YouTube 오디오 번역 — 원본 초 단위에 맞춘 더빙

YouTube audio translation

Turns one video into extra audio tracks in other languages. Everything is free and local except the voices (edge-tts uses Microsoft's online voices). The engine is scripts/dub.py in this skill's folder.

The pipeline: separate voice and music -> transcribe the voice into timed phrases -> you translate each phrase to fit its time -> build (speak every phrase and put it on the original second) -> mix with the music at the exact length of the video -> check.

Rules (each one cost a failed attempt)

1. The length of the result must equal the length of the VIDEO file (--ref video.mp4). Never take it from a raw .aac: ffprobe only estimates its duration from the bitrate and can be 15 s wrong. YouTube rejects a track that does not match the video ("the audio track and the video must be the same length").
2. You translate; do not use an automatic translator. A translator ignores how many seconds a phrase has. Run budget first and write each line within its character budget. Over budget the voice is sped up and sounds rushed; under 80% of the slot the tool slows the voice a little instead.
3. Fix recognition mistakes before translating. Names and terms get misheard (a service called "Klap" was heard as "Clip" and "Клэп"; "экспорт" as "эксперт"). Read the whole source.srt, correct it, and tell the user which guesses you made.
4. Phrases are placed on their original start second, never glued one after another. The tool already does this; do not post-process the audio in a way that shifts phrases.
5. You cannot hear the audio. Verify with check and the numbers, and say plainly that nobody has listened. Ask the user (ideally a native speaker) to listen. Never write that it "sounds natural".
6. Ask before installing anything. Never overwrite or delete the user's files; work in a separate folder (for example work/ next to the video).

1. Check the setup (once)

python3 --version                       # 3.10 or newer
ffmpeg -version | head -1               # 4.4 or newer
python3 -c "import numpy, soundfile, edge_tts, faster_whisper, pyrubberband; print('ok')"
audio-separator --version               # only for the separation step
rubberband --version                    # optional: nicer speed-ups (otherwise ffmpeg atempo is used)

Missing pieces:

• Python packages: pip3 install -r requirements.txt
• ffmpeg: brew install ffmpeg (macOS) / winget install ffmpeg (Windows) / apt install ffmpeg (Linux)
• rubberband (optional): brew install rubberband / apt install rubberband-cli
• the separator (large, install on its own): pip3 install "audio-separator[cpu]". If import scipy crashes on macOS (a dlopen error), pip3 install scipy==1.14.1 fixes it.

The first transcribe downloads a ~1.5 GB speech model; the first separate downloads a ~70 MB model. Tell the user, and make sure there are several GB of free disk space.

2. Separate the voice from the music

python3 scripts/dub.py separate "/path/video.mp4" -o work

Gives work/vocals.wav and work/music.wav. If the video has no music, skip this step and mix with silence (use --music with any silent file) - or simply build and deliver dub_only.wav.

3. Transcribe

python3 scripts/dub.py transcribe work/vocals.wav --lang ru -o work/source.srt

--lang is the language of the ORIGINAL speech. Read work/source.srt completely and correct mistakes in place (keep the timings). Phrases come out about 2-7 seconds long.

4. Translate

python3 scripts/dub.py budget work/source.srt --lang fr

It prints, for every phrase, the seconds it has and the maximum number of characters (about 15 chars/s for French and English, 14 Spanish, 13 Hindi and German, 5 characters/s for Chinese). Write the translation yourself as a JSON list of strings in the same order, one per phrase, e.g. work/lines.fr.json:

["First phrase translated.", "Second phrase translated.", "..."]

Translate the meaning, not word by word; shorten freely (drop fillers, repeated words); keep brand names as they are written. For Hindi, write foreign brand names in Devanagari so the voice pronounces them. Digits stay digits. Keep a consistent form of address ("vous", "ustedes", ...). Phrases may be re-split across neighbouring lines when a sentence crosses a boundary.

python3 scripts/dub.py srt work/source.srt work/lines.fr.json -o work/fr.srt

fr.srt is also the subtitle file for the same language (original timings) - give it to the user.

5. Build, mix, check

Pick a voice: edge-tts --list-voices | grep fr- (examples: fr-FR-HenriNeural, es-MX-JorgeNeural, hi-IN-MadhurNeural, zh-CN-YunxiNeural, en-US-GuyNeural, de-DE-ConradNeural).

python3 scripts/dub.py build work/fr.srt --ref "/path/video.mp4" --voice fr-FR-HenriNeural -o work/fr --loudness-ref work/vocals.wav

Read the summary. If it lists phrases that need more than 1.25x or that spill into the next one, shorten those lines in lines.fr.json, rebuild fr.srt and run build again: only the changed phrases are spoken again.

python3 scripts/dub.py mix work/fr/dub_only.wav --music work/music.wav --ref "/path/video.mp4" -o work/fr/final
python3 scripts/dub.py check work/fr/dub_only.wav work/fr.srt --lang fr --ref "/path/video.mp4"

mix writes final.wav and final.m4a and prints their length next to the video's; they must agree to within a few hundredths of a second. check re-reads the dub with a speech recognizer (error rate: a few % to ~15% is normal; digits, brand names and spelling variants count as errors; Chinese is scored per character).

Repeat steps 4-5 for every language; steps 2-3 are done once.

6. Report to the user

Tell the user, per language: the files (final.m4a, <lang>.srt), the voice used, the length against the video, the error rate, how many phrases were sped up and by how much at most. Say what you could not check: the sound itself, and the translation if you cannot judge the language well. List the guesses you made when fixing recognition mistakes.

Uploading to YouTube is up to the user: YouTube Studio -> Subtitles -> add the language -> add the audio track (and the .srt as that language's subtitles). Do not upload anything on the user's behalf.

If something goes wrong

• TTS failed for phrase N / retry: the network dropped; run build again, finished phrases are kept.
• A phrase is silent: its text was empty; check the line in the JSON.
• The voice sounds rushed: the line is over its budget; shorten it.
• YouTube says the lengths differ: run ffprobe on the video and on final.m4a and compare; rebuild with the VIDEO as --ref. The .m4a length is rounded to the AAC frame (about 0.02 s).
• Windows/Linux: not tested by the author; the code is cross-platform, but report problems honestly.