Skip to content

Cho-SungHyun/Music-Genre-Classification-using-Late-Fusion

Repository files navigation

Music Genre Classification using Late Fusion CNN with Multiple Spectral Features

This repository hosts the core research and architectural frameworks for my publications in IEEE Access (2024) and IEEE ICCE-Asia (2022). The focus is on leveraging Information Fusion strategies to achieve state-of-the-art accuracy in music genre classification.


🏆 Key Research 1: Comprehensive Survey & Late Fusion Excellence

"A Short Survey and Comparison of CNN-Based Music Genre Classification Using Multiple Spectral Features" Published in IEEE Access (Vol. 12, 2024)

  • Objective: To validate the superiority of Late Fusion strategies over existing CNN models.
  • Large-Scale Benchmark: Conducted an empirical study on 12 public datasets (including Ballroom, ISMIR04, and GTZAN).
  • Outcome: The devised Late Fusion CNN outperformed all compared methods, proving that fusing high-level feature maps from parallel branches is more effective for acoustic signal processing.
  • 🔗 Read Full Paper on IEEE Xplore

🔬 Key Research 2: Multi-Spectral Input Architecture

"Effective Music Genre Classification using Late Fusion Convolutional Neural Network with Multiple Spectral Features" Published in IEEE ICCE-Asia 2022

  • Focus: Practical implementation of a multi-input CNN using STFT, Mel-Spectrogram, and MFCC.
  • Strategy: Developed a late-stage integration layer that captures unique temporal and spectral nuances from each feature type independently before final classification.
  • 🔗 Read Full Paper on IEEE Xplore

🛠 Technical Core

  • Fusion Strategy: Late Fusion (parallel CNN branches with integrated decision layers).
  • Feature Extraction: Advanced signal processing using STFT, Mel-Spectrograms, and MFCC.
  • Experimental Rigour: Validated across 12 diverse datasets to ensure model generalisation and robustness.

💡 Why this is relevant for Scalable ML:

My research in Late Fusion is directly applicable to real-world challenges where decision-making requires integrating heterogeneous data sources (e.g., combining user metadata with behavioral signals in Ad-Tech). Managing 12+ datasets also demonstrates my ability to handle complex data pipelines and ensure model performance at scale.

About

SOTA Music Genre Classification using Late Fusion CNN. Evaluated on 12 public datasets. Published in IEEE Access (2024).

Topics

Resources

License

Stars

Watchers

Forks

Contributors