facebookresearch / audioseal

Localized watermarking for AI-generated speech audios, with SOTA on robustness and very fast detector
MIT License
466 stars 56 forks source link

Add additional audio attack methods for enhanced evaluation #57

Closed hastagAB closed 3 days ago

hastagAB commented 2 weeks ago

Introducing an additional set of audio attacks to enhance the capabilities of AudioSeal for evaluating the robustness of audio watermarking techniques.

The following effects have been added:

  1. Pitch Shift: Changes the pitch of the audio without altering the speed, enabling evaluation under shifted tonal characteristics.
  2. Reverse: Reverses the entire audio signal, simulating playback in reverse.
  3. Clipping: Clips the audio signal at a specific threshold to simulate distortion due to high volume levels.
  4. Time Stretch: Changes the duration of the audio without modifying the pitch, useful for testing stretched audio scenarios.
  5. Tremolo: Applies amplitude modulation at a specified frequency, creating a tremolo effect.
  6. Flanger: Introduces a delayed version of the signal with periodic modulation, creating a swirling effect.
  7. Distortion: Amplifies and clips the audio signal, introducing harmonic distortion.
  8. Bit Crusher: Reduces the bit depth of the audio signal, adding digital distortion for lo-fi quality simulation.
  9. Vocoder: Modulates the audio signal with a carrier frequency, creating a robotic sound effect.
  10. Ring Modulation: Multiplies the audio signal with a sine wave, resulting in a metallic sound effect.
  11. Granulate: Breaks the audio into overlapping grains, creating a stutter-like granulated effect.
antoine-tran commented 3 days ago

Duplicate https://github.com/facebookresearch/audiocraft/pull/507