VoiceStudio: Open-Source Voice Cloning to Try Before You Pay for ElevenLabs

Key takeaways

• VoiceStudio is an open-source (AGPL-3.0) desktop app for voice cloning, voice design, video dubbing, transcription, and audiobooks. It runs entirely on your own computer.
• It clones a voice from a few seconds of audio. The default voice model, OmniVoice, supports over 600 languages.
• Setup is one line: curl -fsSL https://voicestudio.sh/install | sh on macOS and Linux, or irm https://voicestudio.sh/install | iex in PowerShell on Windows.
• It connects to Claude Code over MCP. While the app is running, the address is http://localhost:3900/mcp/. Below is a ready-to-use prompt that turns an article into an audiobook in your voice.
• Mind the license: the app is open source, but the default voice model is not licensed for commercial use (CC-BY-NC). No client work, ads, or audiobooks you sell. A paid license doesn't change that.
• Only clone your own voice, or voices you have written permission to use.

What this page is

This is the page behind the reel. Comment VOICE and you land here: the VoiceStudio setup steps, 3 rules for a good reference clip, how to connect it to Claude Code with MCP, a ready-to-use prompt that turns an article into an audiobook in your voice, and a license note on what you can and can't do.

Everything here comes from the VoiceStudio repository, its MCP docs and license notice, voicestudio.sh, and the OmniVoice pages, all checked on October 18, 2026. The project moves fast. If something here disagrees with the official pages, trust the official pages.

What VoiceStudio is

VoiceStudio is an open-source voice app with over 74,000 stars on GitHub. Here's what's in it:

• Voice cloning: makes a copy of your voice from a short clip and reads any text in it.
• Voice design: creates a new voice that doesn't belong to a real person. You pick gender, age, pitch, whisper, and accent.
• Video dubbing: translates a video and voices it to match the timing.
• Dictation and transcription: turns speech into text.
• Audiobooks: voices long text chapter by chapter.
• Local API and MCP: AI agents like Claude Code can use the app themselves.

The default engine is OmniVoice. You can pick other engines too.

The app doesn't need an account or the cloud. VoiceStudio's site says it only goes online to check for updates and download models. Usage analytics are only sent if you agree. Your voice stays on your computer.

What you need

• Windows 10/11, macOS, or Linux.
• A GPU (optional): fast on NVIDIA GPUs and Apple Silicon Macs (M1 and later). Without a GPU, it runs on the CPU, but slower.
• About 8 GB of RAM. The site says that's enough to start.
• Disk space: the default voice model is about 2.3 GB. Without a GPU, setup also installs a CPU package of about 5 GB.

Two exceptions. On Intel Macs, only the interface runs; to generate audio, you connect to VoiceStudio on another computer. Windows on ARM is experimental and runs on the CPU only.

Setup

Pick one of three ways.

macOS and Linux: open a terminal and run:

curl -fsSL https://voicestudio.sh/install | sh

Windows: open PowerShell and run:

irm https://voicestudio.sh/install | iex

With an installer: if you'd rather not run a command, download one from the Releases page: .dmg for Mac, .exe for Windows, .AppImage or .deb for Linux.

After installing, open the app and download the voice model first. It's about 2.3 GB. If you try to generate speech before it's downloaded (especially through Claude Code), the request can time out while the model downloads.

A note: the one-line commands download a script from the internet and run it. Look through the repo before you run one. If you're not sure, use the installer.

3 rules for a good reference clip

The clone is only as good as the clip you give it.

1. 3 to 10 seconds. That's OmniVoice's recommendation. Longer clips slow things down and can lower the clone's quality.
2. Clean audio. A quiet room, no background music or hum, no echo. Speak close to the mic, in your normal voice.
3. One speaker. Nobody else's voice in the clip.

One more tip: record in the language of the text you'll voice. OmniVoice says a clip in a different language makes the result carry that language's accent.

Connect it to Claude Code with MCP

An MCP is a bridge between Claude and another app. Once it's connected, Claude uses VoiceStudio on its own: it voices text, lists your voices, and transcribes audio.

First: VoiceStudio must be running. MCP is on by default and runs inside the app, so there's nothing else to start. The address is:

http://localhost:3900/mcp/

Keep the trailing /.

Option 1: get a ready-made config from the app. VoiceStudio's Integrations section exports a .mcp.json file for Claude Code. Put it in the folder you work in with Claude Code.

Option 2: one command. Run this in a terminal:

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

Check it: open Claude Code and type /mcp. VoiceStudio should show Connected. Then type:

Check VoiceStudio's status and list my saved voices.

The tools Claude can use:

Connecting from your own computer doesn't need a password. Only turn on the app's Local network sharing if you need to connect from another computer, and only on a network you trust.

Ready-to-use prompt: turn an article into an audiobook in your voice

Before you start:

1. Clone your voice in the app and name the voice profile, for example "My voice".
2. Make sure FFmpeg is installed. It's needed to join the parts into one file. If it isn't, ask Claude: "Help me install FFmpeg."
3. Use your own article, or text you have the rights to. Turning someone else's article into an audiobook and sharing it is a copyright problem.

Then paste this prompt into Claude Code. Fill in the brackets:

Turn this article into an audiobook in my voice: [link to the article or file path]

1. Prepare the text: Take the title, headings, and paragraphs. Remove menus, ads, code blocks, link lists, and image captions. Write out abbreviations, symbols, and numbers the way they're spoken (for example, "20%" becomes "twenty percent").
2. Find the voice: List my voices with list_voices in the VoiceStudio MCP and use the "[voice profile name]" profile. If you can't find it, stop and ask me.
3. Split it up: Split the text into parts of at most 1,000 characters each, without cutting a sentence in half. Each heading starts a new chapter.
4. Voice it: Voice each part in [English] with generate_speech. Save the parts in order to the audiobook/[article name]/ folder (001.wav, 002.wav, ...).
5. Join it: Use FFmpeg to join the parts in order. Leave a short silence between parts and a slightly longer one between chapters. Save the result as one MP3 file. Write the chapter titles and the minute each chapter starts to chapters.txt.
6. Check it: Transcribe one part from the beginning, one from the middle, and one from the end with transcribe, and compare them to the text. If anything was skipped or misread, regenerate that part. When you're done, tell me the total length.

Before voicing everything, voice only the first paragraph and let me listen. Continue once I approve.

Three notes on the prompt:

• The first-paragraph test matters most. If the voice doesn't sound like you or the pronunciation is off, you find out in 20 seconds, not after the whole article is done.
• The 1,000-character limit is on purpose. VoiceStudio's docs say processing takes longer past 1,200 characters. Short parts are faster, and if one breaks, you only redo that one.
• If a word comes out mispronounced, tell Claude: "Spell this word the way it's said and regenerate that part."

If you'd rather not use Claude Code, you can paste the text into VoiceStudio's own audiobook section and voice it there.

License note: what you can and can't do

This part matters, because there are two separate licenses: the app's and the voice model's.

The app (VoiceStudio): open source under AGPL-3.0. You can use the app, including for commercial work. But if you change the app and offer it to others over the internet, you have to share your changed code with them.

The default voice model (OmniVoice): the code is Apache-2.0, but the trained model is licensed CC-BY-NC. The Hugging Face page says that's because of limits in its training data. NC means "non-commercial."

The paid license ($99/year or $299 once): for companies that want to put VoiceStudio inside a closed-source product. It also unlocks the Pro features. But it doesn't give you rights to the model weights or the audio you generate. So even with a paid license, you can't use the default model for commercial work.

If you do commercial work: VoiceStudio supports other voice engines and shows each model's license notes in its model manager. But the app's license notice doesn't clear any model for commercial use; that check is on you. Either pick a model whose license clearly allows commercial use and read its terms yourself, or use a paid service that includes commercial rights.

This isn't legal advice. For anything you'll earn money from, read the licenses yourself or ask a lawyer.

Consent and safety

Cloning someone's voice without permission is wrong, and in many places it's illegal. VoiceStudio's acceptable-use rules say the same:

• Only clone a voice with its owner's permission. Keep proof. For commercial work, get a signed release from every voice owner.
• Impersonating a real person or company, fraud, and trying to get past voice verification are banned.
• Label synthetic audio as synthetic wherever listeners could be misled.
• Don't clone children or people who can't give consent.

And think about this: if it's this easy for you, it's this easy for scammers. Someone calling in a family member's voice and asking for money may not be that person. Agree on a code word with your family that only you know.

FAQ

Is VoiceStudio free?

The app is open source and free to download. There's also a paid, closed-source Pro version. And the default voice model isn't licensed for commercial use. So "completely free" isn't quite right.

Can it replace ElevenLabs?

For personal and non-commercial work, it's a good alternative, and your voice never leaves your computer. For commercial work, the default model doesn't allow it, while ElevenLabs' paid plans come with commercial rights. To use ElevenLabs in Claude Code, see the ElevenLabs MCP setup guide.

Does it copy emotion too?

There's no separate emotion setting. It copies the tone of your clip. OmniVoice lets you put tags like [laughter] and [sigh] in the text, which add sounds like laughing and sighing.

I don't have a GPU. Will it work?

Yes, but slower. For a long audiobook, be patient, or try a short text first.

Does my voice go online?

No. Everything runs on your own computer. The app only goes online to check for updates and download models.

Try it

Install the app, download the voice model, and clone your voice with a 5 to 10 second clip that follows the 3 rules. Then connect it to Claude Code and try the audiobook prompt on a short article you wrote. If the first paragraph sounds like you, Claude handles the rest.

New to Claude Code? The Claude Code course at CodeBaseHub Academy walks through the workflow from the start.