WPI - Computer Science Department , MS Thesis Presentation, Abigail Rachel Albuquerque, ""Detecting Intoxication from Speech using Transformer-based Self-Supervised Learning with Pre-Training "

Abigail Rachel Albuquerque 

MS Student 

WPI – Computer Science  

Friday, December 6, 2024 

Time: 1:00 PM – 2:00 PM 

Location: Campus Center the Chair’s Conference room  

Zoom: https://wpi.zoom.us/j/8328872674 

Advisor: Prof. Emmanuel Agu 

Reader: Prof. Fabricio Murai 

Abstract: 

Alcohol inebriation is a major global cause of fatalities, contributing to over 30% of vehicular accidents. Current methods to prevent Driving Under the Influence (DUI) are often expensive or require external devices. Intoxication detection from speech offers an alternative through continuous monitoring and non-invasive analyses methods. This thesis explores Wav2Vec 2.0, a Transformer-based self-supervised model, for classifying alcohol intoxication from raw audio.  

Challenges addressed by this study include severe class imbalance and inadequate data, which were addressed using resampling techniques, and model fine-tuning after pre-training on larger and more diverse datasets. During rigorous evaluation, the proposed model achieved an Unweighted Average Recall of 73.3%, outperforming state-of-the-art baseline models, highlighting its potential for accurate DUI detection to prevent alcohol-related incidents.