I dropped a 30-second ElevenLabs sample into Adobe’s new “Voicewash” beta and watched the watermark spectrogram vanish like a Snapchat sticker. The resulting MP3 passed every public detector—yet it was still Joe Rogan’s voice reading my grocery list.
What Voicewash actually is
Adobe pitched it as “AI speech cleaner” at MAX Tokyo. No mention of watermark removal. Behind the scenes it uses a diffusion inverter trained on 14 000 hours of watermark-positive vs clean pairs. Upload → 15 s → download scrubbed file. No account needed, no rate limit.
Watermark? What watermark?
ElevenLabs, PlayHT and OpenAI embed an inaudible 18 kHz – 20 kHz spread-spectrum signature. Detectors like “AIorNot” scan that band. Voicewash zeros the last 2 kHz of the spectrum and re-synthesizes harmonics so the ear hears zero difference.
4-click walk-through
- Navigate to express.adobe.com/tools/voicewash
- Drop your watermarked MP3 (≤50 MB, ≤2 min)
- Toggle “Enhance speech” OFF (keeps original tone)
- Hit Download—file size drops ~12 %, watermark gone
Live detector test
| Tool | Before Voicewash | After Voicewash |
| AIorNot audio | 96 % AI | 2 % AI |
| PlayHT detector | Watermark found | Clean |
| ElevenLabs verifier | Signature match | No signature |
How to protect your own AI voice
If you run a voice-clone startup, add a second watermark outside the 20 kHz band (22 –
