Voices & Identity
Watermarking & Provenance
8 min+70 XPStep 1/10
A signature humans can't hear
Audio watermarking embeds a recoverable payload in the waveform below perception threshold. Two families:
- Spread-spectrum / psychoacoustic — hide the signal in frequencies and moments human hearing masks; recoverable by a correlating detector. Classic, computationally cheap.
- Neural watermarks — a trained encoder embeds a pattern the paired decoder extracts; the watermark learns to be inaudible and robust jointly.
The engineering spec is a three-way tradeoff: inaudible, robust (survives recompression, EQ, speed changes), capacious (carries enough bits to be useful). You get to pick two and compromise the third.
insight
A watermark's job isn't secrecy — it's persistence. Everyone can know it's there; the question is whether it survives an MP3 transcode, a speed-up, and a re-record through a phone speaker.