This repository add watermarks to music generated by MusicGen by Meta based on the white-box watermarking algorithm in the paper "A Watermark for Large Language Models".

To learn more about the official code and paper, please see

1. MusicGen's official code: facebookresearch/audiocraft
2. Watermark: jwkirchenbauer/lm-watermarking

1. You can run a quick demo on colab: colab notebook
2. Run demo at watermark/musicgen_watermark_demo.ipynb locally if you have installed dependencies
3. Run more tests at watermark/audio_watermark_demo.py locally

Please see the two official repository first.

1. For MusicGen, don't directly pip install -U audiocraft, which will install the official code. Instead, clone this repository and use python -m pip install -e .
2. For watermarking, you don't need to install special dependencies, just transformers and nltk

The original code can be found in jwkirchenbauer/lm-watermarking and this repository copies them into watermark directory.

This work doesn't modify the original code but add a new class WatermarkAudioDetector in watermark_processor.py, which uses Encodec to convert audio into codebooks rather than uses tokenizers to convert text into token ids.

1. audiocraft/models/lm.py The language model used by MusicGen is modified. Before sampling in _sample_next_token(), there is watermarking processing for the logits. Some necessary variables are set in __init__().

   Pay attention: only one of codebooks is used to generate watermarks (default codebook is the first one).

2. audiocraft/models/musicgen.py Set watermark_mode in set_generation_params() to determine whether to add watermarks.

This work only tests MusicGen and it can successfully add and detect watermarks in my testing cases.

For example, the colab demo uses "Bluesy guitar instrumental with soulful licks and a driving rhythm section" as prompt to generate audio. The first picture is without watermark and the second one is with watermark. We can see it can detect watermark properly.

The quality of watermarked music will be influenced. Some watermarked music sounds as good as the music without watermark, while some doesn't.

In addition, compression model (Encodec) of MusicGen is a lossy method, which means the codebook (token IDs) used to add watermarks won't be the same as that used to detect watermarks. After my experiments, the error rate will be around 25%, which may impact the accuracy of watermark detection, but fortunately, the detection still works if we set appropriate parameters for watermarking.

Suggestions and discussions are welcome. 🙂