How to Translate a Video With Voiceover: Subtitles, AI Dubbing, or Your Own Voice

Most "best AI dubbing" roundups rank voices and promise one-click perfection. In practice the thing that decides whether a dub looks real isn't the timbre — it's sync. Does the translated line land inside the same window where the person on screen is actually talking? Translations usually run longer than the source: the same idea takes more syllables in most target languages. So the voiceover either rushes to keep up or finishes early and leaves the lips moving over silence.

Five features let you judge any translator, not just Kutto. Keep them in mind whenever you test a tool.

• Cost — are you paying for voiceover minutes, or just for subtitles?
• Speed — subtitles are quicker than synthesizing and mixing a dub.
• Sync — does the translated line fit the window of the original phrase?
• Keeping the original audio — is the background ducked, or cut out entirely?
• Voice risk — whose voice is speaking, and is there consent to clone it?

That last one matters more than it looks. Pulling a voice off a person in a clip and laying it over a translation without their permission is a fake, not a feature. It's the line between translating a video and putting someone else's voice in your edit.

This is a comparison by how each approach works, not the results of a single run. On cost: subtitle translation doesn't burn plan minutes, while dubbing is billed in voiceover minutes from your plan (overage comes off balance at the voiceover price, with your consent). Current rates live on the pricing page.

If the video carries its meaning through the picture — an unboxing, a recipe, gameplay, a hands-on tutorial — subtitles often do the whole job. The viewer hears the real voice of the original and reads the translation. It's free: subtitle translation doesn't spend plan minutes. And it's faster than dubbing, because there's no synthesis or audio mixing to wait on.

Here's the flow under the hood: the clip is transcribed, speech is split into phrases on the original timing, each phrase is translated, then the words are laid back out across the phrase proportionally to their length. You get the same video with burned-in subtitles in a clip style, plus SRT and VTT files if you'd rather upload captions to the platform yourself.

• Strong: free, original audio stays in full, fast to produce.
• Watch: the viewer is reading, so attention comes off the frame and onto the text.
• Watch: subtitles sit low in the frame — right where a Shorts or Reels caption and the UI already live.

Dubbing earns its place when the voice matters more than the picture: podcasts, interviews, a piece straight to camera where people listen instead of read. The core technical problem is that the translation runs longer than the source and still has to fit the same stretch of time where the person on screen is talking. "Just read the text aloud" doesn't work — it won't fit.

Kutto's translation and dubbing tool works by phrase window. For each line it builds a window — from the phrase's start to its end plus 0.35s, capped before the next phrase and the end of the clip — and fits the translated voiceover into it.

1. 1

   Fits as-is

   If the synthesized line already fits the window, it's left alone.

2. 2

   Speed up to ×1.15

   If it's a little long, the voice is sped up — up to ×1.15, which still sounds like a person rather than a rushed announcer.

3. 3

   Shorten and re-synthesize

   If speeding up isn't enough, the model trims the translation and re-voices it, up to two passes.

4. 4

   Flag as over

   If it still won't fit, the line is marked as over budget rather than mangled. If the translation is much shorter than the source, the voice is slowed slightly instead.

The speed cap is deliberate: past ×1.25 the voice already sounds like a rushed announcer, so everything beyond ×1.15 is handled by trimming the translation instead of pushing the tempo. If your source language is one that tends to blow up on translation, that's where the trimming does most of the work.

One detail about the original audio: it isn't cut, it's ducked — down to about −22 dB inside the phrase windows where the translation plays. Between phrases the original stays as it was, so the laughter, music, and stingers that make a clip feel alive are still there. There's no separate music track at the source, so inside a phrase window the whole background ducks, not just the speech. That's a trade-off of the approach, not a bug.

Kutto's sync rule is simple: the dubbed line ends no later than +0.5s after the original phrase ends. Plus means the voice runs a touch longer than the lips; minus means it finishes early. The numbers below come from one internal dev run on real clips, measured on 2026-10-02 — four runs on specific videos, stated as that run, not a guarantee for any clip you upload.

No run broke the rule — the latest a voice ever ran past the lips was +0.12s. The early finishes, down to −1.71s, show up when the translation comes out shorter than the source: some languages are more compact, and profanity often drops in translation. The voice doesn't "run away" after the lips close, but on a long line the silence at the end is noticeable. That's a limit of the approach, not a failure.

Pulling a speaker's voice off a clip and laying it over a translation without permission is a fake, so Kutto doesn't do it. Dub voices come from the ElevenLabs library instead. If you want something "close to the original," the path is to describe the voice you want (Voice Design), not to clone a specific person.

Cloning your own voice is a different story — it's your voice and your consent. It makes sense when you run a channel and have a recognizable delivery: your translated videos then sound like you, not like a stranger from a library. For that you need the voice owner's explicit consent; without it, no cloning happens.

The flow is the same for subtitles and dubbing — the fork is only at the very end. You can always fix the translation by hand before the clip goes to render or voiceover.

1. 1

   Upload the clip

   A file, a clip from your library, or a link to YouTube, Reels, or TikTok. Pick your target language.

2. 2

   Fix the translation phrase by phrase

   The clip is transcribed and split into phrases on the original timing. Read the translation by eye — clean up cut-off bits, fix meaning and names.

3. 3

   Pick the output

   Subtitles — burned into the clip plus SRT/VTT, no voiceover minutes spent. Dubbing — the clip with a library voice, original ducked.

4. 4

   Grab the voice track

   After dubbing, a separate voice track (m4a) lands in your library. Keep it if you want to re-mix the translation with your own music.

Checking sync on your own clip takes no tools. Find the two longest lines where the speaker is on camera in a tight shot. Pause the moment the lips close and listen for whether the translated voice is still going. If it keeps talking well past the lips, that line is overloaded — trim the translation by hand. If the voice stops well before the lips close, the translation came out shorter than the source; that's not an error, but on a long line you'll see a pause.

Dubbing stumbles in a few predictable places. On long lines where the translation is much shorter than the source, the voice finishes and goes quiet before the lips do. When there's no separate music track at the source, the whole background ducks inside the phrase window, not just the speech. And an exact clone of someone else's voice without their consent isn't done here — the voice comes from the library.

Run through this before the clip goes live, especially the first time you dub a video.

• Check sync on your two or three longest lines by pausing when the lips close.
• Read the translation by eye for cut-off phrases and broken meaning.
• Decide whether you need subtitles over the dub for silent-feed viewing.
• Check your rights to the source clip if it isn't yours.
• If you dubbed, keep the m4a voice track in case you want to re-mix later.

Once the checklist is clean, send the clip to subtitles for a visual video or to a dub for a talking one — and keep the voice track so you can re-mix later.