Every line lands where the actor said it.

Castix Dub transcribes speech with word timings, tells speakers apart, translates line by line, and places each dubbed line at its original start — never stacked after the previous one — so minute 39 is as in sync as minute one. You review and edit every line before anything is rendered.

Public demo: CPU engines only (local Whisper, free read-aloud voices). It accepts clips up to 90 seconds, 3 jobs per visitor per day, and is not for commercial output. Production keys route to licensed engines.

Try it on your clip

Drop an audio/video file here, or click to choose (≤ 90 s)

Spoken language Dub into Speakers — giving the count is more reliable than detection

What you get

Output	Detail
Transcript	Word timestamps, speaker labels
Subtitles	SRT · WebVTT · TTML (IMSC1) · SCC · JSON, per-language line length and reading speed
Translation	Numbered line-by-line, glossary and context, duration budget per line
Dub	Voice per speaker, one pace per speaker, levelled lines, original video stream untouched
Report	Lines compressed, lines that overran, engines that served the job

Measured, not claimed

1.35×

Hard ceiling on per-line compression. Past it, speech stops being understandable. Lines are never stretched.

λ = 9

Speaker-change penalty tuned on 190 real Persian speakers (CC0). The value that worked on synthetic voices split real people in 12 of 15 files.

—

Common Voice word/character error rates are published per language in the evaluation report, with clip ids so anyone can re-run them.