Best AI for voice generation
ElevenLabs v3 tops this table today, because it is the only frontier model that both lists Persian among its languages and lets you build a new voice with a commercial licence. The thing that surprises most people: the same company faster and cheaper model does not read Persian at all.
- Six models with real evidence
- Persian weighted 35 of 100
- An Iran access column
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Where each tool's score comes from
The ring below is the same set of weights printed in the table headers further down.
- Reads Persian 35
- Voice control 30
- Language count 25
- Access from Iran 10
We chose these weights, and that is the only judgment call in the table. Weight them differently and the order changes.
Today ranking, from the makers own language lists
Each model Persian level was read from the language list its maker publishes, not from listening. A model whose maker publishes no language list at all stays unverified in that column.
| Rank | Tool | Score | Reads Persian 35 | Voice control 30 | Language count 25 | Access from Iran 10 |
|---|---|---|---|---|---|---|
| 1 | ElevenLabs v3 Current pick | 99.0 | 1/2 | 3/3 | 74 languages Source: elevenlabs.io | blocked / no working route |
| 2 | Gemini 3.1 Flash TTS | 80.0 | 1/2 | 1/3 | 77 languages Source: ai.google.dev | blocked / no working route |
| 3 | GPT-4o mini TTS | 55.0 | 1/2 | 1/3 | not verified | blocked / no working route |
| 4 | ElevenLabs Flash v2.5 | 49.4 | 0/2 | 3/3 | 32 languages Source: elevenlabs.io | blocked / no working route |
| 5 | ElevenLabs Multilingual v2 | 48.3 | 0/2 | 3/3 | 29 languages Source: elevenlabs.io | blocked / no working route |
| 6 | Amazon Nova Sonic | 0.0 | 0/2 | 0/3 | 5 languages Source: docs.aws.amazon.com | not verified |
An empty cell means we could not verify that number, not that the tool scored zero.
Behind each number
Every judged score in this table carries a written reason, and the Iran column says how it was checked. Those two are open here. The measurement trail behind each number, which row of which leaderboard and on how many votes, opens under the model it belongs to.
1 ElevenLabs v3
Reads Persian 1/2 Persian is written as fas into the 74 language list for this model in ElevenLabs own documentation. No Persian-specific voice is named, so it does not reach level two.
Voice control 3/3 Voice cloning from a sample, with a commercial licence from the Creator tier upward. The pricing page states instant cloning from Starter and professional cloning from Creator.
Access from Iran blocked / no working route ElevenLabs own terms of use name Iran explicitly: the user warrants they are not located in a country under US economic sanctions, and Iran is written into that list. We read it in the export control clause on that page rather than measuring the network.
2 Gemini 3.1 Flash TTS
Reads Persian 1/2 Persian is on the language list on Google speech generation page. Google names no Persian-specific voice and says the model detects the input language automatically, so it does not reach level two.
Voice control 1/3 Thirty preset voices plus tone control through a text instruction. Building a new voice from a sample is not offered on this page.
Access from Iran blocked / no working route The Gemini API available-regions page lists the countries where the API works, and Iran is not on that list. Read there rather than measured.
3 GPT-4o mini TTS
Reads Persian 1/2 Persian is explicitly listed among the supported languages in OpenAI text to speech guide. No Persian-specific voice is named, so it does not reach level two.
Voice control 1/3 Thirteen preset voices plus tone control through an instruction. Building a new voice from a sample is not in this guide.
Access from Iran blocked / no working route Iran is not on the supported-countries list in OpenAI developer documentation. Read there rather than measured.
4 ElevenLabs Flash v2.5
Reads Persian 0/2 The 32 language list for this model in ElevenLabs own documentation does not include Persian. Arabic and Turkish are there; Persian is not.
Voice control 3/3 Voice cloning from a sample, with a commercial licence from the Creator tier upward. The pricing page states instant cloning from Starter and professional cloning from Creator, with commercial use on the same tiers.
Access from Iran blocked / no working route ElevenLabs own terms of use name Iran explicitly: the user warrants they are not located in a country under US economic sanctions, and Iran is written into that list. We read it in the export control clause on that page rather than measuring the network.
5 ElevenLabs Multilingual v2
Reads Persian 0/2 The 29 language list for this model in ElevenLabs own documentation does not include Persian.
Voice control 3/3 Voice cloning from a sample, with a commercial licence from the Creator tier upward. The pricing page states instant cloning from Starter and professional cloning from Creator, with commercial use on the same tiers.
Access from Iran blocked / no working route ElevenLabs own terms of use name Iran explicitly: the user warrants they are not located in a country under US economic sanctions, and Iran is written into that list. We read it in the export control clause on that page rather than measuring the network.
6 Amazon Nova Sonic
Reads Persian 0/2 Amazon names five languages for this model: English US and UK, French, Italian, German and Spanish. Persian is not among them.
Voice control 0/3 Preset voices from the maker. Building a new voice from a sample is not offered in the Nova documentation.
Not enough verified data to rank
These are real tools we track, but more than half of their criteria have no verified number yet, so ranking them would be a guess.
- Voxtral TTS
How this ranking is calculated
Every criterion below has a weight and a source. Change a weight and the whole table recomputes. There is no hand-placed position anywhere in this hub.
| Criterion | Weight | Evidence |
|---|---|---|
| Reads Persian | 35 | defined scale, with a written reason per assignmentIn this category the output itself has a language, so a model without Persian is zero for Persian text rather than slightly worse |
| Voice control | 30 | defined scale, with a written reason per assignmentVoice work means a specific voice. A handful of presets is not the same as building a new one, and the commercial licence is part of that |
| Language count | 25 | vendor stated specificationCounted only where the maker prints the full list or states a definite number |
| Access from Iran | 10 | our access column, with its method statedLightly weighted because its real home is the AI in Iran page, but not zero |
The trap most people fall into
ElevenLabs has three text to speech models. Persian is on the list of exactly one of them.
v3 carries a 74 language list with Persian written into it as fas. Flash v2.5 carries 32 languages and Multilingual v2 carries 29, and Persian is on neither. The problem is that Flash v2.5 is the model people reach for: it is the fastest at 75ms, its character limit is eight times v3 (40,000 against 5,000), and it is the cheapest.
So somebody signs up, picks the fast model, feeds it Persian, gets a poor result and concludes that ElevenLabs cannot do Persian. The accurate version is that this model cannot.
Why Persian carries a weight of 35
In the coding category the language of the output is not the question. Here the output is a voice, and a voice has a language. A model that reads seventy languages without Persian among them is zero for someone with Persian text, not slightly worse. No other weighting represents that honestly, and any table without this column is the wrong table for this site reader.
The levels were read from each maker own language list, not from listening. So this column reports what the maker has declared, not how good the accent is. The second question needs ears, and we have not used ours yet.
Three criteria deliberately left out
Latency. Only ElevenLabs prints a figure (75ms for Flash, 280ms for the conversational v3). Google and OpenAI publish nothing. A criterion with numbers for two candidates out of six drags the rest down for no reason.
Price. ElevenLabs sells monthly credits, Google and OpenAI bill per token, and Mistral hands you the weights and charges nothing. Those do not go in one column. Any table that puts them there is converting one number into another on assumptions it has not written down.
Preset voice count. Google has thirty and OpenAI thirteen, but ElevenLabs lets you build a new voice. Once you can build one, counting the presets stops meaning anything.
Nobody has reached level two
The Persian scale has three levels and level two means the maker also names a Persian-specific voice. No candidate reaches it today. Google says the model detects the input language automatically and names no Persian voice; OpenAI does not tie its thirteen voices to languages. So every model here reads Persian in a voice that was not built for Persian.
A contradiction that has to be said plainly
The model that came first in this table names Iran in its own terms of use. The ElevenLabs export control clause has the user warrant that they are not located in a country under US economic sanctions, and Iran is written into that list.
We write it because it is true and hiding it helps nobody. What we do not write is how to get around it. We sell no accounts, refer no intermediaries and describe no way past a sanction. If your work is Persian voice and you have no official route, hear it from us that you have no official route.
One model with open weights that cannot do your job
Voxtral TTS from Mistral is the only model in this table whose weights are published, so it runs on your own hardware with no country list in the way. Its licence is CC BY-NC 4.0, and the NC means no commercial use.
In a category whose whole point is commissioned voice work, that is precisely the thing you want to do. So the model that is best on access is the most useless on licence. Mistral publishes no language list for it either, so its Persian column stays unverified and it carries no rank.
When not to take our first pick
- If you work only in English and speed matters, Flash v2.5 from the same maker is the better pick at 75ms, and its lower place in this table is entirely about Persian.
- If you want to host the model yourself, none of the top three rows helps you and you are looking at Voxtral, with its licence caveat.
- If your job is transcription rather than voice generation, this table is not your answer. The Scribe and Voxtral Transcribe models do a different thing.
Where it falls short
- The Persian level was read from the published language list, not from listening. We have not judged accent quality.
- There is no price column at all here, because the three makers units are not comparable.
- Latency is published only for the ElevenLabs models, so it is not in the ranking.
- The Iran column was read from the makers own policy pages, not measured from a connection inside Iran.
Our take
If you have Persian text and you want a specific voice, ElevenLabs v3 is the only option today that gives you both. Just set the model to v3 by hand, because the faster default is the one without Persian.
Questions people actually ask
Does ElevenLabs support Persian
It depends on the model. v3 carries a 74 language list with Persian in it as fas. Flash v2.5 and Multilingual v2 carry 32 and 29 languages and Persian is on neither. If you want Persian you have to set the model to v3 by hand.
What is the best AI voice generator for Persian
By the table on this page, ElevenLabs v3. Three models list Persian (v3, Gemini Flash TTS and GPT-4o mini TTS) and of those three only v3 lets you build a new voice with a commercial licence.
How good is the Persian accent on these models
We do not know and we do not claim to. This table reports whether the maker lists Persian among its languages. Judging the accent needs a listening test we have not run, and when we do it will be written here.
Is there a free or open model for Persian voice
Voxtral TTS from Mistral has open weights, but its licence is CC BY-NC 4.0 and forbids commercial use. Mistral publishes no language list for it either, so we do not know whether it does Persian.
We can do the voice for your video
Motion graphics and voice services