How AI voice cloning works
A voice cloning model has two jobs. First, it listens to your reference clip and turns it into a compact voice fingerprint that captures timbre, pitch range, accent and speaking style. Second, when you give it text, it predicts speech sound by sound in a way that matches that fingerprint.
Modern models need only a few seconds of audio because they have already learned how human speech works from many thousands of speakers. Your clip only has to tell the model which voice to produce.
Listen: reference and clone
Below is an 8-second reference clip, followed by speech generated from it. The second sentence was never spoken by the reference voice.
What decides the quality of a clone
- The reference recording. Noise, echo and music in the clip are reproduced. This matters more than anything else.
- Length. 10–20 seconds is the sweet spot. Very short clips give the model too little to work with.
- Speaking style. The clone imitates how you spoke in the sample. A calm reading produces calm speech.
- The text. Clear punctuation and written-out numbers give the most natural rhythm.
Writing text for natural speech
- Use short and medium sentences. Long sentences are split automatically, but natural breaks sound better.
- Spell out numbers, dates and abbreviations the way they should be read.
- Use commas for small pauses and full stops for longer ones.
- Avoid emoji, URLs and special symbols; they may be skipped or read oddly.
Generation limits
Guests can generate up to 2,000 characters per generation; free accounts can generate 5,000. Longer texts are processed sentence by sentence, with a short natural pause between chunks, and start playing almost immediately.
Using voice clones responsibly
Clone your own voice, or a voice you have explicit permission to use. Never use a clone to deceive, impersonate or defraud, and label AI-generated audio when you publish it. Read our Terms of Use.