Aurosweta
Mahapatra

I am a PhD student at the SMILE Lab, CLSP, Johns Hopkins University, advised by Dr. Berrak Sisman, working at the intersection of speech security, synthesis, and healthcare. My research focuses on a critical gap in modern speech security: current speech deepfake detectors often fail against emotional and expressive synthetic speech because they rely on dataset-specific artifacts rather than true speech structure. In practice, modern text-to-speech and voice conversion systems generate speech that is highly intelligible, speaker-consistent, and context-aware, making them effective for spoofed calls, impersonation, misinformation, and biometric deception, while existing detection models remain vulnerable to these attacks. I address this by designing emotion-aware and prosody-driven models inspired by human perception, which learn more generalizable representations instead of relying solely on classification objectives. Beyond speech security, I also work on speech for healthcare, including depression and Alzheimer's detection, as well as speech synthesis. Before joining Hopkins, I was a research assistant at UT Dallas, where I conducted an in-depth literature review on speech security. I completed my master's at UCLA, working on automatic speech recognition for child speech at SPAPL under the guidance of Dr. Abeer Alwan. I am broadly interested in building reliable and trustworthy speech systems that enable the safe and ethical use of AI in real-world applications.