Percify Voice Studio: Clone a Voice, Then Dub

Voice Studio answers two questions that feel like one:

1. Whose voice says this? Yours, cloned from a sample, or one of the presets.
2. What language does it say it in? The original, or a dub.

They share a screen because they share a constraint: everything downstream — the talking avatar, the podcast, the ad — is driven by the audio track. Get the audio right and the video follows. Get it wrong and no amount of re-rendering the face will save it.

Voice cloning takes a reference recording of you talking. The engine we run, `chatterbox-turbo`, needs at least 5.6 seconds of it.

If your sample is shorter, we do not reject it. We pad it — the clip is repeated until it reaches about 6.2 seconds, and the padded version is what the engine hears. That keeps the feature working on a three-second recording instead of throwing an error at you, and it is genuinely the right fallback.

But it is a fallback, not a feature. A repeated three-second clip contains three seconds of vocal information no matter how many times it loops. The clone will be thinner, flatter and more prone to odd emphasis than one built from a real sample.

So the practical advice is: record fifteen to thirty seconds. Not because we require it, but because the floor is where quality starts, not where it lives. Say something with range in it — a question, a statement, a number — rather than reading one flat sentence.

Cloning costs 5 credits, once, and the voice is then yours to reuse. That is roughly what a couple of images cost, which is deliberate: this is a setup step, not a per-use charge.

There are 52 preset voices on the platform. They are worth using when:

• You are drafting and do not want to spend a clone on a script that may change.
• The video is not about you — a product explainer, a faceless channel, a training clip.
• You want two distinct speakers, which is what Podcast Studio is built around.

They are the wrong choice when the whole point is that it is you. A cloned voice on your own avatar is the difference between a spokesperson video and a stock one, and audiences hear it even when they cannot name it.

Dubbing looks simple: translate the words, speak them in the same voice, keep the video. Two of those three are straightforward. The third is where dubbed video goes wrong, and it goes wrong quietly.

The mouth must be regenerated from the translated audio. If the face was animated from the original English and you then swap in the Spanish track, the timing can look almost right while every mouth shape is wrong — because visemes ↗, the visible shapes that correspond to sounds, differ between languages. Viewers cannot usually say what is wrong. They just find it unconvincing and scroll.

This is the single most common reason a dubbed clip feels off, and it is not a quality setting you can raise. Retrying the same way will produce the same fault. The full diagnosis is in why AI lip sync looks wrong, and the step-by-step version is the dubbing tutorial.

The second thing dubbing changes is length. A sentence that runs four seconds in English can run five and a half in German and three in Japanese. If your video has hard cuts timed to the original, a dub will drift against them. Either leave the edit loose, or accept that the dub is a re-cut rather than a re-voice. Anyone who has worked with subtitle timing standards ↗ already knows this problem by another name.

The engine is copying what it hears, including the room. Three things matter more than your microphone:

1. Soft surfaces. A room with curtains, a sofa or a bed beats a room with a good microphone and bare walls.
2. Consistent distance. Moving closer and further mid-sentence teaches the clone to do the same.
3. Even loudness. If your recording swings wildly, normalise it — ffmpeg's loudnorm filter ↗ does it in one command, and it is the same fix professional audio uses.

What does not matter: expensive hardware, a pop filter, or recording in a cupboard. A phone held at a steady distance in a carpeted room is a good sample.

Everything is one credit meter. There is no separate audio subscription and no per-minute voice bill layered on top:

Video is where the money goes — a talking-head render averages 69 credits at 480p — so in practice the audio side of a project rounds to nothing. The full pricing is one meter across images, voices and video, and API access starts on the Scale plan.

Voice is the layer under everything else:

• Avatar Studio takes the audio and drives a mouth with it.
• Clone Yourself pairs the cloned voice with your cloned face, which is the setup most people should do first.
• Content Replication re-performs a short in your voice rather than the original creator's.
• Realtime Video is the one place voice is not the input — it generates picture, fast.

If you are comparing platforms, dedicated voice tools like ElevenLabs ↗ go deeper on audio alone; the trade is that the voice then lives outside whatever renders your video, and you own the sync problem. Our own dubbing page covers that boundary in more detail, and there is a ranked list of lip-sync tools if the video half is what you are really choosing between.

1. Record 15-30 seconds of yourself talking in a soft room, at a steady distance.
2. Clone it once, for 5 credits.
3. Draft scripts against a preset if they are still changing.
4. When you dub, regenerate the mouth from the translated audio — never from the original.
5. Expect the dub to run a different length than the source, and leave the edit room for it.

You can hear what the output sounds like before spending anything.

• Avatar Studio — give the voice a face.
• Build your digital twin — clone the face as well.
• Dub a video step by step — the walkthrough.
• Dubbing, end to end — the product page.
• Every studio at a glance — the whole product on one page.
• What it costs — one credit meter across audio and video.