I'm Theo Lepage. I hold a Ph.D. in Artificial Intelligence from 
Sorbonne Université and my research focuses on AI/ML for Speech, notably speaker recognition and speech anti-spoofing, with a broader interest in self-supervised learning. Through my work, I aim to develop technologies capable of interpreting all aspects of human conversations.
Research
My Ph.D. thesis focused on self-supervised speaker verification, with the objective of reaching supervised performance while reducing reliance on labeled data. The main contributions were SSPS → a novel latent-space positive sampling (-58% EER for SimCLR), and DINO-WavLM → an iterative pseudo-labeling approach to leverage WavLM without speaker labels (SOTA at 1.06% EER on VoxCeleb1-O). This research is open-sourced as sslsv, a PyTorch toolkit for self-supervised speaker verification.
I am also interested in speech anti-spoofing and deepfake detection. Our WavLM-based countermeasure, submitted at the ASVspoof 2024 challenge, placed 3rd (Track 1, Open).
Experience
- Developed self-supervised models for speaker verification, with DINO-WavLM achieving SOTA on VoxCeleb (1.06% EER on Vox1-O)
- Proposed SSPS, a latent-space positive sampling for SSL frameworks, mitigating extrinsic variability (-58% EER for SimCLR)
- Built and maintained sslsv, an open-source PyTorch toolkit for training and evaluating self-supervised speaker models
- Published 8 papers in leading speech venues including Interspeech, IEEE TASLP, and Speech Communication
- Developed deep learning models (CNN with self-attention) for end-to-end MR image enhancement (denoising & super-resolution)
- Contributed to Holovibes, real-time digital holography software for retinal blood flow analysis → 20× input throughput (10,000 FPS)
Education
Thesis: Self-Supervised Learning for Speaker Recognition • Supervised by Reda Dehak and Thierry Géraud @ LRE-EPITA
Major: AI/ML for Computer Vision • Research student • Teaching assistant (C, Unix) • International section • Exchange semester @ CSUMB
