Review flags on the timeline
Translation, speech recognition and timing problems, coloured where they happen.
Dubbing is not switched on yet. You can look around and create your account now.
DUBBE dubs your video (a vlog, a lecture, a film) into another of 24 languages (Korean ↔ English tested most), keeps the background sound, and marks the doubtful sentences, so a person can fix them before anyone watches.
Made for creators · films & series · lectures · e-books, read aloud by Text to Speech
Built on open models
Korean read aloud, dubbed into English with no human correction. Each line shows the original, the dub, and the flag DUBBE put on it when it was unsure.
Whisper writes the speech, a second recogniser listens again, and sentences the two hear differently are marked for you. Speakers are told apart by their voices.
MADLAD translates each sentence; a judge model scores it, and numbers and your glossary terms are checked. Doubtful sentences get a flag on the timeline.
Each speaker gets a voice; a sentence that comes out wrong is heard back and said again; and each one starts where the original was said — with the music and sounds of the original kept underneath.
Read Korean sentences → English — click a coloured block to jump to that sentence. Speech: Google FLEURS dataset, ko_kr dev split, CC BY 4.0 — 12 read sentences joined with pauses. Dubbed by DUBBE without any human correction.
#4 · 가나 안에는 큰 숲이 없었기 때문에 나무가 부쩍 비쌌다
Because there were no big forests in Ghana, the trees were expensive.
A real mistake in this dub: the reference says "Canaan". The reviewer fixes it, and the re-dub never re-translates it.
Every language below can be the original or the dub. Korean ↔ English is tested in depth; the others are measured on read speech (FLEURS) — see the results.
Every sentence looks the same until you have heard it.
5 of 15 on the example clip. Then one final listen — flags miss confident mistakes.
Translation, speech recognition and timing problems, coloured where they happen.
Music and effects duck only while the new voice speaks.
SRT and WebVTT, timed to the dubbed video.
…forests in Ghana…
Corrections are never re-translated; re-dubs use no minutes.
Every sentence starts where it was said — measured on three clips.
Slower, softer, a question, a pause — per sentence. DUBBE also tags what it hears in the original (questions, emotion, laughs, sighs); you keep or untick them.
Time, cost and every check, as a page you can keep.
One card per sentence of the example. Some sentences marked OK are still wrong (#5's names, for one) — flags miss confident mistakes.
30 s to 1½ min of work per video minute. The thirteen steps of every dub, as they run:
MP4, MOV or MKV in one of 24 languages.
Thirteen steps: it tells the speakers apart, gives each a voice, and checks every sentence.
Video, subtitles in both languages, a transcript by speaker, a report.
Every sentence is matched to a speaker by voice; each speaker gets a different voice, and the transcript names them. Similar voices can be taken for one person — split them in the editor.
A speaker can be cloned from the video only after you confirm they agreed. The copy stays in that project, or private in your voices — never published.
Questions, emotion, laughs and sighs in the original become tags on the sentence, shown dashed in the editor so you keep the right ones.
A semester of 30 lectures × 75 min is 2,250 minutes.
Measured 29 s per video minute into English on an RTX 4060 (77 s into Korean); GPU at an assumed $0.35/hour. ElevenLabs ≈ $0.50–0.60/min from its credit prices; human dubbing $80–250/min is an unverified industry figure. Bars on a log scale. During the beta DUBBE is free (15 min/month).
9.1%
character error rate of speech recognition on 22 minutes of read Korean (FLEURS, one pass)
29 s
of processing per video minute into English on one laptop GPU (Korean: 77 s)
0.00 s
largest timing delay on three test clips (one with picture slow-down)
24
languages, in any pair — each with 26 voices, picked per speaker by pitch
No lip sync; people in your video are cloned only with their consent; Korean ↔ English is tested in depth, the other 22 languages only on read speech (FLEURS). DUBBE is built for lectures, where one wrong sentence matters.
Free: 15 minutes a month, review editor included. No card.
Start free