Prooftone masters your audiobook to retail spec in one shot — then hunts down every TTS glitch, stray breath, volume dip and splice click, word by word, so you upload with proof instead of hope.
MP3 or WAV, straight from ElevenLabs, your DAW, or your phone-in-a-closet setup. Nothing to configure.
A speech model re-reads every word with confidence scoring, diffs it against your manuscript, scans for splice clicks at sample level, and profiles loudness second by second.
RMS to −20 dB, peaks capped, room tone added, 192 kbps CBR out — plus a spec report and a clickable list of every suspect moment with spectrogram + slow-mo playback.
Built from the exact pipeline used to ship a real 5¾-hour AI-narrated audiobook — 30 files, three audit passes, every garble found and cut.
High-pass → compress → RMS to target → peak limit → head/tail room tone → 192 CBR / 44.1 kHz mono. Passes the seven retail spec checks, every time.
RMS −23…−18 · peak ≤ −3 dB · floor ≤ −60
Word-level confidence scoring catches TTS gibberish, warbled words, stray breaths, hallucinated phrases that aren't in your book, and “slash” read out loud.
Upload your book text — Prooftone flags anything the narration says that your manuscript doesn't, and knows your intentional jokes from real defects.
See volume dips a meter never shows. Localized quiet passages get flagged with exact in/out points and a suggested ramped boost.
Click any flag: spectrogram of that exact moment, plus normal and 60 % speed playback. Confirm with your ear in seconds, not by scrubbing.
Zoom to any timestamp, drag the handles over a garble, snip. The two halves are welded with a 25 ms crossfade and the seam is checked for clicks at sample level — an edit nobody can hear.
cut → crossfade → verify → undo
Unambiguous garbles are cut for you. Anything fused inside a word you need is deliberately held back and flagged — because a bad cut mangles the narration, and that is worse than a garble.
Hours = hours of finished audio processed per month. A typical audiobook is 6–12 hours.
“I found garbles in chapters I'd already ‘fixed’ — including a hallucinated sentence that was never in my book. Caught before a single listener heard it.”
— David B., indie author and Prooftone’s maker, on shipping their first audiobook with itIt was built on AI narration. TTS engines hallucinate phrases, warble words, read “/” as “slash”, and randomly drop volume — Prooftone's detection was tuned on exactly these failure modes across a real shipped audiobook.
Prooftone masters to the retail spec used by ACX and checks all seven gates: RMS −23 to −18 dB, peaks ≤ −3 dB, noise floor ≤ −60 dB, 192 kbps CBR MP3, 44.1 kHz, consistent channels, and head/tail room tone. You get a per-file report you can keep as proof.
Prooftone runs on your own Mac, so your unreleased narration never leaves your computer — nothing is uploaded anywhere. Your manuscript text is used only for the diff, and your audio is never used to train anything.
Only the duration of audio you run a scan on. Re-downloading masters, re-viewing reports, and zooming into flags are free. Re-scanning a fixed chapter counts only the chapter, not the book.
Yes — for separable defects (inserted gibberish, stray breaths, spoken “slash”, dead air) Prooftone proposes silence-anchored cut points with micro-crossfades and can apply them for you, then re-verifies the splice at sample level. Defects fused inside a needed word get flagged for a quick re-render instead, so we never mangle your narration.
Prooftone is in private beta on macOS. It already shipped a 5¾-hour AI-narrated audiobook through Audible’s retail checks; the four plans above open at public launch.
Mastering and the garble scan run locally. Unreleased narration never leaves your machine.
Thirty chapters, three audit passes, every fix re-verified — zero splice clicks introduced.
Demo Tape stays free. Indie, Author Pro and Studio open when the beta does.