Update README.md
Browse files
README.md
CHANGED
@@ -9,4 +9,100 @@ tags:
|
|
9 |
- watermark
|
10 |
---
|
11 |
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
- watermark
|
10 |
---
|
11 |
|
12 |
+
# WavMark
|
13 |
+
> AI-based Audio Watermarking Tool
|
14 |
+
|
15 |
+
- âš¡ **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.
|
16 |
+
- 🙉 **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/](https://wavmark.github.io/).
|
17 |
+
- 😉 **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.
|
18 |
+
- 🤗 **Huggingface Spaces:** Try our online demonstration: https://huggingface.co/spaces/M4869/WavMark
|
19 |
+
|
20 |
+
## Installation
|
21 |
+
```
|
22 |
+
pip install wavmark
|
23 |
+
```
|
24 |
+
|
25 |
+
## Basic Usage
|
26 |
+
The following code adds 16-bit watermark into the input file `example.wav` and subsequently performs decoding:
|
27 |
+
```python
|
28 |
+
import numpy as np
|
29 |
+
import soundfile
|
30 |
+
import torch
|
31 |
+
import wavmark
|
32 |
+
|
33 |
+
|
34 |
+
# 1.load model
|
35 |
+
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
|
36 |
+
model = wavmark.load_model().to(device)
|
37 |
+
|
38 |
+
# 2.create 16-bit payload
|
39 |
+
payload = np.random.choice([0, 1], size=16)
|
40 |
+
print("Payload:", payload)
|
41 |
+
|
42 |
+
# 3.read host audio
|
43 |
+
# the audio should be a single-channel 16kHz wav, you can read it using soundfile:
|
44 |
+
signal, sample_rate = soundfile.read("example.wav")
|
45 |
+
# Otherwise, you can use the following function to convert the host audio to single-channel 16kHz format:
|
46 |
+
# from wavmark.utils import file_reader
|
47 |
+
# signal = file_reader.read_as_single_channel("example.wav", aim_sr=16000)
|
48 |
+
|
49 |
+
# 4.encode watermark
|
50 |
+
watermarked_signal, _ = wavmark.encode_watermark(model, signal, payload, show_progress=True)
|
51 |
+
# you can save it as a new wav:
|
52 |
+
# soundfile.write("output.wav", watermarked_signal, 16000)
|
53 |
+
|
54 |
+
# 5.decode watermark
|
55 |
+
payload_decoded, _ = wavmark.decode_watermark(model, watermarked_signal, show_progress=True)
|
56 |
+
BER = (payload != payload_decoded).mean() * 100
|
57 |
+
|
58 |
+
print("Decode BER:%.1f" % BER)
|
59 |
+
```
|
60 |
+
|
61 |
+
|
62 |
+
|
63 |
+
|
64 |
+
## Low-level Access
|
65 |
+
|
66 |
+
```python
|
67 |
+
# 1.load model
|
68 |
+
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
|
69 |
+
model = wavmark.load_model().to(device)
|
70 |
+
|
71 |
+
# 2. take 16,000 samples
|
72 |
+
signal, sample_rate = soundfile.read("example.wav")
|
73 |
+
trunck = signal[0:16000]
|
74 |
+
message_npy = np.random.choice([0, 1], size=32)
|
75 |
+
|
76 |
+
# 3. do encode:
|
77 |
+
with torch.no_grad():
|
78 |
+
signal = torch.FloatTensor(trunck).to(device)[None]
|
79 |
+
message_tensor = torch.FloatTensor(message_npy).to(device)[None]
|
80 |
+
signal_wmd_tensor = model.encode(signal, message_tensor)
|
81 |
+
signal_wmd_npy = signal_wmd_tensor.detach().cpu().numpy().squeeze()
|
82 |
+
|
83 |
+
# 4.do decode:
|
84 |
+
with torch.no_grad():
|
85 |
+
signal = torch.FloatTensor(signal_wmd_npy).to(device).unsqueeze(0)
|
86 |
+
message_decoded_npy = (model.decode(signal) >= 0.5).int().detach().cpu().numpy().squeeze()
|
87 |
+
|
88 |
+
BER = (message_npy != message_decoded_npy).mean() * 100
|
89 |
+
print("BER:", BER)
|
90 |
+
```
|
91 |
+
|
92 |
+
|
93 |
+
|
94 |
+
|
95 |
+
|
96 |
+
## Thanks
|
97 |
+
The "[Audiowmark](https://uplex.de/audiowmark)" developed by Stefan Westerfeld has provided valuable ideas for the design of this project.
|
98 |
+
## Citation
|
99 |
+
```
|
100 |
+
@misc{chen2023wavmark,
|
101 |
+
title={WavMark: Watermarking for Audio Generation},
|
102 |
+
author={Guangyu Chen and Yu Wu and Shujie Liu and Tao Liu and Xiaoyong Du and Furu Wei},
|
103 |
+
year={2023},
|
104 |
+
eprint={2308.12770},
|
105 |
+
archivePrefix={arXiv},
|
106 |
+
primaryClass={cs.SD}
|
107 |
+
}
|
108 |
+
```
|