Interests
My main interest concerns applied machine/deep learning to several types of data, especially speech signal and natural language. I’m however also considering other types of time series data, such as system logs to detect system failure for predictive maintenance. The most important aspect that I want to improve in these models is weakly supervised learning, i.e., how to go beyond the traditional supervised/unsupervised dichotomy to better exploit unlabeled raw data and transfer relevant information to our target tasks.
I’m also deeply concerned about ethics in machine learning and AI, but not necessarily in the standard sense that is given to the word ethics: of course, I strongly support experimental ethics and I’m insisting on the difficulty to perform fairplay experimental validation and on reproducible experiments in all of my courses; but I’m also thinking about means to protect citizen’s privacy thanks to physical anonymity.
Some achievements
| Topic | Achievement |
|---|---|
| Sentiment recognition | First transfer model between sentiment and dialog act, which handles sentiments in the contexts of dialogs in social media |
| Machine learning | Formal derivation of an approximation of the true classifier risk to train a model in an unsupervised way |
| Speech alignment | Fast and efficient aligner JTrans |
| Speech recognition | Robust ASR, Multi-band ASR, hybrid neural-HMM ASR |
Courses
- I’m giving the Machine Learning course of at Ecole des Mines in Nancy.
Software
- JTrans: fast text-to-speech alignment for up to 2 hours files (for French only)
Projects
- LUE IMPACT OLKi
- ANR LEAUDS
- CPER LCHN
- CPER MISN
- ANR ORFEO
- ANR ContNomina
- EQUIPEX ORTOLANG
- INRIA ARC Rhapsodys
- FP7 Amigo
- FP6 OZONE
- FP6 MIAMM
Current and past Ph.D. students
- Guillaume Le Berre
- Hoa Thien Le
- Christian Gillot
- Sébastien Demange
- Pavel Kral