BRENNEN

Industries / Media, arts and recreation / Media and entertainment / Publishing

Audiobooks

Audiobook publishing records books read aloud and sells or lends the recordings, which once came on records and tapes and now arrive mostly as downloads and streams.

5

Milestones achieved, from 1934 to 2026.

1

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1934

    Talking books for blind readers

    By spring 1934, 5,000 machines for playing recorded books had been ordered in 17 US states, so the Library of Congress released a $10,000 federal appropriation and the American Foundation for the Blind (AFB) began recording books on records on its behalf. The first titles included the Four Gospels, the Psalms, the Declaration of Independence, the US Constitution and works by Shakespeare. AFB signed a contract with the Authors' Guild and the National Association of Book Publishers that let it record books without paying royalties, for a nominal $25 a title, as long as the records were used only by blind readers.

    LessonThe first audiobooks needed a rights deal as much as a machine.

    afb.orgaudiopub.org

    Credit

    American Foundation for the Blind; Library of Congress

  2. 2009

    Downloads pass CDs

    Audiobooks moved from tapes and discs to files over about a decade. By the Audio Publishers Association's history, Audible brought out the first digital audio player in 1997, a deal with Apple put audiobooks on iTunes in 2003, CDs replaced cassettes as the leading format in 2004, and in 2009 digital downloads passed CDs as the most popular audiobook format.

    audiopub.org

    Credit

    Audible; Apple

  3. 2017

    An exclusive deal ends

    Apple bought the digital audiobooks for its iTunes Store only from Audible, an Amazon subsidiary, and Audible had agreed not to supply other digital music platforms. After a complaint from the German Publishers and Booksellers Association, Germany's competition authority, the Bundeskartellamt, and the European Commission examined the deal, and the two companies dropped the clauses with effect from January 2017.

    LessonThe authority's president, Andreas Mundt, said the deal hurt publishers' sales because, apart from Audible, they had few other buyers for digital audiobooks.

    bundeskartellamt.de

    Credit

    Börsenverein des Deutschen Buchhandels; Bundeskartellamt; European Commission

  4. 2023

    Machine voices go on sale

    Apple began selling audiobooks read by synthetic voices it calls "digital narration", starting with fiction and romance titles read by voices named Madison and Jackson. Smithsonian magazine, reporting the Guardian's story, said Apple approached independent publishers, covered production costs and paid authors royalties. Apple says the service is free and combines speech synthesis with linguists, quality control specialists and audio engineers. At the time Audible required human narration, while Google Play already offered AI narration tools.

    smithsonianmag.comauthors.apple.com

    Credit

    Apple

  5. 2026

    AI titles, little revenue

    The Audio Publishers Association's sales survey, run by the research firm Toluna, put audiobook sales revenue at $2.43 billion in 2025, 9% more than in 2024, with more than 750,000 active titles, up 43%. More AI-narrated audiobooks were published, yet they brought in 0.03% of sales revenue. Its consumer survey, run by Edison Research at SSRS, found that 16% of audiobook listeners had heard an AI-voiced audiobook and that willingness to try one fell from 70% in 2025 to 61% in 2026.

    LessonMore AI titles did not bring matching sales; in 2025 they earned a tiny share of revenue, and fewer listeners said they would try one.

    audiopub.org

    Credit

    Audio Publishers Association; Toluna; Edison Research at SSRS

Watch for

  1. Listeners rate AI cast higher

    In a blinded test, more than 1,000 US fiction audiobook listeners heard samples of a new science fiction thriller, either from a professional human narrator or from Spoken's AI "Multi-Cast" narration, which gives each character a different synthetic voice, unedited. The AI version scored higher on overall favorability (61% against 53%) and engagement, the human narrator scored higher on exposition without several characters, 61% of those who heard the AI version thought it was human (against 65% for the human version), and purchase intent was 46% against 49%, which the release calls statistically comparable.

    Achieved whena group not paid by a narration company repeats the blind test on several books and genres and publishes its data.

    podnews.net

    Claimed by

    Spoken, with Edison Research at SSRS, July 2026

Open

  1. 01

    Count the narration mistakes

    A word-level accuracy count for AI and human narration of the same texts. Take 50 public-domain books that exist both as volunteer human readings on LibriVox and as AI readings in the open-licence collection Microsoft, MIT and Project Gutenberg released in 2023; transcribe every recording with a speech-to-text model; align each transcript with the book's text; and list every word left out, added or changed. A person listens to a random sample of the flagged spots to separate narration mistakes from transcription mistakes and to mark mispronunciations.

    The brief
    Credit · unclaimed
  2. 02

    Read-along for LibriVox books

    A free library of read-along ebooks, in which each sentence lights up as the narrator speaks it, made from LibriVox recordings and Project Gutenberg texts. The files use EPUB 3 Media Overlays, the open standard that links each passage of an ebook to its place in a recording, and are made by forced alignment, software that finds where each sentence falls in the audio.

    The brief
    Credit · unclaimed
  3. 03

    A blind test, many books

    An independent listening test of synthetic against human narration across many books. Choose excerpts from 30 public-domain books of different kinds (dialogue-heavy fiction, exposition, poetry, non-fiction), each in a LibriVox human reading and in two or three synthetic readings made from the same text with stock voices only, never a copy of a real person's voice. Listeners rate each excerpt without knowing how it was made; the plan is registered before the test and all audio and ratings are published.

    The brief
    Credit · unclaimed

Open 01 · the brief

Count the narration mistakes

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A word-level accuracy count for AI and human narration of the same texts. Take 50 public-domain books that exist both as volunteer human readings on LibriVox and as AI readings in the open-licence collection Microsoft, MIT and Project Gutenberg released in 2023; transcribe every recording with a speech-to-text model; align each transcript with the book's text; and list every word left out, added or changed. A person listens to a random sample of the flagged spots to separate narration mistakes from transcription mistakes and to mark mispronunciations.

Why it is still open

The 2023 paper behind the collection, which says it released more than 5,000 audiobooks totaling about 35,000 hours, calls the books high-quality and human-quality but reports no count of reading errors. In March 2026 the linguist Mark Liberman compared The Economist's AI-read and human-read versions of a handful of articles by ear and noted a few parsing and emphasis errors and a repetitive sentence melody, without a count. Spoken's 2026 listening study measured listeners' ratings, not accuracy. SwanBench-Speech (May 2026, accepted at ACL 2026 Findings) scores AI speech models on 1,101 set long-form samples in 17 settings, audiobooks among them, with seven automatic measures; it tests models on its own passages, not finished audiobooks against human readings of the same books.

arxiv.orglanguagelog.ldc.upenn.eduarxiv.org

Built with what exists

LibriVox recordings, which are in the public domain, from more than 21,000 completed projects; the 2023 open-licence AI collection; Project Gutenberg texts; a speech-to-text model such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") or the open Whisper model; Python to align texts and count differences.

librivox.org

  1. First step

    Pick five books found in both collections, transcribe one chapter of each version and compare both with the printed text.

  2. Done looks like

    50 books with at least 500,000 narrated words on each side; omission, insertion and substitution rates per 10,000 words for human and AI narration, with confidence intervals; 1,000 flagged spots checked by ear, with the share that were real narration errors; all alignments published.

Open 02 · the brief

Read-along for LibriVox books

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A free library of read-along ebooks, in which each sentence lights up as the narrator speaks it, made from LibriVox recordings and Project Gutenberg texts. The files use EPUB 3 Media Overlays, the open standard that links each passage of an ebook to its place in a recording, and are made by forced alignment, software that finds where each sentence falls in the audio.

Why it is still open

The tools exist. Storyteller, an open-source platform by Shane Friedman, syncs audiobooks and ebooks that a person already owns into Media Overlay books, one private library at a time. The open-source tool syncabook makes such books too, but its companion collection, synclibrivox, holds files for 11 LibriVox books and was last updated in June 2020, against more than 21,000 completed LibriVox projects. This research found no larger public collection.

gitlab.comraw.githubusercontent.comgithub.comlibrivox.org

Built with what exists

Storyteller's forced aligner, or syncabook and its aligner afaligner; LibriVox MP3 files; Project Gutenberg texts; EPUBCheck, the standard validator for EPUB files; an ebook reader that plays Media Overlays to check each book by ear.

  1. First step

    Rebuild the 11 synclibrivox books with syncabook, then time 100 sentences by hand to set a starting accuracy.

  2. Done looks like

    1,000 books that pass EPUBCheck; in a hand-timed sample of 2,000 sentences across 50 books, 95% highlighted within half a second of when the narrator starts them; all files released openly.

Open 03 · the brief

A blind test, many books

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An independent listening test of synthetic against human narration across many books. Choose excerpts from 30 public-domain books of different kinds (dialogue-heavy fiction, exposition, poetry, non-fiction), each in a LibriVox human reading and in two or three synthetic readings made from the same text with stock voices only, never a copy of a real person's voice. Listeners rate each excerpt without knowing how it was made; the plan is registered before the test and all audio and ratings are published.

Why it is still open

The largest blind test found, Spoken's 2026 study with Edison Research, used samples of one science fiction thriller and Spoken's own system, and the human narrator did better on exposition without several characters. The best-known academic test of voices for long listening, by Mozilla and Carnegie Mellon University (CHI 2020), played one article read by 27 voices, synthetic and human, to 50 listeners each, and found some synthetic voices rated above some human readers; those voices are now several generations old, and the text was an article, not a book. The Audio Publishers Association reports that only 16% of audiobook listeners have heard an AI-voiced audiobook. This research found no independent test across several books and genres with published data.

podnews.netblog.mozilla.orgaudiopub.org

Built with what exists

LibriVox recordings; commercial text-to-speech services with stock voices and an open-weight voice model; a survey platform that pays participants, such as Prolific; the Open Science Framework to register the plan; Python or R for the statistics.

  1. First step

    Pick 10 passages that have LibriVox readings, make two synthetic versions of each, and run a pilot with 100 listeners.

  2. Done looks like

    At least 1,000 listeners and 30 books; ratings and the share of AI excerpts judged human, by genre, with confidence intervals; the plan registered before data collection; all audio and responses released.