How little audio input does AI need to clone your voice accurately?

Less than you'd expect: Google Omni's avatar setup builds an accurate voice clone from a short calibration where you read only double-digit numbers aloud — no sentences, no paragraph. The model infers your intonation, pause timing, and syllable handling from the numbers alone, and in first-hand testing the voice replica came out stronger than the face replica.

To clone your voice in Google Omni, you complete a single calibration session: face the camera and read double-digit numbers out loud. That's the entire audio input — the setup never asks for a sentence or a read-aloud passage. From that number-reading alone, the model extracts the prosodic fingerprint of your speech: how you handle intonation, where you pause, and how you shape syllables.

This is a different design choice from the industry norm. Most voice-cloning setups ask for natural, phonetically varied speech — full sentences with a spread of vowels and consonants — on the assumption that richer input produces a better clone. Omni's avatar pipeline instead treats numbers as sufficient raw material for prosody, and the observed results back that up: across 30+ test outputs evaluating the model, the voice replication consistently outperformed the facial replication. A decent face replica, an incredible voice replica.

Two practical notes for using the clone accurately once it's built. First, keep generated speaking clips short: testing found 6–7 seconds is the ceiling where lip sync stays consistently reliable for a single speaker in frame, so script your avatar's lines in beats under that length rather than one long monologue. Second, this capability is currently unique — Google Omni is the only AI video model offering native avatar generation with voice and face replication built in, so if you're evaluating where minimal-input voice cloning actually works today, this is the reference point.

The low input requirement is exactly what makes the feature consequential: when a voice clone costs one number-reading session instead of a recording setup, avatar-driven talking-head content becomes viable for anyone with a camera and a few minutes.

Watch some of these to see what works for you:

It was just double digit numbers that I was saying to the camera — it didn't ask me to say a sentence, read out a paragraph.

— invideo's creative team, on testing Google Omni's avatar calibration