Przejdź do treści
Biblioteka skilli
Marketing i treści · Konwersja i formaty

dub

/dub

Tłumaczy scenariusz i nakłada na oryginalny lektor nowy głos w innym języku, sekcja po sekcji.

Repozytorium
dub-agent ↗
Licencja
MIT
Ostatnia zmiana repozytorium
10.11.2026

Ocena i bezpieczeństwo

90/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść25 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Tłumaczy scenariusz i nakłada na oryginalny lektor nowy głos w innym języku, sekcja po sekcji.

Oryginalny opis (EN)

Translate a script and dub it onto the original voice-over with the `dub` tool, timed section by section, using a client profile. Use when the user asks to dub, voice or re-voice a script or video into another language, or to re-time or re-mix an existing dub.

Jak działa

  1. Before you start
  2. Steps
  3. Reading align's output
  4. Checking the result
  5. Report
  6. Stop and ask only if
Sekcje z nagłówków SKILL.md, w oryginale · 65 linii, 1 085 słów
Podgląd SKILL.md
---
name: dub
description: Translate a script and dub it onto the original voice-over with the `dub` tool, timed section by section, using a client profile. Use when the user asks to dub, voice or re-voice a script or video into another language, or to re-time or re-mix an existing dub.
---

# Dub a script

Turn a script plus its original voice track into a translated voice-over that starts each section where the original speaker did, mixed over the original music.

## Before you start

1. **Find the profile.** The user names it, or it's in `DUB_PROFILE`, or the current folder has a `profile.toml`. If none of these, ask which profile to use. Don't guess.
2. **Read the whole profile:** `profile.toml` (languages, label, voice), its `CLAUDE.md` if there is one, and the glossary and style guide (`glossary.md` and `style-guide.md`, or wherever the profile's `CLAUDE.md` says they are). The profile's rules win over anything in this skill.
3. **Find the inputs.** In the audio folder: `<NAME>.VO.wav` (the original voice), optionally `<NAME>.BGM.wav` (the music bed), and the script. `<NAME>` comes from the VO's file name; `<LABEL>` is the profile's label. If the VO is missing, you can still translate and read the text, but stop before `align` and say there was nothing to time it to.
4. **Check the tools:** `dub --help`, `ffmpeg -version`, `whisper-cli --help`. If one is missing, say which and how to install it (README, "Install"), and stop.

## Steps

Run the whole chain without stopping between steps, unless a step below says to stop.

1. **Translate** the script into the profile's target language, if it isn't done yet. Follow the glossary and style guide. When a key term isn't in the glossary, use your best rendering and list it as a proposed glossary entry in your report. Don't add it to the glossary yourself.
2. **Write the voice text:** what the voice should actually say. Follow the style guide's voice-text rules (no headings or production cues; references, numbers and acronyms written the way they're spoken). One paragraph per block, with a blank line between. Save it as `<NAME>.<LABEL>.tts.txt` next to the audio.
3. **Read it:** `dub tts --profile <PROFILE> <NAME>.<LABEL>.tts.txt <NAME>.<LABEL>.mp3`
4. **Write the anchors:** `<NAME>.anchors.txt`, one line per paragraph start, in order: the first 3–5 words of the translated paragraph, then `|`, then the first 3–5 words of the original paragraph it pairs with.
   - Take the original's words from the script, not from memory.
   - Pick words that appear only once nearby. A word that's repeated close by (a title, a name) can match the wrong spot; add the next word or two to make it unique.
5. **Align:** `dub align --profile <PROFILE> <NAME> <NAME>.<LABEL>.mp3 <NAME>.VO.wav <NAME>.anchors.txt [--bgm <NAME>.BGM.wav]`. Read its output and fix problems before building (next section).
6. **Build:** `dub build --profile <PROFILE> jobs.json` (`jobs.json` is in the audio folder).
7. **Check the result** (below), then **report**.

## Reading `align`'s output

Each anchor prints a line with its time in both reads and two match scores (dub, original):

| You see | It usually means | Do this |
|---|---|---|
| `<-- check this one` (a score under 0.75) | The words weren't found as written: a typo, or Whisper heard them differently | Look at the transcript around that time with `dub analyze <audio> --lang <code>`, change the anchor to the words Whisper actually heard, and run `align` again |
| `merged … (same start)` | Two anchors snapped to the same pause; often a heading straight before its first paragraph | Fine if they really are back to back. Otherwise the second anchor matched too early; make it more specific. Either way, list it in the report |
| One `merged … (too fast otherwise)` | That section would need more than the profile's `max_tempo` speed-up | Fine on its own: it's joined to the section before, so that part of the dub has one sync point fewer. List it in the report |
| Several `merged … (too fast)` in a row | An anchor before them landed on the wrong spot | Find it: its time jumps compared to its neighbours. Fix it and run `align` again |
| `WARNING section … needs speeding up … past max_tempo` | The last section (nothing after it to merge with), or one cut short by a silence, is too long for its slot. `build` will cap its speed, so it runs long | Check that section's anchor and the one after it. If they're right, build anyway and report it with how long it runs over, so the user can decide whether to re-read it faster with `dub tts --speed` |
| `WARNING a Ns silence … add an anchor right after it` | A song or clip in the original falls in the middle of a section | Add an anchor for the paragraph that starts after the silence |

Re-run `align` until no anchor is flagged and any merges are explained. Don't move or drop an anchor just to make a merge or warning go away: fix only anchors that matched the wrong words. If the same anchor still fails after two attempts, leave it, build anyway, and say so in the report.

## Checking the result

- `build` prints each section's start and tempo. Note any section that `ends Ns early` (over ~1 s), any `WARNING capped at max_tempo` and any `WARNING … overlaps`.
- Confirm the length and loudness line: the dub must be as long as the VO, and its loudness close to the VO's.
- Spot-check the timing: transcribe the dub with `dub analyze <NAME>.<LABEL>.dub.wav --lang <target>` and compare a few section starts with the original's. They should be within about half a second.

## Report

Keep it short, in this order:

1. **Outputs:** the paths of `<NAME>.<LABEL>.dub.wav` and, with a BGM, `<NAME>.<LABEL>.mix.wav`.
2. **Timing:** how many sections; every merge, with its anchor and reason (same start or too fast); any section capped at `max_tempo` and how long it runs over; any that end more than ~1 s early (the user may want those re-read slower or faster with `dub tts --speed 0.7`–`1.2`); and any overlaps. Report these even when the dub sounds fine: each one is a trade-off the user should see.
3. **Anything you fixed or couldn't fix:** anchors you changed and why, warnings left.
4. **Proposed glossary entries,** if any, for the user to approve.

## Stop and ask only if

- the profile or the audio folder can't be found,
- a tool is missing,
- ElevenLabs returns an error (quote it; don't retry more than once),
- or the profile's rules contradict each other.

Never commit, push or upload anything as part of a dub. The outputs stay next to the inputs for the user to check.

Wersje

RepozytoriumWynikLicencja
Ta strona · gabrielmlow21/dub-agent90MIT
WEBGRS/openfilm-dub100MIT

Migawka bazy z 13.11.2026 · wersja paczki 472e17710537 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100youtube-dubbing
    Dubbinguje film z YouTube lub lokalny na inny język, zachowując głos mówców, z napisami i gotowym pakietem do publikacji.
  • 100wjs-dubbing-video
    Generuje dubbing TTS zsynchronizowany z docelowym SRT, aby wideo mówiło w wybranym języku: Volcano dla chińskiego, edge-tts dla innych, opcjonalnie z wieloma mówcami.
  • 85taste-dub
    Odwzorowuje estetykę, typografię, tokeny kolorów i komponenty serwisu dub.co przy budowie stron i interfejsów.
  • 100dub-webhooks
    Odbiera i weryfikuje webhooki Dub, sprawdza podpis dub-signature i obsługuje zdarzenia link.created, lead.created oraz sale.created.
  • 90wazza-dub
    Przygotowuje wersje wideo w innych językach z dubbingiem i napisami, a potem sprawdza każdą.
  • 100language-swap
    Tłumaczy i dubbinguje wideo na inny język z zachowaniem głosu każdego mówcy, synchronizacją obrazu i dźwięku oraz domyślnym lipsyncem, także z dwujęzycznymi napisami.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.