Music Genre Classification App

A research-based full-stack web application that classifies music genres using both machine learning and deep learning models. The project compares traditional algorithms like SVM and k-NN with CNN and CNN+BiLSTM architectures on two benchmark datasets: GTZAN and FMA-Small.

---

Highlights

• Dual-model evaluation: ML vs DL for music genre classification
• Deep Learning Models: CNN and hybrid CNN + BiLSTM with SE block and Attention
• High Accuracy: Up to 93% classification accuracy on FMA-Small
• Feature Extraction: MFCCs and Mel Spectrograms
• User Interface: Django-based frontend with registration, login, and history tracking
• REST API: For real-time genre predictions

---

Datasets

• GTZAN Dataset: 1000 tracks across 10 genres
• FMA-Small: 8000 30-second high-quality audio clips
• All data is preprocessed into .wav format and converted to 128x128 Mel spectrograms.

---

Technologies Used

Backend

• Scikit-learn
• TensorFlow / Keras

Frontend

• HTML/CSS/Bootstrap (with Django templates)
• JavaScript

ML & DL Models

• Logistic Regression, k-NN, SVM
• CNN
• CNN + BiLSTM + Attention + SE Block

---

Parameters

Input Representation

• Audio Length: 30 seconds
• Sampling Rate: 44,100 Hz
• Feature Type: Mel-spectrogram
• Mel-Spectrogram Shape: 128 (mel bands) × 128 (time frames)
• Color Channel: Grayscale (1 channel)
• Final Input Shape: (128, 128, 1)

---

CNN Model (FMA-Small)

• Total Layers: 5 (Convolution + Pooling + Dense)
• Dropout Rate: 45%
• Activation Functions: ReLU (hidden), Softmax (output)
• Loss Function: Categorical Crossentropy
• Optimizer: Adam
• Output Classes: 8 genres
• Best Accuracy: 92%

---

CNN + BiLSTM Model (GTZAN & FMA-Small)

• CNN Layers: 4 blocks with increasing filters (32 → 256)
• LSTM: Bidirectional LSTM with 128 units
• Attention: Applied after BiLSTM
• Dropout Rate: 50%
• Output Classes:
  • GTZAN: 10 genres
  • FMA-Small: 8 genres
• Best Accuracy:
  • GTZAN: 91%
  • FMA-Small: 93%

---

Classical Machine Learning Models

• Input Features: MFCC vectors or flattened spectrograms
• Best Hyperparameters:
  • SVM: RBF kernel, gamma = scale
  • k-NN: 5 neighbors, Manhattan distance, distance-based weight
  • Logistic Regression: C=10, L2 penalty, saga solver
• Output Classes: 8 or 10 (based on dataset)
• Accuracy Range: 57%–76%

Experimental Results & Performance

The performance of the models was evaluated through a rational analysis using Precision, Recall, and F1-Score metrics.

Table II: GTZAN Dataset Results

The comparative analysis on the GTZAN dataset demonstrates the superior stabilization and performance of deep learning architectures.

Model Category	Model	Precision	Recall	F1 Score
Traditional ML	k-NN	0.74	0.74	0.74
	SVM	0.74	0.75	0.74
	Logistic Regression	0.75	0.76	0.75
Deep Learning	CNN	0.91	0.91	0.91
	CNN + BiLSTM	0.91	0.91	0.91

_{Figure 1: Confusion Matrix of CNN + BiLSTM model trained on GTZAN Dataset (10 genres).}

Table III: FMA-Small Dataset Results

On the FMA-Small dataset, the CNN + BiLSTM hybrid architecture proved to be the most effective model for capturing complex data patterns.

Model Category	Model	Precision	Recall	F1 Score
Traditional ML	k-NN	0.51	0.49	0.48
	SVM	0.62	0.61	0.61
	Logistic Regression	0.57	0.57	0.57
Deep Learning	CNN	0.92	0.92	0.92
	CNN + BiLSTM	0.93	0.93	0.93

_{Figure 2: Confusion Matrix of CNN + BiLSTM model trained on FMA-Small Dataset (8 genres).}

Training Settings

Dataset Processing

• Datasets Used:
  • GTZAN (10 genres × 100 tracks)
  • FMA-Small (8 genres × 8000 samples)
• Preprocessing Steps:
  • Convert audio to .wav
  • Normalize audio levels
  • Extract 128×128 Mel-spectrograms
  • Store as .npy or tensor images
• Data Augmentation (for deep models):
  • Time shifting
  • Pitch shifting
  • Additive Gaussian noise

---

Training Configuration

Setting	Value
Epochs (CNN)	50
Epochs (CNN + BiLSTM)	50
Batch Size	32
Optimizer	Adam
Learning Rate	0.001
Loss Function	Categorical Crossentropy
Evaluation Metrics	Accuracy, Precision, Recall, F1 Score
Validation Split	20% (Stratified)
Learning Rate Scheduler	ReduceLROnPlateau (patience=3, factor=0.5)
Model Saving	ModelCheckpoint (save_best_only=True)
Early Stopping	EarlyStopping (patience=7, restore_best_weights=True)

---

Model Saving & Evaluation

• Saved Format: .h5 (Keras HDF5 format)
• Model Selection: Automatically saved best model during training
• Callback Tools Used:
  • EarlyStopping for preventing overfitting
  • ModelCheckpoint for best-model saving
  • ReduceLROnPlateau for adaptive learning rate
• Final Evaluation: Performed on held-out test set
• Visualization: Confusion matrix plotted for all experiments