Here are some codes for paper "An Initial Investigation of Neural Replay Simulator for Over-The-Air Adversarial Perturbations to Automatic Speaker Verification".

• adversarial attack scripts: ./attack
• physical vocoder (the UNet model to simulate over-the-air process, specified in paper): ./phys_vocoder
• model checkpoints (the pretrained models specified in paper): ./pretrained_models
• utils for audio clipping (how we collect real-world recording datasets and align frame to frame): ./audio_clipper
• generate PGD adversarial examples: ./attack.py and init_attack.py
• test attack accuracy on ASV models: ./test.py and init_test.py

First, convert the device recording to .wav format with 16000 resample rate using ffmpeg:

ffmpeg -i "input.wav" -osr 16000 output.wav

Set up the source audio and the folder you want to recover into in the script audio_clipper/recover_pulse.py.

Then, set the offset parameter to the observed first pulse point, as shown in the image, the first pulse is between 44-45 seconds, so set it to 44 * 16000: