ElevenLabs for Free on Your Own PC: VoiceStudio Full Setup (Windows + Mac)

Two voices read the same sentence. One costs $22 a month. The other was made on a laptop, for free, with the internet switched off. The free one comes from VoiceStudio, an open-source app on GitHub that does most of what ElevenLabs does, on your own computer, with no credits and no monthly bill.

We set it up from zero on a 24 GB Apple-chip Mac in October 2026, running VoiceStudio v0.5.7. Then we cloned a voice (3 out of 3 word-perfect), made a 495-word audiobook in 13.5 seconds, dubbed a video into Spanish and made Claude talk. Here is everything, plus the one setting you must change before you use it on anything that makes money.

Why 49,000 people starred it in a month

On GitHub's trending page for the month, VoiceStudio is up 49,116 stars in thirty days, and about ninety-five thousand in total. Its description says why: "the open-source, fully local ElevenLabs alternative."

ElevenLabs is the company behind a huge share of the AI voices you hear online. Its Creator plan is $22 a month, and that buys you roughly two hours of speech. Need more, like a whole audiobook? That's the $99 plan (see ElevenLabs pricing).

VoiceStudio does the same jobs on your own machine: clone a voice, design a voice from one sentence, dub videos, narrate audiobooks and transcribe, in 646 languages. Your audio never leaves your computer.

Will it run on your PC?

• Mac with an Apple chip: yes. That's what we used.
• Windows or Linux with an NVIDIA graphics card: yes, and that's the fastest.
• No graphics card at all: since this week's update, yes. It's slower, but it works.
• Phones: there are early apps for iPhone and Android.

The starter pack is about 5 GB, and the bigger voice model adds about 5 GB more. Memory matters too: on our 24 GB Mac, with many apps open, the big model ran out of room on long text, so we used small models for long jobs.

Setup on Windows and Mac

On the GitHub page, click Releases and download the installer for your system from the v0.5.7 release. On a Mac, drag it into Applications. On Windows, run the setup file. This version is signed and approved by Apple, so on a Mac it opens like any normal app. If you prefer the terminal, the README has a one-line install for Mac Terminal and one for Windows PowerShell.

Open the app, go to Settings, then Models, and pick a pack. We chose Balanced: the voice engine plus speech-to-text (the part that turns spoken audio into written words), 4.8 GB. It downloaded in about 2 minutes 15 seconds, and you only do this once.

You can also let AI do it. The README has a ready-made prompt you paste into Claude Code, and it installs the app and checks that it works.

The setting to change before you monetize

This is the part nobody explains. Every voice model inside VoiceStudio comes with its own license, and the app shows it on each model's panel. The default model, OmniVoice, is licensed for non-commercial use only (CC-BY-NC). Fine for your kids' bedtime stories, not for a monetized channel or client work.

So open the model list and pick one marked for commercial use. We used VoxCPM2. Its panel says commercial use allowed, Apache 2.0, no acceptance needed. It's a 4.62 GB download and took us about 3 minutes 20 seconds. Kokoro and MOSS-TTS-Nano are also Apache-2.0 on their model cards. Every VoiceStudio voice in the video comes from models like these.

One honest note: the app still marks commercial use of the generated audio as "Not verified", and its own docs say it hasn't legally cleared any model for you. Read the license yourself before you sell anything. This is not legal advice.

Clone a voice (3/3 word-perfect)

For our test we cloned the channel's narrator, an ElevenLabs voice we pay for every month. Open Voice cloning and add a short recording. The app asks for five to fifteen seconds of clean speech, and three rules make or break it:

1. A quiet room.
2. One person talking.
3. No music behind.

Then type exactly what's said in the clip, so the app can match each sound to each word. Now type anything you want it to say and hit generate.

The very first run took about 3 minutes, because it loads the model. After that, about 21.2 seconds for an 8.4 second sentence. That's slower than real time on our Mac: fine for a script, too slow for a live call. We tested it on the hard stuff, a price, a date and a time in one sentence ($1,299.99, October 23rd 2026 and 9:30 AM). Across three takes, three out of three were word-perfect.

Design a voice from one sentence

No recording? Describe the voice instead. We typed: elderly man, British accent, low pitch. About twenty seconds later, it spoke. Nobody's voice was copied, so for a narrator on your videos or app this is the safest option of all.

A free audiobook in 13.5 seconds

Long text is where paid voice apps get expensive, and it's where the big model struggled on our Mac. The trick is to switch to a tiny model called Kokoro, which is also free for commercial use. We pasted the first 495 words of a public-domain short story, O. Henry's The Gift of the Magi. Kokoro turned it into 2 minutes 51 seconds of audio in 13.5 seconds.

A full novel, around 450,000 characters, would use almost four months of Creator plan credits. Here it's free.

Dub a video into Spanish (plus a Hindi test)

We gave it 30 seconds of one of our own videos. It pulled the voice away from the background music in 9 seconds, wrote out every word, translated it into Spanish in 2.7 seconds, and spoke it back in a matching voice in 33 seconds. The background sound stays underneath.

One catch. The free offline translator is very literal, and it translated our product's name, PHP, to "Shell", word for word. The fix: before you hit translate, correct any names in the transcript box.

Hindi works with the bigger model, one sentence at a time, and it's slow on our Mac: about 2 minutes per sentence. A mix of Hindi and English, the way millions of people actually talk, only half works. The Hindi words were right, but the English words came out with a heavy accent.

Make Claude talk (MCP, one line)

VoiceStudio has a built-in MCP server. MCP is just a plug that lets Claude use other apps. Add this one line in your terminal and Claude Code can speak with any voice you've saved:

claude mcp add --transport http voicestudio http://localhost:3900/mcp/

We asked Claude to write a short bedtime story about a robot who learns to grow a garden and read it out loud with our saved voice. In under a minute, it wrote the story, called VoiceStudio and saved the audio file in our folder.

Does it really work offline?

We ran a VoiceStudio voice engine inside a sandbox that blocks every internet connection. To prove the block worked, we tried to reach Google from inside it, and it failed. The voice still came out in 4 seconds, word for word.

When you are online, the app does check GitHub for updates and Hugging Face for models at startup. That isn't your audio or text, but it isn't silent either.

Our results at a glance

Should you cancel ElevenLabs?

Our honest scorecard: VoiceStudio wins on cost, privacy, unlimited length and languages. ElevenLabs still wins if voice is your business. Every paid plan includes a commercial license, the most expressive voices are there, and there's nothing to install. If you just want great voices without a monthly bill, VoiceStudio is it.

The catch: pick a family safe word

This matters more than any app. If a free tool can copy a voice from a few seconds of audio, someone can copy yours from a single voice note. Scammers already call parents pretending to be their child in trouble.

So tonight, pick a family safe word, something only your family knows. If anyone ever calls asking for money in a loved one's voice, ask for the word. And only ever clone a voice with that person's permission.

The answer, and your free setup card

The blind test from the start of the video: A was VoiceStudio, the free one made on our laptop, and B was ElevenLabs. The whole setup is on one page in the free PARROT Setup Card, including the setting to change before you monetize. Comment PARROT on the video, or grab it below.

How we tested: VoiceStudio v0.5.7 on a 24 GB Apple-chip Mac, October 2026, one run each. Windows steps follow the official release and docs and were not run on camera. VoiceStudio by debpalash (AGPL-3.0). Independent, not sponsored by or affiliated with VoiceStudio or ElevenLabs. Make Viral Reels is our own product.