audiobook_production v1.0.0

name: "audiobook_production" description: "Produce audiobooks end to end: narration pacing and session practice, chapter mastering, ACX retail specs, and QC listens that catch mouth clicks. Use when recording, editing, mastering, or delivering an audiobook for Audible, ACX, Findaway Voices, or Authors Republic." metadata: { "includeInPrompt": true }

Audiobook Production

Purpose

Take a manuscript from raw narration to a retail-ready audiobook that passes distributor review on the first submission. Covers narration practice (pacing, session management, mic technique), editing (mouth clicks, breaths, pacing fixes), mastering to ACX specs, and the QC listen that catches what specs can't.

Workflow

1. Pre-production

• Manuscript finalized and proofread. Record from a clean script, never a draft.
• Build a pronunciation guide: every proper noun, foreign word, and technical term, with the intended reading. Verify each one before the first session; consistency across chapters matters more than perfection.
• Plan chapter files: one chapter or section per final file, under 120 minutes each. Record opening and closing credits as their own files.
• Budget time: a common narrator rule of thumb is roughly 8,000-9,000 words per finished hour of audio. Plan pickup time after editing, not just recording time.

2. Recording

• Record lossless: WAV or AIFF, 44.1 kHz or higher, 16-bit or 24-bit. Never record straight to MP3; encode only at the end.
• Set input gain so the loudest passages peak around -6 dB. Headroom is cheap, clipping is fatal.
• Mic 6-12 inches away, fixed position for the whole book. Mark the spot (tape on the floor) so every session matches.
• Record in 1-2 hour sessions with breaks. Voice drift across sessions is the most common consistency killer; shorter sessions plus a matched setup beat long heroic ones.
• Monitor on headphones and listen back immediately after each take. Catch a mic bump or a misread now, not in the edit.
• Capture a minute of clean room tone at the start of each session. You will need it for head/tail fills and edit patches.
• Use punch-and-roll (re-record the sentence in place) for mistakes instead of stopping and starting new files. Fewer files, cleaner session.

3. Narration pacing

• Pace serves the text: quicker through action, slower for dramatic weight and for ideas the listener needs to absorb.
• Pause grammar: a beat between sentences, a longer pause at paragraph breaks, a full stop for scene changes. If sentences run together in non-fiction, ideas lose impact; if pauses drag, the book feels slow.
• Keep natural breathing. Cut unusually loud breaths, gasps, and awkward pauses; leave the natural ones that belong to the performance.
• For character voices: subtle shifts in pitch, pace, and placement carry characters better than broad accents. Stay consistent: a character who sounds one way in chapter 3 must sound that way in chapter 30.

4. Editing

• Remove mistakes, false starts, and restart takes. Smooth every edit with a short crossfade so no hard cuts hit the ear.
• Mouth clicks, lip smacks, and saliva noise: fix them individually or selectively. The goal is that the listener stops noticing the recording, not that the voice sounds processed. See references/narration_editing.md for the click-removal pass.
• Apply noise reduction lightly, if at all. Heavy noise reduction hollows the voice and fails the ear before it fails the spec.
• Use gentle compression to even out perceived loudness while keeping the performance's dynamics. EQ serves the voice, it does not replace it: cut rumble and boxiness, add presence only if the recording needs it.

5. Mastering to spec

Target the ACX submission specs (full table in references/acx_specs.md):

• MP3, 192 kbps or higher, constant bit rate (CBR), 44.1 kHz
• RMS between -23 and -18 dB
• Sample peak no higher than -3 dB
• Noise floor no higher than -60 dB RMS
• One chapter per file, under 120 minutes, section header read aloud
• Room tone at the head and tail of each file (roughly 0.5-1 s head, 1-5 s tail), free of clicks and pops
• All files mono or all files stereo, never mixed; mono is the recommended path
• Opening and closing credits as separate files; retail sample 1-5 minutes

Typical ffmpeg mastering chain for a recorded WAV at sane levels:

bash
# Measure first, then apply measured values (two-pass loudnorm, linear)
ffmpeg -i chapter01.wav -af loudnorm=I=-20:TP=-3.0:LRA=11:print_format=json -f null - 2> loud.json
# Read measured_I, measured_TP, measured_LRA, measured_thresh, offset from loud.json,
# then apply with those values filled in
ffmpeg -i chapter01.wav -af loudnorm=I=-20:TP=-3.0:LRA=11:measured_I=-XX:measured_TP=-YY:measured_LRA=-ZZ:measured_thresh=-TT:offset=OO:linear=true -ar 44100 -ac 1 chapter01_master.wav
ffmpeg -i chapter01_master.wav -codec:a libmp3lame -b:a 192k -ar 44100 chapter01.mp3

Then run the bundled spec checker on every file before upload:

bash
python3 bin/acx_check.py chapter01.mp3

It verifies codec, bitrate, sample rate, RMS band, peak ceiling, noise floor, duration, channel layout, and flags missing head/tail room tone. Exit 0 means pass; nonzero lists the failures. Fix the master, not the MP3.

6. QC listen (the part specs can't do)

Specs catch numbers. Ears catch the rest. Listen to every file start to finish while reading along with the manuscript, and log each issue as chapter, timecode, issue. Catch list:

• Missing, added, or swapped words and sentences
• Mispronunciations or pronunciation drift across chapters
• Dialogue in the wrong character voice
• Mouth clicks, lip smacks, pops, plosives that survived the edit
• Background sounds: HVAC cycling on, chair creaks, page turns, phone buzz
• Strange pauses, rushed passages, rhythm breaks
• First and last 5 seconds of each file: room tone present, no truncated fades, no cut-off words
• File naming and ordering; opening credits first, closing credits last

Hand the timestamped list to the narrator for pickups, then re-listen to the fixed sections. A fresh pair of ears on the final QC pass catches what the editor's brain auto-corrected.

Output Contract

• Deliverables: one MP3 per chapter/section plus opening credits, closing credits, and a 1-5 minute retail sample, all passing bin/acx_check.py.
• Naming: zero-padded sequence so files sort in order (e.g. 00_opening_credits.mp3, 01_chapter01.mp3). Three digits once chapter counts pass 99.
• A pickup log (chapter, timecode, issue) for any QC findings, resolved before delivery.

Operating Rules

1. Never fake a spec pass: if the checker fails, remaster the WAV. Re-encoding an MP3 to "fix" levels degrades quality and usually still fails the noise floor.
2. Do not invent distributor specs. Verify against references/acx_specs.md, which cites primary sources. If a platform's requirements changed, re-verify, don't guess.
3. Mouth clicks and breaths get surgical treatment, not a blanket de-breather. Over-processing is audible on headphones and reads as amateur.
4. Consistency across the whole book beats perfection in one chapter. If session 4 sounds different from session 1, match it before mastering.
5. Record the pickup, don't patch around it. A clean re-read of the sentence beats any amount of editing a bad take.