ESTUDO DE EXTRATORES DE CARACTERÍSTICAS PARA CLASSIFICAÇÃO DE SONS AMBIENTAIS

Abstract

The soundscape surrounding us goes far beyond spoken language or structured music; it is composed of a myriad of environmental sounds, ranging from a dog barking to machinery noise. Teaching computers to listen to and interpret these events—a capability known as acoustic perception—is fundamental for the advancement of modern technologies, such as intelligent security systems and autonomous vehicles. However, environmental audio is complex and unstructured, making feature extraction a critical challenge. This work presents a comparative study of feature extractors for environmental sound classification, aiming to analyze how these techniques map and separate audio events into distinct classes within the computational space. Using public datasets consolidated in the state-of-the-art, the class separation and distribution were visually evaluated, and the extracted attributes were used to train a Multilayer Perceptron (MLP) neural network. The results demonstrate the effectiveness of the proposed approach, achieving an accuracy of 77.50% on one of the tested datasets, highlighting the potential of spectro-temporal representations for the evolution of automated acoustic recognition.