Missing dependency

[ChenghaoMou]

It seems that this package requires setuptools. Using uv will lead to PerthImplicitWatermarker being None with an import error No module named 'pkg_resources'.

[notdaniel]

Same here. Unclear if due to the same problem, but if I get Perth to install by preinstalling setuptools into the uv venv, it still fails to install Perth's dependencies, including numpy, librosa, etc.

[Lawrence-Liu]

same here. I am trying to run chatterbox, but failed due to perth.PerthImplicitWatermarker() is None. After digging a bit, I think it's the same issue, and I am using uv as well

[skyne98]

Same here

[notdaniel]

I really wanted to toy around with this watermarking, so actually I did attempt to pull this all apart and repackage it properly, but that was... a bit more than I thought.

The subdirectories have no __init__.py in them, so none of them are packages, which they are getting around by doing arbitrary path inserts and manual locating/importing of modules using the deprecated pkg_resources. The given dependencies aren't actually versions compatible with each other. The project is tagged as supporting down to python 3.8 but the minimum with what they are using is I believe 3.9 or 3.10. Finally, there's no .gitignore so there's some __pycache__ etc. in the repo. I've sorted most of this out, but it's now borderline unrecognizable and I don't know if they'd appreciate that very much (or my converting it to uv). But I suppose no harm in trying, right?

(If you just want to get it to run, use a python 3.10 venv, ignore the requirements, and install torch, torchaudio, numpy, librosa, soundfile, matplotlib, and pyyaml. I think that's the point where it was runnable.)

[riqiang-dp]

Same error here using poetry to manage the dependencies

[notdaniel]

Okay, so I have gotten Perth packaged properly and seemingly working across platforms. Unfortunately, after doing that, I came to realize that the packaging was just one unfinished piece on top of an unfinished project. But it's pretty usable now.

When I first got it to run properly using their model, I noticed the watermarked output from a stereo music input was in mono. Turns out their model only works on mono sources right now, so I fixed that by applying the watermark to each channel separately and then putting it back together. But after I did that, the outputs were occasionally clipping and/or had noticeably different dynamics, so I did what I could to compensate for that, and the results are very close to the input, often indistinguishable, but also often very slightly different. Therefore, it is perhaps not yet suitable for how I imagine most people would expect to use a watermarking tool. It's a good start, I just hope they look at this and continue on with it. It is very possible that when used strictly on mono sources, it works perfectly and sounds exactly the same! But I don't really have a use-case for watermarking just mono audio sources, so I did not attempt to find out.

If you want to try Perth, you can use uv to install it directly from my GitHub branch.

• Install: uv pip install "git+https://github.com/notdaniel/Perth@8e0cf31f406690ffc256b9fd732b8cc4be40a7a6"
• Run Perth's CLI: uv run python -m perth.cli.watermark_cli {YOUR_AUDIO_FILE.wav} --method perth --device cpu

If you want to use the GPU feature with CUDA on Windows, you can, but the packaging isn't set up yet for that. After installing the package, run the following:

• uv pip uninstall torch torchaudio
• uv pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128 (replace with your required CUDA flavor if needed)

You should then be able to run it with --device cuda.

Hope this helps.

[manmay-nakhashi]

@notdaniel feel free the raise a pr to this repo i'll merge it.

[notdaniel]

@manmay-nakhashi Excellent, will do!