NodeTool Cloud is in alpha. Try it →

All recipes
Recipe·5 workflows

Multilingual Video Dubber

One presenter clip, spoken in a second language, with a lip-synced cut, subtitles, and a back-translation to check it.

For: Anyone re-releasing a recorded talk, demo, or ad into another market.

Download all 5 workflowsDownload NodeTool

What it produced

Not a mockup. The chain below was run end to end and this came back. The list under the caption is every model that ran: green where it is the model the workflow names, reached through whichever provider this render held a key for, blue where a better one was chosen deliberately, and amber where a different model had to stand in — each with the reason.

Output from running the Multilingual Video Dubber recipe
The same take, before and after the lip-sync step, with the Spanish the translator produced. The presenter is generated too — the recipe assumes you bring your own footage, and this render had none to bring.
Produced by
  • kie:kling-2.6/image-to-videostands in for fal_ai:fal-ai/ltx-2.3/image-to-video/fast — no provider reachable from this render serves LTX-2.3
  • openrouter:openai/gpt-5-minithe model openai:gpt-5-mini names, billed through OpenRouter
  • replicate:inworld/realtime-tts-1.5-maxstands in for openai:tts-1 — no OpenAI key on this machine, and neither OpenRouter nor AtlasCloud serves TTS
  • replicate:inworld/realtime-tts-1.5-maxas the workflow ships it
  • replicate:sync/lipsync-2the model fal_ai:fal-ai/sync-lipsync/v2/pro names, Sync's own lipsync-2, hosted on Replicate

The spine is transcribe, translate, revoice. What makes it usable outside the languages your team reads is the third step: a back-translation printed next to the localised line, so the person approving the cut can see what was actually said.

The last two steps are a fork, not a sequence. A voice-over cut ends at the revoiced audio; an on-camera cut carries on into lip-sync so the presenter's mouth matches the new delivery.

Keys this chain needs

Read out of the workflows themselves, so this list is what the graphs actually call. You bring the keys and pay each provider directly — NodeTool takes no cut and adds no markup.

The chain, in order

  1. Card art for the Transcribe a Clip template
    01

    Get the script back out of the footage

    In: the source clip. Out: the spoken words as text. The audio is extracted on your machine and only the audio is sent to a speech-to-text model.

    gpt-4o-mini-transcribe
    Transcribe a Clip — see the graph
  2. Card art for the Localise a Script and Revoice It template
    02

    Translate and voice in one pass

    In: the transcript and a target language. Out: audio in that language. A translated document still leaves the recording to do, which is why this step ends in sound rather than text.

    gpt-5-minitts-1
    Localise a Script and Revoice It — see the graph
  3. Card art for the One Tagline, Six Markets template
    03

    Read back what you just shipped

    In: the lines that carry the message — the claim, the offer, the call to action. Out: each one in six languages with a back-translation beside it. This is the review gate for a language nobody in the room speaks.

    gpt-5-mini
    One Tagline, Six Markets — see the graph
  4. Card art for the AI Spokesperson template
    04

    Make the mouth match, for an on-camera cut

    In: the presenter clip and the translated script — the script, not the audio from step two, because this workflow voices the line itself before it redrives the mouth. Out: a take that looks recorded in the new language.

    inworld/realtime-tts-1.5-maxfal-ai/sync-lipsync/v2/pro
    AI Spokesperson — see the graph
  5. Card art for the Subtitle Text from a Recording template
    05

    Ship the version people read

    In: the localised audio. Out: transcript broken into subtitle-length lines. The line-length rule is what separates a subtitle file from a wall of text.

    gpt-4o-mini-transcribegpt-5-mini
    Subtitle Text from a Recording — see the graph

What it does not do

  • Lip-sync redrives the mouth. It does not change gesture, gaze, or anything a presenter does with their hands, so a take with heavy pointing at on-screen text will still read as dubbed.
  • Timing is not stretched to match the original. A language that runs longer than the source will run longer in the cut.
  • The back-translation is a check on meaning, not on register. A native reviewer is still the last word before a paid campaign.

Running it

  1. 01

    Install Studio

    The desktop app is free and runs on your own machine. No account is needed to open a workflow.

  2. 02

    Import the bundle

    Open the command menu and choose Import Workflow as Bundle. All the workflows in this recipe land in your library at once.

  3. 03

    Add your keys

    Paste each key above into Settings. Studio stores them in your OS keychain and sends them only to that provider.

  4. 04

    Run the chain

    Work down the list. Each step takes what the one before it produced, so you can stop and change your mind at any point.

Other recipes

Viral Video Ad Engine

A vertical product ad, plus the hook lines and thumbnails to test it against.

E-commerce SKU Visual Factory

One packshot per SKU becomes the whole channel set: cutout, studio scene, seasonal relight, turntable clip, print master, and the listing copy.

Storyboard to Trailer

A logline becomes a beat sheet, a numbered shot list, a cut teaser, and a score under it.