Music Genre Detection

This project involves the classification of music genres using the GTZAN dataset. Various machine learning algorithms and neural networks are employed to achieve the best accuracy.

Dataset

The dataset used is the GTZAN Music Genre Classification Dataset, containing audio tracks categorized into different genres.

Project Overview

• Data Cleaning and Preprocessing: Handled missing values, encoded labels, and normalized the data.
• Data Visualization: Visualized waveforms of each genre to understand data distribution.
• Model Training and Evaluation: Trained and evaluated several machine learning models:
  • Logistic Regression: 52.33%
  • K-Nearest Neighbors (KNN): 70.67%
  • Decision Tree: 62.00%
  • Random Forest: 78.00%
  • CatBoost: 83.33%
  • XGBoost: 77.33%

• Neural Network Implementation: Trained a neural network for 100 epochs, achieving a test accuracy of 75%.

• Model Comparison: Compared all models based on accuracy. The CatBoost Classifier performed the best with an accuracy of 83.33%.

Model Saving and Loading

The best performing model, CatBoost Classifier, was saved using pickle for later use:

import pickle as pkl

# Saving the model
with open('cbc_model.pkl', 'wb') as file:
    pkl.dump(cbc, file)

# Loading the model
with open('cbc_model.pkl', 'rb') as file:
    cbc_pkl = pkl.load(file)

# Predicting with the loaded model
predictions = cbc_pkl.predict(x_test)

Key Insights

• Spectrograms: Used to visualize audio frequencies over time, providing more information for neural network training.
• Convolutional Neural Networks (CNNs): Leveraged for their efficiency in pattern recognition within spectrograms, improving genre classification accuracy.

GitHub Repository

Explore the source code and contribute to the project on GitHub: View on GitHub

Libraries Used

• Pandas
• NumPy
• Scikit-learn
• Matplotlib
• Seaborn
• CatBoost
• XGBoost
• TensorFlow
• librosa