◆ Fantasy Audiobook Studio

Improvements

The skills get better in four ways. Two depend on the author's ear, and two are automatic. Every lesson ends up as a rule in a skill, so the next book starts with it already in place.

  1. The author reports"This sounds wrong." It becomes a training case with before/after audio; the author picks the fix.
  2. The automatic listenerSpeech recognition transcribes every chapter and compares it with the script, and turns up problems nobody reported.
  3. Rule testsEvery rule has a before/after test that must pass before publishing, so a new rule can't undo an old fix.
  4. Rule proposalsWhen a pattern repeats, Claude drafts the rule; the author approves it. Taste stays with the author.

1 · Training cases

The original text, what was heard, why it happened, and audio of the before and each attempted fix. Newest first.

08

"want it" runs together in the American voice

Second-Hand Dragons "I want it on the record" heard as "I wanted on the record" in af_heart, 4× (found by listen-check)

Rule:

avoid "want it" + vowel in af_heart; rephrased to "Let the record show that…"

fixed in the book; rule proposal P3 07

Too many clipped lines of dialogue

too many two- and three-word lines of dialogue; it feels lazy (author, whole books)

Rule:

clipped lines ≤ ~1 in 7 per chapter, no two in a row unless the scene should snap; rewrite filler, keep decisions/confessions/punchlines

rule (scanner + drafting checklist); existing chapters not revised 06

Two-word paragraphs can come out garbled

ch06 "A wall." heard as "A while." (found by listen-check)

Rule:

tiny standalone paragraph → beat or merge (proposal P2)

fixed in the second edition: "She sat." merged, "A wall." She didn't make it a question. (P2 itself still proposed) 05

Short lines get an emotion beat

short lines like "Can I?" carry no emotion by ear

Rule:

add a small physical beat that shows the emotion (≤ ~8 words); pace it in fast exchanges

fixed (rule; ch01 done, ch02–13 in the speech-flow pass) 04

A one-word hedge boxed in commas sounds choppy

ch01 "not softness, exactly, but attention" sounds choppy

Rule:

no commas around a one-word hedge before but

fixed (applied to ch01) 03

No breath before the result in a long "so … that"

ch01 "so much … who they were that by the time…" runs together

Rule:

comma before that in a long so … that

fixed (applied to ch01) 02

Chapter 1 speech-flow check (first run of the speech-flow skill)

ch01 speech-flow run: commas, unclosed dashes, uncued italics, long sentences, far referent

Rule:

see 02-ch01-flow.md; checks live in speech-flow

fixed (applied to ch01) 01

Short question sounds cut off

ch01 "Can I?" sounds cut off, not like a question

Rule:

short question with no main stress → (+2) on the key word + ??; better: give it a tag or beat in the manuscript

fixed

2 · Automatic listener

The latest listen-check report for each chapter. A flag means "worth a listen", not a confirmed error: each one is double-checked with a second model, then by ear.

ChapterLengthCheckedResult
ch018:162026-10-192 to check 2 dismissed by the second model
  • 7:24 swapped: expected “id”, heard “i”
  • ¶21 pace slow: 2.03 words/s (chapter 3.16)
ch027:122026-10-191 to check 2 dismissed by the second model
  • 0:56 swapped: expected “this”, heard “the”
ch037:072026-10-193 to check 7 dismissed by the second model
  • 4:56 swapped: expected “this”, heard “the”
  • ¶0 pace slow: 2.24 words/s (chapter 3.24)
  • ¶38 pace slow: 2.0 words/s (chapter 3.24)
ch046:562026-10-19clean 5 dismissed by the second model
ch057:572026-10-191 to check 3 dismissed by the second model
  • 1:57 swapped: expected “it”, heard “at”
ch067:082026-10-194 to check 12 dismissed by the second model
  • 3:24 swapped: expected “this”, heard “the”
  • 4:35 swapped: expected “this”, heard “the”
  • ¶6 pace slow: 1.57 words/s (chapter 3.03)
  • ¶38 pace slow: 2.0 words/s (chapter 3.03)
ch077:592026-10-191 to check 5 dismissed by the second model
  • ¶14 pace slow: 2.1 words/s (chapter 3.27)
ch087:272026-10-191 to check 3 dismissed by the second model
  • 6:34 swapped: expected “this”, heard “the”
ch097:532026-10-19clean 3 dismissed by the second model
ch108:182026-10-191 to check 2 dismissed by the second model
  • 6:14 swapped: expected “from”, heard “for”
ch119:352026-10-191 to check 4 dismissed by the second model
  • 1:53 swapped: expected “wholl”, heard “will”
ch128:092026-10-193 to check 3 dismissed by the second model
  • 5:38 swapped: expected “bought”, heard “bot”
  • 5:40 swapped: expected “bought”, heard “bot”
  • ¶10 pace slow: 1.98 words/s (chapter 3.13)
ch137:062026-10-191 to check 5 dismissed by the second model
  • ¶8 pace slow: 2.11 words/s (chapter 3.19)
ch016:162026-10-16clean 1 dismissed by the second model
ch027:542026-10-16clean 2 dismissed by the second model
ch038:172026-10-16clean 5 dismissed by the second model
ch048:332026-10-161 to check 3 dismissed by the second model
  • ¶35 pace slow: 2.16 words/s (chapter 3.14)
ch057:532026-10-16clean 3 dismissed by the second model
ch069:102026-10-16clean 2 dismissed by the second model
ch078:082026-10-161 to check 5 dismissed by the second model
  • 1:20 swapped: expected “it”, heard “at”
ch089:522026-10-16clean 6 dismissed by the second model
ch096:592026-10-161 to check 4 dismissed by the second model
  • ¶22 pace slow: 2.05 words/s (chapter 3.09)
ch107:202026-10-16clean 3 dismissed by the second model
ch117:172026-10-16clean 2 dismissed by the second model
ch128:122026-10-16clean
ch1310:522026-10-16clean 8 dismissed by the second model
ch145:002026-10-161 to check
  • ¶12 pace fast: 4.0 words/s (chapter 3.04)

3 · Rule tests 11/11 passing

Run at every build. The scanner must flag the before text and leave the after text alone.

FromCheckMust flagMust not flag
✅case 01bare short line"Can I?""Can I?" He sat up a little.
✅case 02comma-before-andShe finally looked up at him, and allowed herself the smallest flicker of kindness.She finally looked up at him and allowed herself the smallest flicker of kindness.
✅case 02italics-without-cueMira had spent three years teaching herself to hear it as *I know, love. I know.*Mira had spent three years teaching herself to hear it say: *I know, love. I know.*
✅case 02italics-without-cueMira nodded slowly and stamped the file. *No appeal filed.* Case closed.Mira nodded slowly and stamped the file: *No appeal filed.* Case closed.
✅case 02interrupter-without-commasa sugar bowl that had by all accounts been a very sweet mana sugar bowl that had, by all accounts, been a very sweet man
✅case 02unclosed-dashThe chair creaked again — longer this time, a low settling groan that might have meant anything, or nothing, and Mira had spent three years listening to it.The chair creaked again — longer this time, a low settling groan that might have meant anything, or nothing. Mira had spent three years listening to it.
✅case 02long-sentenceShe logged it, tagged it for collection, and moved it gently to the shelf behind her desk, where it joined a coat rack, a sugar bowl that had, by all accounts, been a very sweet man, and a rather handsome mantel clock that hummed faintly when it was pleased.She logged it, tagged it for collection, and moved it gently to the shelf behind her desk. There it joined a coat rack, a sugar bowl that had, by all accounts, been a very sweet man, and a rather handsome mantel clock that hummed faintly when it was pleased.
✅case 03so…thatPeople spent so much of their lives arguing about who they were that by the time they died, most of them were relieved.People spent so much of their lives arguing about who they were, that by the time they died, most of them were relieved.
✅case 04hedge boxed in commasthe thing that passed for kindness in this job — not softness, exactly, but attention.the thing that passed for kindness in this job — not softness exactly, but attention.
✅case 05bare short line"I know.""I know." She didn't look up.
✅case 07clipped exchange"Was it you?" "Yes." "Why?""Was it you?" "Yes. I signed for her myself." "Why would you do that?"

4 · Rule proposals

Patterns Claude has noticed across two or more training cases or listen-check findings, drafted as a rule, a scanner check and a test. A proposal becomes a rule only when the author approves it.

# Pattern Seen in Draft rule Status
P1 Short lines get both a flat ending (no main stress) and no emotion cases 01, 05 Merged: the "bare short line" check covers both; the listen-check stress check confirms it in the audio approved (merged into case 05)
P2 Paragraphs of one to three words, dialogue or narration, standing alone. Kokoro has too little context and may slur them ("A wall." → "A while.") case 05; case 06 (listen-check, ch06) Extend the scanner's "bare short line" check to narration paragraphs of ≤ 3 words. Fix by merging into the neighbouring paragraph, or with an emotion beat if it's dialogue. Test: flag She sat., leave She sat. He did not open a file. proposed
P3 In af_heart, a verb ending in t + "it" before a vowel ("want it on") is flapped into one word and heard as a past tense ("wanted on") case 08 (listen-check, 4 hits in one book) When listen-check hears "-ed" where the script has "X it", rephrase. A scanner check could flag "want it/get it/put it" + vowel for American voices proposed