search

facebookresearch / audioseal

Localized watermarking for AI-generated speech audios, with SOTA on robustness and very fast detector
MIT License
423 stars 51 forks source link

@pierrefdz I think this is an issue that will be become bigger as more people test Audioseal. #24

Closed KabaTubare closed 5 months ago

KabaTubare commented 6 months ago

@pierrefdz
Hello, I encountered the same issue. I tried five watermarked audio generated from acoustic models (FreeVC, KNN-VC, and my own model), and 4 out of 5 did not meet the watermark with threshold ranging from 0.1 to 0.9, except for the audio from my own model (although the quality of this audio is quite poor). Here's a reproduction: link audio: https://drive.google.com/drive/folders/18u6WO-F__uo7a7fS6bFFNXtli5LN6V-L?usp=sharing link colab: https://colab.research.google.com/drive/1CLtUtYiVGgkCCrbB4Ejkx6hoTSyWacnQ?usp=sharing

Originally posted by @huutuongtu in https://github.com/facebookresearch/audioseal/issues/14#issuecomment-1989863765

pierrefdz commented 6 months ago

Hi @KabaTubare,

Again, watermarking methods are "active". The detector will only detect the watermark in audios where you embed the watermark. Have you embedded the watermark on audios you tested?

PS: If you're interested in synthetic voice detection (without watermarking), you can look at "passive" methods like https://github.com/ghua-ac/end-to-end-synthetic-speech-detection which are agnostic (but have lower performance).