~/guides/SPEAK.md

Guides / AI tools

Clone your own voice from 3–10 seconds of audio: free, open-source VoiceStudio, then Claude Code reads in your voice

VoiceStudio is a free, open-source (AGPL-3.0) voice-cloning app built on OmniVoice (600+ languages). Install, record, generate, connect Claude Code via MCP.

Watch the reel on InstagramComment #SPEAK to get this guide

With a few seconds of audio, AI can say anything in your voice. This is the full version of the SPEAK reel: what VoiceStudio is, how to install it, how to record your sample, how to switch languages, and the best part, connecting it to Claude Code over MCP so it writes a script and reads it out in your voice. Two things up front: only clone your own voice, or a voice whose owner has agreed in writing; and we did not manage to produce a working clone ourselves. I'll explain why below.

01What you get: a few seconds of audio becomes a you that speaks 600+ languages

Reel opening: the voicestudio.sh Voice clone card and the OmniVoice demo table, generated column ringed in gold
The reel opens on the official site and the official demo (captured 23 Oct 2026), not on our own clone.

VoiceStudio is a desktop app that clones voices, designs new voices, dubs videos, makes audiobooks and transcribes speech. It's open source on GitHub under AGPL-3.0, and the desktop app is free. The official site says it can "Mirror a voice from three seconds".

voicestudio.sh screenshot: Mirror a voice from three seconds highlighted, the Pro plan below
The official site (captured 23 Oct 2026): clone a voice from three seconds. Below it is the paid Pro version.

Its default engine is OmniVoice, open-sourced by the k2-fsa team (Apache-2.0). The official README says it supports over 600 languages and recommends a 3–10 second reference clip. VoiceStudio itself says 646 languages. So "3 seconds" in the reel is the website's claim; when you actually record, 3–10 seconds as the README says is the safe choice.

Reel frame: the VoiceStudio GitHub page, star count circled in red, tags 81.5k stars and AGPL-3.0
The GitHub page in the reel: 81.5k stars when we filmed; the API said 82,018 on 23 Oct 2026, and it keeps climbing.

To be honest: we tried it ourselves. On a computer with no dedicated graphics card (GPU), we installed the underlying model, OmniVoice, with Python and cut a 5-second sample from our own channel's voiceover. The 3.1 GB model downloaded and loaded fine, but the first 20-character line was still generating after 12 minutes, so we stopped it. That's why this guide has no clone of our own: the steps and numbers come from the official pages, and I've noted the source for each.

  1. What to prepare: a computer, disk space and one clean recording.
  2. Install the desktop app (Windows / macOS / Linux), or install with one command.
  3. Record 3–10 seconds of your own voice and save it as your voice.
  4. Type, pick a language, generate.
  5. Connect Claude Code over MCP so it reads your scripts in your voice.
  6. Read the warnings: consent, licensing, speed and security.

023 things to prepare first

  1. A computer. An NVIDIA GPU (Windows / Linux) or an Apple Silicon Mac is fastest. It also runs without a dedicated GPU: the site says about 8 GB of RAM is enough to start, but it's much slower. On an Intel Mac only the app UI runs, connected to a backend on another machine.
  2. Disk space. The official Windows guide says to keep about 10 GB free for the app, its Python environment and the models.
  3. 3–10 seconds of your own voice recorded somewhere quiet. A phone voice-memo app is fine; the next section shows how.

03Step 1: install the desktop app

  1. Go to the Releases page on the official GitHub and download the build for your system: .dmg for macOS, .exe for Windows, .AppImage or .deb for Linux.
  2. Or install with the official one-liner (the two blocks below). The installer keeps your settings, projects and models.
  3. On first launch the app sets up Python, FFmpeg and the models it needs itself; you don't need any developer tools.
  4. Open Voice cloning, and when it asks you to install a model, install it.
macOS / Linux terminal: one-line install (official README)
curl -fsSL https://voicestudio.sh/install | sh
Windows PowerShell (a normal, non-admin window): one-line install (official README)
irm https://voicestudio.sh/install | iex

If you already use Claude Code, the project also gives you an "ask your agent to install it" prompt. It follows the official agent guide: it detects your hardware, asks before downloading models, and finishes with a test generation.

Paste into Claude Code (official wording)
Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md

04Step 2: record 3–10 seconds of your own voice

Screenshot of the OmniVoice README Tips with Use a 3–10 seconds reference audio clip highlighted in yellow
The official OmniVoice README (captured 23 Oct 2026): use a 3–10 second clip; longer is slower and may sound worse.
  1. Find a quiet room, switch off fans and air conditioning, and hold the phone about a hand's width from your mouth.
  2. At your normal speaking pace, read the text below (with your own name and details). It takes about 5–8 seconds.
  3. Listen back. If there's noise, breathing or background music, record again. Clean matters more than long.
  4. In VoiceStudio, open Voice cloning, add the recording and save it as your voice.
A sample you can read (replace the [ ] parts)
Hi, I'm [your name]. Today I want to try letting AI read [something you do often, e.g. my weekly work notes] in my own voice.

05Step 3: type, pick a language, generate

  1. In Voice cloning, pick the voice you just saved.
  2. Type what you want it to say.
  3. Pick a language and generate. The first run is slower because the model has to load.
  4. Listen. If it doesn't sound like you, try a cleaner sample first; that usually helps more than changing settings.
Reel frame: a globe in a blue room with language tags (English, 日本語, Español) and the OmniVoice README
Same voice, in English or Japanese too. OmniVoice officially lists 600+ languages, Apache-2.0.

06Step 4: connect Claude Code so it reads scripts in your voice

This is my favourite part. VoiceStudio has a built-in MCP server, so AI agents such as Claude Code and Cursor can call it directly: generate speech, clone voices, transcribe audio and list your voices. The official docs say the server runs whenever the app is open; there's nothing extra to start.

VoiceStudio docs/mcp.md screenshot: MCP server — let agents speak in your voice, Claude Code highlighted
The official MCP doc (captured 23 Oct 2026): tools such as generate_speech, clone_voice and transcribe.
  1. Open VoiceStudio and leave it running. The MCP address is http://localhost:3900/mcp/ (keep the trailing slash).
  2. In the app's Integrations, find Claude Code, export the .mcp.json and put it at the top level of your project folder.
  3. Open Claude Code in that folder. The first time, it asks whether to allow this project's MCP server; check that it's VoiceStudio, then allow it.
  4. Type /mcp in Claude Code. If VoiceStudio shows as connected, you're done.
docs/mcp.md screenshot: Integrations highlighted, then Claude Code (.mcp.json), Cursor and Codex CLI
Where the official doc says to export it: Claude Code (.mcp.json) under Integrations.

If you'd rather not use the exported file, add it with Claude Code's own command in a terminal. The header at the end is the "client id" from the official doc, which later lets Claude Code always use one particular voice.

Terminal: add it by hand (Claude Code's official MCP command format)
claude mcp add --transport http voicestudio http://localhost:3900/mcp/ --header "X-OmniVoice-Client-Id: claude-code"
Reel frame: a dark server room, a terminal under the MCP doc typing claude › generate_speech("今天的
An illustration from the reel: Claude Code calling generate_speech. We drew this; it isn't a real run.
Once it's connected, tell Claude Code (replace the [ ] parts)
Write a [30-second] [product intro / weekly work update] about [your topic].
Then read it out with VoiceStudio's generate_speech, using [the name of your saved voice], in [English].
Finally, tell me where the audio file is.

07Good to know

Reel frame: a red room, a sign reading clone only your own voice, a stranger and a red infringement stamp
Only clone your own voice. Never clone someone else's without their clear consent.
  1. Only clone your own voice, or a voice whose owner has agreed in writing. The official site is blunt: only with the owner's say-so; impersonation and fraud are off the table. Generating with someone else's voice can violate their rights and is an easy tool for scams.
  2. If you get a voice message that sounds like someone you know asking for money, hang up and call them back on a number you already have. Cloning a voice really does take only seconds now.
  3. It's slow without a GPU. The project says the CPU is fully usable, just slower; our own CPU machine didn't finish a 20-character line in 12 minutes.
  4. The free part is the desktop app (AGPL-3.0). Pro and Lifetime are a separate closed-source build with a paid commercial licence: the site lists $99/year or $299 once. The README also notes that models, voices and generated output have their own licence terms, so read the official licence notes before any commercial use.
  5. MCP has no password. The official doc says the MCP transport isn't authenticated, so don't expose it to the open internet; to use it from another machine, stay on a trusted local network or use an encrypted tunnel.
  6. Only install from known sources: github.com/debpalash/VoiceStudio, voicestudio.sh and github.com/k2-fsa/OmniVoice.
  7. We haven't fully tested these steps ourselves (CPU too slow, no working clone). The content comes from the official pages, checked on 23 Oct 2026; versions, star counts and button names may change, so follow the official docs if you get stuck.

Sources

Related guides