Pixazo’s AI VFX Studio is an AI video translator that runs end to end: give it a video you already have and it will translate video dialogue into another language, clone the speaker’s own voice from that same clip, speak the translated line, and re-time the mouth to the new audio. It is free to start — a 7-day trial with no card and 100 bonus credits. In a measured English-to-Spanish run on 29 July 2026 a five-second, single-speaker clip took about fifteen minutes, and both clips are on this page with sound.
Illustration
Measured run · 29 July 2026
It keeps the picture and the speaker’s own voice, replaces the language, and re-times the mouth to the new audio — the two clips below are the genuine output of one run on 29 July 2026, and because the evidence is audio you have to press play.
The clip file is about five seconds long; the transcriber reported 5.68 seconds of speech. We print both exactly as they came back and do not reconcile them.
| Stage | What the agent called | What came back | Measured time |
|---|---|---|---|
| 01 Transcribe | pixazo_transcribe |
“Detected English, 5.68 seconds, 1 speaker”, plus the timed transcript “Welcome to our community. We’re so glad you’re here.” | 13 s |
| 02 Translate | the agent itself — no separate translation model | “Bienvenidos a nuestra comunidad — nos alegra mucho que estés aquí.”, phrased to fit the spoken duration it had measured | not timed separately |
| 03a Pull the source audio | pixazo_extract_audio |
The source track, reused as the voice-clone reference | 9 s |
| 03b Clone the voice | pixazo_add_voice |
A voice built from the clip’s own speech — no separate voice sample was supplied | ~10 min (the bottleneck) |
| 03c Speak | text-to-speech in the cloned voice | The translated line in that voice | 42 s |
| 04 Re-sync the mouth | pixazo_lip_sync, sync_mode “remap” |
The picture time-fitted to the new audio | 3 min 09 s |
| End to end ≈ 15 minutes · whole session $1.88, including generating the source clip · which vendor model performed the text-to-speech or the lip-sync was not disclosed by this run, so this page does not name one. | |||
These two clips are the only genuine dub output on this page. Every other image and loop here is an illustration generated to explain a stage, and every one of them carries a visible “Illustration” label and says so in its alt text.
Four stations
Four stations run in order — transcribe, translate, voice, then lip sync — and each one hands its output straight to the next, which is why one instruction to the agent is enough to get a finished dub back. That chain is what separates an AI video language translator from a voiceover: the old picture is not left to drift, the mouth is rebuilt against the new spoken track.
Station 01 · Transcribe
13 s
Illustration
The agent listens to the clip you gave it and writes down what is actually said, with timings attached.
Station 02 · Translate
no separate call
Illustration
There is no separate translation model in the chain — the agent phrases the new line directly.
Station 03 · Voice
~10 min
Illustration
The source track is pulled out, used as the clone reference, and then made to say the translated line.
the bottleneck
Station 04 · Re-sync
3 min 09 s
Illustration
Remap re-times the footage against the translated take so the mouth lands where the new words do.
What runs where
AI VFX Studio is the only place that runs the whole chain: the playground has no speech-to-text and no translation model, so driving models there by hand covers the voice and the re-sync only, and you must bring your own translated script.
Give AI VFX Studio a clip and a target language and it walks all four stations itself. In the playground you are opening one model at a time, so the transcript and the translation have to arrive already written — there is nothing there that can produce either.
Illustration
Illustration — generated to picture the re-sync stage, not dub output
| Job | AI VFX Studio (one instruction) | Playground, models driven by hand |
|---|---|---|
| Speech-to-text transcript | Covered. Covered | Not available. Not available — bring your own transcript |
| Translation | Covered. Covered — the agent writes it itself | Not available. Not available — no translation model here |
| Voice-clone reference | Covered. Covered — pulled from the clip’s own audio | Not verified. Not a verified capability here — the measured run never cloned a voice from a reference in the playground |
| Cloned-voice speech | Covered. Covered | Covered. Covered — Chatterbox v1, 4 credits |
| Lip re-sync to new audio | Covered. Covered | Covered. Covered — MultiTalk, from 104 credits |
| Separating dialogue from music and effects | Not available. Not available | Not available. Not available |
| Subtitle or SRT file out | Not available. Not available — the timed transcript is printed on screen, not exported as a subtitle file | Not available. Not available |
| Replacing text burned into the picture | Not available. Not available | Not available. Not available |
Measured, not estimated
Because cloning the voice is the bottleneck — roughly ten of those fifteen minutes go into building a voice from the clip’s own speech, and every other station finishes in seconds.
Bar widths are each station’s share of about 850 seconds of measured stage time, counting the roughly ten-minute voice clone as ten minutes flat; the translate step was not timed separately, so it is not in that total. This is one measured run on 29 July 2026, not a benchmark — a longer clip, more speakers or a different language pair will not produce the same numbers.
The voice, by hand
Open the Chatterbox v1 station in the playground at 4 credits, paste in the translated line you already have, and you get the spoken audio without running the rest of the chain. Used this way it is an AI voice translator for video rather than the full chain — the playground cannot transcribe or translate for you, so the script has to be yours. Your 100 signup credits cover this station many times over.
One model, one line of text, one take — useful when the translation already exists and all you need is the speech.
What it does. It speaks a line of text you provide, at 4 credits a run. It covers the same job as station 03 — speaking a line — as a model you run yourself, so you can iterate on a take without re-running a transcript or a re-sync. Which model ran inside the chain was never disclosed, so that is a comparison of jobs and not an identification.
What it will not do. It will not transcribe your clip and it will not translate anything — the playground has no speech-to-text and no translation model, so the translated script must arrive already written. It also does not touch the picture; the mouth is still saying the original words until you run a re-sync. Cloning a voice from a reference you upload is not something the measured run tested here, so this page does not claim this station does it.
Open the voice stationThe checkpoint
You do, and it is the only checkpoint there is — the agent prints the transcript and the translated line on screen before it generates a single second of audio, because nobody has read-checked either and no accuracy figure has been published for the transcript or the translation.
The only checkpoint
Illustration
Check it against what is actually said in the picture. The transcriber is the first thing in the chain, so a word it mishears becomes a word the translation is built on.
Check the meaning, and check it can be said inside the duration the transcriber measured. The agent phrases to fit that duration, which is a performance decision you are signing off on.
Everything after this point inherits whatever you approved, in a cloned voice. The voice station and the re-sync do not re-read the script; they carry it forward exactly as approved.
Who it’s for
Anyone who already owns finished footage in one language and needs the same speaker saying it in another. As an AI translator tool for video editing and video content teams, it fits course and tutorial teams, support and onboarding owners, product marketing, creator channels, internal comms, and localisation leads testing a language before commissioning a studio dub.
You start from a finished lesson recording with one instructor talking to camera, and you want the same instructor in a second language.
You start from a short walkthrough clip in your help centre where the voice matters more than the visuals.
You start from an approved, locked cut and you need a second language without re-opening the edit.
You start from your own to-camera footage, where the voice is yours to clone and permission is not a question.
You start from a leadership or policy message recorded once, and it has to reach teams who do not share that language.
You start from one representative clip and want to hear a language before you commission a studio dub for the whole library.
Holds / breaks
It holds up on a single speaker framed front-on and holding reasonably still, and it does not hold up on profile framing, on frames with two people talking, or under heavy motion blur — remap re-times the picture to the new audio, it does not repaint a face.
Holds
Illustration
One face, facing camera, mouth unobstructed — the re-sync has a clear reference to work from.
Breaks
Illustration
In profile the mouth is barely in frame, and motion blur destroys what is left of the reference.
Breaks
Illustration
With two speakers in one frame the re-sync has no way to know whose mouth to drive.
What it does. It re-syncs a mouth in existing footage to an audio track you hand it, from 104 credits a run. It covers the same job as station 04 — re-timing picture to new audio — as a model you run yourself, so you can re-time a clip against a take you produced elsewhere. Which model ran inside the chain was never disclosed, so that is a comparison of jobs and not an identification.
What it will not do. It will not translate, transcribe or produce the voice — the audio has to exist first. It also will not rescue framing it cannot read: profile shots, more than one speaker in frame, and heavy motion blur stay unreliable, and no setting changes that.
Open the lip-sync stationThe Holds verdict is what the measured run showed, on one front-facing single-speaker clip. The two Breaks verdicts are known limits of remap re-timing, not results from this run — it contained no profile framing, no second speaker and no heavy motion blur. Nothing here is a scored test set, and there is no published pass rate.
Honest limits
Four things flatly: it cannot pull dialogue back out of an already-mixed track, it cannot replace text burned into the picture, it cannot promise reliable lip-sync on profile framing, multi-speaker frames or heavy motion blur, and it cannot give you a guaranteed accuracy figure for the transcript or the translation.
The agent named only Pixazo’s own tool names. The catalogue lists several speech and lip-sync families, but this run proves none of them, so this page does not claim which vendor model spoke or re-synced the clip above.
| What you need | Why it matters | Who has to agree |
|---|---|---|
| Permission to clone the speaker’s voice | A cloned voice is that person’s voice; consent to appear is not consent to be re-voiced | The speaker |
| A licence that covers redubbing | A licence to use a clip does not extend to putting new words in its speaker’s mouth | Whoever owns the footage |
| Sign-off on the translated line | The agent’s translation is fitted to the measured duration, so it is a phrasing choice, not a literal gloss | Whoever owns the message |
| Costs from the verified run: $1.88 for the whole session including generating the source clip; by hand, 4 credits at the voice station and from 104 credits at the lip-sync station. No per-minute or per-language pricing is published for a dub. | ||
Your first dub
Four steps: upload a front-facing single-speaker clip, read the transcript and translated line the agent prints, approve them, then leave it alone while the voice clone runs — budget about fifteen minutes for a short clip.
Illustration
One front-facing clip, one instruction, one read-through of what it prints — then about fifteen minutes of waiting, most of it the voice clone.
One speaker, facing camera, mouth unobstructed. That framing is what the re-sync stage needs later, so it is worth getting right before anything else runs.
One instruction is enough; the agent chooses the stations. You are not picking models or wiring a chain together yourself.
The transcript and the translated line appear before any audio exists. This is your only checkpoint — read both properly.
Roughly ten minutes on a short clip, unattended; that pause is expected, not a stall. The re-sync that follows took 3 min 09 s in the measured run.
FAQ
These ten cover timing, voice samples, language pairs and direction, what a dub leaves untouched, cost and the free tier, permission, using a YouTube upload as the source, and what the run did not disclose.
About fifteen minutes end to end for a five-second, single-speaker clip in the measured run on 29 July 2026, and roughly ten of those minutes were the voice clone alone.
Transcribing took 13 seconds, pulling the source audio 9 seconds, speaking the translated line 42 seconds and re-syncing the mouth 3 minutes 9 seconds. A longer clip, more speakers or a different language pair will not produce the same numbers.
No — the run cloned the voice from the clip’s own speech, with no separate sample supplied.
The source track is pulled out with pixazo_extract_audio and reused as the clone reference, which also means a noisy or music-heavy original gives the clone a worse reference to work from.
The agent writes the translation itself rather than calling a separate translation model, and the verified run was English to Spanish.
Because there is no separate translation model in the chain, there is no published list of supported pairs to quote, and no accuracy figure for the translation either. Note also that the playground has no speech-to-text and no translation model at all — only AI VFX Studio runs the whole chain.
It replaces the whole audio bed and re-times the picture, and it touches nothing else.
It cannot separate dialogue from music and effects inside an already-mixed track, and it cannot replace text burned into the picture — titles, lower thirds and hard captions stay in the original language.
The whole verified session came to about $1.88, and that figure includes generating the source clip, not just the dub.
If you drive the stations by hand instead, Chatterbox v1 is 4 credits and MultiTalk is from 104 credits per run. There is no published per-minute or per-language dub price.
Yes — cloning a real person’s voice needs that person’s permission, and a licence to use footage does not extend to dubbing it.
Clear both separately before you dub anything you did not shoot yourself.
Not disclosed: the agent named only Pixazo’s own tool names — pixazo_add_voice, pixazo_lip_sync.
The catalogue lists several speech and lip-sync families, but this run proves none of them, so this page does not attribute either step to a vendor.
Either direction. The transcribe stage detects the source language itself — in the measured run it reported “Detected English” without being told — so you name the language you want out and it works from whatever went in.
The run on this page went English to Spanish. Spanish to English, or any other pair, is the same instruction with a different target language; coverage of the target is set by the speech model that speaks the line, not by Pixazo.
You need the video file, not a link. Nothing in the measured run imported from a URL, so export the clip from YouTube — or use your own master — and upload that.
A single speaker facing camera is what the re-sync stage handles best, so a talking-head upload dubs far more cleanly than a fast-cut montage.
Yes, free to start: a 7-day trial with no card required and 100 bonus credits, which is what the pricing page states. Those credits go a long way at the voice station — Chatterbox v1 costs 4 credits a line.
Be realistic about the whole chain, though. The lip re-sync station starts at 104 credits, so a complete end-to-end run needs more than the signup grant. The measured run on this page cost about $1.88 in AI VFX Studio, and current plan details are on the Pixazo pricing page.
Deepak Joshi
Content Marketing Specialist · Pixazo
Deepak documents Pixazo’s speech, voice-cloning and lip-sync models by running them rather than reading their spec sheets. The dub on this page is his own test run in AI VFX Studio on 29 July 2026 — the transcript, the translation, the step timings and the $1.88 session cost are all copied from that run, and the two clips are its actual output. Where the run did not reveal something, such as which vendor model spoke the line, this page says so instead of guessing.
Explore
The ones that make or extend the footage you are dubbing — start upstream if you still need the video, and use the directory if you are looking for a different job entirely.
Illustration
Re-time a mouth in footage you already have to an audio track you already have — station 04 on its own.
Illustration
Work on the voice side by itself, when the speaker’s voice is the thing you need and the picture is not.
Illustration
Make the footage in the first place. The English clip at the top of this page started here.
Illustration
Start from a still instead of a shoot, then bring the result back here to dub it.
Drop one front-facing, single-speaker clip into AI VFX Studio and approve the transcript and translation it prints — that is exactly how the two clips at the top of this page were made.
Open AI VFX Studio
