factor → swap speaker → recombine (plus watermark)
target ref (5 s)
alice_5s.wav
dry, close-mic, single speaker
speaker encoder
ECAPA / WavLM-SV
192-d embedding
cloning TTS
F5 / XTTS / VALL-E 2
text + speaker_emb → mel
watermark
SilentCipher
~32 bits, inaudible
text: "add milk to list"
source speech
bob saying "turn off"
any speaker, any text
content extractor
ASR PPG / WavLM frames
speaker-invariant representation
resynthesizer
KNN-VC / Diff-HierVC
content + alice_emb → wav
output
alice saying "turn off"
content preserved
legal and ethical stack (required 2026)
consent gate signed record + speaker-hash match before inference
watermark SilentCipher or PerTh; ~32-bit payload, MP3-robust
detector pair AASIST / RawNet2 — ship with the generator
audit log tamper-evident record of every synthesis request