Comparative Study of Audio Features for Intelligent Sound Recognition

Abstract

The accurate classification of environmental and urban sounds is an important problem due to its relevance in applications such as surveillance, smart cities, healthcare, and multimedia retrieval. This work explores how different audio features affect classification accuracy in machine learning systems. The proposed solution employs a Convolutional Neural Network (CNN) to classify urban sound events using various audio feature representations. The main research contribution lies in the comparative analysis of multiple audio features and their effectiveness in improving CNN-based classification. This study uses the UrbanSound8K dataset, which contains 10 common urban sound classes. Key audio features MFCCs, log-Mel spectrograms, and Chroma are extracted and used to train separate CNN models. The performance of each feature type is then evaluated and compared. Results show that MFCCs and log-Mel spectrograms significantly outperform Chroma features in classification accuracy. MFCCs provide a compact and informative representation of audio signals, while log-Mel spectrograms capture time-frequency characteristics effectively.  Models trained with these features exhibit robust performance across varied acoustic conditions. The results indicate that appropriate feature selection can lead to improved model generalization and efficiency. In conclusion, this study emphasizes the critical role of audio feature selection in urban sound classification and provides practical guidelines for developing efficient machine listening systems.