Producing an audiobook with a cloned voice

Updated October 28, 2026 · 9 min read

A ten-hour audiobook is not ten hours of work. It is thirty to forty hours in a booth, plus retakes, plus editing, plus proofing against the manuscript — and that is for a narrator who has done it before. Most authors who skip the audiobook edition do not skip it because they dislike the format. They skip it because the production cost is larger than the royalty they expect.

Cloning changes one variable in that equation: who performs the reading. It does not remove the work. It moves the work from the booth to the proofing pass, and it changes the cost from a per-finished-hour rate to a per-byte charge.

Where the time actually goes

• Performance — reading 90,000 words aloud at a sustainable pace, with consistent energy across sessions recorded weeks apart.
• Retakes — a misread word, a truck outside, a dry mouth. The ratio of good audio to recorded audio is rarely better than two to one.
• Editing and mastering — removing breaths, matching room tone between sessions, hitting loudness targets the store requires.
• Proof-listening — following the manuscript while listening, at 1.25× or 1.5×, catching the places where the reader said the wrong word and nobody noticed.

Cloning removes the first two and most of the third. The fourth stays, and it is the part that cannot be delegated to a model, because it is a judgement about whether the audio matches the book.

What cloning changes, concretely

You record once — a minute of yourself talking naturally, not reading — and the model learns your voice. Then the manuscript is read by that voice, chapter by chapter, as fast as the generation runs. Re-recording a paragraph you fluffed on page 212 costs nothing but the credits for that paragraph, and it happens at midnight if you want it to.

What you lose: the interpretation. A human narrator decides that this sentence is wry and that one is tired. A cloned voice reads what the text says, directed by the style tags you attach. For non-fiction, technical books, memoir written in a plain register, and anything where the author's voice is the value, that trade is usually worth making. For literary fiction with six characters and an unreliable narrator, it usually is not — see where AI narration actually falls down.

The workflow, in the order that avoids rework

1. Prepare the manuscript as plain text

Strip chapter decoration, drop the cover blurb and copyright page (they belong in the store listing, not the audio), and decide how chapter headings should be spoken. Do this before the first generation — re-cutting 30 chapters because the front matter was wrong is the most common beginner mistake.

2. Build a pronunciation list

Character names, place names, invented terminology, and homographs. Read, live and lead are the three that break most often in English, because the model has to choose a pronunciation from context and sometimes chooses wrong. Anything that must be said a specific way goes on the list, and the list gets tested with one paragraph before the chapter runs.

3. Clone the voice

A minute of conversational speech, quiet room, no music. Cloning is 100 credits, once, and the model stays in your list for every future project.

4. Generate in chapter-sized chunks

Not one enormous request. Chunking means a fix costs one chunk, and it means a failure halfway through chapter 19 does not cost you chapters 1 to 18.

5. Proof-listen against the manuscript

At 1.25× to 1.5×, with the text open. This is where the real hours go — expect roughly real-time for a first pass. You are listening for wrong words, flat delivery on a line that needed emphasis, and mispronunciations the list missed.

6. Re-render only what failed

Then assemble, normalise loudness, and check the store's technical requirements before uploading.

What it costs per finished hour

Speech is charged at 4,200 credits per million UTF-8 bytes. English is one byte per character, and a finished hour of narration runs about 9,000 words or 54,000 characters. The arithmetic, so you can check it against your own manuscript:

Credits are prepaid, $1 buys 50 credits, packs start at $9, and unused credits do not expire — which matters when the next book is eight months away. Non-Latin scripts cost three bytes per character, so Chinese or Japanese narration runs about three times the figures above for the same number of characters. Full line items are on the pricing page.

Before you produce thirty hours: check the store's policy

Audiobook retailers and distributors each set their own rules on synthetic narration, and several require disclosure or restrict which programmes accept it. Those rules change, and they differ by store and by territory. Confirm the current policy of the specific store you intend to publish on before you generate the whole book. Producing first and discovering the policy afterwards is an expensive way to learn it.

The QC checklist that catches most failures

• Numbers and dates. "1984" as a year versus a title. "3/4" as a fraction versus a date.
• Proper nouns. Every name on the pronunciation list, plus every name that was not on it.
• Homographs. Read, live, lead, bass, wind, tear, wound.
• Abbreviations at sentence start. "Dr." and "St." are the usual casualties.
• Chapter boundaries. The last line of chapter 4 and the first line of chapter 5, in one sitting — that is where a chunk boundary hides.
• Emphasis on negations. A flat "did not" sounds like "did". Worth a targeted re-render with a style tag.
• Loudness across chapters. Generate months apart and the levels drift; normalise at assembly, not per chapter.

Directing a voice that cannot take direction

Two levers do most of the work. Punctuation is prosody: a comma is a pause, a period is a fall, a dash is an interruption, and an em dash used as a beat sounds different from a full stop. Editing punctuation in the script is faster than fighting the model. Style tags set the register — calm, urgent, wry — without changing whose voice it is. Use them at chapter level for consistency, and only reach for line-level tags when a specific sentence needs it.

Questions people ask before starting

Related: how voice cloning works and what a clone costs, recording a sample that clones well, and where AI narration still fails.