Music Genre Classification – GTZAN Dataset

Overview

This project implements a music genre classification pipeline using the GTZAN dataset (1,000 tracks, 10 genres).

The system combines audio feature engineering and supervised machine learning models, achieving up to 93% accuracy through ensemble methods.

Developed as part of a Deep Learning course project.
This repository includes only the music classification task.

Objective

• Extract meaningful audio features from raw waveform signals
• Compare classical ML models and neural networks
• Evaluate ensemble learning strategies
• Build a demo music genre classifier

Dataset

GTZAN Dataset

• 1,000 audio tracks
• 10 genres
• 30 seconds per track
• WAV format (22.05 kHz, mono)

Genres: Blues, Classical, Country, Disco, Hip-hop, Jazz, Metal, Pop, Reggae, Rock

Preprocessing Pipeline

• Audio segmentation: 12 overlapping 3-second segments per track
• Resampling and padding
• Standardization (StandardScaler)
• 80/20 train-test split

Feature Engineering

Extracted features include:

• MFCC (Mel-Frequency Cepstral Coefficients)
• Spectral Centroid
• Spectral Bandwidth
• Spectral Rolloff
• Spectral Contrast
• Chroma Features (12 pitch classes)
• Zero Crossing Rate (ZCR)
• Root Mean Square Energy (RMS)
• Tempo (BPM estimation)

Models Evaluated

Model	Accuracy
SVM	0.91
Random Forest	0.88
KNN	0.89
XGBoost	0.89
Neural Network	0.92
Ensemble (soft voting)	0.93

The ensemble model combining SVM, RF and KNN achieved the best overall performance.

Tech Stack

• Python
• librosa
• scikit-learn
• XGBoost
• TensorFlow / Keras
• NumPy / Pandas

Key Takeaway

Carefully engineered audio features combined with ensemble learning can achieve competitive performance without requiring deep end-to-end CNN architectures.