🎵 Music Genre Classification with CNNs

This repository implements a music genre classification system using deep learning (Convolutional Neural Networks) on audio spectrograms. It explores multiple model versions with enhancements in architecture, preprocessing, and evaluation.

---

📂 Repository Structure

├── Colab_version.ipynb           # Final/Colab-optimized version
├── improvedmusicgenre.ipynb      # Enhanced version with augmentations and callbacks
├── musicgenre.ipynb              # Base version with basic CNN pipeline
└── README.md                     # Project documentation

---

🧠 Objective

The goal is to classify audio samples into one of 10 music genres (Blues, Classical, Country, Disco, HipHop, Jazz, Metal, Pop, Reggae, Rock) using mel spectrograms as input features for CNN-based models.

---

📦 Dependencies

Make sure to install the following libraries:

pip install numpy pandas matplotlib seaborn librosa scikit-learn tensorflow

---

🎼 Dataset

Dataset used: GTZAN Genre Collection

• Format: .wav files, 22050 Hz, 30 seconds each.
• Location: Place the dataset inside C:/Datasets/Data/genres_original/ or change the path in the notebooks accordingly.

---

🧰 Feature Extraction

All audio files are converted to Mel Spectrograms using librosa:

S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128)
S_DB = librosa.power_to_db(S, ref=np.max)

• Shape standardized to (128, 660) for CNN input.
• Skips any problematic or corrupted files with error handling.

---

🏗️ Model Architectures

1. Basic CNN (musicgenre.ipynb)

• Simple 3-layer CNN
• Uses mel spectrograms directly
• Accuracy ~75%

Conv2D → MaxPool → Dropout → Flatten → Dense

---

2. Improved CNN (improvedmusicgenre.ipynb)

• 4 Convolutional blocks with BatchNorm and Dropout
• Data augmentation using noise and pitch shifts
• Regularization (L2) on dense layers
• Callbacks: EarlyStopping, ReduceLROnPlateau, ModelCheckpoint

Conv2D → BN → Conv2D → MaxPool → Dropout (×4)
→ Flatten → Dense (L2) → Dropout → Output

---

3. Colab Version (Colab_version.ipynb)

• Final cleaned and modular version
• Wrapped in functions like load_dataset_enhanced(), train_model_enhanced(), etc.
• Training history and confusion matrix visualizations

---

🏃 Running the Project

1. Extract Features and Load Data:

X, y, label_encoder = load_dataset_enhanced("C:/Datasets/Data/genres_original")

2. Train/Test Split:

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

3. Train Model:

model, history = train_model_enhanced(X_train, y_train, X_test, y_test, label_encoder)

4. Evaluate Model:

evaluate_model(model, X_test, y_test, label_encoder, history)

---

📊 Results

• Accuracy: Up to 92% with the enhanced model.
• Precision and Validation Accuracy monitored during training.
• Confusion matrix and learning curves plotted using matplotlib and seaborn.

---

📈 Visualizations

• Training vs Validation Accuracy
• Training vs Validation Loss
• Confusion Matrix
• Classification Report

---

🧪 Enhancements Implemented

• 🧹 Dataset cleansing and error handling
• 📈 Stratified train-test splitting
• 📊 Visualizations and evaluation metrics
• 📦 Model checkpoints and adaptive learning
• 🔄 Augmented dataset for better generalization

---

🔮 Future Improvements

• Add real-time inference with microphone input
• Use pretrained models (e.g., VGGish or YAMNet)
• Implement genre probability ranking
• Deploy using Streamlit or Flask for interactive web app

---

📝 Acknowledgements

• GTZAN Genre Dataset
• LibROSA
• TensorFlow, Keras, NumPy, Seaborn, Matplotlib

---

📚 References

• Music Genre Classification with CNN
• LibROSA Documentation
• TensorFlow Keras API