Machine Learning Approaches in Song Recognition: Feature Extraction and Preprocessing Techniques for Enhanced Accuracy

Accurate song recognition plays a vital role in music streaming platforms, copyright enforcement, and personalized user experiences. However, existing systems often struggle with challenges such as variations in audio quality, background noise, and differences in file formats. This work proposes a feature extraction and preprocessing framework designed to improve recognition accuracy using open-source tools. The method was evaluated on a dataset of 88 songs spanning multiple genres, specifically curated to test its effectiveness. Key audio features—including MFCCs, spectral centroid, chroma features, spectral contrast, and tonnetz were extracted and normalized to ensure consistency across inputs for machine learning models. Visual tools like correlation heatmaps and feature importance analysis provided deeper understanding of feature relationships, aiding in reducing redundancy and optimizing feature selection. The results revealed distinct audio patterns that contribute to robust song identification, while error-handling strategies addressed file compatibility issues. The framework offers a reproducible solution and lays the groundwork for scalable, real-time, and culturally diverse music recognition systems in future developments.