msc speech & language processing · university of edinburgh · 2025–2026
At Edinburgh, my work is focused on speech, language, and machine learning for conversational systems. The MSc has covered speech processing, NLP, machine learning, computer vision, and the mathematical foundations that connect them.
Before starting the MSc, I spent a year working for dataannotation.tech, training and interacting with frontier models across difficult coding, mathematical, and scientific reasoning tasks, and encountering many programming languages in the process. In my spare time I delved deeply into calculus, linear algebra, and probability, and pursued coding projects.
My dissertation is supervised by Sarenne Wallbridge and co-supervised by Catherine Lai. It is on turn-taking prediction in conversational AI. Current dialogue systems often rely on silence thresholds to decide when a speaker has finished. Human conversation uses richer cues: prosody, syntax, timing, breath, and interactional context.
The project uses sparse autoencoders to study hidden representations in dialogue models and identify features associated with turn boundaries. The key question is whether those features are causally involved in the model’s predictions. I am using activation patching to test whether intervening on candidate features can change turn-boundary predictions, rather than treating the features as correlational explanations only.