Audio Classification Research
I participated in the 2024 RECSEM REU program in computational science research at the University of Tennessee. Through this program I conducted research on audio classification techniques under the mentorship of Dr. Kwai Wong, a now retired researching professor with UTK's Innovative Computing Labratory and with Oak Ridge Labs. I designed and conducted my own research project optimizing a convolutional neural network (CNN) model to classify a number of audio datatsets of ranging types of audios. The purpose of this project was to compare Mel-Scale and STFT spectrogram feature extraction methods in a standardized CNN model to aid in the efforts to increase efficacy and modularity of audio classification models. Click on the image below for further information on the process and outcome of this research.
I was confident in my finalized model (that was modified from a Github repository cited in the paper) as the model accuracy was comperable to existing models in papers condicting siminlar research. I concluded that the STFT spectrogram features resulted in higher effocacy of the audio classification model in all datasets I tested as the discrepancy between model accuracy between training and validation data was much smaller for the STFT features than for the Mel-Scale features.
I am still working on perfecting this research and comparing my results to those run on a pre-existing CNN model, ResNet18. Once I collect these final results I plan on submitting this paper to a number of conferences and undergraduate research publication journals with the hopes of it helping to advabce the world of machine learning, specificaly audio classification.
