How to Dub a Video With AI in 2026: Translate, Voice, Lip-Sync

AI video dubbing replaces the speech in a video with another language: the script is translated, an AI voice reads it, and, if needed, the speaker's lips are re-synced to the new track. In the IIshki studio this takes two or three steps: ElevenLabs Voice or MiniMax voice the translated text in dozens of languages, and for a face on screen you feed that track into Kling Avatars v2 with a frame from the original — you get a talking speaker in 720 or 1080, up to 5 minutes long. If the face is small or off screen, a plain voice-over without lip-sync is enough. No VPN needed, pay with tokens.

Types of dubbing and which one you need

Before you dub a video with AI, decide what the result should be. It sets how many steps and models you need.

• Voice-over. New speech plays over the clip, the original is muted or cut. Lips don't match, but viewers don't mind when the speaker is small or absent: screencasts, product reviews, vlogs with B-roll.
• Full dub. Original speech is replaced, music and ambience stay. Technically it's the same voice track laid over a clean background track.
• Lip-synced dub. The speaker's mouth moves with the new speech. Needed for talking heads: addresses, expert videos, ads with a close-up face.

How to dub a video with AI: 5 steps

Step 1. Prepare the translated script

Transcribe the speech with timestamps, at least per paragraph. Translate the way a native speaker would say it, not word for word: adapt idioms, units and jokes. Watch the length: if a line doesn't fit the original pause, shorten it now instead of speeding up the voice later. Club members can draft the translation with ChatGPT in the Agent section.

Step 2. Voice the translation

Open ElevenLabs Voice or MiniMax, paste one segment and pick a voice close to the original in gender, age and pace. There is no voice cloning, so choose by ear: try two or three voices on the same line. In ElevenLabs the v3 engine understands emotion and pause tags — handy for matching rhythm to the picture:

[calm] Today I'll show you how to build a shelf in ten minutes. [pause] All you need is a screwdriver and a level.

[excited] This is our new backpack — and yes, it really fits under an airplane seat!

For MiniMax write the same text without tags and set the pace with punctuation: a period gives a pause, a comma a short stop. More on voices and engines in the MiniMax voice-over guide.

Step 3. Choose voice-over or lip-sync

If the speaker's face is small, covered or on screen for a couple of seconds, stop at the voice: put the new track under the video in your editor and remove the original speech. If the face is large and on screen for long, go to step 4.

Step 4. Lip-sync the speaker in Kling Avatars v2

1. Grab a frame where the speaker faces the camera, mouth open or relaxed, face not covered by hands or a mic. At least 300 px per side, jpg, png or webp up to 10 MB.
2. Upload the frame to the Avatar portrait slot in Kling Avatars v2 and the voice from step 2 to the Audio slot.
3. Pick quality: 720 for social media, 1080 for YouTube and websites. The prompt is optional; if you use it, describe behavior, not the words:

The speaker calmly talks to camera, slight nods, natural expressions, the camera is almost still.

The model generates a new video where the person from the frame speaks your track. Background and pose come from the still, so pick a frame that looks like the whole original shot. More tips on portraits and audio in how to make a talking photo.

For short clips up to 30 seconds you can try Seedance 2.5: References mode, Edit task, the source as @Video1 and the new track as @Audio1. It rebuilds the whole scene and doesn't work with every face, so Kling Avatars v2 is the safer choice for talking heads.

Step 5. Mix the audio and assemble

Put the new speech on its own track and the original music and ambience under it at 15–20% volume (if there is no clean background track, generate new music in ElevenLabs Music). Check the joins between segments: pauses between lines should land on edit points. If the source was low-res, run the finished video through Topaz Video Upscale.

How to check a dub before publishing

Watch the finished video end to end, not segment by segment, and run a short checklist:

• Meaning. Ask a native speaker to listen to at least a minute — translation errors are easier to hear than to read.
• Terms and names. Brands, products and people should sound the same in every segment; if the voice reads them differently, spell them in the script the way they are pronounced.
• Rhythm. A line shouldn't start before the speaker opens their mouth or run past a cut.
• Loudness. Speech sits above the music across the whole video, with no jumps between segments.
• Lips in close-ups. On lip-synced segments check the "p", "b" and "m" sounds — that's where drift shows first.

Common dubbing mistakes

Where dubbing pays off fastest

• Tutorials and courses. One lesson, many languages: a voice-over is usually enough, lip-sync only for the intro with the teacher's face.
• Ads and product videos. Short close-up clips are the best case for Kling Avatars v2: 15–30 seconds, several languages from one shoot.
• Expert videos and addresses. A 1–3 minute talking head fits into a single Kling Avatars v2 generation.
• YouTube and Shorts. Split a long video into blocks; dub Shorts with lip-sync and the long version with a voice-over.

If dubbing is part of a bigger pipeline, build it on the Canvas: voice, avatar and upscale nodes are wired together, and editing the text re-runs only the step that changed. If you need to transfer one person's motion to another rather than translate a video, see cloning a video with AI.

What's next

Start with one 20–30 second segment: translate it, voice it with two voices, keep the better one and run it through Kling Avatars v2. Once the voice-and-frame formula works, the rest follows the same recipe. Compare voice engines in the best AI voice-over tools and other ways to voice a video in the best lip-sync tools.