A free AI model can now copy someone's voice in Indonesian

The short version

• Edge0 has released a voice model that copies a voice from one recording and its transcript, in 23 languages including Indonesian and Malay.
• Any company can download it and use it free for business, with no vendor to ask for permission.
• Quality in Indonesian is uneven, but you should already plan for convincing fake voices and for cheaper voice-overs.

What was released

The new model is called Audio8-TTS Realistic Air. Give it a clean voice recording and the exact words spoken, and it can read new text in that voice. It can also produce speech with no sample at all.

The release adds 12 languages to the 11 in the earlier preview. Indonesian was missing before. Malay and Thai are also among the additions, along with Vietnamese.

The model is small enough for a company to run on its own equipment. Its makers tested it on an NVIDIA graphics chip, the kind of processor AI runs on. An online demo also lets anyone try it, subject to a queue.

Why it sounds more natural

Most voice models are trained to pronounce words correctly and to sound like the target speaker. This team added a third goal: to sound natural to human listeners.

To do that, they trained a separate AI system to judge which recordings sound more human. In a blind listening test, people then rated the final version well above the team's earlier versions.

How good it is in Indonesian

The makers report results on 2 test sets, and the gap between them is large. The model got about 2% of Indonesian words wrong on one set and about 14% on the other.

Indonesian had little training material to learn from. Chinese and English each made up about 30% of it, while every other language had 3% at most. The team says error rates for the newly added languages remain substantially higher than for the 9 best-tested ones.

The makers warn of mispronunciations and of speech that stops early. They tell users to listen before using any output in a finished product. They also say their tests would not catch audio flaws such as buzzing.

What it means for your company

For marketing teams, this lowers the cost of testing. A brand could quickly draft Indonesian and Malay voice-overs for a campaign, then pay a professional to record the version that wins. You could also update phone menus and service recordings whenever a product changes.

The risk sits on the other side. A clip from a podcast or an earnings call may be enough raw material to imitate a senior leader. Because the model is downloadable, you can't count on a vendor's safety checks to stop this. Anyone can run it without such checks.

The makers' own advice is short: "Obtain permission to clone a voice and disclose synthetic audio where appropriate." That binds only the people who choose to follow it. So your defense needs to assume some won't.

What to do next

• Find every process where a voice alone is treated as proof, such as phone approvals for payments or call center identity checks. Add a second check that a copied voice can't pass.
• Tell finance staff and executive assistants that an urgent call in a familiar voice is no longer evidence. Agree on a call-back rule.
• Get written permission before copying any voice, and check that talent and ambassador contracts cover AI copies.
• If you test it for content, have a native speaker review every Indonesian script. Label the audio as AI-made where appropriate.
• Ask your legal team how Indonesia's personal data rules apply to the voice recordings you hold.