Music Genre Classification using Spectrogram Images

This repository presents a comprehensive framework for hierarchical music genre classification using spectrogram images and multiple machine learning and deep learning approaches. It includes implementations with EfficientNet, Audio Spectrogram Transformer (AST), Custom Convolutional Neural Network (CNN), and traditional machine learning models. All models utilize multi-modal spectrogram features (mel, CQT, chroma) and support multi-level genre classification.

---

Problem Statement

Goal:
To classify music tracks into genres and sub-genres by learning from spectrogram images generated from audio signals. The challenge involves:

• Handling hierarchical genre classification (coarse to fine: classic/non-classic → genre → sub-genre).
• Effectively processing multi-modal spectrogram images.
• Building models that generalize well to unseen music and provide interpretable predictions.

---

Approach Overview

1. Data Source

• Uses the ccmusic-database/music_genre from HuggingFace.
• Each sample contains mel, CQT, and chroma spectrogram images, and three levels of genre labels.

2. Models Implemented

• EfficientNet_model.ipynb: Music genre classification using a pre-trained EfficientNet-B0 backbone.
• AST_model.ipynb: Classification using the Audio Spectrogram Transformer (AST), leveraging transformer attention for spectrogram images.
• CustomCNN_model.ipynb: Custom CNN with residual and attention blocks, designed for multi-modal spectrogram input.
• ML_model.ipynb: Traditional machine learning pipeline with PCA for dimensionality reduction and gradient boosting classifiers.

3. Workflow

• Data loading & preprocessing
• Multi-level label mapping and encoding
• Model training, validation, and testing
• Advanced techniques: SpecAugment, label smoothing, cosine LR scheduling, mixed precision
• Evaluation with metrics and confusion matrices

---

Hierarchical Genre Labels

Level	Description	Example Classes
First Level	Binary	Classic, Non-classic
Second Level	9 genres	Symphony, Opera, Solo, Chamber, Pop, Dance/House, Indie, Soul/RnB, Rock
Third Level	16 sub-genres	Symphony, Opera, Solo, Chamber, Pop vocal ballad, Adult contemporary, Teen pop, Contemporary dance pop, Dance pop, Classic indie pop, Chamber cabaret and art pop, Soul/RnB, Adult alternative rock, Uplifting anthemic rock, Soft rock, Acoustic pop

---

Results Table

Model	Overall Accuracy	Classic/Non-classic	2nd-level Genres	3rd-level Sub-genres	Notable Features
EfficientNet-B0	96.9%	100.0%	95.9%	94.9%	Transfer learning, multi-input
AST Transformer	89.3%	99.3%	86.3%	82.3%	SpecAugment, transformer attention
Custom CNN	91.1%	99.8%	90.1%	83.5%	Residual, SE blocks, label smoothing
ML Model (GBM + PCA)	73.3%	96.0%	65.0%	59.0%	Feature engineering, memory-efficient

---

Repository Structure

File	Description
EfficientNet_model.ipynb	EfficientNet-based deep learning workflow
AST_model.ipynb	Audio Spectrogram Transformer (AST) workflow
CustomCNN_model.ipynb	Custom CNN architecture for music genre classification
ML_model.ipynb	Traditional machine learning pipeline
README.md	Project overview and results

---