About
About this tool
Check a chapter, correct its levels and verify the exported file against the measurable ACX limits. Narrator Check helps you move from a recording to a consistent delivery file, with your audio on your own device.
Made for the recording workflow
Check individual chapters, compare a whole book, prepare room tone and export a delivery file. Local processing keeps unreleased recordings on your own device. The three checking aids use a Google sign-in to start a session; their models then work in your browser.
The book workspace brings chapters, pickup review and the complete delivery package together. Save its metadata and notes to your account; recordings remain local. The developer API runs checks and fixes for scripts and production pipelines. It is opt-in and keyed, with Starter and Pro plans. The developers page states the current beta pricing and limits; choosing a plan does not upload files from the browser tools.
Why recordings stay local
Not as a promise, but as an architecture: the tools on this site never upload your audio: every check and fix runs in this tab. Your file is read, measured, corrected and encoded by code running on your own machine, and the finished MP3 is handed to you out of your browser’s memory. You can disconnect from the network after the page loads and everything still works, including the download.
The developer API is the only part of this site that processes files anywhere else, and only the files a key holder sends to it on purpose. Local check/fix functions never call its upload routes; checking aids call it to start a signed-in session, carrying no text and no audio. The workspace uses account routes only when you explicitly save or reopen project metadata, including measurements and individual review notes; recordings and the full manuscript remain local.
It matters more here than it would elsewhere. An unreleased manuscript is usually somebody else’s property under an agreement you signed, and “I uploaded it to a website” is not a defence.
What it will not do
- It will not tell you your audiobook will be accepted. It proves five measurable things; ACX’s reviewers judge pronunciation, extraneous sounds, credits and consistency, and no measurement here can stand in for a person listening.
- It will not rescue a noisy room. When your speech is not far enough above the room, no amount of processing produces a clean file, and the settings that make the meter pass make the recording breathe. The tool says so and offers no download, which is the opposite of what a tool with something to sell would do.
- It will not split a file over 120 minutes. Where a chapter breaks is an editorial decision.
How the numbers were checked
The five thresholds are read from ACX’s own published requirements, most recently on 9 August 2026. They are quoted, with that date, on the requirements page, so you can see exactly what this tool believes and when it last looked.
The engine was built against synthetic test signals and then run against real narration — public-domain LibriVox chapters from four different narrators and rooms. That found six defects the synthetic tests had not, including a limiter that was not limiting. It is the reason this page can say the tool works rather than that its tests pass.
What it is built on
The MP3 encoding and decoding is done by rusty_mp3 0.1.1 by Mata Network, written in Rust and compiled to WebAssembly, with a four-line patch that removes its use of the system clock so the build stays reproducible. It is used under the Apache License 2.0, and the full licence text ships with this site: read the Apache-2.0 licence.
EPUB and DOCX containers are read with fflate 0.8.2 under the MIT License, and PDFs are read with Mozilla PDF.js 6.2.108 under Apache License 2.0. Both are served from this site, run locally and receive no file over a network. Read the bundled fflate licence and PDF.js licence. The measurement engines, corrections and pages are this site’s own code, with no framework runtime.
The three checking aids — hear your manuscript, the pronunciation guide and proof-listen — run the Kokoro speech model, eSpeak NG and OpenAI’s Whisper in your browser through ONNX Runtime. They are there to check a text or a recording, never to narrate one. Every component, its licence and the source offer for eSpeak NG are on the licences page.
Not affiliated with ACX
ACX, Audible and Amazon are not connected to this tool and have not endorsed it. Their names appear here because describing whose published requirements are being measured is the only way to describe them.