No filming:An AI avatar or clone yourself
    Versely
    Guides

    Dub a UGC ad without a new mouth

    Lay a voice-over bed when the face is not the product. Lipsync only when the mouth has to match.

    Versely Team••7 min read

    A voice-over bed is enough when the mouth is not the product. You need lipsync when visemes have to match the new language. AI dubbing is a chain: transcribe, translate, speak, fit. Dub a video runs that chain (elevenlabs clones delivery without retargeting the mouth; heygen adds lip-sync, video-only, 8-minute cap). Most UGC ads are a face plus a pack. Decide which of those the market has to believe.

    This is the mouth-or-not pick on a locked UGC spot. It is not “dub first, then caption the new language” — that order-of-operations page already exists. Here the question is whether you owe a new mouth at all.

    Decide whether the mouth is the product

    Write the answer before you pick an engine.

    The mouth is the product when:

    • The ad is a talking-head close-up and the market will watch lips.
    • You are selling the person (founder, creator, identifiable talent).
    • The first second is a spoken confession on a big face.
    • You already shipped a lipsynced original and a loose VO would look like a recut.

    The mouth is not the product when:

    • The plate is pack, hands, pour, texture, a screen.
    • The face is a small PiP or a nod.
    • The original was already a VO over B-roll.
    • You will crop to the SKU for this language.

    Procedure:

    1. Open the locked English (or source-language) master. Pause on the first second and on the line that carries the claim.
    2. Circle the frames where a mute viewer would notice a mismatch: teeth, visemes on a close-up, a laugh.
    3. If you circled nothing, you want a bed: translate the script, TTS or cloned VO, replace video audio.
    4. If you circled a close-up, you want a mouth pass: heygen on dub_video, or a dedicated video-to-lipsync row after a new wav.
    5. If you circled a mix — close-up hook, then pack demo — split the file. Mouth pass on the hook clip, bed on the demo, then merge. One engine on a whole spot that is 80% B-roll is how you pay for visemes nobody sees.

    Do not dub a miss. Picture lock first. A recut of the master is a recut of every language. Source media must already live on Versely storage. YouTube links are not an input.

    The AI dubbing tool is the same work as the editing task. Cost scales with length and engine. estimate_cost before a long file. Then still wait for lock.

    Lay a voice-over bed when the face is not the job

    Bed means: new language in the speakers, old picture, no viseme retarget.

    Procedure:

    1. Get a script. Get a video transcript on the master if you do not already have the words. Edit names, numbers, the SKU, the claim. This is where quality is won. Idioms and offers fail here, not in the vocoder.
    2. Translate that script. Human pass on the offer. A faithful translation that does not fit the gap will be rushed later. Cut the target-language sentence to the hole you have.
    3. Speak it. Options:
      • dub_video engine elevenlabs (default): cloned delivery, audio or video, up to 30 minutes, optional trim, no mouth retarget. Keep background if that background is the mix you are shipping.
      • Or TTS in the target language (Gemini 3.1 Flash TTS, 4 credits) and replace. You lose the original timbre unless you clone first.
    4. Fit the wav to the picture. If the line overruns, cut words. Do not speed to cartoon.
    5. Replace the master’s audio with the new bed, or take the elevenlabs output as the file.
    6. Watch with sound off, then on. Mute still has to work via burned type (next pass). Sound-on should feel like the same spot, not a radio read on a stranger’s face — and if the face is small, it will.

    When the original mouth is still visible and clearly saying English, a Spanish bed will read as VO, not as a performance. That can be honest: pack demo, English lips tiny in a corner, Spanish bed. If it looks like a bad dub, you answered the mouth question wrong. Go to a mouth pass or crop the face out.

    translate_audio_only on the same dub_video tool is the lighter cousin: translated audio, no cloned-performance pipeline. Still a bed. Still later. Still wasted on a miss.

    Keep background audio only if it is the bed you want in this market. A smash English-language lyric under a Spanish VO is two languages.

    Lipsync only when visemes must match

    Mouth pass means the picture changes.

    Doors:

    • dub_video engine heygen: video-only, lip-sync included, 8-minute cap, fewer languages, no trim range. Glued to this face performance. A new take has new visemes; the old dub does not transfer.
    • Video-to-lipsync rows on a new target-language wav: Kling Lipsync 7 credits, 4K listed; VEED Lipsync 4 credits. Plate + audio in. New mouth out.

    Procedure:

    1. Confirm the face is large enough that a bed would look wrong. If you had to zoom in to decide, crop instead of lipsync.
    2. Lock the cut. HeyGen will retarget this timeline.
    3. Set target_lang and engine. num_speakers if it is a two-hander. Do not set two speakers on a one-person UGC.
    4. Submit. Poll get_dub_status. This is async.
    5. QC the mouth on the claim sentence, not only on the greeting. Names and numbers. A viseme miss on the price is a miss.
    6. If the face is a generated synthetic, you still owe whatever disclosure the market requires. A new mouth does not change that.

    Do not lipsync a pack-only plate. There is no mouth. Do not lipsync a 15-second spot because “that is what dubbing means.” Dubbing means the language changed. Lipsync is optional last stage.

    If the original was silent Turbo plus TTS, you do not “dub” silence. Translate the script, new TTS, replace (and lipsync if the face is the product). Dub tools want a spoken original.

    Caption the language you now hear

    Captions transcribe. They do not translate. After the bed or the mouth pass, burn the new soundtrack.

    Procedure:

    1. Picture lock on the dubbed (or replaced) child.
    2. Add captions to video with the language code that matches what is spoken now (es-MX, not “Spanish vibes”). 165 codes. BASIC 1× unless you are testing DYNAMIC 2×.
    3. Spot-check the offer, the SKU, and any legal overlay. Auto-transcribe will not invent a disclaimer you still owe in this market.
    4. If on-screen English type remains (price cards, pack text you burned as overlay), that is a graphics problem. Dubbing audio does not translate pixels. Forced narrative on the pack is a different page. Either recut the overlays or leave trademarked letters and translate the claim elsewhere.
    5. Watch muted in the target language as a person who does not hear. Then watch with sound.

    Do not reuse the English caption file with a shifted clock. The new language does not fit the old cues.

    Archive the source master, the translated script, and the dubbed child. A recut of the English is a recut of this stack.

    Pick bed or mouth, then run the language in Versely. Bring a locked UGC file and a target language. Do not bring “make it match” if there is no mouth on screen.

    FAQ

    Is a cloned-voice bed “not a real dub”?

    It is a dub without a mouth pass. AI dubbing always translates and speaks. Lipsync is the optional last stage. Use it when lips are the product.

    When do I use HeyGen versus ElevenLabs on this catalog?

    ElevenLabs: bed, up to 30 minutes, trim, audio or video, no viseme retarget. HeyGen: video-only mouth pass, 8-minute cap, fewer languages. Pick from the mouth test, not from the brand name.

    Should I caption English and upload that file to a Spanish slot?

    No. Caption the Spanish audio you actually laid in. Transcription is not translation.