Free AI Voice Cloning in 12 Indian Languages: Maya 2 Native Turns a Ten-Second Hindi Clip Into Tamil Narration, and You Still Need a Native Ear
Record ten seconds of yourself speaking Hindi. Type out exactly what you said. Paste in a paragraph of Tamil. A few seconds later your voice is reading that Tamil paragraph aloud, in a language you may not speak a word of.
That is what Maya 2 Native, a voice model Maya Research put on Hugging Face on October 26 (the card dates the release a day earlier), will do for free. It covers Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. It ships under the Apache 2.0 licence, which means you can use the voices it makes in paid work. And it runs in a browser tab on a free demo page, no install needed.
For anyone who narrates, dubs or makes explainer videos for Indian audiences, this is a real tool, not a lab curiosity. It also hands you a problem the model cannot solve for you. If the voice is yours and the language is not, who checks that it said the right thing?
What it actually does
Maya 2 Native is a text-to-speech model: you give it text, it gives you speech. It comes with 20 built-in voices, ten female and ten male, each with a name (Ahana, Ritvik, Sreekar and so on) and a short description of how it sounds. Any voice can speak any of the 12 languages, though each was trained mostly on a few of them, and the model card warns that languages outside a voice's training set sound less native.
The part that matters for creators is cloning. You give the model a short clip of a real person speaking, between 3 and 10 seconds of clean audio, plus a word-for-word transcript of that clip. From then on it can read new text in something close to that person's voice.
The clip and the output do not have to share a language. The project's own example registers a Hindi recording and then generates Tamil. So a Marathi creator can clone their own voice and publish the same tutorial in Telugu and Bengali, without hiring three narrators.
It also streams, which means it starts handing back audio in small pieces before the whole sentence is finished. On a single data-centre graphics card serving 224 voices at once, the card reports the first sound arriving in about a tenth of a second for each one. That matters if you are building a live tool. For recording narration, it mostly means you are not waiting.
Why the licence is the headline
Plenty of voice cloning tools exist. Most of the good ones are closed services that charge per character and keep the model on their servers. The open ones often carry non-commercial licences that rule out the paid jobs creators actually take.
Maya 2 Native's weights and code are Apache 2.0. One piece it depends on, Kyutai's Mimi audio codec, is under a Creative Commons attribution licence (CC BY 4.0), which allows commercial use as long as you credit it. In plain terms: you can narrate a client's product video with it, sell a course voiced by it, or build it into an app, as long as you respect the credit and the conditions below.
Those conditions are about people, not money. The card and the code both say to clone only voices you have permission to use. And the model adds no audio watermark, so nothing in the file marks it as synthetic. That puts the honesty burden on you. If you publish a cloned voice, say so.
Put this into practice: a dubbing test in 20 minutes
You can run the whole thing in the free Maya 2 Native Studio Space on Hugging Face, published under Maya Research's own account. A second free copy runs in the hugging-apps Space.
1. Record the reference. Use a decent microphone in a quiet room. Speak one or two natural sentences in your strongest language, 5 to 10 seconds long. No music, no second voice, no echo. The card says cloning gets worse with noise, extra speakers or a wrong transcript, and all three are easy to avoid at this step.
2. Write the transcript exactly. Every word, in the language you spoke, matching what is on the recording, including the little restarts. The Space refuses to clone without it. This is the step people rush, and the card names an inaccurate transcript as one of the main reasons a clone drifts.
3. Pick the languages. Set the reference language to what you spoke and the output language to your target. The Space caps each request at 1,200 characters, roughly a minute of speech, so work paragraph by paragraph.
4. Generate the same paragraph three ways. Once in your own language, once in the target language with your cloned voice, and once in the target language with a built-in voice that lists that language among its strongest. Now you have a control. If the built-in voice sounds clearly more natural than your clone, your audience will hear that too.
5. Spell out the hard bits. The card says URLs, symbols, dates, long strings of digits and sentences that mix two languages still trip the model. Write "twenty twenty-six" instead of "2026" if the reading comes out wrong, and spell out web addresses the way a person would say them.
6. Send it to a native speaker. This is the step the rest depends on, and it gets its own section.
If you outgrow the demo, the GitHub project runs on Linux with an NVIDIA graphics card that has at least 8 GB of memory, and a community conversion, maya2-native-mlx-8bit, packs it into a 3.1 GB file for Apple Silicon Macs. That Mac version is unofficial and publishes no speed figures, so treat it as an experiment.
The listener problem
Here is the uncomfortable part. The quality numbers on Maya's card are good. Its overall rate of misread characters is around 3 percent, from about 1.5 percent in Hindi to 5.4 percent in Assamese on Maya's main judge, and it picks the intended voice out of the 20 almost every time. But those scores come from Maya's own automatic judges: speech recognition models that transcribe the output and compare it to the script, and a voice-matching model that checks who it sounds like. The card says plainly that there is no human listening study and no head-to-head comparison with other voice tools.
A transcription model can tell you the right words came out. It cannot tell you that a Tamil listener heard a slightly Hindi rhythm under your voice and stopped trusting the video. It cannot tell you a word landed with the wrong stress, or that the tone was flat in a line that needed warmth.
There is one outside signal. The card reports that VoiceArena, a public site where listeners vote between voice models, listed Maya 2 Native second among real-time models for Hindi when Maya checked on October 27. The card also admits the ranking range was second to fourth and the entry is not tied to a specific version. Take it as encouraging, not settled.
So the position I would take is simple. Use Maya 2 Native to make the draft. Pay a native speaker of each target language to listen to the whole thing before it goes out. Ten minutes of their time per video is cheaper than a hundred comments telling you the narration sounds wrong, and it is the only quality check that matches how your audience will hear it.
Where it breaks
Assamese is the weak language. None of the 20 built-in voices has an hour of Assamese training audio, and the card says Assamese output can drift toward a Bengali-sounding voice. If Assamese is your market, test hardest there.
Very short lines wobble. One- and two-word prompts are less stable in voice identity. Button labels and single-word callouts may sound like a slightly different person. Generate them inside a short sentence and trim.
Emotion is a best effort. You can add cues like [laughs], [sighs] or [whispers], and the card says these were not the focus of this release. Expect them to work sometimes.
The clone is close, not identical. The card says clones approximate the reference voice and that similarity and accent vary. People who know you well will notice.
It can cut off or stumble. The card warns the model sometimes makes sampling errors, and its safety guards against looping can end repetitive text early. Listen to the ends of long paragraphs.
Running it yourself means Linux and NVIDIA. The official code needs Linux and an NVIDIA card. One voice stream holds about 7.7 GB of graphics memory and briefly peaks at 9.2 GB while loading. Windows and Mac users are on the free Space or the unofficial Mac port.
Nothing marks the audio as synthetic. No watermark is added, and the server has no built-in login. If you put it online for a team, lock it down yourself.
What this changes
For a long time, reaching a second Indian-language audience meant a second narrator, a second recording session and a second invoice. That cost kept a lot of good tutorials in one language. A free, commercially usable model that carries your voice into eleven more languages removes most of that cost on the first draft.
It does not remove the listener. The model was graded by other models, and your viewers are not models. Record your ten seconds this week, generate one paragraph in a language you know well and one you do not, and play both for someone who grew up speaking the second. Their face will tell you more about this tool than any number on the card.