Two things called by one name
Dubbing means the speech is translated and re-recorded in the target language, and the finished audio track is laid over the video. The picture stays exactly as it was.
Lipsync does more: it modifies the picture so that the speaker's mouth fits the new audio track. That is a separate operation, a separate service and separate money. The market looks like this: some providers only dub, some only lipsync, and some do both in one pass. The documentation of the largest voice provider states outright that their dubbing does not include mouth-movement matching — the most expensive misunderstanding in this field.
The chain your video goes through
Machine dubbing is four steps, and it loses something at each of them.
- Transcription. The machine writes down what is said. This is where names, technical terms and misheard passages come in.
- Translation. The transcript becomes target-language text. Length is decided here — and length is what breaks the next step.
- Speech synthesis. A voice reads the target-language text, typically the cloned voice of the original speaker.
- Alignment. The finished audio is timed to the original picture. With lipsync, this is where the picture changes too.
Where the quality is lost
The most common complaint is not pronunciation but that "something is off". Usually that something is emotion. There is a measurement for it: according to a 2026 study, cascade-type systems — where speech becomes text and text becomes speech again — score around 1.7 on a five-point scale for preserving emotional content. The meaning gets through; the delivery does not.
The reason is simple: in the middle step the audio turns into text, and text does not carry that the speaker was irritated, that they smiled, or that they paused for effect. The re-recorded version is therefore technically correct and emotionally flat. Where that matters you need a human voice actor — or at least human post-production on the audio.
The timing, which is the hard part
The second recurring failure is length. A sentence in the target language is rarely as long as the original, and the picture is fixed. Providers pick between two bad options:
- Fixed timing: the audio speeds up or slows down to fit. The result is intelligible but unnaturally paced.
- Flexible timing: the audio runs at its own pace and slides into the next sentence. The result sounds natural but drifts.
- Length-aware translation: the translation model writes a sentence that fits in the first place. According to a 2025 study this brings roughly a 16 percent relative improvement.
The third one is the good answer, and it is available at more and more providers. If there is a switch for it, turn it on — and if there is not, you can do it yourself by editing the transcript: shorter sentences, less filler.
Your language: what works and what does not
The good news is that dubbing has the widest coverage of the whole chain. The dubbing service of the large voice provider handles more than a hundred languages, and the large avatar platforms are in the same range. If your language is spoken by any meaningful number of people, dubbing almost certainly covers it.
The trap is that the dubbing language list and the speech-synthesis language list are not the same list. At that very same provider, the best-known model advertised as "multilingual" covers just 29 languages — a fraction of what the dubbing pipeline handles. And there are providers that will dub into a language but do not offer it for reading a script aloud as a standalone job. Before you pay, find your language on the specific model you are going to use, not on the company's marketing page.
What it costs
- Machine dubbing runs around two dollars per minute, per target language.
- Lipsync is dearer and varies more: one and a half to ten dollars per minute, depending on the model you pick.
- A ten-minute video dubbed into three languages is therefore roughly sixty to seventy dollars — without lipsync.
- With lipsync the same job can go over three hundred dollars.
That is still a fraction of a studio dub, but it is not the "couple of dollars" many people expect.
The quality check you cannot skip
1. REVIEW THE TRANSCRIPT FIRST, before any translation.
Names, brand names and technical terms break here,
and from there they spread into every language.
2. LISTEN to the target-language audio in full, with no picture.
Without the picture you hear the wrong stress and the drift.
3. WATCH picture and sound together, and on mute too.
With lipsync, muted playback shows where the mouth does not fit.
4. CHECK every number and every proper noun.
This is where machine dubbing fails most often.
5. LABEL it if you have to. Cloning your own voice for dubbing
needs no label on several platforms — someone else's voice does.
Common misconceptions
- "Dubbing and lipsync are the same thing." They are not. Two services, two prices, and in some places you only get one of them.
- "If the company supports my language, all its models do." The dubbing and the text-to-speech language lists differ — this is the single most common trap.
- "Machine dubbing replaces the voice actor." For informational and educational content, yes. Where the performance is the product, no: the emotion is measurably lost.
- "Set it up once and it just runs." Reviewing the transcript is manual work on every video, and that step decides the quality of the whole chain.
What to take away
First decide whether you need lipsync at all — often you do not, and then the cost drops to a third. Then check your language on the specific voice model you will be working with. And go through the transcript by hand before you have anything translated: the quality of the whole chain is settled there. On the craft of writing the script, see AI Voiceover, a companion post; on the legal side of voice cloning there is Voice Cloning and Consent, which goes through the permissions; and for talking-head video, Avatar Video for Business picks up the thread.