Clone any voice for free with AI, in minutes, from a short voice clip. In this step-by-step tutorial you'll use OmniVoice, an open-source AI voice cloning model, on a free Google Colab GPU, with nothing to install. One sample clip becomes the same voice in English, Spanish, French, Arabic, Hindi and Japanese, and the model supports 600+ languages.
Chapters: 0:00 Hear it first: one voice, six languages 0:45 What is voice cloning? 1:06 Meet OmniVoice (free, 600+ languages) 1:37 How to prepare your reference clip 2:01 Important: clone only voices you have permission to use 2:20 Open the notebook 2:34 Turn on the free GPU (T4) 3:11 Step 1: Install OmniVoice 3:21 Step 2: Load the model 3:46 Step 3: Upload your reference voice 4:04 Step 4: Generate your first clone (English) 4:42 The same voice in Spanish 5:17 How to download your voices 5:30 French and Arabic 6:16 Hindi and Japanese 6:46 600+ languages 6:56 Five extra tips 7:38 Quick fixes if something goes wrong 7:51 Where to get the code 8:07 Next video: Voice Design
How to use the notebook: 1. Open the Colab link and choose File → Save a copy in Drive (the shared notebook is view-only). 2. Runtime → Change runtime type → T4 GPU → Save, then click Connect. 3. Run the cells in order with the play button: Step 1 (install), Step 2 (load the model), Step 3 (upload your clip), Step 4 (generate). 4. To say something new, change the text between the three quote marks in a Step 4 cell and run it again. 5. Download your files before closing the tab. Colab deletes them when the session ends.
Your reference clip: 10 to 30 seconds, clean audio with no music or background noise, one person speaking, saved as .wav, .mp3 or .flac. The clone keeps the accent of your reference clip.
Please only clone a voice that is yours or one you have clear permission to use. The OmniVoice project forbids impersonation and fraud. The voice in this video is a sample voice used only for the demo.
Notes: - Free Colab GPU time is limited and can vary by account. - Voice quality varies by language, so always listen to the result.
This video provides a comprehensive tutorial on using OmniVoice, a free, open-source AI model that allows users to clone a voice into over 600 languages using a short audio sample. The process is demonstrated using Google Colab to leverage free GPU acceleration, making high-quality voice cloning accessible without expensive hardware or paid subscriptions.
This is my reference voice.
Right now, somewhere in your house, cold air is being pumped
Now listen to what a free AI model does with it.
Welcome to PexBell. This is your own channel.
Hola a todos mis seguidores de YouTube. Bienvenido a este tutorial.
Seguro que te llamarán a la canasta.
Namaste, amigos.
Adiós a todos.
Ese es el título.
Hello, everyone.
My name is...
Same voice, six languages.
And the only thing I gave it was this one 24 second clip.
No paid subscription, no expensive computer.
It runs on Google's free GPU right in your browser.
In the next few minutes,
I'll show you exactly how to clone a voice into 600 plus languages.
Let's go.
Here's the idea. You give the AI a short recording of a voice.
After that, you type any text in almost any language,
and it speaks that text in that
same voice. For a creator,
that means dubbing your videos for new audiences,
fixing a mistake in a voiceover without re-recording,
or narrating in languages you don't speak
The model we will use is called OmniVoice.
It's open source, so it's free to use,
and the license allows commercial use.
It supports more than 600 languages,
which the project says is the broadest coverage
of any zero-shot voice model.
Zero-shot just means it doesn't need to be trained on your voice first.
One short clip is enough.
And we will run it on Google Colab, which gives you a real
graphics card, a GPU for free in a browser tab.
You don't install anything on your computer.
The quality of your clone depends mostly on your reference clip.
So let's get that right first.
Four rules.
One, about 10 to 30 seconds long.
Mine is 24.
Two, clean audio.
No music, no echo, no background noise.
Three, one person speaking in a natural voice.
Four, save it as a WAV, MP3 or FLAC file.
One more important thing.
Only clone a voice that is yours or one you have permission to use.
The OmniVoice project explicitly forbids using it for impersonation
or fraud, and so do I.
In this video, I'm using a sample voice just for the demo.
The notebook link and a written guide are on pexbell.com,
linked in the description.
Open the notebook, and you'll see four steps.
Everything is already written for you.
You only press play buttons.
First, we turn on the GPU.
Click Runtime, then change runtime type.
Choose T4 GPU.
If you see CPU selected, change it.
The CPU works too, but it is far slower.
The GPU is what makes this fast.
Click Save.
Now click Connect.
Wait a few seconds until you see a green tick.
That tick means Google has given you a computer with a GPU,
and it's connected.
If you open the Resources panel,
you'll see roughly 12 GB of RAM,
15 GB of GPU memory,
and lots of disk space.
All free.
Step 1. Install.
Click the play button on the left of the first cell.
It installs the OmniVoice package.
For me this took about 6 seconds.
Step 2. Load the model.
Same thing. Click play.
Now it downloads the AI brain.
About 13 files and a little over 3 gigabytes.
And loads it onto the GPU.
You'll see some yellow warnings.
Ignore them.
They're harmless.
Wait until the output says model is ready.
That took about a minute for me.
And you only do it once per session.
Step 3. Upload the voice.
Click play on the third cell,
then click choose files and pick your audio.
I'll choose explainer.mp3.
When the upload finishes you'll see
a green successfully uploaded message.
That's it.
The AI now knows which voice to copy.
Step 4 is where the magic happens.
Each demo has a block of text between triple quotes.
That's what the voice will say.
Demo A is English. I'll click play.
After a few seconds the clone appears. Let's listen. Now let's
make it say something of our own.
I'll replace the text with Welcome to PexBell. This is your
own channel. Thanks for watching the tutorial.
Run it again. Welcome to PexBell. This is your
own channel. Thanks for watching the tutorial.
That's your sample voice saying a brand new
sentence. Change the text to anything you like as often as you like.
Now the part that surprises
Same voice, but this time the text is in Spanish.
You don't have to tell the model the language.
It detects it automatically.
Hola a todos mis seguidores de YouTube.
Hoy estamos probando la clonación de voz en español con OmniVoice.
¿Qué les parece cómo suena mi voz en otro idioma?
Perfect Spanish in the same voice.
And it saves the file for you.
Open the files panel on the left and you'll see clonedenglish.wav
and clonedspanish.wav.
To keep a file, click the three dots
on the audio player and choose Download.
Or, in the Files panel,
click the three dots next to the file and choose Download.
Either way works.
Next, French. About 6 seconds to generate roughly 9 seconds of audio.
Bienvenue dans ce tutoriel.
C'est incroyable de voir à quel point cette intelligence artificielle
peut cloner ma voix et s'exprimer parfaitement en français.
Then, Arabic. Also about 6 seconds.
A tip here. The clone keeps
the accent of your reference clip.
My reference speaks English,
so the Arabic has a slight English accent.
If you want a native-sounding Arabic voice,
use a reference clip recorded by an Arabic speaker.
Hindi works the same way.
Namaste dosto, aaj hum isse itool ki madad
se apni awaaz ko Hindi mein badal rahe hain.
Yeh sachmuch jadu jaisa lakta hai.
Same tip, for a Hindi accent, start from a Hindi reference.
And finally, Japanese.
Mina-san, konnichiwa. Watashi no aatarashi jinko chino,
karon boyse, youkoso. Nihongo no hibiki wa ikagadishoka.
Six languages,
And these six are just a sample.
one voice, no cost.
There are more than 600 supported languages.
The full list is on the Omni Voice GitHub page,
linked in the description.
A few extras worth knowing.
Reference text.
If you also type what your reference clip says,
the clone can be more accurate.
If you don't, the tool transcribes it for you automatically.
Speed.
You can make the voice faster or slower with the speed setting.
Expression.
You can add tags like laughter
or sigh in square brackets inside your text
to make the voice react.
Quality versus speed.
A higher step count gives better quality but takes longer.
And in the next video, I'll show you voice design,
where you create a brand new voice
just by describing it.
Like female, low pitch, British accent, with no reference clip at all.
If something goes wrong, here's the quick fix list.
Most problems come down to a noisy reference
clip, a missing GPU, or a collapsed session that timed out.
All of those are one-click fixes.
That's it. One short clip, a free GPU,
and your voice in more than 600 languages.
You can save a lot of time and a lot of money this way.
The notebook, the full written guide,
and the language list are all linked in the description,
on paxwell.com.
If you hit any issue, tell me in the comments and I'll help.
In the next video, we'll explore OmniVoice's voice design
and more of its features.
If you liked this one, subscribe so you don't miss it.
Thanks for watching and I'll see you in the next one.