Skip to content

AI Voice Cloning (runs on your GPU)

Record a few seconds of your voice and let an AI model on your own computer speak any text in it.

Checking your browser for WebGPU…

1. Your voice sample (5–20 seconds)

or
Upload a voice clip
WAV, MP3, M4A, OGG, WebM or a video; 5–20 s of one person speaking clearly

Tip: read this aloud in a quiet room, about 15 cm from the microphone: "Hello! I am recording a short sample of my voice. The weather is lovely today, and I am going to read a few sentences clearly and naturally."

2. Text to say in that voice

3. Options

Runs on your own device with WebGPU/WebAssembly. Your voice and text are never uploaded; the model is downloaded once from Hugging Face and cached in your browser.

Report a problem

Don't include personal or sensitive information.

What happened, and what did you expect? (10–2000 characters) 0 / 2000

Your browser and system, screen size and recent error messages on this page. Never what you typed or your files. Privacy

What is sent
 

This voice cloning tool copies the sound of a voice from a short sample. Record or upload 5–20 seconds of your own speech, type what you want to hear, and the AI speaks the text in a voice that sounds like yours: the same pitch, tone and accent. It uses VoxCPM2, an open model from OpenBMB with 2 billion parameters that speaks 30 languages, including English, Khmer, Chinese, Thai, Vietnamese, Japanese, Korean, French and Spanish, and produces clear 48 kHz audio.

Unlike other voice cloning websites, nothing is sent to a server. The model is downloaded once (about 1.75 GB) and then runs on your own graphics card through WebGPU in your browser, so your voice sample and your text stay on your device. There is no account, no credit system and no watermark. Please use it responsibly: clone only your own voice or a voice you have permission to use.

How to use the AI Voice Cloning tool

  1. Read the consent note, then press "Record my voice" and speak for 5–20 seconds, or upload a clean clip of one person talking.
  2. Optionally type the exact words of the sample in "Transcript" for a closer copy, and tick the box confirming it is your voice or you have permission.
  3. Type the text the cloned voice should say. Long text is split into sentences automatically.
  4. Pick a quality and press "Clone voice and speak". The first time, the model downloads (about 1.75 GB) and is saved in your browser.
  5. Listen as each sentence is ready, then download the result as WAV or MP3.

Why use ToolOrca AI Voice Cloning

  • Zero-shot cloning from a 5–20 second sample: no training, no waiting
  • Record in the browser or upload WAV, MP3, M4A, OGG, WebM or video
  • 30 languages including Khmer, English, Chinese, Thai and Vietnamese
  • Runs on your own GPU with WebGPU: no upload, no account, no watermark
  • Model is cached after the first visit, so the next run starts in seconds
  • Optional style hints such as "cheerful" or "slower"
  • Stop button, live playback per sentence, WAV and MP3 download

Hardware requirements

Voice cloning needs WebGPU, which is built into recent Chrome and Edge on Windows, macOS, ChromeOS and Linux, and into Safari 26 on Mac and iPad. The model needs about 2–3 GB of free GPU memory: a computer with a dedicated graphics card (NVIDIA, AMD, Intel Arc) or an Apple Silicon Mac (M1 or newer) with 8 GB of memory or more works well; integrated graphics on recent laptops also work but are slower. The first visit downloads about 1.75 GB, so use Wi-Fi.

Most phones do not offer WebGPU yet, and the ones that do usually have too little memory for a 2-billion-parameter model, so use a computer for cloning. Without WebGPU the tool can fall back to the CPU (WebAssembly), but this is very slow: expect several minutes for a single sentence.

On a GPU, speed depends on the card: a fast graphics card or an Apple M-series chip produces roughly real-time speech or faster with the "Balanced" setting; slower integrated graphics may need two to four seconds per second of audio. Choose "Fast" quality to roughly halve the time.

How to get the most natural copy

Record in a quiet room without music or echo, about a hand width from the microphone, and speak naturally in the language you want to generate. Eight to fifteen seconds is ideal; only the first 12.8 seconds are used. Typing the exact transcript of the sample switches the model to "continuation" mode, which copies rhythm and pronunciation more closely. Keep sentences of the target text reasonably short; the tool splits long text for you.

The model speaks each sentence independently, so the voice may vary slightly between sentences, and numbers or unusual names are sometimes misread: write numbers as words if a sentence comes out wrong, and generate it again (every run is a little different).

Use voice cloning responsibly

Cloning someone's voice without permission can be illegal and can hurt people. Only clone your own voice or a voice whose owner has clearly agreed. Do not use cloned speech to impersonate anyone, to trick people (for example in calls or messages asking for money), to spread false statements, or to bypass voice-based security. When you publish cloned audio, say that it was generated by AI.

About the model

VoxCPM2 is a tokenizer-free, diffusion-autoregressive text-to-speech model by OpenBMB, released under the Apache 2.0 licence. This page runs a 4-bit quantized ONNX export of it with ONNX Runtime Web, plus the model's own audio encoder to read your sample. The quantized version is close to, but not identical to, the full model; it sometimes slips on numbers.

Frequently asked questions

Is my voice uploaded anywhere?

No. The recording, the transcript and the text are processed on your own device. Only the model files are downloaded (from Hugging Face) and saved in your browser's storage.

Can it clone a voice in Khmer?

Yes. VoxCPM2 supports 30 languages including Khmer, Thai, Lao, Burmese, Vietnamese, Chinese, English, Japanese and Korean. Record your sample in the language you want and type the text in that language.

Is it legal to clone a voice?

Cloning your own voice, or a voice whose owner has given permission, is fine. Cloning someone else without consent, impersonating people or using cloned speech to deceive can be illegal. Please read the consent note next to the record button.

Why does it not work on my phone?

The model needs WebGPU and about 2–3 GB of graphics memory. Most phones lack WebGPU or enough memory. Use Chrome or Edge on a computer, or Safari 26 on a Mac.

Why is the first run slow?

The first time, your browser downloads the model (about 1.75 GB) and prepares it for your GPU. It is cached, so later visits skip the download and start much faster.