Skip to main content

AI-Based Audio Watermarking Tool

Project description

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}
}

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

wavmark-0.0.3.tar.gz (564.5 kB view details)

Uploaded Source

Built Distribution

wavmark-0.0.3-py3-none-any.whl (13.9 kB view details)

Uploaded Python 3

File details

Details for the file wavmark-0.0.3.tar.gz.

File metadata

  • Download URL: wavmark-0.0.3.tar.gz
  • Upload date:
  • Size: 564.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/4.0.2 CPython/3.9.6

File hashes

Hashes for wavmark-0.0.3.tar.gz
Algorithm Hash digest
SHA256 64e5773d7c1bf07d31fe92375fa921f6e5c511a560a8a523c449626adcdeaa0f
MD5 4670d23289e385b43958ccf499130a38
BLAKE2b-256 264631acbe081fe7822a26af90e7e9f2160daa01bcf78cc94b2f46132c9b9adc

See more details on using hashes here.

File details

Details for the file wavmark-0.0.3-py3-none-any.whl.

File metadata

  • Download URL: wavmark-0.0.3-py3-none-any.whl
  • Upload date:
  • Size: 13.9 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/4.0.2 CPython/3.9.6

File hashes

Hashes for wavmark-0.0.3-py3-none-any.whl
Algorithm Hash digest
SHA256 f77c7234e66b6983273fa7bd7509acaa9195cd936fd1c178b46c8e996d7e8f02
MD5 f1cbebe8320d986f17c1862b57cdd506
BLAKE2b-256 46ddb929dff997313091c60a93df26b82f400267c17a425bd1b4666f1761ad34

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page