6 beatra-ai

ai-multilingual-dubbing Skill

根据准备好的脚本为视频、新品发布、在线课程、培训库、创作者内容和国际营销活动生成多语言配音音频。这个 AI 多语言配音工作流会统筹每个市场和分片,帮你选出适配当地语言的声音,试音校准发音和节奏,并按语言交付可审听的旁白。可用于 AI 视频配音、AI 视频翻译音频、视频本地化配音、培训视频本地化、全球营销旁白、多语言文字转语音,或中英日配音。收到配音后逐

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。

查看源码

技能指令原文(SKILL.md)

AI Multilingual Dubbing

Create organized, loudness-consistent localized narration or speech-converted
tracks for every target market from approved scripts or authorized source
recordings. Reuse the message, destination, languages, voice direction,
segment boundaries, and timing targets already visible in the conversation. Move
by the shortest route to a representative pilot per locale, then produce the
remaining approved cells.

Start from source transcripts or approved spoken scripts

Build a segment × locale production matrix. Each cell records the locale and
normalized BCP-47 tag, approved text (or source conversion cell), protected
terms, pronunciation notes, exact voice_id, delivery direction, format,
timing target, and status. Use natural segments such as a scene, lesson,
chapter, ad, or publishable message—never a mid-sentence split.

  • If the source exists only as audio or extracted video audio and you need a

base script with timestamps, check beatra.models.list with
{"capability": "speech_to_text"}, present a six-field production card, and
after approval call beatra.audio.transcribe (upload local media via
scripts/mcp_client.py upload first; pass audio, not source, plus
optional subtitles: true, diarize, num_speakers, tag_audio_events,
language). Keep the returned transcript strictly as draft text until the
user approves localized spoken copy.

  • When localization or adaptation is requested, preserve product names, claims,

numbers, units, URLs, and calls to action, and preserve simplified/traditional
characters and dialect wording unless asked. Do not claim professional or
native-language review unless the user supplies that fact.

  • Never manually strip / pause tags or invent bracketed vocal-effect cues. When

dramatic pacing or breathing cues help a localized performance, inspect
{"capability": "speech_script"}, approve its six-field card, and run
Smart Voice Director (beatra.speech.prepare_script), passing
scripts[i].synthesis unchanged plus a fresh client_request_id (and fixing
any synthesis: null from warnings). Re-calling prepare_script within
48 hours with exact same input/instructions/language/context changing
only voice or model recompiles at zero extra charge (balance cover & max-2
concurrency limits still apply; never call prepare_script generally free).

Read speech and audio guide, [matrix
design](references/matrix-design.md), and [locale readiness and
quality](references/locale-readiness-and-quality.md) before casting or sizing
batches.

Cast voices, clone authorized talent, or convert existing speech

Reuse an accepted voice_id after verifying its live card via
beatra.voices.list. Search candidates using all 11 capability filters
(category, language, gender, age_group, use_case, emotion,
supports_context, supports_direction, supports_performance_tags,
supports_dialogue, supports_speech_to_speech), confirm status == "ready",
and audition free with preview_url or beatra.voices.preview
({"voice_id": "..."}). Never pay to audition voices.

  • Authorized voice cloning (beatra.voices.clone): confirm plain speaker

consent BEFORE any scripts/mcp_client.py upload. If a sample carries
background noise or music, offer paid pre-cleaning first via
beatra.audio.process (operation: "isolate_voice", source field,
4.6 s–5 min, <= 100 MB, checked via {"capability": "audio_voice_isolation"}).
Inspect
beatra.models.list({"capability": "voice_clone", "include_activation_required": true})
and choose by capability only—never
mention upstream vendor names or internal IDs. Use immediate-ready routes
(auto or expressive multilingual) directly here; if the user picks a
two-step trial-then-activate route (pending_activation) or wants to edit or
delete slots (voices.activate, voices.update, voices.delete), verify
scripts/mcp_client.py tools and hand that slot lifecycle step to
voice-cloning-studio or beatra-ai-voice-studio. Show the six-field clone
card before calling voices.clone.

  • Speech-to-speech delivery transfer (beatra.speech.convert): when

preserving source delivery timing and cadence across a target voice, confirm
rights to transform the source recording, check
{"capability": "speech_to_speech"} and verify target voice has
supports_speech_to_speech: true, approve its six-field card, and pass
audio (not source) with optional remove_background_noise.

For TTS synthesis, default to model: "auto" (only when every live
voice-compatible candidate supports the normalized BCP-47 primary tag, mapping
nb -> no, tl -> fil), speed 1.0, mp3 (or wav/flac for studio
editing), no emotion, and unchanged pitch/volume/sample rate. Call
beatra.models.list ({"capability": "text_to_speech"}) before deciding models
or estimating beatra_weighted_characters (Han = 2, others = 1). Never
hardcode credit figures. Always run free beatra.speech.validate
(estimated_duration_seconds) before each paid beatra.speech.synthesize. Use
subtitle_enable: true (NEVER subtitles) when free aligned SRT / VTT
captions are needed for editing or lip-sync alignment.

Confirm pilots, full scope, and loudness finishing

Planning, copy drafting, voice search, voices.preview, speech.validate, and
estimates are free. Before any paid step (audio.transcribe, prepare_script,
isolate_voice, voices.clone, speech.convert, speech.synthesize, or
normalize_loudness), present its six-field production card: Route & Model,
Exact Scope, Price & Basis, Duration / Quantity, Pay Mode & Call Count, and
If stop (stating clearly what free assets stay usable at zero extra cost).

For multi-cell localization, pilot one high-risk segment per locale first
(names, numbers, offers), compare returned duration against target timing, and
approve the remaining matrix card explicitly afterward. When unifying output
volume across languages (e.g. -14 LUFS), inspect
{"capability": "audio_process"}, approve its finishing card, and run
serialized beatra.audio.process (operation: "normalize_loudness", source
field, <= 60 min / <= 300 MB / >= 1 s output, 50/day cap) one track at a time.
Never claim finishing is free or present it as a mastering service.

Create one stable opaque client_request_id per logical paid cell, submit once,
and poll only that task_id via beatra.tasks.get (stopping active wait at 30
min if no deadline_at is set). The bundled client handles non-billable
installation registration automatically.

Use the bundled MCP Client

Use only this Skill's bundled scripts/mcp_client.py for every remote MCP
operation (call passing one JSON object on stdin, or upload for local
media). Do not configure or call a host Beatra Connector, and do not use
REST/OpenAPI as a fallback. For exact commands and troubleshooting, use
Bundled MCP Client diagnostics.

Recover the task, handle balance, and deliver audio

Never make beatra.wallet.get a mandatory gate before paid submits and never ask
users to confirm they have enough credits. Call read-only beatra.wallet.get
when asked about balance/fit and beatra.wallet.ledger when asked what was
charged; relay top_up tiers in returned order verbatim. On
insufficient_balance, relay Beatra's message and exact topup_url verbatim,
wait until the user confirms topping up, and replay the exact frozen payload
with the same client_request_id.

If a task ID is lost, paginate beatra.tasks.list across next_cursor and
inspect candidates via beatra.tasks.get. Any changed text, voice, language,
model, control, or LUFS target creates a new card, approval, and
client_request_id. Deliver returned URLs, artifact_id, MIME type, duration,
optional subtitles, resolved model, and billing.net_charged_credits grouped
by locale and segment without claiming listening review or exact lip-sync
unless verified. See recovery and delivery.

References by task

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.