Most voice cloning services work like this: you upload your recording, a GPU server in a data centre processes it, and you pay for the computing time with a subscription or credits. Easy Voice Clone turns that around. The neural network is downloaded to your browser once, and from then on your own device does the work.
That's the reason the service can be free, and the reason your recordings stay private. Here is what actually happens behind the Load model button.
What downloads when you press “Load model”
The voice model isn't one big file. It is a set of smaller models that each do one job, plus the built-in preset voices. For the English model, the download is roughly:
| Part | What it does | Size |
|---|---|---|
| Main speech model | Predicts the next piece of speech from the text and the voice | ≈ 76 MB |
| Built-in preset voices | Ready-made voices such as Alba and Jean | ≈ 52 MB |
| Audio decoder | Turns the model's output into an actual sound wave | ≈ 23 MB |
| Audio encoder | Listens to your sample and creates the voice profile | ≈ 21 MB |
| Text conditioner | Turns your text into a form the model understands | ≈ 16 MB |
| Flow model | Refines each step of generated speech | ≈ 10 MB |
| Tokenizer and settings | Splits text into word pieces | < 1 MB |
| Total | ≈ 200 MB (shown as 189 MB in the Studio, which counts in binary megabytes) |
The models are compressed (“quantised”) versions that trade a tiny amount of quality for a big reduction in size and much faster generation on ordinary processors.
Where everything is stored
- Model files go into your browser's Cache Storage. Each file is saved the moment it finishes downloading, so an interrupted download only has to fetch what's missing.
- Your cloned voices, meaning the compact voice profile plus your original reference clip, go into the browser's IndexedDB database for this site.
- Preferences such as your last language and voice go into local storage.
None of this is sent to our server. The only things our server receives from the Studio are anonymous counters: that a generation happened and how many characters it had, never the text or audio.
How to check it yourself
You don't have to take our word for it. In Chrome or Edge on a computer:
- Open the Studio and press F12 (or right-click → Inspect).
- Open the Network tab and generate some speech. You'll see no request carrying your text or audio, just a small counter request.
- Open the Application tab. Under Cache storage you'll find the model files, and under IndexedDB your saved voices.
How to delete it
- A single voice: click the ✕ next to it in the Studio sidebar.
- The model files: click Delete downloaded models in the Studio. Your voices are kept.
- Everything: clear this site's data in your browser settings (in Chrome: the padlock or tune icon in the address bar → Site settings → Delete data).
Because the data lives only in this browser, it doesn't follow you to another device or browser. Keep your original recording if you'd like to recreate a voice elsewhere.
Why the Studio is faster than the home page
Modern processors have several cores. To use more than one of them from a web page, the browser requires a security mode called cross-origin isolation. The dedicated Studio page runs in that mode and can use up to 8 threads, which typically makes generation 2–4 times faster. The home page demo runs on a single thread so that it stays compatible with everything else on the page. If your browser can't use the faster mode, the Studio switches to the compatible mode automatically.
What kind of device you need
| Device | What to expect |
|---|---|
| Recent laptop or desktop | Model starts in a few seconds from cache; speech is generated faster than it plays |
| Older laptop (4+ years) | Works; starting the model and generating take longer. Close heavy tabs |
| Recent phone | Works; generation is slower than on a laptop, so keep texts shorter |
| Older or low-memory phone (under 4 GB RAM) | May struggle to start the model. The Studio warns you when it detects limited memory |
You also need about 200 MB of free storage per language, and a recent version of Chrome, Edge, Firefox or Safari.
The trade-off, honestly
Running AI on your device has one real cost: the first download. On fast Wi-Fi it takes well under a minute; on slow mobile data it can take several minutes. We've made it as painless as we can. The Studio shows the size, speed and time remaining, warns you about slow or metered connections and low storage, retries automatically if the connection drops, and resumes by itself when you come back online. After that first time, it loads from your device in seconds, even offline.
For a free, private voice clone, we think that's a fair trade. Try it in the Studio.
Clone your voice for free
Studio-quality voice cloning in 15 seconds, right in your browser. No sign-up required.
Open the Voice Cloning Studio