ai-audiobook-narration Skill
把定稿书稿或课程文字,变成由同一旁白朗读的有声书。这个 AI 有声书生成与配音工作流会整理章节、统一人名和专业术语的读法、把控长篇节奏、协助挑选合适的旁白音色,并在继续生成全书前先产出代表性试听样本。适用于书稿转有声书、分章有声书、课程配音、长文转语音和章节试听,并提供最新的价格估算、按顺序交付音频、一致的旁白音色和针对片段的精修。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
AI Audiobook Narration
An audiobook is not a long TTS clip. It is a structured deliverable of many
chapterized files. Open on one of three scenes: fiction; nonfiction or course;
or sample-only. Route M4B packaging, translation, and multi-track DAW mixing
out. Keep listing covers (beatra.images.generate), narrator cloning from an
authorized sample, native multi-character chapter dialogue scenes
(beatra.speech.synthesize_dialogue), optional chapter SFX/ambience
(beatra.audio.generate_sfx), and chapter loudness/trim/fade finishing
(beatra.audio.process) in this package.
Turn final manuscript or course text into ordered chapter audio. Work as an
audiobook producer: preserve the manuscript's meaning, shape long-form pacing
with Smart Voice Director (beatra.speech.prepare_script), direct a
consistent narrator or chapter cast, prove the direction with a representative
pilot, and expand only after that pilot is accepted.
Start with the smallest complete intake
Reuse everything already supplied. Before narration, obtain only the hard
inputs that are still missing:
- final readable text for at least the pilot;
- chapter or section order and the target BCP-47 language or dialect;
- a pronunciation table extracted from the full manuscript for recurring
names and terms, confirmed by the user before it is frozen;
- a narrator or character sample when the user wants custom cloned voices; and
- title and tone when the user wants a listing cover.
Ask for a missing hard input rather than inventing it. Create a
chapter-and-segment ledger. Preserve intended listening order, label every
segment, and split at chapter, section, scene, paragraph, or sentence
boundaries. Never split mid-sentence. If beatra.speech.validate or synthesis
refuses the text as too long, follow the refusal message: split at natural
chapter, section, or sentence boundaries and submit the parts in order. If a
dialogue request is refused as too long, split it at scene or turn boundaries as
the refusal message says. Remove only page-only material the user does not want
spoken. Follow
chapter production for text preparation,
dialogue scenes inside chapters, casting, and the pilot boundary.
Discover, audition, update, clone, and select model paths
Use beatra.voices.list (bilingual search, limit up to 500, cursor,
facet filters, supports_dialogue=true for cast scenes, inspecting all 11
capabilities.* metadata fields) and free beatra.voices.preview
(preview_url, kind) so the user can audition candidates at zero charge before
freezing opaque, ready voice_id handles. Never submit a display name or guess
from memory. Use free beatra.voices.update (display_name,
make_default, gender, age_group, use_case, persona, language) to
record character/series notes on a saved voice or pin the book's recurring
narrator (make_default=true).
When a narrator or character sample is supplied, do not default to cloning
without explicit user intent and full confirmation discipline from
speech, voices, cloning, and audio finishing:
- Confirmation #1 — Consent (
consent_attested: true): require explicit
user confirmation in the current conversation of voice ownership or
permission before uploading (scripts/mcp_client.py upload). Access to a file
is never consent.
- Sample noise clean (paid): if the sample has background music or noise,
offer paid beatra.audio.process (operation="isolate_voice", source,
{"capability": "audio_voice_isolation"}, 4.6 s – 5 min, <= 100 MB) on
its own card first.
- Clone model choice: select the clone model as described in
speech, voices, cloning, and audio finishing
§5. When a cloned voice will speak in a multi-character dialogue scene,
select the clone card that states multi-speaker dialogue; a single-narrator
book keeps the default. Then show the production card with the live clone
price; after the user approves it, submit once and handle any refusal per
billing, errors, and recovery.
- Confirmation #2 — Activation (
beatra.voices.activate): when a clone
result or a listed voice has status pending_activation, follow §5 Step E of
speech, voices, cloning, and audio finishing
before using the voice.
- Confirmation #3 — Two-step permanent deletion (
beatra.voices.delete):
when deleting a clone, call step 1 (voice_id only), display
paid_slot_warning, irreversible_notice, and 15-minute expiry, wait for
explicit approval, then call step 2 with voice_id and confirmation_token.
Before quoting models or checking language support, call beatra.models.list
with the exact capability (text_to_speech, text_to_dialogue, speech_script,
voice_clone, sound_effects, audio_process, audio_voice_isolation, or
image generation for covers). For an explicit single-speaker model, require an
available live card in voice.compatible_models supporting the requested
BCP-47 language. For auto, collect all live available voice-compatible
candidates, compare the primary BCP-47 language against each candidate's
constraints.supported_languages (documented aliases only), and retain auto
only when every candidate supports the language.
Default to model: "auto", format: "mp3", speed: 1.0, volume: 1.0,
pitch: 0, subtitle_enable: true (never pass subtitles — keeps aligned
JSON/SRT/VTT subtitles and dialogue segments[] free), no emotion, and no
explicit sample rate unless required. Choose the pilot where the book is
hardest: dialogue and names for fiction; dense terminology and numbers for
nonfiction or courses. See
performance, cost, and quality.
Direct pacing and chapter dialogue scenes
- Smart Voice Director (
beatra.speech.prepare_script,{"capability": "speech_script"}):
compile long-form pauses, narrative beats, and pronunciation cues from plain
manuscript text + instructions without changing spoken
manuscript wording and without hand-writing braced tags. Inspect directed
and scripts[] (fix via warnings if synthesis is null), then pass
ready scripts[].synthesis unchanged + fresh client_request_id. Respect max
2 concurrent tasks; within 48 hours on identical wording/instructions/
language/context, calling prepare_script again changing only voice (and
model) reuses direction at zero cost (balance must cover admission).
- Dialog scenes inside chapters (
beatra.speech.synthesize_dialogue): when
fiction or dialogue chapters use several voices with
capabilities.supports_dialogue == true, run free
beatra.speech.validate_dialogue and generate cohesive multi-cast scenes
with beatra.speech.synthesize_dialogue (when the refusal message says a
dialogue is too long, split it at scene or turn boundaries with
previous_text / next_text). Keep single-narrator continuity if
preferred, or fall back to segmented beatra.speech.synthesize when the cast
is refused or specialty dialects apply.
Golden path
Build the chapter ledger and confirm pronunciations. Browse or audition voices
(beatra.voices.list, beatra.voices.preview, optional beatra.voices.update,
or authorized beatra.voices.clone + beatra.voices.activate when pending).
Optionally run Smart Voice Director (beatra.speech.prepare_script) and always
run free preflight (beatra.speech.validate / beatra.speech.validate_dialogue).
Present the pilot production card and wait for listening review before quoting
remaining chapters in incremental batches. Quote any listing cover
(beatra.images.generate), chapter stingers/ambience (beatra.audio.generate_sfx),
or normalize_loudness chapter finishing (beatra.audio.process) separately
on current cards — never call finishing steps free and never auto-retry paid
calls.
Price and confirm
Free tools: beatra.voices.list, beatra.voices.preview, beatra.voices.update,
beatra.models.list, beatra.speech.validate, beatra.speech.validate_dialogue,
beatra.voices.delete step 1, beatra.tasks.get, beatra.tasks.list,
beatra.wallet.get, and
beatra.wallet.ledger. Keep wallet.get and wallet.ledger strictly optional
read-only lookups when the user asks about balance or charges — never make
wallet.get a mandatory gate before submit, and never ask users to verify they
have enough credits first.
Paid tools (beatra.speech.prepare_script, beatra.speech.synthesize,
beatra.speech.synthesize_dialogue, beatra.voices.clone,
beatra.voices.activate, beatra.images.generate,
beatra.audio.generate_sfx, beatra.audio.process) require live pricing and
never reuse remembered credit figures or hardcoded numbers. Count weighted
characters per live card (beatra_weighted_characters: Han = 2, others = 1),
or use free beatra.speech.validate / beatra.speech.validate_dialogue.
A clone approval never authorizes the pilot, a pilot never authorizes remaining
chapters or a cover, and cover/SFX/finishing approvals never authorize speech.
Present a production card with scope/counts, frozen voice_id(s) and controls,
live estimate or auto range, fresh client_request_id(s), what is kept on
stop, and verbatim top-up URL handling on insufficient_balance. Wait for
explicit approval before executing.
Execute, finish, and recover without duplicating paid work
Invoke every remote Beatra tool only through bundled scripts/mcp_client.py
(tool name as CLI arg, JSON on stdin). Never configure a host Beatra Connector
and never use REST/OpenAPI as a fallback. See
Bundled MCP Client diagnostics.
Submit each approved request once with its opaque client_request_id (<= 60
min / 300 MB / 50 calls/day, one operation per call for
beatra.audio.process normalize_loudness -14 LUFS, trim, fade; prompt
duration_seconds0.5–30.0+ optionalloop=true,model: "auto"for
beatra.audio.generate_sfx) and poll beatra.tasks.get until terminal.
If a task_id is lost, paginate beatra.tasks.list by exact capability
(text_to_speech, text_to_dialogue, speech_script, voice_clone,
sound_effects, audio_process, audio_voice_isolation), follow every
next_cursor, and verify candidates with beatra.tasks.get. Replay the same
client_request_id only for field-for-field identical JSON when creation was
genuinely unknown. Never auto-retry; changed arguments mint new IDs and new
cards. Follow delivery and recovery and
speech, voices, cloning, and audio finishing.
Review and deliver what actually returned
Deliver every succeeded chapter/scene with task_id, audio URL, artifact_id,
duration_seconds, actual mime_type, actual sample rate, task.links.assets,
free aligned output.subtitles (and dialogue segments[]), resolved model,
usage, and returned billing facts in listening order. Do not claim to have
auditioned audio if the host cannot play it. Revise only the smallest failed or
corrected segment under its own card and ID. Preserve multi-track mixing, M4B
assembly, and distribution requirements honestly when routing finishing steps
out.
Account balance
Answer balance or charge questions with read-only beatra.wallet.get or
beatra.wallet.ledger. Relay any top_up tiers verbatim in listed order
without ranking or selecting one.
Installation and account operations
Follow speech, voices, cloning, and audio finishing,
installation and authentication,
installation registration,
tasks and results,
billing, errors, and recovery,
and uninstall and disconnect.
Updates and security
This Skill runs only the files installed with it and leaves them unchanged while
it runs; it never downloads or executes other code. It talks to Beatra's official
services over HTTPS, and a file you choose to upload goes only to the one-time
upload address Beatra issues for it.
Once a day at most, the bundled client reads Beatra's public version list. When
the store or tool you installed from has a newer version, it prints one line with
the command that updates it there. Updating stays your choice, through that store
or tool. See updates and security.