Mohammad Omar Khursheed

MATS Research Scholar

I grew up in Saudi Arabia and studied computer science at Aligarh Muslim University, India. I completed my MS at UMass Amherst with Mohit Iyyer and Brendan O'Connor, studying gender bias in narrative tropes. I spent five years as an Applied Scientist at Amazon building wakeword models for Alexa, shipping efficient speech systems to hundreds of millions of devices. I'm currently on leave to pursue research interests in AI safety as a scholar at MATS.

Research Interests

Speech & Efficient ML

Efficient Deep Learning Architectures
Speech Recognition & Keyword Spotting
Self-supervised Representations
Model Compression

AI Safety

Computational Social Science
Model Behavior Analysis
Mechanistic Interpretability

Publications

Small Footprint Slimmable Networks for Keyword Spotting

Akhtar, Khursheed, Du, Liu · IEEE ICASSP

Latency Control for Keyword Spotting

Jose, Wang, Strimel, Khursheed, et al. · INTERSPEECH

Tiny-CRNN: Streaming Wakeword Detection

Khursheed, Jose, Kumar, et al. · IEEE ASRU

Analyzing Gender Bias within Narrative Tropes

Gala*, Khursheed*, Lerner, O'Connor, Iyyer · EMNLP Workshop

Projects

Hansard NLP Explorer

GitHub

1.2M debates 6.0M speeches 200+ years

Socio-technical analysis of the UK Parliament's Hansard Debate Corpus (1803-2005).

Key Findings

92.9% accuracy classifying 2,808 suffrage speeches via LLM
5,138 arguments extracted across 9 taxonomy categories
95% drop in suffrage speeches 1913-1915 (WWI impact)

Writing

A GSM8K Study Comparing Pseudo Labeling for SFT and GRPO

GRPO on 1K labeled matches SFT on 5K real labels. Pseudo-labels don't help.