Welcome to my page!
This site contains information regarding my research and some personal projects. Here’s a subset of my research interests:
- Machine learning
- Speech processing
- Disentangled representations
- Multiple instance learning
- Computer vision
I act as principal research scientist at Ubisoft in the La Forge lab. I work there since 2017. I lead a group of resarchers applying the latest techniques in machine learning, speech, signal processing, computer vision & graphics, animation to video games.
Before that, as a PhD student, I was affiliated with two labs:
|Dec 22, 2023
We are excited to share our recent work on monocular 3D face reconstruction. We introduce MoSAR, a new method that turns a portrait image into a realistic 3D avatar.
From a single image, MoSAR estimates a detailed mesh and texture maps at 4K resolution, capturing pore-level details. This avatar can be rendered from any viewpoint and under different lighting condition.
We are also releasing a new dataset called FFHQ-UV-Intrinsics. This is the first dataset that offer rich intrinsic face attributes (diffuse, specular, ambient occlusion and translucency) at high resolution for 10K subjects.
Check out the project page!
|Oct 27, 2023
Our paper EDMSound: Spectrogram Based Diffusion Models for Efficient and High-Quality Audio Synthesis has been accepted for presentation at the NeurIPS Workshop on ML for Audio. This work has been done in collaboration with colleagues from Rochester University.
In this paper, we propose a diffusion-based generative model in spectrogram domain under the framework of elucidated diffusion models (EDM). We also revealed a potential concern regarding diffusion based audio generation models that they tend to generate duplication of the training data.
Check out the project page!
|Sep 21, 2023
Our paper “Rhythm Modeling for Voice Conversion” has been published in IEEE Signal Processing Letters. We also released it on Arxiv.
In this paper we model the natural rhythm of speakers to perform conversion while respecting the target speaker’s natural rhythm. We do more than approximating the global speech rate, we model duration for sonorants, obstruents, and silences.
Check out the demo page!
|Jul 15, 2023
Ubisoft had published a blog page describing our system for gesture generation conditioned on speech.
This system was presented in “ZeroEGGS: Zero-shot Example-based Gesture Generation from Speech” and showcased on 2 minute papers.
MoSAR: Monocular Semi-Supervised Model For Avatar Reconstruction Using Differentiable ShadingArXiv, 2023
EDMSound: Spectrogram Based Diffusion Models for Efficient and High-Quality Audio SynthesisIn NeurIPS Workshop: Machine Learning for Audio, 2023
Measuring Disentanglement: A Review of MetricsIEEE Transactions on Neural Networks and Learning Systems, 2022
Rhythm Modeling for Voice ConversionIEEE Signal Processing Letters, 2023
A Comparaison of Discrete and Soft Speech Units for Improved Voice ConversionIn ICASSP, 2022
Daft-Exprt: Robust Prosody Transfer Across Speakers for Expressive Speech SynthesisIn INTERSPEECH, 2022
Multiple instance learning: A survey of problem characteristics and applicationsPattern Recognition, 2018
ZeroEGGS: Zero-shot Example-based Gesture Generation from SpeechComputer Graphics Forum, 2023