AI voice cloning from 3 seconds: Test results
The video tests whether AI can copy a voice from just three seconds of audio. It creates clones from three seconds, ten seconds, thirty seconds and two minutes of one speaker’s recordings, then compares them with a separate recording of the real voice.
Every clone scores 0.96 on a speaker-recognition model, compared with 0.98 for another real recording. More audio does not improve the score in this test, but one voice and one model do not establish whether the clones can fool family members.
The video also examines impersonation scams, limits of human detection and bank voice ID. It closes with safeguards: a family secret word, callbacks to known numbers and alternatives to voice-based login.
Key points
- Models trained on other people’s speech can use a short sample to identify which voice to imitate.
- The test creates four clones from samples ranging from three seconds to two minutes, with each clone reading the same lines.
- All four clones score 0.96 on the speaker-recognition model, versus 0.98 for a separate real recording.
- The result covers one voice, one test and one model; it does not prove that a clone would fool family members.
- Reported scams, listener studies and voice ID bypasses illustrate the risks of trusting a familiar-sounding voice.
- The video recommends a family secret word, callbacks to known numbers and asking banks about alternatives to voice ID.
Chapters
The description and key points were written with AI from the video's transcript and checked against it, then reviewed by Vishal on . The video plays from YouTube. How we write and check our work.
