calesthio/generative-media-skills/audiobook-production

Audiobook production

An audiobook is not a long TTS clip. It is a structured deliverable of many chapterized files, each of which must pass an automated loudness/noise gate, carry the right metadata, and sound like the same performer that opened Chapter

1. The problems that dominate this work — voice drift over ten hours, a proper

noun mispronounced 40 times, a footnote that makes no sense read aloud, a file rejected for a -58 dB noise floor, a platform that silently bans AI narration — do not exist in short-clip synthesis. This skill is about managing hour-scale production, not about generating one good sentence.

This is a craft skill, provider-neutral. Generative voice tools are referenced by capability (pronunciation lexicon, SSML support, seed/consistency control, per-chapter regeneration), not by brand. Specific tools and retail platforms are named only as illustrative examples or as dated policy facts.

Labels used below:

• [Fact] — documented in a primary/official source, cited.
• [Standard] — an established industry technical standard.
• [Heuristic] — a production judgment that experienced producers use; not a rule.
• [Policy — dated] — a volatile platform policy verified on the stated date.

1. Decide the production route first

Three routes exist, and they diverge on cost, rights, quality ceiling, and where you can sell. Pick before touching the manuscript, because the route changes how you prepare it.

1. Human narration — a person performs the book. Highest quality ceiling,

required by some retailers, needs a performer and studio-grade audio.

1. Author/producer-driven generative narration — you supply the text to a

TTS system, tune pronunciation and pacing, and master the output yourself. You control every file and can distribute the finished audio wide.

1. Platform auto-narration — a retailer generates the audiobook from your

ebook inside their walled system (e.g., Amazon's Virtual Voice, Google Play Books auto-narration, Apple Books digital narration). Lowest effort, least control, and distribution is often tied to that platform.

The rest of this skill mostly serves route 2 (the one where an agent does real production work) and gives route-3 platform facts in §10.

2. Prepare the manuscript for narration

Print text is written to be seen. Narration text must work when heard, with no page to glance back at. Preparing the manuscript is where most audiobook quality is won or lost, and it is entirely upstream of the voice tool.

Front and back matter — decide, don't default

[Heuristic] Treat each front/back-matter element as a separate short segment and make an explicit include/omit decision:

• Title page / copyright — usually spoken briefly. Accessibility-focused

production (see §12) narrates copyright, title, and section labels in full; commercial retail production often compresses them. [Fact] Accessibility guidance is to begin each section by speaking its title ("Copyright", "Chapter One"). (NNELS accessibility guidelines, verified 2026-07-10)

• Dedication, epigraph, acknowledgements, also-by lists — short standalone

segments if kept; frequently omitted from commercial audio. Decide per title.

• Opening/closing credits — retail platforms expect them (see §9). Opening

credits state title, author, narrator; closing credits signal finality ("You have been listening to…" then title/author/narrator, then "The End").

Footnotes and endnotes — the hardest call in nonfiction

[Heuristic] A footnote read aloud mid-sentence derails the listener because there is no visual marker for "this is an aside." For each note, choose one of:

• Fold into the sentence — rewrite so the note's essential fact becomes part

of the spoken prose.

• Read in place, clearly bracketed — introduce it audibly ("In a footnote,

the author adds…") so the listener knows it is an aside.

• Batch at end of chapter/section — collect notes and read them together.
• Drop — citation-only notes (page numbers, ibid., bare URLs) usually add

nothing in audio and are cut.

For scholarly/accessibility editions the default flips toward including notes and bibliographies so the audio carries all the print information. (NNELS, verified 2026-07-10)

Tables, figures, charts, images — rewrite or cut

[Heuristic] A narrator cannot read a chart. Flag everything that depends on the reader seeing it — tables, graphs, figures, image captions, sidebars, complex formatting. For each: extract the essential finding as one or two spoken sentences ("Table 3 shows sales roughly doubling each year from 2020 to 2024"), or cut it. This is why platform auto-narration explicitly warns against image/table-heavy books and cookbooks/coloring books being unsuitable. (Amazon KDP Virtual Voice eligibility, verified 2026-07-10)

Inline text that must be spoken out

Build a pre-pass that resolves anything ambiguous when voiced:

• Numbers — "$1,200" → "twelve hundred dollars" or "one thousand two hundred

dollars" (pick per context); "1990s" → "nineteen nineties"; "III" → "the third". TTS number handling is inconsistent; normalize deliberately.

• Abbreviations/acronyms — decide spell-out vs. say-as-word ("NASA" as a

word; "e.g." → "for example"; "St." → "Saint" or "Street" depending on use).

• URLs and emails — almost always rewritten ("visit the site linked in the

book description") rather than read character by character.

• Symbols — %, &, #, ° must be expanded.

Dialogue attribution

[Heuristic] In print, "she said angrily" tells the reader the tone. In audio the performance should carry the tone, so heavy adverbial tags can feel redundant — but for generative narration you often cannot rely on the voice to act, so you may need to keep attribution the human ear would find obvious, or add bracketed performance cues the tool can act on where the tool supports them. Whichever way, make sure every line of dialogue can be attributed by ear: a long unbroken back-and-forth with no tags leaves the listener unsure who is speaking.

3. Cast the narration

Single narrator vs. multi-voice vs. full cast

[Fact/Heuristic] Two base classes: solo (one voice performs the whole book, voicing all characters) and multicast (two or more voices). Full cast is the cinematic extreme — a distinct actor per character, dialogue tags stripped, sometimes sound design added. (Swift Publishing, Spines, verified 2026-07-10)

Decision drivers:

• Point of view. First-person narration usually points to a single

narrator (the listener bonds with one voice = the protagonist). Third-person ensemble stories tolerate or benefit from multicast. [Heuristic]

• Genre. Nonfiction, memoir, and self-development are overwhelmingly single

narrator — it reads like a lecture or a personal address. SFF and heavily dialogued fiction gravitate to full/dual cast. [Heuristic, supported by practitioner consensus above]

• Cost and assembly. [Fact] Full-cast raises cost and, more importantly,

editing burden — you cast, direct, and splice many performers' files together. (Spines, verified 2026-07-10)

• Listener preference. [Fact] An Audio Publishers Association figure cited

widely holds that a majority of listeners enjoy the experience more with distinct character voices — but this is about differentiation, not necessarily separate actors; a skilled solo narrator supplies it too. (Reported via Spines, verified 2026-07-10 — treat the exact percentage as secondary.)

For generative production, "full cast" is cheap to attempt (assign different synthetic voices to characters) but expensive to make good: you now have multiple voices that each must stay consistent, plus the assembly problem. Many platform auto-narration tools support adding character voices within one book. (Google Play Books auto-narration, verified 2026-07-10)

Character differentiation without caricature

[Heuristic] The goal is that a listener can tell who is speaking, not that every character is a cartoon. For a solo narrator (human or synthetic):

• Differentiate by pitch, pace, and energy first, accent last. Accents drift

and offend more easily than a slightly lower, slower register does.

• Keep each character's voice reproducible: write down the parameter settings

or the descriptive anchor for each character so Chapter 12's dialogue matches Chapter 2's. With generative voices this means locking a voice/seed/style per character and logging it.

• Avoid demographic caricature — do not signal a character's ethnicity, age, or

gender through stereotype. Restraint reads as skill.

4. Hour-scale synthesis: the problems short clips never have

This is the technical core that distinguishes audiobook production from TTS. A 100k-word book is roughly 8–12 hours of audio. (Fish Audio, verified 2026-07-10)

Voice drift

[Observation, well-documented] Across many hours the timbre, energy, and even pace of a synthetic voice can shift — the tone that worked in Chapter 1 is subtly different by Chapter 15 — because chunks are generated independently and random sampling nudges each one. (Fish Audio, verified 2026-07-10)

Mitigations:

• Fix the seed / consistency control where the tool exposes one, so sampling

is deterministic across chunks. [Heuristic, tool-dependent]

• Use larger chunks rather than sentence-by-sentence generation; more context

per generation reduces inter-chunk timbre jumps. (tts-audiobook-tool, Fish Audio, verified 2026-07-10)

• Anchor voice settings once and reuse them for every chapter — never let the

tool re-pick a voice per session.

• Spot-check across the book — listen to the first minute of Chapter 1, a

middle chapter, and the last chapter back-to-back specifically for drift, not content.

Chunking strategy

[Heuristic] Chunk on semantic boundaries (sentence/paragraph), not fixed character counts, so a chunk never cuts mid-clause. Keep chunks as large as the model reliably handles (a common working figure is on the order of ~80 words per segment for models that stay accurate at that length; smaller for models that degrade on long input). (Qwen3 long-form TTS, verified 2026-07-10) Match chunk size to the specific model's stability, not a universal number.

Per-chapter architecture

[Heuristic] Generate and store one chapter per file/session. This gives QC checkpoints and — critically — means fixing Chapter 15 does not force regenerating the whole book. (tts-audiobook-tool, Fish Audio, verified 2026-07-10) It also maps directly onto retail chapterization (§9).

Consistent pacing