This page reports 1,046 of the 1,152 judgments the design specifies (53 of 128 pairs complete). The study is still running; the page will be updated as v1.1 when the design completes, and the per-judgment data file is the snapshot behind these numbers. Intervals are reported throughout and no claim here rests on the order of voices whose intervals overlap.
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: Fish's s2.1-pro output at default settings, every clip brought to the same loudness (-26.1 LUFS) with a single gain, nothing else touched.
All 32 voices, ranked, with their clips
Voice names are the uploaders' own 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 998.
Question Among voices in the Fish Audio voice library, which do ordinary listeners prefer for an audiobook, and which properties of a voice (gender, age, delivery, Fish's quality check, popularity) predict that preference?
Study 32 English narration voices were drawn from the library, stratified so that gender, age, usage and Fish's source-audio quality check are each spread rather than left at the library's skew, with real-person and character clones screened out. Every voice read one 73-word first-person nonfiction passage with Fish's s2.1-pro model at default settings, and every clip was brought to the same loudness. Each voice met eight others in a connected design of 128 pairs; 72 paid US listeners made 998 forced-choice judgments, no listener hearing any voice more than twice. The design is the one used in Study 004 on ElevenLabs, with the same passage.
Findings “Old” finished first, winning 75.9% of its pairings (Wilson 95% CI 63 – 85) and ranking first in 47% of listener-resampled replicates. “Warm Royal” (74%) and “Sleepless historian” (69%) follow; the top eight cannot be ordered among themselves. “Young Narrative Storyteller” was last at 21%, in the bottom five in 99% of replicates. Across properties: popular voices were preferred (a high-usage voice beat a low-usage one in 60% of 625 mixed pairs, CI 56 – 64; +0.12 log-odds per tenfold usage, CI +0.04 to +0.19); voices that passed Fish's source-audio quality check won only 41% of pairs against unchecked voices (CI 37 – 45), but the badge sits almost entirely on low-usage voices and its coefficient with usage in the model is -0.09 (CI -0.40 to +0.21), so the check neither helps nor hurts once popularity is accounted for; female voices were slightly preferred (54% of mixed pairs, CI 50 – 58); age and the delivery tag made no detectable difference.
Scope One neutral nonfiction passage of about 30 seconds, one model at default settings, one render per voice, clips loudness-matched, 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
- “Old” is the best-supported first place: 44 wins of 58 (75.9%), first in 47% and top-five in 95% of 3,000 replicates. “Warm Royal” (74%, first in 32%) is close behind; the top eight overlap and are not ordered.
- The bottom 8 are clearly below even: Jessica (37%), Calm Conversational Voice (33%), Friendly Young Male Narrator (32%), Old Indian Woman (31%), Soft Dreamy Storyteller (29%), Deep Melodic Baritone (28%), Gus (27%), Young Narrative Storyteller (21%). “Young Narrative Storyteller” lost 53 of 67 pairings.
- Popularity predicts preference. The opposite of Study 004, where ElevenLabs usage predicted nothing. Here the usage coefficient is +0.117 log-odds per tenfold (CI +0.04 to +0.19), positive in 100% of replicates; a very-high-usage voice beat a low-usage one in 108 of 161 pairs (67%). Mean win rate rises across the usage bands: 43%, 49%, 58%, 57% from low to very high.
- Fish's quality badge does not predict preference. Voices whose source audio passed Fish's quality check won 40.5% of the 570 pairs against unchecked voices (CI 37 – 45) and averaged a 44% win rate against 55%. That raw gap is a confound, not a verdict: the check is recent (the passed voices here were uploaded between May 2026 and November 2026), so 10 of the 14 passed voices are in the low-usage band against 1 of 18 unchecked ones, and with usage in the model the badge coefficient is -0.09 (CI -0.40 to +0.21). The badge says the upload was clean; it says nothing about whether listeners will like the voice.
- Female voices were slightly preferred: 54.4% of 568 mixed-gender pairs (CI 50 – 58), +0.24 log-odds (CI +0.03 to +0.47). Study 004 found no gender effect on ElevenLabs. Small, and at the edge of what this sample can resolve.
- Age and delivery tags made no detectable difference. Young -0.08 (CI -0.27 to +0.10), old +0.13 (CI -0.10 to +0.36) against middle-aged; the "expressive" delivery tag -0.12 (CI -0.40 to +0.16). Study 004's "old is disfavoured" result did not recur here. Side of presentation did not matter (47.2% left).
- Level was matched, so the Study 004 loudness caveat does not apply. Fish ships these voices 19 dB apart (-29.9 to -11.0 LUFS); every clip was brought to -26.1 LUFS before any listener heard it (§4.3). Residual RMS differences after matching span 1.8 dB; the clip with the higher residual RMS won 53.2% of pairs (CI 50 – 56), and a voice's shipped level correlates with its strength at r = 0.16. Loud delivery is not why anything here won.
1What we tested
Fish Audio's voice library is a different kind of catalogue from ElevenLabs'. Nearly every voice is a user-uploaded instant clone; tags and descriptions are largely generated from the audio rather than written by the uploader; titles are whatever the uploader typed; and the top of the library by usage is announcers, game characters and clones of real people. What it does offer are gender and age tags, a descriptive vocabulary (calm, warm, expressive, dramatic), a usage count, a small "Fish Official" set, and, recently, a source-audio quality check that marks an upload as passed, unchecked or failed. Anyone choosing a narrator from the library is choosing on those, or on a preview.
This study asks the same two things of that catalogue that Study 004 asked of ElevenLabs'. First, which of a broad sample of its narration voices do listeners actually prefer for an audiobook. Second, whether any of the catalogue's own signals predict that preference: does the quality badge mean a better voice, does popularity, does the age or gender tag, does a calm or expressive delivery tag. The 32 voices were drawn to spread every property (§4.1) rather than at random, and real-person and character clones were screened out, because a random draw from this library would be mostly unusable for a book.
The passage is the same 73-word first-person nonfiction paragraph as Study 004 (§4.2): plain prose with no humour, grief or jargon, so that no voice is rewarded for acting. Every voice read it with Fish's default production model, s2.1-pro, at default settings. One thing changed from Study 004 in response to its readers: the clips were loudness-matched before anyone heard them (§4.3). The design, the passage, the voice list and the analysis were written down before any judgment was collected.
The study was prompted by a Fish Audio community member who asked, after Study 004 was published, for the same test on Fish's library. Woodstitch ran it at its own expense: no credits, payment or input were taken from Fish Audio, who did not see the voice list or the results before publication (§6).
2Results
2.1 Ranking
Table 1 ranks the 32 voices by win rate over all 998 judgments; Bradley-Terry strength, which adjusts for the opponents each voice happened to face, gives nearly the same order and is shown alongside. “Old” won 75.9% of its pairings and is first in 47% of replicates; “Warm Royal” is first in 32%, and no other voice in more than 7%. The eight voices at the top (“Old”, “Warm Royal”, “Sleepless historian”, “Lily”, “alex”, “Sarah”, “Wise Mature Business Narrator”, “mck”) have overlapping intervals and the study cannot order them. The middle of the table is a crowd near even. The bottom separates: 8 voices have intervals entirely below even, and the bottom five finish in the bottom five in at least 54% of replicates each.
The spread is wider than on ElevenLabs. Study 004's table ran from 24% to 66%; this one runs from 21% to 76%, with 7 voices under 35%. A library of unscreened uploads has a longer tail of voices listeners do not want to hear, even after the obvious junk is removed by hand.
Side of presentation did not matter: the left-hand clip was chosen on 47.2% of judgments (CI 44 – 50). 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 60 pairings (eight pairs, about 8 judgments each).
Table 1. All 32 voices ranked by win rate. Tags are Fish's library tags (delivery is the calm/expressive grouping of the descriptive tag, §4.1). Check is Fish's source-audio quality check. Usage is the library's lifetime generation count in four bands. 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. LUFS is the integrated loudness of the clip as Fish shipped it, before matching.
| # | Voice | Tags | Check | Usage | W–L | Win rate | 95% CI | Strength | 95% CI | P(top 5) | P(bottom 5) | LUFS |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Old | male · old · calm | Unchecked | high | 44–14 | 75.9% | 63 – 85 | +1.21 | +0.59 to +1.99 | 95% | 0% | -16.6 |
| 2 | Warm Royal | female · middle-aged · expressive | Passed | mid | 43–15 | 74.1% | 62 – 84 | +1.08 | +0.55 to +1.73 | 92% | 0% | -13.9 |
| 3 | Sleepless historian | male · middle-aged · calm | Unchecked | very high | 44–20 | 68.8% | 57 – 79 | +0.63 | +0.09 to +1.24 | 68% | 0% | -19.2 |
| 4 | Lily | female · young · mixed | Unchecked | mid | 39–19 | 67.2% | 54 – 78 | +0.76 | +0.07 to +1.64 | 58% | 0% | -24.9 |
| 5 | alex | male · young · calm | Unchecked | very high | 35–18 | 66.0% | 53 – 77 | +0.82 | +0.21 to +1.56 | 49% | 0% | -18.1 |
| 6 | Sarah (official) | female · young · calm | Unchecked | very high | 41–23 | 64.1% | 52 – 75 | +0.38 | -0.14 to +0.98 | 39% | 0% | -29.2 |
| 7 | Wise Mature Business Narrator | female · old · calm | Passed | low | 40–23 | 63.5% | 51 – 74 | +0.77 | +0.19 to +1.47 | 34% | 0% | -14.3 |
| 8 | mck | male · old · calm | Unchecked | mid | 38–24 | 61.3% | 49 – 72 | +0.45 | -0.10 to +1.11 | 22% | 0% | -22.5 |
| 9 | British | female · middle-aged · calm | Unchecked | high | 37–27 | 57.8% | 46 – 69 | -0.18 | -0.73 to +0.37 | 8% | 0% | -23.2 |
| 10 | Dorothy (official) | female · old · mixed | Unchecked | low | 33–25 | 56.9% | 44 – 69 | +0.05 | -0.51 to +0.59 | 7% | 0% | -19.0 |
| 11 | Noel | male · young · calm | Unchecked | high | 36–28 | 56.2% | 44 – 68 | +0.43 | -0.12 to +1.08 | 7% | 0% | -28.4 |
| 12 | Deep | male · middle-aged · calm | Unchecked | high | 36–28 | 56.2% | 44 – 68 | -0.27 | -0.84 to +0.25 | 4% | 0% | -21.1 |
| 13 | FEMALE NEW | female · middle-aged · mixed | Unchecked | high | 35–28 | 55.6% | 43 – 67 | +0.31 | -0.31 to +0.94 | 4% | 0% | -18.7 |
| 14 | Mysterious Fantasy Narrator | female · young · expressive | Passed | low | 33–29 | 53.2% | 41 – 65 | -0.02 | -0.56 to +0.54 | 3% | 0% | -14.7 |
| 15 | Abby: English Woman (US) - Audiobook | female · middle-aged · mixed | Unchecked | very high | 34–30 | 53.1% | 41 – 65 | +0.28 | -0.29 to +0.90 | 2% | 0% | -29.1 |
| 16 | Calm Female Voice | female · young · expressive | Passed | high | 34–31 | 52.3% | 40 – 64 | +0.16 | -0.41 to +0.71 | 1% | 0% | -15.3 |
| 17 | sleeeppp | male · old · calm | Unchecked | high | 35–32 | 52.2% | 40 – 64 | -0.02 | -0.68 to +0.61 | 1% | 0% | -23.2 |
| 18 | 08-RVV_HOUSEMAID_01 | female · old · calm | Passed | low | 33–31 | 51.6% | 40 – 63 | +0.29 | -0.25 to +0.89 | 2% | 0% | -13.2 |
| 19 | story | male · middle-aged · calm | Unchecked | mid | 30–30 | 50.0% | 38 – 62 | +0.24 | -0.34 to +0.86 | 1% | 0% | -21.4 |
| 20 | Mature Authentic Narrator | male · old · calm | Passed | low | 29–29 | 50.0% | 38 – 62 | +0.03 | -0.63 to +0.70 | 2% | 1% | -17.8 |
| 21 | POLO | female · young · expressive | Passed | very high | 34–36 | 48.6% | 37 – 60 | -0.20 | -0.88 to +0.41 | 1% | 1% | -11.0 |
| 22 | Narrative Male Voice | male · middle-aged · calm | Passed | low | 29–34 | 46.0% | 34 – 58 | +0.10 | -0.47 to +0.72 | 0% | 1% | -17.9 |
| 23 | Old woman | female · old · mixed | Unchecked | mid | 28–35 | 44.4% | 33 – 57 | -0.75 | -1.42 to -0.17 | 0% | 3% | -29.9 |
| 24 | horror | male · middle-aged · calm | Unchecked | very high | 27–34 | 44.3% | 33 – 57 | -0.02 | -0.62 to +0.54 | 0% | 4% | -18.6 |
| 25 | Jessica | female · middle-aged · calm | Passed | low | 24–41 | 36.9% | 26 – 49 | -0.13 | -0.71 to +0.44 | 0% | 19% | -14.4 |
| 26 | Calm Conversational Voice | male · middle-aged · calm | Passed | mid | 21–42 | 33.3% | 23 – 46 | -0.81 | -1.56 to -0.21 | 0% | 38% | -28.7 |
| 27 | Friendly Young Male Narrator | male · young · calm | Passed | low | 22–46 | 32.4% | 22 – 44 | -0.57 | -1.32 to +0.09 | 0% | 47% | -17.2 |
| 28 | Old Indian Woman | female · old · calm | Unchecked | mid | 17–37 | 31.5% | 21 – 45 | -0.56 | -1.24 to +0.05 | 0% | 54% | -26.8 |
| 29 | Soft Dreamy Storyteller | female · young · calm | Passed | low | 17–42 | 28.8% | 19 – 41 | -0.73 | -1.54 to -0.09 | 0% | 73% | -21.6 |
| 30 | Deep Melodic Baritone | male · old · calm | Passed | low | 19–48 | 28.4% | 19 – 40 | -1.25 | -1.91 to -0.73 | 0% | 76% | -22.1 |
| 31 | Gus | male · young · expressive | Unchecked | mid | 17–46 | 27.0% | 18 – 39 | -0.81 | -1.44 to -0.30 | 0% | 82% | -21.3 |
| 32 | Young Narrative Storyteller | male · young · expressive | Passed | low | 14–53 | 20.9% | 13 – 32 | -1.64 | -2.38 to -1.13 | 0% | 99% | -15.1 |
2.2 What predicts preference
Each judgment was modelled as a logistic choice on the difference between the two voices' properties: gender, age, Fish's quality check, the delivery tag and log usage, with an intercept for side. Table 2 gives the coefficients with listener-resampled intervals. Two coefficients exclude zero, usage and gender; the rest sit across it.
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) | +0.243 | +0.03 to +0.47 | 98% |
| Young (vs middle-aged) | -0.084 | -0.27 to +0.10 | 20% |
| Old (vs middle-aged) | +0.129 | -0.10 to +0.36 | 87% |
| Quality check passed (vs unchecked) | -0.091 | -0.40 to +0.21 | 27% |
| Expressive delivery tag (vs calm or mixed) | -0.117 | -0.40 to +0.16 | 20% |
| Library usage, per ×10 generations | +0.117 | +0.04 to +0.19 | 100% |
| Left-hand side (intercept) | -0.120 | -0.31 to +0.06 | 10% |
Usage is the result that most separates this library from ElevenLabs'. On ElevenLabs, usage was a null: the most-rendered voices were the ones that were early and well placed. On Fish, usage tracks preference: +0.117 log-odds per tenfold generations (CI +0.04 to +0.19), and the four usage bands line up in order (Table 3). One reading is that on Fish usage still reflects people trying voices and keeping the ones that sound good, rather than early voices locked in by default placement. Another is that popular voices are popular partly because they are louder or more produced; level is matched here, so that at least is not it. The study cannot tell these apart.
The quality check is the result that matters most for anyone picking a voice, and it needs care. Fish's check marks whether an upload's source audio was clean enough to clone well, and the natural assumption is that a passed voice is a better voice. In this sample, passed voices did worse on their face: a 44.3% mean win rate against 54.9%, and 40.5% of the 570 mixed pairs. But the check is recent and has been run mostly on recent, little-used uploads: the passed voices in this sample were uploaded between May 2026 and November 2026, the unchecked ones mostly in 2025, and 10 of the 14 passed voices are in the low-usage band against 1 of the 18 unchecked. Put usage in the model and the badge's coefficient is -0.09 (CI -0.40 to +0.21): no evidence either way. The honest statement is that the badge certifies the upload, not the voice, and a listener choosing by it would have done no better than one ignoring it.
Gender is the one tag with a signal: female voices won 54.4% of mixed pairs (CI 50 – 58), +0.24 log-odds. Study 004 found none on ElevenLabs (49% in mixed pairs, published figure), so this is either a feature of these 16 female voices or a small real preference for the passage; it is not large enough to choose by. The two Fish Official voices, Sarah and Dorothy, finished 6th and 10th, and won 61% of their 122 pairs against community uploads; two voices is a description, not a result.
Table 3. Mean win rate and mean strength of the voices in each group. Groups of two voices (Fish Official) describe those two voices, not the group.
| Group | Voices | Mean win rate | Mean strength |
|---|---|---|---|
| Female | 16 | 52.5% | +0.09 |
| Male | 16 | 48.1% | -0.09 |
| Young | 11 | 47.0% | -0.13 |
| Middle-aged | 11 | 52.4% | +0.11 |
| Old | 10 | 51.6% | +0.02 |
| Quality check passed | 14 | 44.3% | -0.21 |
| Unchecked | 18 | 54.9% | +0.16 |
| Delivery tag: calm | 21 | 50.2% | +0.04 |
| Delivery tag: expressive | 6 | 46.0% | -0.24 |
| Delivery tag: mixed | 5 | 55.5% | +0.13 |
| Fish Official | 2 | 60.5% | +0.22 |
| Community upload | 30 | 49.6% | -0.01 |
| Usage low | 11 | 42.6% | -0.28 |
| Usage mid | 8 | 48.6% | -0.05 |
| Usage high | 7 | 58.0% | +0.23 |
| Usage very high | 6 | 57.5% | +0.31 |
2.3 Level
Study 004 left ElevenLabs' clips at their shipped levels and found that the louder clip of a pair won 54% of the time, which readers rightly questioned. This study matched them. Fish ships these 32 voices across a 19 dB range (-29.9 to -11.0 LUFS integrated), wider than ElevenLabs' 15 dB; each clip received one static gain (-15.1 to +3.8 dB) to bring it to -26.1 LUFS with a true-peak ceiling of -1 dBTP, no compression or limiting (§4.3). What remains after matching is a 1.8 dB spread in plain RMS, which the loudness measure weights differently from a meter; the higher-RMS clip won 53.2% of pairs (CI 50 – 56), and residual RMS correlates with strength at r = 0.23. The shipped level, which listeners never heard, correlates with strength at r = 0.16. Level is reported per voice in Table 1 as a property of what Fish delivers, not as something the listeners judged.
3Interpretation and limitations
The defensible findings are the shape of the table and the two signals. “Old” and “Warm Royal” are the best-supported choices for this kind of passage; the bottom 8 should be avoided for it; popularity is a usable signal in this library where it was not in ElevenLabs'; and the quality badge is not. The gender result is small and worth a replication before anyone acts on it.
What this study does and does not show
It shows which of 32 screened library voices lay listeners preferred reading one neutral nonfiction paragraph with loudness matched, and which of the library's own signals track that preference.
It does not rank the library: 32 voices were sampled from an 843-voice gated pool, 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. It measures one render of each voice with one model at default settings; Fish's API exposes no seed, so a second render of the same voice would differ slightly. It says nothing about ElevenLabs versus Fish: the two studies share a passage and a design but not pairs, so "Fish's best against ElevenLabs' best" is not answerable from them.
Specific limitations:
- Sample. The design specified nine judgments per item (1,152); this interim snapshot reports the 1,046 collected so far, 53 of 128 items complete, and will be updated when the design completes. 3 listeners whose sign-in locations included a country outside the United States were excluded (48 judgments), the rule used in Study 004.
- One passage, one model, one render. Every voice read the same 73 words through
s2.1-proat default settings. A different text, model (Fish's expressive Drama 3 was not used) or draw could reorder the middle of the table. - Tags are mostly machine-generated. Fish's gender and age tags are inferred from the audio and were wrong often enough that the sample was checked by hand; the delivery grouping (calm / expressive / mixed) is a coding of Fish's descriptive words. "Old" means the tag said old.
- The quality check is confounded with usage by the library's history, not by the design; the model separates them only as far as 14 passed voices allow.
- Screening was manual. Real-person and character clones were removed by a title and description screen and a hand pass; what counts as a usable narration voice is a judgment, made before any data.
- Forced choice, no tie. A listener with no preference contributed a coin flip, which widens intervals but does not bias them.
- Listeners. 72 raters contributed a median of 16 judgments (maximum 16); 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 58 s against two clips of roughly 25 seconds.
4Methods
4.1 Pool and sample
The Fish Audio voice library was listed on 12 November 2026 through its model API, English voices with narration-type tags or titles (narration, storytelling, audiobook, narrator, calm, warm, conversational): 4,992 voices, 4,831 English-only. A gate kept voices that were English-only, tagged for gender and age, carried a narration-type tag and a plausible English description of the voice, and had not failed Fish's quality check, and removed real people and characters by a title and description screen: 843 voices. A stratified oversample of 136 was screened by hand and 32 drawn to balance gender (16/16), age (11 young, 11 middle-aged, 10 old), usage band by lifetime generations (11 under 100, 8 at 100–5,000, 7 at 5,000–50,000, 6 above 50,000), quality check (14 passed, 18 unchecked) and delivery tag (21 calm, 6 expressive, 5 mixed; the narration pool skews calm). Two Fish Official voices were included as the curated cell. Every voice in the pool is an instant clone; Fish has no studio tier.
4.2 Passage
One paragraph, 73 words, written for Study 004 and reused unchanged: first-person general nonfiction in plain prose, about 30 seconds at a narration pace, with nothing in it to act.
4.3 Rendering and level
Each voice rendered the passage once through the Fish Audio API with s2.1-pro, WAV output, API defaults (temperature 0.7, top-p 0.7, text normalisation on, no prosody block); identical for all 32, so that the only difference between clips is the voice. Shipped loudness ranged from -29.9 to -11.0 LUFS. Each clip was then brought to a common integrated loudness with a single static gain measured by BS.1770 (ffmpeg ebur128), no compression or limiting; the common target was lowered to -26.1 LUFS so that no clip would exceed -1 dBTP. Gains applied ranged from -15.1 to +3.8 dB. Listeners heard the matched clips; the shipped level is reported per voice as a property of the library.
4.4 Design
As Study 004: 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). 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, quality check passed, expressive delivery, log10 usage) with an intercept for side, intervals from the same listener bootstrap; the level covariate used in Study 004 is dropped because the clips are matched, and level is reported as a residual check (§2.3). 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 study007_judgments.csv (998 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, the normalisation log and the analysis code are available on request.
6Funding and competing interests
This study was funded by Woodstitch, which ran it and sells listening studies of this kind. It was prompted by a public request from a Fish Audio community member for the ElevenLabs study to be repeated on Fish's library; Woodstitch paid for its own renders and took no credits, payment or direction from Fish Audio, which was not consulted and did not see the voice list or the results before publication. Woodstitch has no relationship with Fish Audio or ElevenLabs beyond being an API customer of both, nor with any of the voice uploaders. Its interest is in the answer, not in any particular voice or vendor winning.
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.
- ITU-R BS.1770-5. Algorithms to measure audio programme loudness and true-peak audio level. 2023.
- Fish Audio voice library. https://fish.audio/. Listed 12 November 2026.
- Hawkes T. Which ElevenLabs voice is best for audiobooks? Woodstitch Research, 11 November 2026. woodstitch.com/research/which-elevenlabs-voice-is-best-for-audiobooks
8Version history
- 1.0 (interim) — 18 November 2026. Interim snapshot, 998 judgments analysed of 1,046 collected toward 1,152.