Best AI voice generator for audiobooks: what a book costs
The useful thing about an AI voice generator for audiobooks is not the voice — it is that the bill can be knowable from the manuscript. Two of the three speech models in the Actionway catalog price by the character, so with either of those a book that has been written has already been costed. We ran the same 373-character passage through two models from a Codex session: $0.02 for 373 characters at $0.002 per 40 characters (text) each, about twenty seconds of audio, and the amounts matched the catalog to the cent. Both clips are below.
連接你的 AgentOne passage, two default voices
Audiobook narration is a different job from character dialogue. It runs long, it has to stay consistent for hours, and the read wants to be level rather than performed. So the test is not "which model does the best accent" — it is what a plain narrative paragraph sounds like with nothing tuned, because that is what you will be listening to for four hours.
你這樣問
The harbour had emptied by the time she reached it. Rope lay coiled on the stones where the last boat had been tied, and the water moved without hurry against the wall. She set the case down, straightened, and looked back once at the road she had come down. Nobody was following. That was the part she had not planned for, and it unsettled her more than pursuit would have.
373 characters, billed in whole blocks of 40, so ten blocks. Sent to both models with no voice chosen and no speed, pitch, or emotion set.
MiniMax TTS · 20.4 s of audio · $0.02 for 373 characters at $0.002 per 40 characters (text)
VoxCPM · 21.0 s of audio · $0.02 for 373 characters at $0.002 per 40 characters (text)
They are close, and for narration that is the point rather than a disappointment. MiniMax reads a little quicker and flatter; VoxCPM sits slightly slower and softer, and takes about half a second longer over the same words. Either would carry a chapter. Neither needed a voice picked from a library to be listenable, which matters when the alternative is auditioning voices before you know whether the project is affordable at all.
Pricing a whole book before you start
This is the part that changes how you plan. On the two character-priced models the charge follows the character count, so the manuscript already contains the number. A chapter of 20,000 characters is $1.00 for 20000 characters at $0.002 per 40 characters (text). A 300,000-character book, split into fifteen chapters of that size, is $15.00 for 15 runs of 20000 characters at $0.002 per 40 characters (text) — and that total is arithmetic you can do before recording anything, not an invoice you discover afterwards.
| Job | Characters | Cost at catalog prices |
|---|---|---|
| The passage above | 373 | $0.02 for 373 characters at $0.002 per 40 characters (text) |
| One chapter | 20,000 | $1.00 for 20000 characters at $0.002 per 40 characters (text) |
| A fifteen-chapter book | 300,000 | $15.00 for 15 runs of 20000 characters at $0.002 per 40 characters (text) |
For running time, our passage gives a rate to extend: 373 characters produced 20.4 seconds, or roughly eighteen characters a second. On that basis a 300,000-character book lands near four and a half hours of finished audio. Treat that as an estimate — it comes from one passage at one model's default speed, and a different book, a different model, or a changed speed setting will move it.
What to decide before the first chapter
Is the text final?
The charge follows the characters you send, so an edited paragraph is a new call. Narrate after the copy edit, not before — at these prices a re-run is cheap, but a whole re-record of a changed book is still fifteen chapters of calls.
One call per chapter, not per book
Each call rounds up to whole blocks of 40 characters, so many tiny calls waste a little on every rounding. Chapter-sized calls keep the rounding negligible and give you files you can re-run individually.
Listen to one chapter before committing
A voice that is fine for twenty seconds can wear over an hour. Run one full chapter, listen to the end of it, and only then price the rest.
Keep the audio and the text together
Name the output files after the chapters. When chapter nine changes, you want to re-run one call rather than work out which of fifteen files is stale.
Running it from Codex
你這樣問
Narrate chapters/ch-01.txt and save the audio next to it, then tell me what the whole book would cost at that rate.
The agent lists the speech models with their prices before it runs anything; picking one is the approval. The mp3 lands in your project and the arithmetic for the rest comes back in the same reply.
Codex has no speech model of its own; the Skill installed with the CLI is what adds these services to the session through the voice generation capability. Every run is priced before it starts, and if the wallet cannot cover it the agent stops before anything is charged. The same session can also transcribe existing audio if you are working the other direction, and the character voice guide covers the short-line case where the rounding actually matters.
Set it up once
npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.codex/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.codex/skills/actionway"You need Node.js 20 or newer and an Actionway account. Claude Code users copy the same folder into ~/.claude/skills/actionway instead, and both routes reach the same catalog and the same wallet — the get started page lists the steps per client side by side.
常見問題
How much does it cost to narrate a book with AI?On MiniMax TTS and VoxCPM it is priced per 40 characters of script, so the manuscript gives you the number up front: a 20,000-character chapter is $1.00 for 20000 characters at $0.002 per 40 characters (text), and fifteen of them come to $15.00 for 15 runs of 20000 characters at $0.002 per 40 characters (text). Seed Audio is the exception — it settles on the seconds of audio returned, so its amount is known only after the run.
Which voice model is best for audiobook narration?Listen to the two clips above — same passage, same price, no settings tuned. MiniMax reads quicker and flatter, VoxCPM slower and softer. For narration, pick the one you can still stand at the end of a chapter, not the one that sounds best in twenty seconds.
Can Codex generate audiobook narration?Not on its own; it has no speech model. With the Actionway Skill installed it calls one — MiniMax TTS and VoxCPM at $0.002 per 40 characters (text), Seed Audio settled per second of audio — and saves the mp3 into your project, so a chapter file in, an audio file out, in the same session.