Hello! I am a second-year PhD student at the University of Edinburgh researching structured world representation in brains and vision-language models (VLMs). I am supervised by Benjamin Peters and Frank Keller. My work is funded by the UKRI AI CDT in Designing Responsible Natural Language Processing. Previously, I completed a BSc. (Hons) in Computer Science & Cognitive Science at the University of Toronto as a Lester B. Pearson International scholar.
In my spare time, I also like to make art and music.
Affiliations: Institute for Language, Cognition & Computation, EdinburghNLP (previously): CL&NLP group, CoNSens Lab
While the capabilities of frontier AI systems have improved rapidly, our understanding of the mechanisms underlying their behaviour remains limited; particularly, how such mechanisms align or diverge from those in humans. The construction of structured representations of the world from noisy sensory observations is a fundamental aspect of cognition, determining how an agent perceives, plans and reasons in its environment. I am interested in understanding how humans and AI systems construct such representations, and how structure within these representations can be characterised and comparitively analysed across modalities.
The predominant method for scoring the quality of automatic speech recognition (ASR) transcripts when ground-truth labels are not available is to predict the word error rate (WER) from the corresponding audio segment. We propose WAV2LEV, a novel paradigm for WER estimation which predicts the underlying sequences of Levenshtein edit operations (substitutions, deletions, insertions and matches) from which the WER can be computed. This approach offers more fine-grained token-level error estimation in comparison to previous work without compromising on performance for WER estimation. To support this investigation, we present Mini-CNoiSY (Miniature Clean-Noisy Speech from YouTube), a bespoke 354-hour noisy speech corpus which ensures confidence in ground-truth labeling and captures a diverse range of noise artifacts which degrade ASR performance. Our results show that WAV2LEV achieves near state-of-the-art performance for the task of WER estimation with a root mean square error (RMSE) of 0.1488 and a Pearson correlation coefficient (PCC) of 89.71%, while generating predictions of ASR error that are more informative and fine-grained than that of direct WER estimators.