Date of Award

1-25-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Mariofanna Milanova

Abstract

Audio classification plays a crucial role in interpreting and understanding soundscapes, enabling applications like voice assistants, sound event detection, and music analysis. However, deploying deep learning models for audio classification on edge devices presents significant challenges. These models often require substantial computational resources and memory, which are limited on edge devices. Balancing performance, efficiency, and accuracy remains a key hurdle in this field. In this research, we explore various deep learning architectures, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Audio Spectrogram Transformer (AST), for the purpose of audio classification on the ESC 50 and Audio Set datasets. The feature extraction process employs Mel spectrograms, leveraging their compatibility with advanced image classification algorithms. Subsequently, model compression techniques such as XNOR networks, network pruning, and quantization are applied to reduce model size. The study reveals that a weight-pruned model achieves an accuracy of 89%, which, while slightly lower than the 91% accuracy attained by the uncompressed CNN, demonstrates a balance between efficiency and performance. This finding underscores the potential of model compression in facilitating efficient deep learning applications on resource-constrained devices without compromising accuracy.

Download

Share

COinS