Deep Learning for Music Genre Classification: A Systematic Method Collection and Experimental Framework

This GitHub repository (maintained by furkan-ersoz) focuses on Music Genre Classification (MGC). By systematically collecting, organizing, and experimenting with various deep learning methods in this field, it provides reusable technical references for audio classification research. It does not pursue new model innovations; instead, it establishes standardized experimental benchmarks to help researchers compare the performance of different methods, promoting reproducibility and experience accumulation in the field.

Music genre classification is a classic multi-class problem at the intersection of audio signal processing and machine learning (input audio → output genre labels). Core challenges include: ambiguous genre boundaries and difficulty in processing high-dimensional audio data. Traditional Music Information Retrieval (MIR) relies on handcrafted features (MFCC, chroma, etc.), while deep learning methods automatically learn features, reducing reliance on experts.

The repository adopts a "systematic experiment" methodology: 1. Collect various architectures (CNN to capture local time-frequency patterns, RNN/LSTM to model temporal dependencies, CRNN hybrid architecture, Transformer self-attention); 2. Standardized process: unified preprocessing, consistent dataset division, multi-dimensional evaluation metrics (accuracy/F1, etc.), and complete hyperparameter records to ensure reproducibility.

The experiments use public benchmark datasets such as GTZAN, FMA, and MagnaTagATune. Features include: Mel spectrograms (mainstream time-frequency representation), raw waveforms (end-to-end learning), and handcrafted features (as baseline comparison).

1. There is no absolute optimal architecture; task characteristics, data scale, etc., need to be balanced; 2. Data quality is more critical than model complexity; 3. Standardized code and hyperparameter records solve the reproducibility problem.

Technical applications include: personalized music recommendation, automatic music library management, improved copyright authorization efficiency, and music education and research assistance.

Developers should note: choose librosa/torchaudio libraries; optimize data loading efficiency; handle class imbalance; lightweight models (distillation/quantization) to adapt to mobile deployment.

This repository is a model of open-source research, providing complete learning resources for beginners and an experimental framework for researchers. In the future, with the development of technologies such as self-supervised learning and multi-modal fusion, the MGC field will continue to evolve.