Skip to main content
Skill: heygen-translate · Invoke: /heygen-translate [video_url_or_path] [--to language] · Source
heygen-translate localizes a finished video into 175+ languages. The presenter keeps their face, their voice is cloned into the target language, and their lips re-sync to the new audio — viewers see the same person speaking natively. You supply a source video and target languages; the engine handles transcription, translation, voice cloning, lip-sync, and optional burned-in captions. This is not new-video generation. The performance, framing, and brand assets in the original are preserved — translation rides on top of what’s already there.

​
When to Use

  • Localize an existing video — “translate this to Spanish”, “dub this into Japanese”, “I need this in 10 languages for a launch”.
  • Same presenter, another language (not a new presenter — that’s heygen-video).
  • Podcast / audio-only translation — “translate this podcast, keep my video”.
  • High-stakes content where you want to review/edit subtitles before final render (the proofreads workflow).
Not for: creating new videos from scratch (use heygen-video and write the script in the target language), avatar creation (use heygen-avatar), or text-only translation. Use this skill only when a source video already exists.

​
Workflow

Four phases. Discovery is the only place the skill asks questions; the rest is silent until delivery.
1

Discovery

Gathers the minimum: source video (URL, file path, or HeyGen asset_id) and target language(s) — asked open-ended, no picker. Confirms speaker count, content type, captions, and duration flexibility with smart defaults.
2

Pre-flight (silent)

Validates the language string against the canonical list, routes the source (public URL passes through; local files upload, max 32 MB), and classifies the content into one of five profiles to set the right flags.
3

Submit + Poll (silent)

Submits one job per language (batchable), then background-polls — 30s for the first 3 min, then 60s. Most finish in 5–15 min; some take 30+. Surfaces only on completion, hard failure, or a stall.
4

Deliver

Posts one message per completed language with the video URL, duration, mode, and caption status. Delivers each as it finishes rather than waiting for the whole batch.

​
Content Profiles

The skill picks one profile from the source and sets flags accordingly — you rarely need to specify:
ProfileUse whenKey flags
Talking head (default)One person to camera, clean audioprecision, speech enhancement, captions, dynamic duration
Podcast / audio-onlyStatic or absent visualtranslate_audio_only, speech enhancement
Music / high-soundtrackBackground music interferes with speechdisable_music_track, speech enhancement
Multi-speakerTwo+ distinct speakersTalking-head defaults + speaker_num (required — don’t guess)
Corporate / brandedBrand voice, glossary disciplineTalking-head defaults + brand_voice_id; consider proofreads
Always precision mode unless the user explicitly asks for speed. Always confirm speaker count for non-obvious cases — wrong speaker count is the #1 quality killer, causing voices to bleed across speakers.

​
Proofread Workflow

For high-stakes content, run a proofread session so the user can review and edit translated subtitles before the engine commits to a final render. Default ON for long videos (>3 min), corporate/branded content, high-stakes legal/medical/educational, and languages the user reads natively. 
proofreads create → poll → fetch SRT → edit (glossary, register, names) →
upload edited SRT → generate final → poll → deliver
This gives you an editable SRT you can also use as a sidecar caption file — preferable to burned-in captions when you might want to restyle them later.

​
What to Know Before Translating

Translation can’t improve on the source. Muffled audio, fast cuts, heavy face occlusion, or low resolution all degrade lip-sync and voice quality. Warn the user upfront in Pre-flight — never after a bad result. Lip-sync is best on stable, front-facing, well-lit shots ≥720p with minimal cuts.
en→zh, en→ja, en→ko run ~30% shorter; de→en, ja→en run longer; en→ar/he expand. Dynamic duration matters most for these — without it, en→zh sounds artificially slow. Regional variants matter too: Spanish (Spain) ≠ Spanish (Mexico); “Chinese (Mandarin, Simplified)” ≠ Cantonese ≠ Taiwanese Mandarin.
Register-heavy languages (Japanese keigo, Korean honorifics, German Sie/du, Thai) default to neutral-formal — proofread if you need to match a conversational source. RTL languages (Arabic, Hebrew, Urdu, Persian) render captions right-to-left and can collide with source lower-thirds; consider audio-only or caption styling review.
Translations bill by source duration. A 5-min video × 5 languages = ~25 billable minutes, each rendering in 10–20 min. Surface source-minutes × language-count and an honest render-time range — don’t quote dollar figures (they vary by plan).

​
Example Prompts

PromptWhat happens
”Translate this video to Spanish: [URL]“Single-language dub with voice clone + lip-sync.
”Dub this launch video into French, German, and Japanese.”Batch translation, one job per language, delivered as each completes.
”Translate this podcast but keep my video.”Audio-only mode — returns a translated MP3 track.
”Localize this for our enterprise launch — I want to review the subtitles first.”Proofread workflow with editable SRT before final render.

View the full SKILL.md

Includes the full proofreads playbook, language-locale guide, asset routing, and failure-mode decoder.