AI voice, TTS and dubbing
Text to speech means a model reads your text aloud; dubbing means the voice on an existing video is replaced with a new one in another language. For Persian, the one thing that actually decides the outcome is whether your chosen model lists Persian among its languages, and at one well known vendor the model everybody reaches for first does not.
- Lesson 9 of 10
- Intermediate
- Free, no signup
Character ceiling in one request, from the vendor own page
The column is the character ceiling. The three chips below are three text to speech models from one vendor.
- Eleven v3, has PersianIts language list writes Persian with the code fas. A ten minute script becomes several files here.
- Multilingual v2, 29 languagesAll 29 languages are named on the page and Persian is not among them.
- Flash v2.5, 32 languagesThe same 29 plus Hungarian, Norwegian and Vietnamese. The fastest and cheapest, and without Persian.
This column measures length ceiling, not quality: the tallest row here is the model without Persian. The numbers are the ceiling of one request, not of a project; longer text you cut yourself, on a sentence boundary.
Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.
TTS and dubbing are two different jobs
Text to speech takes one input: text. You hand over sentences and get back an audio file. Timing, pauses, and which shot the voice sits under are none of the model business.
Dubbing is a different job with three stages, only one of which is text to speech: transcribing the original speech, translating it with sentence length in mind, and then producing a new voice that fits inside the same time window. The third stage is where it gets hard, because a Persian translation of an English sentence usually comes out longer, and the voice has to fit a frame the original speaker mouth already set.
So if you are making your own video and you wrote the script, you do not want dubbing; text to speech is enough, and it comes with one real advantage: you can rewrite the script for the ear and set its length yourself. If you are bringing somebody else video into Persian, that is dubbing, and you should expect to shorten sentences to make them fit. That is a translation decision, not a software setting.
There is a third job people also call voiceover and should not: voice changing, taking a recording of one person and covering it with another voice. That one takes audio rather than text as its input, and it raises the question of consent earlier than the other two.

New narration or dubbing? It depends whose script it is
Both routes are correct, for two different situations. Choosing the wrong one means fighting a timing problem you never had to have.
The script is someone else: dubbing
- Three stages: transcribe, translate, generate the voice
- A Persian translation usually comes out longer than the original
- You have to shorten sentences to fit the same window
- If the original speaker voice is rebuilt, consent enters the picture
The script is yours: new narration
- You set the sentence length yourself and fit it to the picture
- The script can be rewritten for the ear before generation
- The same script doubles as the subtitle text later
- One stage fewer: no transcription needed
Neither route solves the absence of a Persian voice. That is a model limitation, and the next section opens it.
The trap that catches Persian users
ElevenLabs has several text to speech models. Persian is on the list of only one family.
On the day this lesson was written we re-read both of the vendor own pages, and the evidence does not get cleaner than this. The text to speech capability page says the multilingual v2 models support 29 languages and then names all 29: English, Japanese, Chinese, German, Hindi, French, Korean, Portuguese, Italian, Spanish, Indonesian, Dutch, Turkish, Filipino, Polish, Swedish, Bulgarian, Romanian, Arabic, Czech, Greek, Finnish, Croatian, Malay, Slovak, Danish, Tamil, Ukrainian and Russian. Persian is not on that list. The next sentence on the same page says Flash v2.5 supports 32 languages, being all languages from the v2 models plus Hungarian, Norwegian and Vietnamese. Three languages were added and none of them is Persian, so Persian is not on the Flash list either, and we know that from the definition of the list rather than from a guess.
By contrast, the Eleven v3 language list writes Persian out explicitly with the code fas. The data layer of the AI section of this site recorded that on 28 August, and it was still there today.
Now why this is a trap rather than a simple difference. The same vendor models page introduces Flash v2.5 with three properties, all three attractive to an editor: ultra low latency at around 75ms, a ceiling of 40,000 characters in one request, and fifty percent lower price per character on the API. Against those three, the model that has Persian carries a 5,000 character ceiling, an eighth as much. So someone signs up, makes the most reasonable choice available, feeds it Persian, gets a poor result, and concludes that this vendor cannot do Persian. The accurate sentence is: that model cannot.
Our position, and the lesson that outlives this one vendor: choose the model from its language list, not from its speed or price badge. In no other category is that rule this necessary, because in translation or code a bad output looks bad immediately; here the output is a fluent voice reading words wrong, and someone who does not speak the language cannot tell.
What a language list does not tell you
A language list is a vendor claim, not a measurement. It says we support this language and says nothing about how the accent comes out. That distinction is worth stating plainly, because any table carrying this column is tempted to read it as quality.
Two specific things a language list cannot tell you. First, whether a Persian voice exists. The voice ranking in the AI section of this site has three levels for Persian, and level two means the vendor also announced a voice built for Persian; no candidate has reached it. Which means even the model that lists Persian reads it in a voice built for another language. The vendor own guide says that for the most natural result you should choose a voice whose accent matches your target language and region, and when no Persian voice exists that sentence is itself the answer.
Second, names and numbers. No language list tells you whether a model reads an Iranian city name or a family name correctly. There is a practical note in that same vendor documentation that few people read: the multilingual v2 model does a better job of normalising numbers, and their own recommendation for low latency applications is to have a language model normalise the text before it reaches the speech model. Translated into practice for Persian: write out numbers, dates and abbreviations as words yourself, then send. That is one line of work and it rescues more output than any setting.
And one thing we do not know and will not claim to: how it actually sounds. We have not run these models on one identical Persian script and compared them. When we do, we will publish the number with its method.
Voice cloning: consent, licence, and a line that does not move
Building a new voice from a recording of a person is technically easy today and just as heavy ethically and legally. Our position is one sentence and it is not negotiable: never clone anyone voice without their explicit, written consent, even when the voice is publicly available and even when the work is internal.
The reason is not only the law. A voice, like a face, is identity, and unlike a face it also works as identity verification on a phone call or a voice message. A short sample built for one project still exists after that project ships, and you do not control it. If you are doing work in which a real person voice gets generated, get consent, write the scope of use into that consent, and delete the files when the project ends.
And a contradiction that has to be said plainly. The model that ranks first in the voice ranking of the AI section of this site names Iran in its own terms of use: the ElevenLabs export control clause asks a user to warrant that they are not in a country under US economic sanctions, and Iran is written into that list. We write this because it is true and hiding it helps nobody. What we do not write is how to get around it. We do not sell accounts, we do not refer intermediaries, and we do not explain sanction workarounds.
There is an alternative route, and it has to be stated with its condition attached: Mistral Voxtral TTS publishes open weights, so it can run on your own hardware with no country list in the way. Its licence, though, is CC BY-NC 4.0, and NC means no commercial use. So in work whose whole nature is custom voiceover for a client, the model that is best on access is the least usable on licence. For practice and a personal project, yes; for an invoice, no.
Fill this sheet in before you build a new voice
What you must have before you start
- Written consent from the owner of the voice, naming the project
- The scope of use, spelled out: where, for how long, which languages
- The licence of the model you are using, read rather than assumed
- A decision about when the sample files get deleted
Things that do not count as consent
- The voice was available on the internet
- It is internal work and will not be published
- They said it was fine, out loud
- It is only a mockup and will not be in the final
This sheet is not legal advice and does not replace a contract. What it does is stop you making undeliverable decisions on the spot and without thinking.
Voicing a video, in the order that saves the job
The order matters more here than the choice of model, and the common mistake is always the same one: the video gets edited first and the voice is hunted for afterwards.
Write the script for the ear, not for the eye. Short sentences, one idea per sentence, no parentheses and no subordinate clauses folded into the middle. A text that is read and a text that is heard are two different things, and the model will not tell them apart for you.
Then normalise the script: write numbers, dates and abbreviations out as words. A year, a percentage and a set of initials are three things a model may read differently every time, and ten minutes of work takes that out of the hands of chance.
Now cut the script into pieces that fit under your model character ceiling, and put the cut on a sentence boundary rather than inside a sentence. With a 5,000 character ceiling a ten minute script becomes several files, and if a cut lands mid sentence the melody of that sentence breaks at the join. This is the one place where the character ceiling genuinely hurts you, and the fix is not technical, it is an editorial decision.
Generate, then listen, and regenerate only the sentences that were read wrong. Regenerating the whole script hoping for a better result is the same mistake people make with images, and it does not work here either.
Finally, lay it against picture. If the video was generated, keep the shots silent and build the whole audio bed in one pass; the lesson on AI video generation explains why. And if the output is also getting subtitles, that same normalised script is your subtitle text and there is nothing left to transcribe; the lesson on subtitles and captions starts exactly there.
The fast path, with AI
The fast path here is not a better model. It is two moves you will almost never see in a Persian tutorial: rewrite and normalise the script for the ear with a text model before it ever reaches the voice model, and mark the cuts under the character ceiling on sentence boundaries in the same pass. The second move is what turns several files into one continuous narration. A fast, cheap text model is enough for this; our current pick lives in <a class="text-link" href="/en/ai/">the AI section</a>, and take the voice model from <a class="text-link" href="/en/ai/audio/">the voiceover ranking</a>.
- Pick the model from its language list first, not from a speed badge. Write down that model character ceiling; the rest of the path is built on that number.
- Give the recipe below to a text model with your own script and ceiling. Part two of the output, the list of things the model is likely to misread, is the most valuable thing you get back.
- Generate the pieces in order, all with one voice and one setting. Do not change the voice halfway; two pieces in two similar voices sound worse than two pieces in two clearly different ones.
- Listen, and regenerate only the sentences that were read wrong. If a name comes out wrong twice, write its phonetic spelling into the script rather than generating again.
- At the end, check two things yourself: that names and numbers were read correctly, and that if a real person voice is involved you hold their written consent.
Copy-ready recipe
Role: a spoken-copy editor preparing a script for a text to speech model.
Script language: {your language}
My model character ceiling: {e.g. 5000}
Tone: {e.g. calm and explanatory, or high energy and short}
Raw script:
{paste your script here}
Return exactly these four parts:
1. The script rewritten for the ear. Break long sentences, remove parentheses
and mid-sentence clauses, and write every number, date and abbreviation out
as words. Do not change the meaning and do not add anything.
2. A list of what the model is likely to misread: proper nouns, place names,
an English word inside a sentence in another language, numbers. For each
one, propose a phonetic spelling and say why it is risky.
3. Cut the rewritten script into pieces that each stay under the ceiling
above. Every cut must land exactly at the end of a sentence. Give each
piece a number and its character count.
4. One sentence: which part of this script a person has to read before any
audio is generated, and why.
Make no claim about the quality or the accent of any model.Before you trust the output: This path does not solve three things, and all three are a person job. First, accent: normalising the script stops numbers and names being misread, it does not fix the accent, because no Persian-specific voice exists today. Second, the consent line: no recipe substitutes for permission to use a person voice. Third, licence: a model with open weights does not necessarily permit commercial work, and you have to read its licence yourself. Fast does not mean unsupervised.
AI in this kind of work
This whole lesson is about working with AI, so this section gathers what the text scattered: which model announced what, and where each stands from Iran. Our position in one sentence: <strong>in this category the language list is the most important column and every other criterion comes after it, because a model without Persian scores zero for Persian text rather than slightly worse.</strong>
Tools that actually help
- Eleven v3 The only top-tier model here that writes Persian into its language list with the code fas, plus new-voice creation with a commercial licence on the paid plans. Its ceiling is 5,000 characters in one request, an eighth of the fast model from the same vendor. The ElevenLabs terms of use name Iran in their export control clause, and we do not explain workarounds.
- Eleven Flash v2.5 It is here because it is the common wrong choice: the fastest at around 75ms, a 40,000 character ceiling and fifty percent cheaper per character on the API. Its 32 language list defines itself as the v2 list plus Hungarian, Norwegian and Vietnamese, so Persian is not in it. A good pick for English work and not a pick at all for Persian.
- Voxtral TTS The only model here with published weights, so it runs on your own hardware with no country list in the way. Its licence is CC BY-NC 4.0, and NC means no commercial use, which is exactly what a voiceover artist wants to do. Mistral publishes no language list for it either, so we make no claim about its Persian.
- ElevenLabs The vendor page, for when you want the whole family of its models side by side and need to see which model carries which capability. It is where the finding in this lesson was first recorded, on 28 August, and we read it again today and it was still there.
Where it backfires
Three specific risks here, and only one of them is technical.
First, the silent failure. A model that does not carry your language in its list does not return an error; it hands you a fluent voice reading words wrong. If you do not speak the language yourself, you will not see that failure. The only defence is reading the language list before choosing the model.
Second, consent. Building a person voice without their explicit permission, even for a mockup and even internally, is a red line. A voice works as identity verification on a voice message and a phone call, and a sample once built exists after the project ends.
Third, a licence you did not read. Open weights means you can run it, not that you can sell it. Voxtral TTS is the plain example. Before you invoice any work, read the model licence yourself.
And on access from Iran we suggest no way around anything. To see where each tool stands, read AI in Iran and the buying guide.
Sources: ElevenLabs: models, character limits and latency ElevenLabs: text to speech supported languages ElevenLabs: terms of use, export control Mistral: models overview, Voxtral TTS licence
Where this advice stops
This lesson did not run any of these models on one identical Persian script and no number in it comes from a listening test of ours; everything you have read comes from the vendors own lists and documentation. Three things are deliberately absent. Any judgement about Persian accent quality, because no vendor publishes a figure and we have not measured one. Any price column, because these vendors units are not comparable: one sells monthly credits, one bills by tokens and one only publishes weights. And any route to access from Iran. One more boundary: this lesson is about generating speech, not about transcribing it; if your work is transcription, the subtitles lesson is your answer rather than this one.
From our own work
This finding was first recorded in the data layer of the AI section of this site rather than in a blog post: in the voiceover category the criterion of whether a model reads Persian carries 35 of 100 weight, the heaviest in that table and unlike anything in the other categories. The reason is written on that same page: in code or images the output language is not the question; here the output is language. On the day this lesson was written we re-read both vendor pages and the finding was still standing, with one piece of evidence that was not in the original record and that we found today: the 32 language list for Flash v2.5 defines itself as all languages from the v2 models plus Hungarian, Norwegian and Vietnamese. Which means the absence of Persian from that list is no longer the result of counting a long list, it falls out of the definition of the list. The same check takes a minute for any other tool: open its supported languages page and look for the name of your language rather than for the total count.
Real follow-up questions
Which AI is best for Persian voiceover?
In the voiceover ranking in the AI section of this site, Eleven v3 scores highest, because it is the only top-tier model that both lists Persian and offers new-voice creation with a commercial licence. Just be careful to select that model by hand; the faster default is the one without Persian.
Why did my Persian output come out badly?
Check the model before the settings. If the model you picked does not carry Persian in its language list, no setting rescues it and that same model gives you the same result at any setting. If the model was right and names and numbers still came out wrong, normalise the script: write numbers, dates and abbreviations out as words.
Can I clone a well known voice?
No, and that answer has nothing to do with whether a sample is available. A voice is identity and must not be rebuilt without the explicit written consent of its owner, even for a mockup and even when it will not be published. If you want a similar voice, use the stock voices and state plainly that the voice is synthetic.
How many audio files will a ten minute video take?
It depends on your model character ceiling rather than on the length of the video. With a 5,000 character ceiling a ten minute script becomes several files and you have to place the cuts on sentence boundaries. With a 40,000 character ceiling that same script is one file, but that model has no Persian; that contradiction is the subject of this lesson.