360 Secure SkillHub
360 安全技能中心

webnovel-serial-audio

核心功能是为网络连载小说生成声线统一的有声内容,支持后续章节沿用同一旁白(可自定义克隆或选用库内声音)。典型场景为连载小说的有声化制作,输入需提供章节文本、章节顺序、目标语言及专有名词发音表(可选旁白样本),输出按章节排序的MP3音频,支持局部内容修正,按合成块计费。

beatra-ai
beatra-ai
热度 20.98
最近更新 2026-09-07 01:00:36
分类:创作依赖:代码执行 / MCP 服务 / 读写文件应用场景:故事创作 / 文件转换

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add webnovel-serial-audio
Metadata
name
webnovel-serial-audio
description
Turn a serialized web novel into chapter-by-chapter webnovel audiobook audio with one consistent narrator. This serialized webnovel audio studio and web novel narration shop records the current chapter first, then continues later chapters in the same voice so listeners can follow a serial novel TTS library as new chapters land. Use it for chapter audiobook production, webnovel chapter audio, serial fiction narration, and ongoing web novel listening.

Webnovel Serial Audio

Turn finished chapter text into serial listening audio. One narrator. Finish the current chapter, then continue when the next chapter is ready.

Scope and adjacent routes

Use this Skill when a webnovel needs chapter-serial audio with one consistent narrator, and later chapters should continue in that voice.

Route a one-shot literary book to ai-audiobook-narration. Route course lectures to course-narration-studio. Route personal IP episodes to voice-clone-series-studio.

Inputs and defaults

The hard inputs are:

  • final readable text for at least the current chapter;
  • chapter order or the next chapter number;
  • the target language;
  • a pronunciation table whenever the chapter contains recurring names, terms, or place names.

If those names exist and the table is empty, stop and collect the readings. Do not guess them. A narrator sample is optional. Without a sample, choose a catalog voice with beatra.voices.list. Reuse already-known tone, speed, and format.

Defaults: model: "auto" only when every voice-compatible live card supports the language; format: "mp3"; speed: 1.0. Keep each submitted input at or below 50,000 characters. Split on sentence or section boundaries, never mid-sentence. Write short spoken sentences. Do not send chapter headings, author notes, or bracket labels into input.

Planning, casting, and the pronunciation table are free. Clone, if used, and each synthesis block are paid.

Golden path

Submit every Beatra call through the bundled scripts/mcp_client.py. Do not configure a host Beatra Connector. Do not use REST/OpenAPI as a fallback.

  1. Build the chapter ledger: title, order, block labels, and the pronunciation table. Turn the current chapter into speakable text only. Confirm the table before it is frozen.

  2. If the user brought a narrator sample, confirm consent first: the user must state this is their voice or that the speaker authorized this cloning use. Access to a file is not consent. A suitable sample is currently about 10 to 300 seconds, no larger than 20 MiB, and clean single-speaker speech. Only after consent, upload a local file through the bundled client:

    python3 scripts/mcp_client.py upload ./narrator-sample.wav --mime-type audio/wav
    

    Then present the current beatra.voices.clone admission card before any clone client_request_id — route voice_clone, tool beatra.voices.clone, live price for one successful activated voice, the provisional estimate, the 600-credit gift usually cannot start this clone, and what happens if the balance is short. Do not create client_request_id or submit until the user confirms they have topped up or already have enough credits. Do not offer a free clone. Submit beatra.voices.clone exactly once:

    {
      "sample": {"type": "artifact", "artifact_id": "artifact_..."},
      "display_name": "Serial Narrator",
      "consent_attested": true,
      "model": "auto",
      "client_request_id": "opaque-webnovel-clone"
    }
    

    Poll with beatra.tasks.get until terminal, then freeze the returned voice_id.

  3. Without a sample, select a catalog voice_id with beatra.voices.list. Never put a display name in voice.

  4. Call beatra.models.list with {"capability":"text_to_speech"}. Confirm the voice and the current-chapter estimate, then synthesize only that chapter with beatra.speech.synthesize. Poll each speech task with beatra.tasks.get until terminal.

  5. Deliver the chapter audio. After the user listens, revise only unclear or misread blocks: edit the spoken text, then resubmit those blocks with new request IDs. Do not regenerate the whole chapter.

  6. After the user accepts the chapter, keep the same voice_id for later chapters. Confirm the next-chapter estimate before that paid wave.

  7. Deliver labeled chapter audio in reading order with actual duration, usage, and billing.net_charged_credits. Review names, numbers, and voice consistency.

Decisions that require confirmation

Confirm the pronunciation table, the voice, and the current-chapter estimate before the first paid TTS. A changed chapter text is new paid work for the changed blocks only.

Recovery

Recover a lost create response only with the identical frozen payload and ID. A changed text is a new ID. Recover a lost task ID through beatra.tasks.list and beatra.tasks.get before any resend. Cancel only at the user's request. On 409, keep polling the same task. On insufficient_balance for clone or TTS, keep the top-up URL inside the balance error exact and retry the same frozen ID only after the user says they have topped up.

Account balance

When the user asks how many credits remain or whether a live estimate fits, call beatra.wallet.get. When they ask what was charged, call beatra.wallet.ledger. Both are read-only. Do not invent an account-balance or top-up tool. Do not make wallet.get a required step before every paid submit.

When a model card comes back carrying a top_up block, relay its tiers as the card lists them and in that order. Do not rank them, do not talk one down, and do not pick one for the user. Which tier suits them is their call, made on the wallet page with the whole list in front of them. Never quote a tier from memory.

References by task

Runtime and safe automatic updates

Use or invoke the bundled scripts/mcp_client.py for every Beatra operation. Before ordinary commands it silently checks for a newer release at most once every 24 hours per installation. Silent checks are enabled by default, and a newer release installs without separate confirmation.

The updater accepts only the fixed official discovery address and immutable Beatra CDN path embedded for this package, channel, and locale. It verifies the discovery data, archive, manifest, and every file's size and checksum before replacement. It replaces only package-owned files and rejects redirects, downgrades, wrong package/channel/locale/version data, unexpected URLs, unsafe archives, and files outside the owned destination.

Update checks, downloads, verification, replacement, rollback, and recovery fail open: the current installation remains usable and the user's original command continues. An update failure never authorizes retrying a paid generation. The automatic-update choice persists across later commands for this installation:

python3 scripts/mcp_client.py update --auto off
python3 scripts/mcp_client.py update --auto on
python3 scripts/mcp_client.py update --check