ElevenLabs Needed 200,000 Human-Voiced Audiobooks

ElevenLabs can produce a reading without booking a narrator. For ElevenReader, it licensed 200,000 audiobooks narrated by humans from major publishers. The voice company needed a library.

By May 2024, Audible had marked more than 40,000 books as made with an AI “virtual voice.” Bloomberg reported authors saving hundreds or thousands of dollars per title on narration—and listeners complaining about the influx of AI-made material. The same lower cost that let a book enter the audio market did not make anyone want to finish it.

The audiobook business is crossing from a production gate to a catalog gate: when software can narrate a text cheaply and adapt speech across languages, the harder work is securing permission to use the book and voice, deciding which recordings to admit, and helping listeners identify an edition worth their time. ElevenLabs’ human-voiced licenses make that reversal visible. Generating sound and assembling a listening service are different jobs.

Cheaper narration filled shelves without proving demand

Amazon made eight Polly voices free for Alexa developers in 2018, according to archived coverage. Developers could put speech inside a skill, but a voice available to read words was not itself a license to distribute books. The technical component arrived years before platforms began treating synthetic narration as a way to enlarge an audiobook catalog.

Apple launched an AI-narrated catalog in its Books app in January 2023. Audible’s later virtual-voice count showed what happened when generative AI met a large retail channel: authors could consider audio for titles whose expected sales might not have justified hundreds or thousands of dollars in narration costs. Audible’s 40,000-plus marked titles measured availability, not listening satisfaction. Bloomberg’s account contained both sides of that distinction.

Audible put conversion inside the publisher pipeline

In 2025, Audible partnered with US publishers to turn print books and e-books into audiobooks using more than 100 AI voices across English, Spanish, French, and Italian. Publishers brought the books; Audible supplied a route from text to recordings it could distribute. That arrangement made the publisher’s catalog, rather than an individual author’s recording budget, the unit of expansion.

Other distributors drew the boundary differently. Spotify began accepting ElevenLabs-made audiobooks in February 2025, subject to restrictions, and introduced an ElevenLabs-powered self-publishing tool in May 2026. In October, Amazon expanded its renamed Kindle+ subscription to include AI-narrated audiobooks. An author’s production tool, a publisher-conversion program, and a consumer bundle can all carry synthetic audio, but each operator makes its own decision about which files reach listeners.

More expressive voices multiply the uses that need approval

ElevenLabs says Dubbing v2 preserves a speaker’s emotion, tone, and pacing while synchronizing content across more than 90 languages. Google’s Gemini 3.1 Flash TTS supports more than 70 languages and audio tags for finer speech control. Pocket FM was already working with ElevenLabs in July 2024 to turn scripts into audio. Speech software no longer addresses only the plain reading of a finished English-language book; it can also serve serialized production and language adaptation.

Those capabilities do not establish that a translated audiobook will sound good to its audience, or that a publisher’s agreement permits every adaptation a model can produce. A platform can generate a new-language performance faster than it can secure the rights to distribute one. Publishers must ask who authorized this version of the work, in this voice, for these listeners, rather than whether a model can speak the words.

A narrator’s replica needs a contract, not just a sample

Audible planned to invite selected US narrators to train replicas of their voices and receive royalties for title recordings. Its royalty-sharing plan treated the performer’s permission and payment as parts of production, not matters to settle after a file had been made. Performers had reason to insist on the distinction: reporting in 2023 described people who had signed away voice rights in perpetuity and then faced competition from AI versions of themselves.

SAG-AFTRA’s 2025 interactive-media agreement requires informed performer consent for planned AI use of voice, likeness, and movement. That agreement covers interactive media, not the whole audiobook market; it shows how specifically a consent rule must identify permitted uses. Nor does a rule at one distributor protect every voice. Consumer Reports found flimsy barriers to nonconsensual impersonation at several cloning services, including ElevenLabs.

Audible’s proposed royalties are one form of consent architecture. ElevenLabs’ publisher licenses address another part of the same file: permission to offer the book. The distinction resembles the one in community-data licensing, where access to material and enforceable rights to commercialize it are separate things. A convincing voice cannot supply either permission on its own.

The listening shelf needs a way to refuse books

ElevenLabs did not fill ElevenReader solely with machine readings. Its licensed human-voiced catalog gives subscribers recordings that publishers had already authorized for distribution, and it preserves a place for performances listeners may prefer. That is a commercial answer to the complaints Bloomberg heard in 2024: making another title available does not settle whether its performance belongs in someone’s queue.

Publishers face a more basic screening problem outside licensed services. The New York Times reported growing numbers of pirated, synthetically narrated audiobooks on YouTube; removal was cumbersome enough that some publishers hired technology companies to help. A “virtual voice” label can tell a listener how a book was read. It cannot, by itself, prove that the text was licensed, the voice was used with consent, or the recording deserves a recommendation.

Audible’s shelf could hold tens of thousands of newly voiced books before listeners agreed that they wanted them. ElevenLabs could make the voices and still need publishers’ human recordings. The recording booth had stopped guarding the entrance; the signed rights page and the shelf label remained.