A real AI voice cloner (F5-TTS) cloning a voice from its unprotected vs
Deckrd-salted reference. Speaker-similarity measured by ECAPA-TDNN
(1.0 = identical, ~0.12 = a different person). Voice F0 123.1 Hz.
No protectionthe AI clones you freely
Your voice
The AI's clone
73% you · ECAPA +0.73
perturbation: none
Protection · -29.56 dBsalted reference
Your voice (inaudible)
The AI's clone
18% you · ECAPA +0.18
perturbation -29.56 dB
Protection · -23.97 dBsalted reference
Your voice (audible)
The AI's clone
16% you · ECAPA +0.16
perturbation -23.97 dB
Protection · -18.27 dBsalted reference
Your voice (audible)
The AI's clone
0% you · ECAPA -0.10
perturbation -18.27 dB
Protection · -12.07 dBsalted reference
Your voice (audible)
The AI's clone
0% you · ECAPA -0.11
perturbation -12.07 dB
Protection · -6.15 dBsalted reference
Your voice (audible)
The AI's clone
0% you · ECAPA -0.16
perturbation -6.15 dB
The white line on each bar marks where a completely different person sits
(12% you). When the AI's clone of your protected voice drops past that line, the clone no longer
matches you. Generated by Deckrd Clone Studio.