VoiceStudio: Open Source Voice Cloning for Australian Business

VoiceStudio: Open Source Voice Cloning for Australian Business

VoiceStudio is a free, open source app that clones a voice on your own computer. It's popular, it works, and it has no real consent check. Here's when we would put it inside an Australian business, when we wouldn't, and how to run it safely.

There's a clip doing the rounds on TikTok right now. Someone announces that a free, open source alternative to ElevenLabs has "just been released" and runs entirely on your own computer. Another creator installs it, feeds it 15 seconds of his own voice, and gets a convincing copy back. Comments go wild. Saves go wilder.

Two things about that clip are true and one isn't.

True: the app exists, it's free, and it really does run on your own hardware. True: a short clip was enough for his demo.

Not true: "just released". VoiceStudio has been on GitHub since April 2026. The latest release, v0.5.6, landed on 23 September 2026. When we checked the GitHub API on the morning of 24 October 2026 (Sydney time), it had about 83,000 stars. This is not a secret new toy. Lots of people already have it.

Which is exactly why it's worth a proper look. If you run a business in Sydney, Melbourne or anywhere else in Australia, the question isn't "is this cool?" It obviously is. The question is "would I let this anywhere near my business, and if so, where?"

Here's our answer.

---

What VoiceStudio actually is

Stick to what's verifiable and VoiceStudio is pretty simple to describe.

View on GitHub↗

• Open source, AGPL-3.0 licence. The code lives on GitHub.
• Four main jobs: voice cloning from a short clip, voice design (making up a new voice), dubbing, and transcription.
• Default voice engine: OmniVoice, a model from the open source speech group k2-fsa.
• Local once it's set up. The README says local workflows run on your hardware, online services are optional, and usage analytics need your consent. The models themselves download from the internet the first time you use them, so "offline" only starts after that.
• The installer checks your hardware first. It looks at your operating system, CPU, GPU, memory and disk before it picks what to install.

And the hardware reality, because the demos never mention it:

• Roughly 10 GB of free disk.
• A GPU with about 6 GB of video memory is recommended for the default engine. It runs on a CPU too, just slower.
• Intel Macs get the interface only. They need a remote backend to do the actual work, which rather defeats the "local" point.

Compare that with ElevenLabs, the cloud product everyone holds it up against. When we checked the ElevenLabs pricing page on 24 October 2026, the monthly list prices were Free (US$0), Starter (US$6), Creator (US$22) and Pro (US$99), with first-month discounts running at the time. Your audio goes to their servers. With VoiceStudio it stays on your machine.

What we can't tell you is whether VoiceStudio sounds as good as ElevenLabs. We haven't found an independent side-by-side test, and one creator's 15-second clip judged by ear isn't one. If quality matters for your use, test both yourself on your own voice.

---

What the demos skip

The TikTok version stops at "wow, it sounds like me". The interesting part starts after that.

1. Nothing checks that the voice is yours

The README asks you to clone voices only with permission. That's it. A sentence. There's no built-in consent check in the app.

ElevenLabs does something different for its Professional Voice Clones. You have to read lines aloud, a "Voice CAPTCHA", to prove the voice you're cloning is your own. (Their lighter Instant Voice Cloning is a separate process, and we haven't verified exactly what it checks.) You can argue about how strong that is. But it's a control. VoiceStudio has a polite request.

That isn't a moral failing on the developer's part. It's open source software running on your computer. Nobody is in the middle to say no. It does mean the control has to come from you.

2. Free app, separate model licences

The app is AGPL-3.0. The voice models are not automatically covered by that. The README is clear that model weights, voices and the conditions on generated output have their own licences, and that even a paid app licence doesn't grant those rights.

Speaking of paid, the app has a "Get Pro" button. So "free" is accurate for the open source app, and there's also a paid tier. Fine. Just know which one you're using and read the licence of whatever model you load before you put its output in anything commercial.

3. The install is a script from the internet

The documented one-line install pipes a script from the internet straight into your shell. In the viral demo, the creator didn't even run it himself. He pasted the README's "install with prompt" text into an AI coding assistant and let it do the job.

On your personal laptop, that's your call. On a work machine that holds client files, you've just let an AI agent follow a stranger's instructions with install rights. We put AI agents to work inside Australian businesses for a living, and we'd still want a human to read that script first.

---

Why Australians should care more than most

This isn't abstract.

Scamwatch, the ACCC's scam service, warns that scammers can use a few seconds of a voice recording to create an AI clone (scamwatch.gov.au). A few seconds. Your voicemail greeting is probably longer than that. So is the video on your website, your podcast cameo and that LinkedIn clip from the conference.

And in March 2023, Guardian Australia showed that an AI copy of one of its journalist's voices, plus his customer reference number, was enough to get into his own Centrelink phone self-service account, which used voice to verify identity (The Guardian, 16 March 2023). That was over three years ago, with tools that were harder to use than a free desktop app with a hardware-checking installer.

Put those two facts together and the uncomfortable conclusion is that "sounds like the boss" is no longer proof of anything. Not for a payment request. Not for a password reset. Not for "it's me, just approve it".

---

When we would deploy it, and when we wouldn't

This is AI Lab Australia's recommendation, not a universal rule. We're a forward-deployed engineering shop, secure by default, and this is the line we'd draw today.

AI Lab's take: VoiceStudio is good for narrating your own training videos on one locked-down laptop, with the speaker's OK in writing. We would not use it for anything that answers the phone, and we would not use it anywhere near customer-facing voice verification.

Yes, with conditions

• Internal training and onboarding videos narrated in the voice of someone who has agreed to it, in writing. Say you're a practice manager in Parramatta who records the same induction walkthrough every quarter. Cloning your own voice to re-voice updated slides is a sensible use.
• Drafting voiceovers you'll later re-record or check by ear before they go anywhere.
• Transcription of your own recordings, kept on the same machine.
• One machine, not the whole fleet. A single, locked-down laptop with full-disk encryption, a named owner, and the voice samples stored in one place you control.

No

• Anything that answers or makes phone calls. Receptionist bots, outbound calls, after-hours lines. A cloned voice on a phone line is exactly the thing Scamwatch is warning your customers about. Don't train them to trust it.
• Voice verification of any kind. If your business, or any provider you rely on, uses "my voice is my password", this tool is a reminder to replace that with something else, not to build on it.
• Cloning anyone who hasn't said yes in writing. Staff, clients, a director who's overseas, a voice actor from a video. No.
• Machines that hold client data where you can't control what gets installed or where the files end up.

Is this conservative? A little. But the cost of being wrong here isn't a slightly worse training video. It's a fake phone call in your voice that someone in accounts believes.

---

How to use VoiceStudio safely: a checklist

If you've read this far and still want to try it (fair enough, it's a genuinely useful tool), do it like this.

1. Get written consent from the speaker. An email is fine. Say what the clone will be used for and where the samples will live.
2. Pick one machine. Not a shared PC. Turn on full-disk encryption (FileVault on Mac, BitLocker on Windows) and keep it patched.
3. Check your hardware first. About 10 GB free disk, ideally a GPU with around 6 GB of video memory. Intel Mac? Don't bother locally.
4. Read the install script yourself before you run it, or before any AI coding assistant runs it for you.
5. Check the voice model's licence, separately from the app's AGPL licence, before you use the output commercially.
6. Decide on the online bits. Remote services are optional and analytics need consent. Leave them off unless you have a reason.
7. Lock down the voice samples. One folder, one owner, not synced to a shared drive. Delete them when the project ends.
8. Label the output. Tell viewers the narration is AI-generated. It costs nothing and builds trust.
9. Fix your callback rule. Separately from all this, make sure your team knows: any payment or credential request by voice gets a call back on a number you already have. Always.

---

FAQ

---

The bottom line

VoiceStudio is real, free, popular and useful. It's also a voice-cloning tool with a polite sentence where a consent check should be. Both things are true, and the second one is your job to handle.

Use it for your own voice, on one locked-down laptop, with written consent. Keep it off the phones.

If you're a business in Sydney or anywhere in Australia and you want help putting AI tools like this to work with the controls that make them safe to keep, email us at info@ailabaustralia.com. We'll tell you honestly whether it's worth doing.