prayag-1771/Cough-D ? reverse-engineered prompt

Build me a small research project that reproduces a cough detection study from respiratory audio.

I want a pipeline that can take the original audio data, turn it into training images, split the data in a way that keeps the same person out of both training and test when that is possible, train a simple convolutional model, and compare whether wavelet denoising before making the spectrogram helps or not. It should also let me switch the frequency axis style and compare results across three kinds of labels, prompt based cough labels, human annotated labels, and labels coming from another detector.

Please include the scripts and manuscript sources needed to run the whole thing again, plus a way to collect results from multiple random seeds and summarize the differences clearly. I also want the paper checks and figure generation to work from the repo itself.

If anything depends on the current version of a library or external data, look up the current docs online if you need to and make the project easy to run from the command line.