This project focuses on classifying music/audio files using deep learning techniques.
It extracts audio features and predicts the music category/genre based on trained models.
- Audio feature extraction
- Music / genre classification
- Deep learning based prediction
- Simple and reusable code structure
- Python
- NumPy
- Pandas
- Librosa
- Scikit-learn
- Matplotlib (for visualization)
- streamlit
The deep learning model was trained for 100 epochs and evaluated on a separate test dataset. The training and validation metrics indicate strong learning and good generalization.
🔹 Final Training Results
Training Accuracy: 95.97%
Training Loss: 0.1731
🔹 Validation Results
Validation Accuracy: 88.61%
Validation Loss: 0.3781
🔹 Test Results
Test Accuracy: 90.38%
Test Loss: 0.3003
The model achieved over 96% training accuracy, showing effective feature learning.
Validation accuracy remained close to training accuracy, indicating limited overfitting.
A test accuracy of 90.38% confirms that the model generalizes well to unseen audio samples.
Minor fluctuations in validation accuracy near the final epochs suggest the model is close to optimal convergence.
A dual-branch deep learning model was developed for music genre classification by combining raw waveform and Mel spectrogram features. The waveform-based 1D CNN captures temporal characteristics, while the Mel spectrogram–based 2D CNN learns spectral information. Regularization techniques such as batch normalization, dropout, and L2 penalties help stabilize training and reduce overfitting. The model achieves strong classification performance, demonstrating the effectiveness of jointly learning temporal and spectral audio features, with scope for further improvement through data augmentation and hyperparameter tuning.