Research profiles
Where to verify the publication record
For the cleanest public record, use DBLP first. I will add direct publisher links for accepted work once the final publication pages go live.
Research output
Research that grew out of signal processing and stayed connected to systems.
These papers come from the part of my work that started in speech and signal processing. They also mark the beginning of how I started thinking more seriously about machine learning and systems.
Conference paper
In-Domain Data Augmentation to Enhance Severity Level Classification of Dysarthria from Speech
2024
Presented at IEEE SPCOM 2024. This paper studies dysarthria severity classification in a low-resource setting, where collecting enough labeled speech is difficult and off-the-shelf augmentation choices are not always a good fit.
We explored speaking-rate, pitch, formant, and vocal-tract length perturbation, then looked at how those choices changed classification performance. The combined augmentation setup gave a 42.86% relative improvement over the baseline and became the clearest early example of how I like to work on constrained ML problems.
Low-resource dysarthric speech classification with a 42.86% relative improvement over the baseline.
Journal article · 2026 acceptance
Fusion of data augmentation for improved dysarthria severity classification
Apr 2026
Accepted for publication in the Journal of Signal Processing Systems. This paper extends the earlier SPCOM work by studying how different augmentation methods work in combination rather than in isolation for dysarthria severity classification.
The study evaluates fusion strategies across the TORGO and UASpeech datasets in both speaker-dependent and speaker-independent settings. The strongest results showed relative improvements of 42.86% and 12.22% on TORGO, and 18.78% and 5.29% on UASpeech, depending on the evaluation setting.
A follow-on journal paper that turns one augmentation idea into a broader comparative study across two dysarthric speech datasets.