Skip to content

wavmark/wavmark

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

WavMark

AI-based Audio Watermarking Tool

  • Leading Stability: The watermark resist to 10 types of common attacks like Gaussian noise, MP3 compression, low-pass filter, and speed variation; achieving over 29 times in robustness compared with the traditional method.
  • 🙉 High Imperceptibility: The watermarked audio has over 38dB SNR and 4.3 PESQ, which means it is inaudible to humans. Listen the examples: https://wavmark.github.io/.
  • 😉 Easy for Extending: This project is entirely python based. You can easily leverage our underlying PyTorch model to implement a custom watermarking system with higher capacity or robustness.
  • 🤗 Huggingface Spaces: Try our online demonstration: https://huggingface.co/spaces/M4869/WavMark

Installation

pip install wavmark

Basic Usage

The following code adds 16-bit watermark into the input file example.wav and subsequently performs decoding:

import numpy as np
import soundfile
import torch
import wavmark


# 1.load model
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = wavmark.load_model().to(device)

# 2.create 16-bit payload
payload = np.random.choice([0, 1], size=16)
print("Payload:", payload)

# 3.read host audio
# the audio should be a single-channel 16kHz wav, you can read it using soundfile:
signal, sample_rate = soundfile.read("example.wav")
# Otherwise, you can use the following function to convert the host audio to single-channel 16kHz format:
# from wavmark.utils import file_reader
# signal = file_reader.read_as_single_channel("example.wav", aim_sr=16000)

# 4.encode watermark
watermarked_signal, _ = wavmark.encode_watermark(model, signal, payload, show_progress=True)
# you can save it as a new wav:
# soundfile.write("output.wav", watermarked_signal, 16000)

# 5.decode watermark
payload_decoded, _ = wavmark.decode_watermark(model, watermarked_signal, show_progress=True)
BER = (payload != payload_decoded).mean() * 100

print("Decode BER:%.1f" % BER)

How it works?

In paper WavMark: Watermarking for Audio Generation we proposed the WavMark model, which enables encoding 32 bits of information into 1-second audio. In this tool, we take the first 16 bits as a fixed pattern for watermark identification and the remaining 16 bits as a custom payload. The same watermark is added repetitively to ensure full-time region protection: Illustrate

Since the pattern length is 16, the probability of "mistakenly identifying an unwatermarked audio as watermarked" is only 1/(2^16)=0.000015.

Low-level Access

For a specific watermarking algorithm, there exists a trade-off among capacity, robustness, and imperceptibility. Therefore, a watermarking system often needs customization according to application requirements. The good news is that WavMark is entirely implemented with PyTorch. Here is an example of directly calling the PyTorch model:

# 1.load model
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = wavmark.load_model().to(device)

# 2. take 16,000 samples
signal, sample_rate = soundfile.read("example.wav")
trunck = signal[0:16000]
message_npy = np.random.choice([0, 1], size=32)

# 3. do encode:
with torch.no_grad():
    signal = torch.FloatTensor(trunck).to(device)[None]
    message_tensor = torch.FloatTensor(message_npy).to(device)[None]
    signal_wmd_tensor = model.encode(signal, message_tensor)
    signal_wmd_npy = signal_wmd_tensor.detach().cpu().numpy().squeeze()

# 4.do decode:
with torch.no_grad():
    signal = torch.FloatTensor(signal_wmd_npy).to(device).unsqueeze(0)
    message_decoded_npy = (model.decode(signal) >= 0.5).int().detach().cpu().numpy().squeeze()

BER = (message_npy != message_decoded_npy).mean() * 100
print("BER:", BER)

Thanks

The "Audiowmark" developed by Stefan Westerfeld has provided valuable ideas for the design of this project.

Citation

@misc{chen2023wavmark,
      title={WavMark: Watermarking for Audio Generation}, 
      author={Guangyu Chen and Yu Wu and Shujie Liu and Tao Liu and Xiaoyong Du and Furu Wei},
      year={2023},
      eprint={2308.12770},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}