This is a submission to the HackSMU VII hackathon for the iMasons-Elephant Voices challenge which ran from April 11-12, 2026. The challenge statement was:

Develop a method to remove overlapping noise from elephant recordings (generators, cars, planes) without distorting the elephant call.

• Use spectrograms to visualize sound
• 44 audio recordings with 212 elephant calls embedded in mechanical noise
• Some of the calls overlap with each other too
• Spreadsheet has sound file name, start time, and end time for each call

We implemented a convolutional neural network to denoise elephant rumble audio files, based on existing academic research on the subject.

Considering the short 24-hour time constraint of HackSMU VII, we used off-the-shelf LLMs such as Claude Code and Perplexity to assist with some of the implementation and debugging.