Audiobook Skill

audio-book-skill

Turn a manuscript into long-form narration, locally and reproducibly.

The proven engine is Kokoro-82M (an 82M-parameter open TTS model) run through the mlx-audio CLI on Apple Silicon (MLX). Synthesis costs no per-character fees and needs no network at run time. This skill wraps that engine with deterministic chunking, resume, lossless assembly, and a job manifest so any capable agent can reproduce a render.

When to use

Triggers: "narrate this," "turn this into an audiobook," "read this aloud," "make an audio/spoken version of X," "produce a narration MP3/MP4 from this document."

Not this skill: speech-to-text/transcription; music or background scoring; automatic publishing or upload; thumbnail/artwork generation.

What you get

• scripts/prepare_manuscript.py — source manuscript → narration-safe text (never edits the source; configurable heading/URL/citation/bibliography handling; applies the verified pacing fixes).
• scripts/narrate.py — narration text → ordered MP3 + lossless WAV master
  • manifest.json. Sentence-aware chunking, --resume, verification.
• scripts/assemble_video.py — OPTIONAL still image + MP3 → verified MP4.
• scripts/verify_job.py — re-verify a finished job against its manifest.
• references/FIXES.md — verified vs. suspected quality fixes.
• example/ — an original public-domain example manuscript.
• README.md — install commands, versions, licenses, hardware assumptions.

The one verified command (manuscript → MP3)

After setup (see README), from the skill directory:

# 1) prepare a narration-safe script (does not touch the source)
python3 scripts/prepare_manuscript.py \
    --input example/example_manuscript.md \
    --output out/narration.txt \
    --drop-sources

# 2) synthesize to MP3 (+ lossless WAV master + manifest)
python3 scripts/narrate.py \
    --input out/narration.txt \
    --out-dir out \
    --voice af_heart \
    --lang a

Output: out/narration.mp3, out/narration_MASTER.wav, out/manifest.json, out/segments/. This exact chain was run end-to-end during authoring.

Voices and language (choose deliberately)

Voice id prefixes encode accent+gender; the --lang G2P code MUST match:

Accent is a voice choice, not a post-filter. A British bf_ voice under --lang a will use American phonemization (mismatch). Set --lang b for a British voice. List what's downloaded:

python3 scripts/narrate.py --list-voices

Tested during authoring: af_heart (American, --lang a) and bf_emma (British, --lang b). Others are available; test a short excerpt before a full render. Do not hardcode a voice — pass --voice/--lang per job.

Chunking, order, and joins

narrate.py splits on sentence boundaries and packs whole sentences up to --max-chars (default 600) so a chunk never breaks mid-sentence. Oversized single sentences are split at the nearest clause boundary (; : , —), never mid-word. Segments are assembled in numeric order via ffmpeg stream-copy (no re-encode), and the manifest records chunk order + a duration check that flags any dropped/duplicated/clipped audio.

Resume

--resume reuses existing chunk WAVs only when the input text, model, voice, language, speed, and max-chars all still match the manifest signature. Any change re-renders. Interrupted jobs resume from the last completed chunk.

The manifest

out/manifest.json records: input SHA-256 + size, engine/model/voice/lang/ speed, per-chunk hashes and order, output paths + duration, a verification block, and captured dependency versions. It is the reproducibility record — keep it with the outputs.

Verified quality fixes (see references/FIXES.md)

1. Mid-sentence pauses = surviving soft line breaks in wrapped prose; fixed by unwrapping to full paragraphs (biggest single fix).
2. Abbreviation periods (U.S., e.g., Dr.) read as full stops; converted to spoken forms.
3. Accent is a native voice + matching --lang, not a filter.
4. Sources/bibliography narrated unintentionally; optional --drop-sources relocates them to a description/companion file.
5. Lossless intermediate, single final MP3 encode (no repeated loss).
6. Long renders: background them, poll at 30–60s; don't SIGKILL mid-write.
7. Apple Silicon: stop other large local model servers first (GPU OOM).

Suspected/defensive behaviors (Roman-numeral labels, slash alternatives, the benign phonemizer warning) are documented separately in FIXES.md — do not treat them as guaranteed.

Safety / hygiene

• The source manuscript is never modified; narration is a separate file.
• Do not commit credentials, private manuscripts, generated media, or model weights (see .gitignore). Weights download to the HuggingFace cache.
• No cloud dependency for the proven local workflow.