Researching Drug Addiction Through Naturally Generated Speech and Its Connection to Brain Activity

Abstract: This thesis is a collaboration with an addiction-research neuroscience lab at Mount Sinai, part of our group's broader effort to apply NLP to scientific discovery. Participants with heroin use disorder (HUD) and matched healthy controls watched the same film inside an fMRI scanner and later, in a separate session, freely described what they remembered — naturally generated speech. I ask two questions. First, what signals can NLP extract from the speech alone, and do they align with the leading theoretical model of addiction (iRISA)? Second, can spontaneous speech — a cheap, easy-to-collect modality — serve as a window into an individual's addiction state, examined directly through its relationship to the brain activity recorded during the same stimulus? Two methodological challenges shape the work. Keeping results interpretable and statistically defensible in an era where most NLP pipelines are black boxes — a non-negotiable for findings consumed by the scientific community. And aligning two signals that were never recorded together: linking speech to brain activity required a dedicated cross-modal alignment approach rather than off-the-shelf multimodal tooling. The results show that NLP analyses of free speech recover theory-consistent signatures of addiction and track individual variation in brain responses to drug-related content, establishing speech as a scalable signal for mental-health research