AI Voice for Audiobooks
Published September 2, 2026 · Every claim below was read from the vendor’s own documentation on September 2, 2026. This page is about the constraints that decide a book-length project, not a ranking — for that, see the flagship comparison. No listening test has been run.
The short answer
A book takes weeks, and the model can change underneath you in that time. That is the constraint nobody plans for and the one that ruins a project, because a voice that shifts between chapter 1 and chapter 20 cannot be fixed by regenerating the last chapter.
Amazon publishes the warning itself, which is more than any competitor does: “any updates to the training data and the model could result in a slight variations to the way the voices sound, even in case when their overall quality improves”.
Cartesia sells the fix. It publishes dated model snapshots — “Snapshotted, will never change” — and tells you plainly to use one: “If ensuring consistent behavior is important for your use case, we recommend using a dated version in production.”
Then the mechanics: chapters must fit through the API. OpenAI caps a request at 4,096 characters; Speechify’s streaming endpoint takes 20,000 and names the use case — “Audiobook production: Process full chapters without hitting the 2K character limit”.
Basis: vendor documentation only, checked September 2, 2026 (how we verify).
On this page
The problem: voices drift between model updates
Every other consideration on this page is subordinate to this one, and it is the one least discussed anywhere.
You do not record an audiobook in an afternoon. You record it over days or weeks, in chapters, with revisions. Across that period the vendor may ship a model update. If your voice moves with it, the chapters you recorded before the update and the ones after it were made by different systems.
Amazon states the risk in its own documentation, and deserves credit for it: “A side effect of the technology is that any updates to the training data and the model could result in a slight variations to the way the voices sound, even in case when their overall quality improves”. The grammatical slip is Amazon’s and we reproduce it as published. The substance is a vendor telling you that improvements are also changes.
We have not measured drift, and no vendor publishes a measure of it. What we can tell you is which vendors give you a way to avoid the question.
Who lets you pin a version
Cartesia publishes dated snapshots and recommends them for exactly this. Its model table lists sonic-3.6-YYYY-MM-DD against the description “Snapshotted, will never change”, for “Customers who want to run internal evals before any updates”. And its guidance is unambiguous: “If ensuring consistent behavior is important for your use case, we recommend using a dated version in production. A dated snapshot never changes once released.”
That is the clearest answer to the drift problem published by anyone in this market, and it costs nothing to use.
Resemble pins by design. Its synthesis API “automatically uses the model associated with your voice_uuid” — the voice carries its model version with it. Note that Resemble no longer sells voice generation, so this matters only to existing customers.
Cartesia also migrates cloned voices forward, which is the opposite behaviour and equally documented: “As Cartesia releases newer models, we automatically make your voice available on them, so it keeps working when you upgrade.” Both statements are true because they describe different things — the model alias you call, and the cloned voice you call it with. For a book, pin the first and understand the second.
Murf documents deprecation rather than pinning: “Our legacy Gen2 streaming model is deprecated. We recommend using our latest Falcon 2 model instead. Gen2 will still be available to use via the non-streaming Synthesize Speech” endpoint. A deprecation notice is not version pinning, but it is advance warning, which is more than most offer.
For the rest we found no published policy on what happens to a voice when the model beneath it changes. That is a gap worth raising with a vendor directly before you commit a book to it.
Getting a chapter through the API
A chapter is commonly 3,000 to 6,000 words, which is roughly 18,000 to 36,000 characters. Whether that fits in one request varies by more than a factor of ten.
| Provider | Limit per request | What the vendor says about long text |
|---|---|---|
| ElevenLabs | 5,000–40,000 | Published per model; Multilingual v2 is described as “Most stable on long-form generations” |
| Speechify | 2,000 / 20,000 | Its streaming endpoint is documented for “Long-form text (articles, chapters)” |
| Amazon Polly | 6,000 | Async tasks accept 200,000 characters, of which 100,000 are billed |
| OpenAI | 4,096 | The tightest here. A 30,000-character chapter is at least eight requests |
| Resemble AI | 3,000 | Synchronous cap |
Speechify names the use case directly, which is unusual and useful: “Audiobook production — Process full chapters without hitting the 2K character limit”, with its /v1/audio/stream endpoint documented at 20,000 characters for “Long-form text (articles, chapters)”.
Where you must split, you are introducing seams, and no vendor documents how to make a split inaudible. That is a production problem the tooling does not solve for you.
Book-length jobs
For a whole book, the synchronous endpoint is the wrong tool. Five providers publish an asynchronous path, and three name long-form content in doing so.
- MiniMax names books specifically: “Applicable scenario: Speech generation for long texts, such as entire books.” Its flow is a create-task call returning a
task_id, then a query endpoint. - Azure names the audience: its batch synthesis API “can synthesize a large volume of text input (long and short) asynchronously. Publishers and audio content platforms can create long audio content in a batch. For example: audio bo[oks]”. It accepts up to 10,000 text inputs per job with a job lifetime of up to 31 days. One caution: “The Long Audio API will be retired on April 1st, 2027”, so check which API a tutorial is using.
- Google publishes Long Audio Synthesis, which “asynchronously synthesizes up to 1 million bytes on input” — the largest single job documented here — but it carries Pre-GA terms, offered “as is”. For a commercial book that is a real consideration.
- Amazon offers speech synthesis tasks at 200,000 characters, 100,000 billed.
- ElevenLabs publishes an async endpoint with webhook delivery, restricted to MP3 output only — which matters if your distributor wants lossless masters.
Names, places and invented words
Fiction is full of words no model has seen. A dictionary you define once and apply everywhere is worth more here than in any other use case, because the same invented name recurs three hundred times across a book and must be identical every time.
ElevenLabs and Cartesia both support dictionaries. ElevenLabs uses W3C pronunciation lexicon files — “Pronunciation dictionaries use XML-based .pls files” supporting “both IPA and CMU alphabets” — with the caveat that phoneme tags “only work with eleven_flash_v2 and eleven_v3 models”. Cartesia’s is “a simple search and replace” referenced per request by pronunciation_dict_id, accepting IPA or a “sounds-like” spelling.
OpenAI has no dictionary. Pronunciation guidance goes in the prompt, and its own advice is to “Keep to a short list”. For a novel with a large invented lexicon, that is a structural mismatch rather than an inconvenience. The full comparison of mechanisms is in our guide to how AI voice generators work.
Regenerating one line
You will need to re-record a sentence after a proof pass. Whether the regenerated line matches the audio around it is the question, and no vendor we read guarantees that identical input produces identical output. None documents a seed parameter for text to speech.
The signal runs the other way. Resemble exposes temperature, documented as controlling “generation randomness (0.1–5.0)”, with a default of 0.8 — non-zero by default. Amazon describes its generative engine as rendering “speech token by token” and refers to “the unlikely event of model hallucination”.
The practical consequence for a book: test a single-line regeneration against its neighbours before you commit to a provider, not after chapter 12.
A workflow that survives a long book
- Pin a dated model version if your vendor offers one. Cartesia does and recommends it. If yours does not, ask what happens to your voice at the next model release, and get the answer in writing.
- Build the pronunciation dictionary before chapter 1, not after you notice a name drifting. Confirm it survives a model change.
- Record a reference sample and keep it. Re-render it weekly. If it stops matching, the model moved.
- Use the async endpoint and check its output format. One vendor’s async path is MP3-only.
- Keep chapters within one request where you can, and where you cannot, split at paragraph boundaries you would have paused at anyway.
- Test a mid-chapter regeneration early. Nothing in the documentation promises it will match.
- Check the commercial position before distribution, in our guide to can you use AI voices commercially — four providers never state that you own the audio at all.
Frequently asked questions
Will the voice stay the same across a whole book?
Not guaranteed by anyone. Amazon warns that model updates “could result in a slight variations to the way the voices sound”. Cartesia is the one vendor publishing dated snapshots that “never change” and recommending them where consistency matters.
Can I send a whole chapter in one request?
It depends by more than a factor of ten. OpenAI caps at 4,096 characters, Amazon at 6,000, Speechify’s streaming endpoint at 20,000, and ElevenLabs runs 5,000 to 40,000 by model. A typical chapter is roughly 18,000 to 36,000 characters.
Which providers handle book-length jobs?
Five publish an async path. MiniMax names “entire books”; Azure names audio books and allows 10,000 inputs per job with a 31-day lifetime; Google accepts up to 1 million bytes but under Pre-GA terms; Amazon takes 200,000 characters; ElevenLabs is MP3-only on its async endpoint.
How do I stop character names being mispronounced?
Use a dictionary where one exists. ElevenLabs takes .pls files with IPA or CMU; Cartesia takes IPA or a sounds-like spelling. OpenAI has no dictionary at all and advises keeping prompt-based pronunciation lists short.
If I regenerate one sentence, will it match?
No vendor promises it. None documents a seed for text to speech, and Resemble ships a randomness parameter defaulting to 0.8. Test this before committing a project.
Which one sounds best for narration?
We have not tested that and do not rank it. Vendors position specific models for the job — ElevenLabs calls Multilingual v2 “Most stable on long-form generations” and Amazon sells a Long-form engine — but those are vendor claims, and what each charges for such claims is documented separately.
Sources and what we could not verify
Every claim was read from the vendor’s own documentation on September 2, 2026. Per-request limits and async endpoints are set out with sources in our API comparison; pronunciation mechanisms in our guide to how AI voice generators work; the method is on the methodology page.
What we could not verify
- How much a voice actually drifts between model updates. Amazon states variation occurs; nobody publishes a magnitude, and we have measured none.
- What happens to a voice at model release for most providers. Only Cartesia, Resemble and Murf publish anything approaching a policy.
- Whether any provider is deterministic. None states it; none documents a seed.
- Whether a pronunciation dictionary survives a model change. ElevenLabs gates phoneme tags by model, which suggests the question matters; no vendor answers it directly.
- How to split a chapter without an audible seam. No vendor documents this.
- How any of these sound. No listening test has been run.
Change log
September 2, 2026 — First publication. All claims read from official vendor documentation on September 2, 2026 and dated accordingly.
Published September 2, 2026 · This page shows no audio and reports no listening results, because our benchmark has not run. Model versions and limits change without notice; corrections are recorded with their date on the corrections page. Independence note: this page contains no affiliate links, and no vendor paid for placement or influenced the order — see how we make money and our editorial policy.