Minutes Are Not the Binding Constraint: ElevenLabs' Audiobook-Grade Tier Is Self-Voice-Only, and ACX Prohibits Unauthorized AI Narration
The question
Verbatim: "For a deceased author with only degraded/partial voice recordings, what's the minimum viable clean-audio length/quality for ElevenLabs-class voice cloning to produce audiobook-grade narration?"
Context: the grandma-memoir-republish project is a KDP republish plus an audiobook, and the audiobook leg has been parked pending recovery of the author's voice recordings. This brief was supposed to price the recovery effort in minutes of audio. It ends up re-pricing the whole leg, because the gate turns out to be contractual, not acoustic.
What we already know (from the vault)
- The memoir project has two published titles and a settled rights position — family rights confirmed, one aunt is rights holder of record, and the outreach purpose shifted from negotiating rights to locating surviving manuscript/production files ([[2026-07-21-spanner-books-rights-outreach-draft]], founder correction 2026-07-21 evening). The audiobook is named in that same thread as the family's stated want.
- The author died 2010-09-02, aged 91 ([[2026-07-20-wilson-tree-v0]], obituary read directly). That date sets the media era for any surviving recordings: analog cassette, camcorder tape, VHS, MiniDV, answering-machine microcassette, early digital camcorder audio. It effectively rules out any modern clean-mic source.
- No voice recording of her has been located or inventoried anywhere in the vault. A full-vault grep for cassette / answering machine / voice recording returns nothing related to this project. The gate is still at step zero: nobody has confirmed a recording exists, let alone measured one.
- RDCO already runs ElevenLabs in production for Ray's own synthetic voice (Eric,
voice_id cjVigY5qzO86Huf0OWal) and lists "ElevenLabs to read drafts aloud" as a reusable capability ([[2026-10-05-voice-and-design-packs-spec]]). ElevenLabs' speech-isolation endpoint is already in our tool pool — testing a degraded sample costs us nothing new. - Source separation for exactly this problem is already on our radar: the Voice-Pro pipeline note describes splitting voice from music as a commodity local step ([[2026-05-05-alphasignal-voice-pro-local-dubbing]]).
What the web says
- ElevenLabs Instant Voice Cloning (IVC): docs recommend roughly 1–2 minutes of clean audio, and advise against going much past ~3 minutes because extra audio adds little and can destabilize the clone. The docs are explicit that quality beats length — "Audio quality is the most important aspect to consider when using instant voice cloning. The number of samples is irrelevant; what's important is the total run time" — and that the model reproduces everything it hears, including speed, inflection, accent, tonality and breathing (ElevenLabs IVC docs). Any circulating "30 seconds is enough" figure is a marketing-register claim about instant cloning, not a documented floor for narration quality — keep it out of the decision.
- ElevenLabs Professional Voice Cloning (PVC): "The bare minimum we recommend is 30 minutes of audio," with "closer to 2-3 hours" for the optimal, most accurate clone. The docs contradict themselves on the floor: the step-by-step guide says "we recommend uploading at least an hour of training audio, and ideally as close to three hours as possible," while the FAQ gives a 30–180 minute band. Total runtime matters, not file count; multi-hour material should be split into ~30-minute samples (ElevenLabs PVC docs, fetched 2026-10-25).
- ElevenLabs' stated audio-quality floor (same PVC page): single speaking voice only — "Make sure there's only a single speaking voice throughout the audio"; multiple speakers, background noise and reverb will be replicated in the clone; MP3 at 192 kbps or higher (uncompressed WAV "usually do not improve clone quality"); −23 dB to −18 dB RMS with a true peak of −3 dB; acoustically treated room; singing unsupported.
- The PVC blocker: verification is mandatory before a fine-tuning request is submitted, and the docs state flatly, "Even with their consent, you cannot clone someone else's voice." The only sanctioned path for another person's voice is for that person to create and verify the clone on their own account and share it. A deceased speaker cannot complete a verification step — so the entire 30-minute-to-3-hour tier, the only tier whose output the vendor describes as high-fidelity, is structurally unavailable here on the self-serve product. (That verification is a live read-back is widely reported but I did not verify the mechanism in this pass; the "cannot clone someone else's voice" sentence is quoted directly and is sufficient on its own.)
- ElevenLabs Terms of Use (Last Updated 09 October 2026): §4(b) limits uploads to "the voice you are authorized to share with us"; §4(g) — "You may not provide Input or create Output for which you do not have all the rights necessary" — plus a representation that the content and voice models do not violate anyone's rights or injure any person or entity. §4(b) also gives a deletion right for user voice models. The ToU says nothing about deceased persons, estates or heirs, and contains no AI-output disclosure requirement. So "rights-holder consent" is the operative standard but the vendor does not define what artifact satisfies it for a decedent — that gap is ours to fill.
- ACX is closed to this. ACX's audio submission requirements article (dated Apr 15, 2026) states: "Unauthorized use of text-to-speech, AI, or automated recordings in ACX titles is prohibited," with no disclosure-based allowance described, and a note that Audible is "working to accept third-party TTS content from publishers and creators who are interested" — i.e. a future intake, not a current one (ACX submission requirements). Same page carries the delivery specs: −23 to −18 dB RMS, peak ≤ −3 dB, noise floor ≤ −60 dB RMS, 192 kbps+ CBR 44.1 kHz MP3, room tone ≤ 5s, uniform mono/stereo, ≤120 min per file.
- Wide distributors — unverified. Secondary blogs claim Spotify (via its ElevenLabs intake, reportedly from Feb 2025), Google Play Books, Apple Books, Kobo and Findaway accept AI narration with disclosure, and that Findaway's wide route now runs through a renamed product. One blog claims ACX has allowed AI with disclosure since 2024, which directly contradicts ACX's own page. None of this is sourced to a distributor's own policy page and should not be relied on. Treat as a to-verify, not a finding.
Convergences and contradictions
- Sharp convergence worth keeping: ElevenLabs' recommended input loudness band (−23 to −18 dB RMS, true peak −3 dB) is numerically identical to ACX's output delivery spec. ACX's −60 dB RMS noise floor is therefore the best available proxy for what "audiobook-grade clean" means as a measurable number — and ElevenLabs has no published noise-floor requirement at all. We can borrow ACX's and use it as our acceptance threshold on recovered source audio. A typical consumer cassette or camcorder recording sits far above −60 dB RMS noise floor; that gap, not minute count, is the thing to measure first.
- Contradiction inside the vendor's own docs: the PVC floor is 30 minutes in the requirements/FAQ and "at least an hour" in the guide. Do not quote a single settled number. The honest statement is: 30 minutes is the vendor's stated minimum, 1 hour is the vendor's own recommended floor in a different place on the same page, and 2–3 hours is where the docs say quality actually lands.
- Contradiction between vendor minimum and task fitness: nothing ElevenLabs publishes is a tested claim about audiobook-length narration consistency. "30 minutes works" is a clone-admissibility threshold; sustaining 4–8 hours of listenable narration without drift, mispronunciation, or prosody collapse is a different and unmeasured bar. No source found tests it. Flag every duration number in this brief as vendor guidance, not benchmark.
- Vault vs web: the vault treated "find the recordings" as the gate. The web says the gate is consent plumbing plus distributor policy. Both the vault's stated want (ACX/Audible audiobook) and the vault's implied method (clone her voice) are blocked independently of how much tape turns up.
Synthesis for RDCO
The minute-count answer, tiered — and then the reason it doesn't matter. Given N minutes of recovered audio at quality X, the realistic ceiling is: under ~60 seconds of usable single-speaker speech, or anything whose noise floor is within ~30 dB of the speech, is nothing usable — don't clone it, archive it. Roughly 1–3 minutes of single-speaker audio that survives isolation gets you an IVC clone whose honest ceiling is short memorial clips — a recorded dedication, a paragraph in her own words at the head of a chapter — and nothing longer, because IVC's own docs cap useful input at ~3 minutes no matter how much more you have. Between 3 and 30 minutes you gain sample selection, not model quality: you can pick the cleanest 2 minutes, which is a real improvement, but you are still on the instant tier. 30 minutes to 1 hour clean crosses ElevenLabs' stated PVC minimum while sitting below its own recommended floor; 1–3 hours clean is the only band where the vendor's docs predict a clone good enough to carry a book. Both of those upper bands require PVC, and PVC requires a verification the speaker must perform herself. So the practical ceiling for this project, on the self-serve product, is the memorial-clip tier — regardless of whether we recover two minutes or three hours.
What degradation actually costs, and what pre-processing can buy back. Separate the recoverable from the unrecoverable before anyone estimates effort. Recoverable with standard tooling: steady broadband hiss, 50/60 Hz hum and harmonics, level inconsistency, mild clipping, mild wow/flutter, and — most usefully — crosstalk from other speakers, which source separation and ElevenLabs' own speech-isolation endpoint handle well; this matters because the PVC page's hardest requirement is single-speaker-only, so isolation is what converts a family-dinner tape into an admissible sample at all. Partly recoverable: room reverb — de-reverberation is decent now but leaves artifacts, and since the docs say reverb is replicated into the clone, a half-fixed room still contaminates the voice. Effectively unrecoverable: band-limiting. Answering-machine and telephone audio has no energy above roughly 3.4 kHz, and worn cassette or misaligned-azimuth tape loses the top octaves — that information never existed in the file, and bandwidth-extension tools invent it. For identity cloning specifically, inventing the high-frequency content means inventing part of the thing that makes the voice recognizably hers, which is the one failure mode a family will hear instantly and the one no amount of extra minutes fixes. Also unrecoverable: tape dropout, print-through, and any passage where the noise is louder than the speech. Concrete first action, and it is cheap: digitize flat at 24-bit/48 kHz with no noise reduction on the archival master, then measure three numbers per clip — speech RMS, noise floor in dBFS, and the spectral ceiling. Spectral ceiling under ~4 kHz is a hard stop for narration use. Noise floor worse than ACX's −60 dB RMS after isolation means clip-tier at best. Our ElevenLabs isolation endpoint is already provisioned, so a real test of any one recovered clip is a single call, not a project.
The consent artifact and the distribution route are the real work, and they are both more tractable than the audio. ElevenLabs' ToU puts the whole burden on us — §4(g) requires that we hold "all the rights necessary," while saying nothing about decedents. So the family sign-off this project already requires needs to be a different, larger document than "we approve the republish": a written consent to synthesize the author's voice, executed by the rights holder of record and the surviving children, naming the specific use (audiobook narration and/or memorial clips), the platform, the retention and deletion posture (ToU §4(b) gives a model-deletion right — say we'll exercise it), and the disclosure language that will appear on the finished product. That is a one-page artifact we can draft now, with labeled guesses, before a single minute of tape is found; it also forces the family conversation about whether they want her voice synthesized, which is the question most likely to kill this leg and the one nobody has asked. On distribution, ACX self-serve is out as of its Apr 15 2026 page, which leaves three honest routes: Amazon's own labeled synthetic-voice product (which uses a stock voice, defeating the entire point); a human narrator drawn from the family — a daughter or granddaughter reading her mother's memoir is legally trivial, ACX-eligible today, and arguably a better artifact than a cloned approximation of a woman who died in 2010; or wide-only distribution with AI disclosure, which the secondary sources suggest exists but which nobody has verified against a distributor's own policy page. My read: reframe the leg as human-narrated audiobook with an optional cloned memorial clip as a dedication, and stop treating recording recovery as the blocker. That unblocks the project this quarter instead of parking it on a tape hunt that, even if it succeeds spectacularly, lands in a tier the vendor won't sell us.
Why this is in the vault
This re-prices the audiobook leg of the grandma-memoir-republish project, which has been parked on "find her voice recordings" since the 2026-07 rights thread ([[2026-07-21-spanner-books-rights-outreach-draft]]): the gate is not minutes of tape but ElevenLabs' self-voice-only verification for its only audiobook-grade tier plus ACX's AI-narration prohibition, which together mean the project should proceed as a human-narrated audiobook now rather than wait on recovery. It also gives the family-history project a measurable archival spec (digitize flat 24-bit/48 kHz, measure noise floor against ACX's −60 dB RMS, check spectral ceiling) and names the consent artifact the existing family sign-off gate does not currently cover.
Open follow-ups
- Does ElevenLabs have any non-self-serve route (enterprise, estate, or memorial program) for cloning a deceased person's voice with rights-holder consent? The docs' "even with their consent" sentence governs the self-serve product; an enterprise or legal-review path may exist and was not checked.
- What exactly is the PVC verification mechanism — live read-back of a generated script, government ID, or both? This determines whether any workaround exists at all, and I did not verify it in this pass.
- What do Spotify (via its ElevenLabs intake), Apple Books, Google Play Books, Kobo and the renamed Findaway/wide route actually require disclosed for AI-narrated titles, per their own policy pages rather than blogs? All current evidence is secondary and partly self-contradictory.
- Has Audible's "working to accept third-party TTS content from publishers and creators" intake opened since the Apr 15 2026 ACX page, and does it admit a cloned voice of a deceased author?
- Is there any published benchmark — not vendor guidance — on narration consistency over 4+ hours as a function of PVC training-audio duration? Every duration figure in this brief is vendor guidance with no tested result behind it.
- Does a recording of the author exist at all? Nobody has inventoried family media (cassettes, camcorder tapes, Folk School or church archives, local oral-history collections). This is the one question whose answer is free and which nothing above can proceed without.
- Would the family consent to synthesizing her voice? Never asked. It may be a no, which would make the entire acoustic question moot.
- Does the John C. Campbell Folk School or a regional Appalachian oral-history archive hold a recorded interview with her? A 1990s–2000s oral-history interview would be the single highest-quality plausible source and is a targeted archive query, not a tape hunt.
Related
- [[2026-07-21-spanner-books-rights-outreach-draft]]
- [[2026-07-20-wilson-tree-v0]]
- [[2026-05-05-alphasignal-voice-pro-local-dubbing]]
- [[2026-10-05-voice-and-design-packs-spec]]
- [[2026-10-14-quiverai-tos-commercial-art-license]]
Sources
Vault
- [[2026-07-21-spanner-books-rights-outreach-draft]] — rights settled, audiobook named as family want, files-first correction
- [[2026-07-20-wilson-tree-v0]] — author's death date 2010-09-02, memoir titles and publisher, memoirs-unread status
- [[2026-05-05-alphasignal-voice-pro-local-dubbing]] — source separation as commodity step; RDCO pays for ElevenLabs
- [[2026-10-05-voice-and-design-packs-spec]] — ElevenLabs listed as a reusable RDCO capability
- [[2026-10-14-quiverai-tos-commercial-art-license]] — prior vendor-terms brief, same read-the-actual-ToU pattern
Web (all fetched 2026-10-25)
- ElevenLabs Professional Voice Cloning docs — https://elevenlabs.io/docs/eleven-creative/voices/voice-cloning/professional-voice-cloning
- ElevenLabs Instant Voice Cloning docs — https://elevenlabs.io/docs/eleven-creative/voices/voice-cloning/instant-voice-cloning (IVC figures reached via search excerpt of this page, not a direct fetch)
- ElevenLabs Terms of Use, Last Updated 09 October 2026 — https://elevenlabs.io/terms-of-use
- ACX audio submission requirements, article dated Apr 15 2026 — https://help.acx.com/s/article/acx-audio-submission-requirements
- Secondary/unverified on distributor AI-narration policy (flagged, not relied on): narrationbox.com, inkfluenceai.com, theauthorcentral.com blog posts surfaced via search; no distributor-owned policy page was reached for Spotify, Findaway/INaudio, Apple or Kobo.
Fetch note: two intended sources 404'd (ElevenLabs prohibited-content-policies, legacy acx.com help URL), so the prohibited-use policy proper was never read — the consent standard above rests on the PVC doc and the ToU only.