FINAL PROJECT

Authors: Dipto Biswas (1075562), Ivan Odiel Magtangob (1092635)
Last Updated: December 8th, 2023

FILES:

• Notebook (.ipynb) files for each evaluated model
  • Naive Bayes
  • K-Nearest Neighbours
  • Decision Tree
  • Logistic Regression
  • Stochastic Gradient Descent
  • Support Vector Machine
• audio_embeddings.pt (torch file storing processed audio embeddings)
• README.md

RUNNING

Each model is self-contained in its own Jupyter Notebook file, where dependencies are automatically checked and downloaded.

The Naive Bayes classifier (naivebayes.ipynb) was created first and is responsible for processing the audio dataset into embeddings. These embeddings are then saved locally (audio_embeddings.pt) for other classifiers to access. If audio_embeddings.pt is missing, run Naive Bayes before any other classifier to create it.

Execute each Notebook file by pressing 'Run All' in each Notebook's upper toolbar.

DATASET

The dataset used for this project from is from Huggingface by TrainingDataPro, which contains 80 .wav files (<10s) of conversation of 4 classes: ['euphoric', 'joyfully', 'sad', 'surprised'].

RESULTS SUMMARY

After evaluation, the following mean testing accuracies were obtained:

• Naive Bayes: 30.00%
• K-Nearest Neighbours: 22.50%
• Decision Tree: 20.00%
• Logistic Regression: 28.75%
• Stochastic Gradient Descent: 28.75%
• Support Vector Machine: 27.50%