Research papers and contributions
Submitted to Interspeech 2026
We propose a content-agnostic deepfake speech detection model trained on a curated dataset of 1M+ real-fake speech pairs. By using transcripts from real speech to generate synthetic audio using cutting-edge TTS systems (YourTTS, XTTS-v2, OpenVoice), we eliminate linguistic cues and force the model to rely solely on acoustic artifacts. Our model achieves superior performance compared to prior work and outperforms human listeners across all speech lengths.
University of Michigan — ongoing
Joint-embedding predictive architectures predict in latent space rather than pixel space, which makes the anti-collapse regularizer the central design decision. We evaluate the LeWM world model with free-energy and coding-rate regularizers as replacements for SIGReg, improving hard-protocol (100-step planning) accuracy on a two-room navigation dataset from 16% to 75% with the coding-rate variant.
A free-energy (Gaussianizing) loss that improves reconstruction and classification across CNNs, MLPs, and autoencoders, together with a joint Gaussianizing classifier. Released in both Python and Julia.
I'm looking for full-time robotics software, perception, and ML roles starting after I graduate in December 2026, and I'm always glad to talk about world models, imitation learning, and foundation models for robotics. Feel free to reach out!
Get in Touch