// HOW-TO · AUDIO

How to create an audiobook with OpenAI (2026)

Turn a manuscript into an AI-narrated audiobook with OpenAI's gpt-4o-mini-tts: pick a voice, set a delivery style, chunk each chapter, and stitch the files.

Last verified · 2026-10-13 · by Moe Ameen

OpenAI does not ship an "audiobook" button. What it gives you is a text-to-speech endpoint — its newest model is gpt-4o-mini-tts — that reads text you send it in a chosen voice and returns an audio file. Turning a manuscript into a finished, chaptered audiobook is a workflow you assemble on top of that: clean the text, pick one voice and one delivery style, split each chapter into chunks the API will accept, generate the audio chunk by chunk, and stitch the pieces back into chapters and a single book. The payoff is real — a full-length narration for the price of a few dollars of API usage instead of hundreds to thousands for a studio and a human narrator — but the quality and the legality both depend on getting the details right.

This is the end-to-end pass for a non-fiction or fiction manuscript you already own the rights to. One honest note up front: OpenAI's TTS is an API product billed per use, not something a flat ChatGPT subscription unlocks — ChatGPT's own Read Aloud voice plays text back to you but does not export a file you can sell. If you want a shippable audiobook, you (or a tool you run) call the API. A related primer on the model options sits in [Gemini 3.8 Text-to-Speech](/ai-tools/gemini-3-8-text-to-speech) if you want to compare engines before you commit.

The steps

Clean and segment the manuscript. Start from plain text, one file per chapter, with the running headers, page numbers, and footnote markers stripped out — a TTS model will read every stray artifact out loud. Spell tricky names, acronyms, and foreign words the way they should sound (write "en-vy" style phonetic hints inline where a model reliably trips), and decide how you want numbers, dates, and abbreviations voiced. The text you feed the API is read literally, so the edit you do here is the single biggest lever on how professional the result sounds.
Pick one voice and lock it for the whole book. Call OpenAI's speech endpoint with gpt-4o-mini-tts and audition its built-in voices on a representative paragraph — the docs recommend marin or cedar for the highest quality, alongside options like alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, and verse. Choose one that fits the genre and your narrator persona, then never change it mid-book. A listener forgives an AI voice; they do not forgive the narrator's timbre shifting between chapter three and chapter four.
Write one delivery-style instruction. gpt-4o-mini-tts takes a separate instructions prompt that steers accent, emotional range, intonation, pacing, tone, and whispering independently of the text. Write one instruction that matches the book — "calm, measured, authoritative, unhurried" for non-fiction; "warm and expressive, slightly dramatized dialogue" for fiction — and reuse that exact string on every request. Changing the instruction between chunks is the fastest way to make a book sound like three different narrators.
Split each chapter into chunks under the input limit. A single request is capped at 4,096 characters of input text — OpenAI's API reference states this limit for the speech endpoint, and it applies to gpt-4o-mini-tts as well as tts-1 and tts-1-hd — so a full chapter has to be broken into chunks before you send it. Split on sentence or paragraph boundaries, never mid-word or mid-sentence, so the model's prosody resets cleanly between chunks instead of clipping a word in half. Keep chunks a consistent size so pacing stays even across the chapter.
Generate the audio chunk by chunk. Send each chunk with the same model, voice, and instructions, and request a lossless or high-bitrate format — the endpoint can return MP3 (the default), Opus, AAC, FLAC, WAV, or raw PCM. Generate to FLAC or WAV while you are still editing and down-convert to MP3 only at the final export, so stitching and mastering do not compound compression artifacts. Save chunks in order with zero-padded filenames so the next step concatenates them correctly.
Stitch chunks into chapters, then into the book. Concatenate a chapter's chunks into one file, then assemble the chapters, inserting a short beat of silence (roughly half a second to a second) between chunks and a longer pause at each chapter break so the narration breathes. Tools like ffmpeg do the join losslessly. Give each chapter its own file or marker if you are producing an M4B-style chaptered audiobook, which lets listeners navigate and resume.
Master, add the AI disclosure, and export. Normalize loudness to the target your distributor specifies (many audiobook stores expect around -18 to -23 LUFS integrated with a peak ceiling), trim long silences, and listen end to end at least once — a machine will mispronounce a name no spec check catches. OpenAI's usage policy requires you to disclose to listeners that the voice is AI-generated, so put that statement in your metadata and description. Then export your final masters in the format each store accepts.

Common gotchas

Sending a whole chapter in one request fails — the API caps input per call, so chunk first. Splitting mid-sentence instead of on a sentence boundary makes the prosody clip words; always break on punctuation.
Changing the voice or the instructions string between chunks produces audible tonal drift. Lock both for the entire book and reuse the exact same values on every request.
Treating ChatGPT's Read Aloud as an export path is a dead end — it plays audio back to you but gives you no file to sell. The shippable workflow is the API, billed per use, not a flat subscription.
Skipping the pronunciation pass means the model confidently mispronounces author names, place names, and acronyms. Fix them phonetically in the source text before you generate.
Generating straight to MP3 and re-encoding at each stitch stacks compression artifacts. Work in FLAC or WAV and down-convert once, at final export.
Assuming you can upload an OpenAI-narrated file anywhere. Major stores restrict AI narration or require their own engines — check each distributor's current policy before you produce the whole book.

Legal note

You must own or license the text you narrate, and OpenAI's usage policy requires a clear disclosure to listeners that the voice is AI-generated and not human — put that in the audiobook's metadata and description. Distribution rules vary and change: as of 2026, Audible/ACX standard submissions require human narration and prohibit unauthorized text-to-speech, steering AI narration through Amazon's own labeled Virtual Voice program instead of self-supplied files; Spotify and Google Play Books accept AI-narrated audiobooks with disclosure, though some distributors only ingest audio produced by specific providers. Confirm each store's current policy before you publish, and never clone a real person's voice without written consent.

Where Kompozy fits

OpenAI narrates the book. The problem the day it's done is that an audiobook is a file nobody can find — and [Kompozy](/) is the engine that turns the same manuscript into the launch that sells it. Kompozy is an AI content generation and multi-platform publishing engine, so it does not do the TTS narration (that's the API workflow above); what it does is take the book's text as a source and manufacture the marketing an audiobook launch lives or dies on, at the scale a book actually needs.

The distinct move here is working chapter by chapter. Feed Kompozy a chapter and it generates a Carousel Post that summarizes its argument, a Quote Graphic built around that chapter's strongest pull-quote, and a captioned [Persona Short](/glossary/persona-shorts) where your AI author avatar reads the chapter's hook aloud to camera — the on-screen counterpart to the voice listeners hear in the audiobook, with both governed by one written [Persona Brief](/glossary/persona-brief) so the author's spoken register and marketing voice match. A twenty-chapter book becomes twenty teaser threads instead of one 'my book is out' post, which is the difference between a launch day and a launch week.

Around that core it also drafts the announcement Blog Article and the launch Email Newsletter to your list, and a Marketing Short book trailer. Then [autopilot](/glossary/autopilot) schedules and fans the whole set across the eight social platforms plus blog and email from one queue, behind a per-post review gate so you approve every claim about the book before it ships. If you narrated the audiobook in more than one language, the [content translation workflow](/how-to/set-up-an-ai-content-translation-workflow) keeps the promo in step with each edition. Kompozy will not write your book or narrate it — that's the human and the API's job. It removes the per-chapter production ceiling that otherwise makes a real launch impossible for a solo author. Starter ($199/mo, 5,500 credits) fits one author launching one title; Pro ($499/mo, 18,000 credits) suits a publisher or team running multiple launches; Enterprise is custom for agencies managing many authors' catalogs.

Frequently asked questions

Which OpenAI model is best for making an audiobook?

gpt-4o-mini-tts is OpenAI's newest and most reliable text-to-speech model and the right default for an audiobook — it accepts a separate instructions prompt to steer tone and pacing, which the older tts-1 and tts-1-hd models do not. Within it, OpenAI recommends the marin and cedar voices for the highest quality. Audition a few on a sample paragraph, then lock one voice for the whole book.

Can I make an audiobook with just a ChatGPT subscription?

Not a sellable one. ChatGPT's Read Aloud feature plays text back to you inside the app but does not export an audio file you can publish. Producing an audiobook you can distribute means calling OpenAI's text-to-speech API, which is billed per use (pay-as-you-go), not unlocked by a flat ChatGPT plan. Budget it as API usage, not a subscription.

How much does it cost to narrate a book with OpenAI TTS?

Far less than a human narrator or studio — typically a few dollars of API usage for a full-length book rather than the hundreds to thousands a professional recording costs. Pricing is usage-based and changes, so estimate from your word count against OpenAI's current rates for gpt-4o-mini-tts before committing. The cost scales with how much text you run, so a clean pass that avoids re-generating chapters keeps it low.

How do I keep the AI voice consistent across the whole audiobook?

Pick one voice and one instructions string on day one and reuse both, unchanged, on every single request. Each chunk is a separate API call, so any difference in voice, model, or delivery instruction between calls produces audible drift that listeners notice as the narrator 'changing.' Lock the parameters, chunk on sentence boundaries, and master at a single loudness target to keep it seamless.

Can I sell an OpenAI-narrated audiobook on Audible or Spotify?

It depends on the store and the rules change. As of 2026, Audible/ACX standard submissions require human narration and route AI narration through Amazon's own labeled Virtual Voice program rather than accepting self-supplied AI files; Spotify and Google Play Books accept AI-narrated audiobooks with a disclosure, though some distributors only take audio generated by approved providers. Always disclose that the voice is AI and confirm each platform's current policy first.