• transformers
• vision transformers
• cnn
• convolutional neural networks
• audio tagging
• audio classification
• noise monitoring
• noise tagging
• spectrograms
• deep learning

This thesis investigates audio tagging and noise recognition, an increasingly significant domain due to its broad applications. The study processes and classifies specific noises for various use-cases, highlighting its utility from machinery malfunction detection to urban security. While Convolutional Neural Networks were traditionally employed for such tasks, the trend is now shifting towards Vision Transformers. This work delves deep into both architectures, comparing their merits.
The main objective is to test these models on benchmark and real-world datasets, aiming to gauge the true performance of Vision Transformers in real scenarios. This involved creating custom datasets, developing new training pipelines, and conducting a detailed model comparison.
The findings were notable. Vision Transformers performed exceptionally on benchmark datasets, but their superiority varied across real-world datasets. This highlights the importance of context in model application and the distinct challenges of real-world situations.
Overall, this thesis offers crucial insights into the complexities and opportunities of audio tagging and noise recognition, setting the stage for future research in this field.