Proofing an AI Audiobook: An Acceptance Checklist Before You Upload

Proof an AI audiobook by listening to the entire book once, logging each synthetic-voice error against its chapter, then checking the files against your store's specs and the credits and disclosure metadata. A full listen is the one step you cannot skip, because none of the AI narration tool documentation we read describes a check of the finished audio against your text. The rest of this page is what to verify around it.

Step 1: Listen to the whole book and log by type

The proof-listen means every chapter, start to finish, not a spot check. For a 90,000-word book that is roughly 9.7 hours by ACX's 2013 figure of 9,300 words per finished hour, which is our arithmetic, not a measured effort. Keep a log with the chapter, timestamp, error type and fix. These are the synthetic-voice error types worth logging, offered as a method.

Reading along with the manuscript is one way to spot dropped and repeated words. A pronunciation entry applies one fix to every instance, where editing the text means repeating it.

Accuracy and delivery are different checks

Proofing covers two questions: does the audio match the text, and does it sound right. Audio proofing tool Pozotron draws this distinction on its own blog, says its product flags differences between audio and text, and says it does not judge delivery. That is the vendor's description. We have not evaluated Pozotron, and the post does not address AI narration.

Professional proof-listener Maxwell Cantrell, writing on The Creative Penn in 2020 (the article does not discuss AI narration), suggests following the manuscript as you listen and, in a segment with frequent audio problems, setting the text aside to focus on the noises alone. Both ideas carry over to synthetic voices.

What a useful log looks like

A log entry should let you find the spot again and know what to change. These rows are invented to show the format:

Sort severity with a rule you set before you start. A sensible one is must fix for anything that changes meaning, misnames a character or leaves out words, fix if cheap for awkward pauses and odd stress, and accept for anything you would forgive in a human performance. Pozotron's post makes the related point that a flagged difference is a prompt for a decision, not an automatic correction.

Consistency errors are easier to catch when you check them directly. Google's Pronunciation check plays every occurrence of a word, for example, so you can run your character names before the full listen.

A worked proof-listening schedule (estimate)

Plan about a week of part-time work for a 90,000-word book, on the assumptions below. This is a planning estimate we built from one sourced figure and our own assumptions. It is not a measurement of how long proofing takes.

Assumptions:

• Runtime: 90,000 words ÷ 9,300 words per finished hour = 9.7 hours. The 9,300 figure is from a 2013 ACX post about human narration, so replace it with your file's real runtime as soon as you have one.
• Stop-and-log factor: 1.5 hours at the desk per hour of audio, because you pause to log and replay. This is an assumption with no study behind it.
• Sessions: two a day of 90 minutes each, at normal speed. Each session covers 1 finished hour (90 ÷ 1.5 = 60 minutes of audio).

Listening therefore takes 9.7 × 1.5 = 14.5 desk hours, at 2 finished hours a day.

Time your first chapter and replace the 1.5 with your own number (desk minutes ÷ chapter minutes). The table shows how much that one assumption moves the total:

Listen at normal speed for the accuracy pass. The short sessions are deliberate: Pozotron's post notes that fatigue can make even experienced proofers miss small errors on long listens.

Step 2: Make fixes in the right place

Where you can fix matters. Voices by INaudio's FAQ accepts only unmodified LPF packages from Google Play Books, ElevenLabs and Spoken Press, so all fixes must happen in the generating tool before export. See Voices by INaudio. On Spotify's creation tools, Spotify says the voice can be changed until you submit, and a change after publishing needs a new version, so finish proofing before you press publish.

Each tool also limits what a late fix can do, as read on 2026-11-08:

• Spotify creation tools. Pronunciations lock at Submit and Publish, and each project allows up to 4 generations, counting those after publishing. The EPUB cannot be replaced after publishing.
• Google Play. You can make several versions but only one can be live, and Google's pages do not explain how to edit an audiobook after it is live. See Google Play auto-narration.
• ElevenLabs Studio. A changed paragraph must be generated again, and Lock paragraph protects the ones you have approved. ElevenLabs describes its Stability setting as controlling randomness between generations, so a regenerated paragraph may not match the delivery around it. Re-hear the sentences on both sides of every fix.

Every regenerated passage goes back through the log. A fix that has not been heard is not finished.

Step 3: Check the files against the store's spec

A clean listen is not an accepted file. Specs differ by store, and an AI tool's export tier can affect them. ElevenLabs lists 128 kbps on lower plans and 192 kbps from Creator up, per its pricing page as read on 2026-11-08, but its Studio docs list 128 kbps MP3 for Free, Starter and Creator, and 192 kbps MP3 from Pro up, for newly created projects. Those two pages conflict, so confirm the bit rate of what you actually exported.

ACX publishes a detailed specification, which is a useful benchmark. It applies to ACX submissions, which must be narrated by a human unless otherwise authorized (see ACX and Audible), but you can use it as a benchmark for every route. From the ACX page, which displays April 15, 2026:

• Loudness between -23 dB and -18 dB RMS in each file, with peaks below -3 dB
• Noise floor no higher than -60 dB RMS
• Room tone of 1 to 5 seconds at the start and end of each file
• MP3 at 192 kbps CBR or higher, 44.1 kHz, mono or stereo but not mixed
• One chapter or section per file, each with a section header, and no file longer than 120 minutes
• Opening and closing credits as separate files, and a retail sample of five minutes or less

Other stores differ. Google Play's guidelines accept .mp3, .aac, .flac, .wav, .zip and .lpf with a minimum of five minutes and a maximum of 100 hours, and Apple's specification asks for the same bit rate across tracks. Check each in the audio specs guide, because we have not read every store's spec from a primary page.

None of the tool pages we read promises a loudness target. Studio's export settings include a volume normalization option, but the docs read give no target level, so measure the exported file in an audio editor rather than assuming it passes.

Acceptance by route

The route decides which checks apply. As read on 2026-11-08:

Step 4: Check credits and disclosure metadata

Labelling is part of acceptance. Confirm these before upload:

• Opening and closing credits state the title, author and narrator you are crediting. If you use ACX's rule as a model, the credits must match the metadata and cover.
• Spotify: the "This audiobook uses digital voice narration" option is ticked, and you know Spotify will add a sentence to the description. ElevenLabs' Spotify export lets you edit the narrator credit, which defaults to "Digital Voice" plus the voice name. Titles made with Spotify's own tools carry a "Narrated by Digital Voice" label.
• Kobo: the narrator contributor is a "Synthesized Voice" variant, per Kobo's FAQ as captured on 2026-11-08. A follow-up read that day was blocked.
• INaudio: the narrator credit and disclosure are set automatically, and the title must not be submitted as Human Narrated.
• Apple: the Audiobooks Specification lists an "AI Generated by" contributor role without explanation. It is not an acceptance rule. Apple's formatting guidelines say audiobooks narrated by AI must state this in the title or narrator field and reference it in the description.
• Embedded tags: ElevenLabs Studio can write title, author, ISBN and description into downloaded files. If you filled those in, check that they match your store listing.

The labelling guide collects store wording, and the APA and Audio Publishers Group guidelines suggest "AI Voice" as non-binding best practice.

The sign-off list

Do not upload until each line is true:

1. Every chapter has been heard from start to finish against the final text.
2. Every must-fix entry in the log is corrected and re-heard, with the neighboring sentences.
3. Front matter, back matter and credits are present and in the right order.
4. Chapter order and file count match the table of contents.
5. Format, bit rate, sample rate and channel layout have been read from the exported files.
6. Loudness and peaks have been measured if your store sets limits.
7. Narrator credit and AI disclosure match the rule for each store you are using.
8. Title, author and ISBN in the files match the listing.
9. For INaudio, the LPF package has not been opened, renamed or changed.

When the book passes

Keep your log and the final exported files together. Then return to the full workflow for packaging and upload, or the AI audiobooks hub for the next guide.

Last verified 2026-11-08. Rules and prices change: check the linked primary page before you rely on this.