The short version
- Pick or design a voice and save it, so every chapter uses the same speaker.
- Clean the text: remove page numbers, footnote marks and headers; spell out unusual names.
- Generate one chapter or section per generation, within your plan's length limit.
- Listen through, regenerate any section that sounds off, and join the files in your editor.
Choose the voice once
For anything longer than a few minutes, use a saved voice: a preset, one of My Voices, or a designed voice you saved. A saved voice is the same speaker in every generation, so chapter 12 sounds like chapter 1. Calm, steady voices hold up best over hours; try a few on one page of your real text before committing. See how to design a voice and how to clone one.
How much fits in one generation
| Plan | Characters per generation | About |
|---|---|---|
| Free | 1,000 | 1 minute |
| Starter | 3,000 | 3⅓ minutes |
| Creator | 10,000 | 11 minutes |
| Studio | 20,000 | 22 minutes |
Chinese, Japanese and Korean characters count three times. About 900 characters make a minute of speech, so a 70,000-word novel (around 400,000 characters) is roughly 7–8 hours of audio.
What happens with long texts
Texts over about a minute are generated in segments that end at paragraph or sentence breaks, then joined with natural pauses: about 0.75 seconds between paragraphs and 0.4 seconds between sentences. You hear the first segment within seconds while the rest is prepared. Later segments are usually ready ahead of playback; those are checked by speech recognition before you hear them, and one that skipped or garbled words is generated again automatically. If a part you already heard may contain a slip, the studio offers a free regeneration.
Prepare the text
- Remove page numbers, running headers, footnote markers and hyphenation from PDFs or e-books.
- Use paragraphs: a blank line gives a longer, natural pause.
- Spell out names and terms the way they're pronounced if they come out wrong.
- Numbers, dates and units are read out for you; check unusual formats once.
- Need a fixed silence, for example between scenes? Use
[pause 2s](see pauses and timing).
Levels and audiobook platforms
Every file comes out at a consistent loudness of about −18 LUFS, with peaks kept below about −1 dB, so chapters match each other without extra work. Audiobook platforms have their own rules (ACX, for example, asks for peaks below −3 dB and RMS between −23 and −18 dB): lower the volume by about 3 dB in your editor and check the result with the platform's own tools before uploading. Also check each platform's policy on AI narration; some require disclosure or don't accept it.
A workflow for a whole book
- Generate chapter 1 and listen to all of it. Adjust the voice or text rules now, not after 30 chapters.
- Generate the remaining chapters with the same saved voice. Download WAV for editing.
- Listen through and regenerate any section that sounds off. When the studio flags a take, its retake is free.
- Join the files in your editor, add room tone or music if you like, and export to the platform's format.
Rights
You may use the audio commercially. Make sure you have the rights to the text you narrate. Every file carries an inaudible AI watermark and, in downloads, signed Content Credentials marking it as AI-generated. Full rules in the Terms.