James Imbuido | Data Scientist & AI Engineer

Real-Time Voice Cloning

Clone a voice from a short recording and hear it speak — models run entirely in your browser, and audio never leaves your device.

Record a short voice sample and this page converts it into a speaker embedding, then speaks any text you type in that voice — a three-model pipeline, “Real-Time Voice Cloning” implementation (speaker encoder, Tacotron synthesizer, WaveRNN vocoder), converted to ONNX and executed client-side over WebAssembly.

Privacy by architecture: your recording and the generated audio never leave your device — there is no upload endpoint to send them to. The models (~111 MB, cached by your browser) download only when you record, and synthesis takes tens of seconds per sentence on the single-threaded wasm backend.

Step 1 — Voice sample

Record a short voice sample

Read the sample text below in your normal voice. Recording never leaves your device — audio is processed entirely in this browser tab.

Sample text — read this aloud

“The quick brown fox jumps over the lazy dog while a gentle rain falls on the quiet harbour. Numbers sound different from words, so count them out: three, seven, twelve, forty-five. This voice was recorded in a browser, and it never left this device.”

Aim for 10–20 seconds at a natural pace — varied sentences give the encoder a better picture of your voice than repeating one line.

Step 2 — Synthesize

Make it speak

Available once a voice sample is embedded above.