Shashank Madala

I research speech-based measurement of autism symptom severity at the Yale Child Study Center, working with a longitudinal clinical corpus of 197 children. My primary project addresses a measurement problem rather than an accuracy one: the published open-source severity model summarizes each recording as an ordered set of randomly drawn groups of vocalizations, so the same recording can receive a different score depending only on which groups are drawn and the order they are fed in. I quantified that nuisance variance and designed an order-invariant architecture that removes the ordering artifact by construction. It uses about 3,300 parameters against the released model's 1.2 million, at no detectable cost in accuracy, and returns a single score per recording that does not depend on input order. I am first author on the resulting paper, which was accepted as a technical paper at the 2026 IEEE MIT Undergraduate Research Technology Conference and selected for presentation at MIT in October and publication in the IEEE Xplore digital library.

Alongside this I contribute to the lab's parent–child dyadic synchrony work, analyzing acoustic and behavioral alignment during recorded interactions, and I support preprocessing and annotation workflows on the longitudinal interaction data.