PERSA+: A Deep Learning Front-End for Context-Agnostic Audio Classification

TL;DR

This paper introduces PERSA+, a deep learning front-end designed to filter out irrelevant information from raw audio data, aiming to improve robustness and generalization in real-world audio classification tasks.

Contribution

The work proposes a novel preprocessing front-end that enhances deep learning models' ability to perform reliably across diverse and unpredictable audio environments.

Findings

Improved robustness in real-world audio classification

Enhanced generalization across different audio contexts

Reduced detrimental information before modeling

Abstract

Deep learning has been applied to diverse audio semantics tasks, enabling the construction of models that learn hierarchical levels of features from high-dimensional raw data, delivering state-of-the-art performance. But do these algorithms perform similarly in real-world conditions, or just at the benchmark, where their high learning capability assures the complete memorization of the employed datasets? This work presents a deep learning front-end, aiming at discarding detrimental information before entering the modeling stage, bringing the learning process closer to the point, anticipating the development of robust and context-agnostic classification algorithms.