Master’s Thesis (2020)

Author: Aron Mandrella

About

The aim of the thesis was to find an artificial neural network that would provide high accuracy when used for classification of percussive sounds. Both regular neural networks and convolutional neural networks were tested. Tests were carried out with few various audio representation (various inputs), and with various model training approaches (dropout, batch normalization, stride, max-pool).

• 📜 Master’s Thesis PDF

• 📜 Master’s Thesis PDF (longer version)

• 📦 Output predictions (Google Drive)

🧰 Used tools:

• Python
• TensorFlow 2, Librosa, Matplotlib, NumPy, Pandas, sklearn
• Spyder IDE

🎓 Knowledge I’ve gained or improved:

• Data normalization
• Gradient descent algorithms (GD, SGD, ADAM, RMSProp, parameters meaning)
• Various methods of preventing overfitting (dropout, batch normalization)
• Methods of sound analysis and sound representation (Fourier transform, spectrogram, constant Q transform, cosine transform, mel-cepstral coefficients)
• Model validation techniques (cross-validation, confusion matrix, classification accuracy, etc.)
• Academic methods of statistical analysis of collected data (e.g. box plots, plots, t-sne)