A cloned-voice dub of a throwaway take is a rate leak

Dubbing is a pipeline, not a voiceover swap. dub_video translates, clones delivery, and — on the HeyGen engine — lip-syncs the mouth to the new language. The job is async: you submit, then poll get_dub_status or list_dubs. Cost scales with length and engine. None of that belongs on a take you are still regenerating. A recut of the master is a recut of every language.

Dub a video into another language is the page. The AI dubbing tool is the same work. Source media must already live on Versely storage. YouTube links are not an input.

Two engines, both later

elevenlabs is the default. Audio or video. Optional trim via start/end. Up to 30 minutes. It dubs sound; it does not retarget the mouth.

heygen is video-only, lip-sync included, 8-minute cap, fewer languages, no trim range. That lip-sync is the expensive-looking half, and it is glued to this face performance. A new take has new visemes. The old dub does not transfer.

People dub first because the market list is the brief: "we need Spanish." You need Spanish of the keeper. Ten languages of a miss is ten async jobs, ten polls, ten files you will not publish. estimate_cost and check_credits before a long dub. Then still wait for lock.

Keep background audio only if that background is the mix you are shipping. Dubbing a scratch bed "so we have a sense of the Spanish" is a sense that costs a full job.

Lock the master. Derive the languages.

Picture lock means this timeline is the parent. Picture lock when shots can still regen is the contract. Dub last. Always. If you are still extending, trimming, or swapping VO, you do not have a parent.

Captions, if any, follow the dubbed soundtrack, not the English one. Timing in the new language will not match the old cues. Burn after the dub returns.

If you only needed the words translated and not a cloned performance, that is translate_audio_only on the same dub_video tool — still a later step, still async, still wasted on a miss.

Do not treat polling as a reason to start early

Async is not a queue you should fill with candidates. Starting a HeyGen job on take three while take four generates feels parallel. It is parallel spend. The job that returns will be the one you do not use.

FAQ

Can I dub a range of a longer file to test the engine?

ElevenLabs supports start/end. That is a test on a locked master, not a reason to dub a candidate. HeyGen does not trim; you would cut a section first, which is another reason to freeze the picture.

Does the English caption track survive the dub?

The pixels survive. They will be the wrong language and the wrong timing. Plan a new caption pass on the dubbed file.

Why not regenerate the whole video in Spanish instead?

Sometimes that is the right job — a new generate in the target language. It is a new picture, not a dub. Dub exists to keep this locked picture and change the voice. If the picture is not locked, neither job should run as a "first step."

Is lip-sync required?

Only if the mouth is the shot. Talking close-up: HeyGen, after lock. Voiceover over B-roll: you did not need a dub with visemes. You needed a translated read on a locked cut.