Why do people use STFT as a preprocessing step to using CNN?

Just briefly looking up some research papers on audio data and I have come across some papers that use STFT as a preprocessing step to using CNN. Why is this the case? What are the advantages and disadvantages of doing so?

Some research papers that use STFT;

• https://arxiv.org/pdf/1712.02898.pdf
• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6928168/