🎧 DIY Audiobook · 有聲書自己做
Make your own audiobook, children's story or bedtime radio drama — every character with their own voice,
with music and pictures — by saying one sentence to Claude Code.
"Make a children's story of The Lion and the Mouse in Traditional Chinese, with pictures and gentle music."
…and you get a finished episode: the story adapted for listening, a different voice for every character,
soft music that ducks under the dialogue, illustrations timed to the story, a video, a cover, subtitles,
chapters and upload text. Free voices and free music out of the box — no account, no API key.
Plug in MiniMax, OpenAI, Suno or your own ComfyUI when you want more.
English · 繁體中文
🎬 Watch / listen — every file made by this repo:
Kids · 〈獅子與老鼠〉 video ·
English · The Speckled Band ·
繁體中文 · 〈紅髮會〉 ·
or watch them all on the website
It started as the production line behind 枕邊偵探社 (Pillow Detective Agency),
a YouTube channel of Sherlock Holmes and other classic mysteries as bedtime radio dramas — nearly 200 episodes produced.
Every default here (pacing, gaps, loudness, how loud the music sits) was tuned by ear over those episodes.
Quick start (5 minutes)
You need: Claude Code, Python 3.9+, ffmpeg.
git clone https://github.com/mickyyang-1407/diy-audiobook ~/.claude/skills/diy-audiobook
pip install -r ~/.claude/skills/diy-audiobook/requirements.txt # free voices, music check, covers
brew install ffmpeg # or: apt install ffmpeg / winget install ffmpeg
python3 ~/.claude/skills/diy-audiobook/scripts/doctor.py # shows what your computer can do
Then just ask Claude Code:
Make a 10-minute children's story of "The Tortoise and the Hare" in English, with pictures.
把〈小紅帽〉做成給孩子聽的有聲書，要有插圖和輕音樂。
Make an audiobook of "The Speckled Band" — one voice per character, no music.
把〈紅髮會〉做成睡前廣播劇，慢一點、輕一點。
What you can choose
| free, zero setup | with an account / key | your own |
|---|
| Voices | Microsoft Edge voices | MiniMax (best Mandarin), OpenAI | — |
| Music | generated music-box lullaby | Suno (step-by-step guide + prompts) | any audio file |
| Pictures | — | OpenAI, MiniMax · or local ComfyUI (free) | any picture, any app |
| Style | kids · bedtime · audiobook | | |
Guides: paid voices / API keys (中文) ·
music & Suno · pictures: local or cloud
What you get
| file | |
|---|
build/<slug>.m4a | the episode, −14 LUFS, ≤ −1 dBTP (YouTube / Spotify / Apple ready); --nomusic version too |
build/<slug>.mp4 | 1080p video: the pictures timed to the story, soft crossfades |
build/cover.jpg | 1280×720 thumbnail with the title and a hook |
build/subtitles.srt | sentence-level captions, timed from the actual audio |
build/description.md | title, chapters from the scene breaks, credits |
How it works
story ──▶ fetch ──▶ Claude adapts ──▶ Claude casts ──▶ check ──▶ synth ──▶ mix ──▶ video ──▶ qc ──▶ package
(book, (Gutenberg, for the ear voices, (gate) voices + music + pictures
file) URL, file) pictures, music + cover
Claude does what language models are good at — adapting a story for the ear, deciding who speaks,
casting voices, writing picture and music prompts. Scripts do what models are bad at — counting quotes,
aligning paragraphs, measuring loudness, trimming silence, ducking music, timing pictures, catching
glitches. Nothing reaches a paid service until check.py passes.
Without Claude Code
Every step is a plain script, every file Claude writes is simple JSON:
python3 scripts/check.py examples/red-headed-league && python3 scripts/synth.py examples/red-headed-league
python3 scripts/mix.py examples/red-headed-league --nomusic && python3 scripts/qc.py examples/red-headed-league
python3 scripts/selftest.py # offline self-test
Why the defaults sound the way they do
Full list in references/lessons.md:
- TTS pads ~0.4 s of silence on both ends of every clip. Untrimmed, a line of dialogue and the "he said"
after it sit a full second apart. Clips are trimmed, then exact gaps inserted.
- Different TTS voices differ by up to 9 dB. Each voice is measured over all its lines and matched to the narrator.
- Never run dynamic loudness normalisation on a mix with music — it pumps the music up in every pause.
The voice is mastered first; music goes in afterwards, 16.5 LU under, ducked 3:1 while someone speaks.
- AI music sometimes freezes on one chord for minutes.
music.py check finds those stretches before your
listeners do.
- Image models have no memory between pictures — recurring characters are described identically every time.
- Model self-checks miss whole untranslated paragraphs. Deterministic checks don't.
Copyright
Use texts that are public domain where you publish (author dead 70+ years covers most countries; some use
50) or your own, and translate from the original — existing translations have their own copyright.
Music and pictures must be yours, generated under terms that allow your use, or licensed. Examples:
Conan Doyle (1891–92, Project Gutenberg) and an Aesop fable retold for
this repo.
Repo layout
SKILL.md the instructions Claude Code follows (the whole flow, step by step)
scripts/ doctor · fetch · check · synth · music · draw · mix · video · cover · qc · package · try_voice · selftest
voices/ voice pools per backend, language and style
workflows/ ComfyUI workflow template (bring your own, too)
references/ translation · speaker tagging · music & Suno · pictures · lessons from nearly 200 episodes
docs/ API key guides (English, 中文) + the website (GitHub Pages)
examples/ three complete episodes: kids (中文), bedtime (中文), English
Contributing
Voice pools for more languages, more TTS / image backends and better QC checks are all welcome.
Keep the two principles: validate before you spend, and measure, don't guess.
MIT © Micky Yang
