Personal homepage

I am a scientist researcher at CNRS, working on automatic speech recognition.

I belong to the PAROLE research team in the LORIA UMR 7503 French laboratory.
My detailed CV can be accessed on the left column.


Talking to computers is an old dream, but despite all the efforts that have been spent on this research area, nowadays achievements are extremely far from what was expected.
We can find today speech recognition systems mainly in mobile phones, for example to find a phone number by saying the name of someone, or in automatic call routing applications, or yet in dictation systems.
However, all these applications suffer from serious drawbacks: need to train the system for a long time, lack of robustness to noise, very limited domain, small vocabulary, ...
... And above all, the main limitation of the computer is that it simply can not associate any meaning to the speech sounds it tries to transcribe, and we now know that higher-level information such as semantic or syntactic is important to better transcribe, to choose between several alternative words that have similar phonemes: this is something that we are doing (unconsciously) all the time !

I have worked for a long time on improving robustness of speech recognition systems to noise by adapting acoustic models to noise,  or by isolating the noise contribution in narrow frequency bands, for example in the multi-band architecture or with missing data recognition.

Another research area on which I have worked for several years is to study and design speech and multimodal interaction systems for ambiant intelligence platforms, also known as ubiquitous or pervasive computing. In such a context, speech interactions are rarely direct human-machine interactions, but are generally implicit (or transparent) , i.e., are derived from observations of the user.
I have particularly worked on the concept of implicit speech interactions, which consists in computing context-related information from the speech input and the conversations between users. Autonomy and permanency requirements of ubiquitous platforms also impact the way we interact with computers.

More recently, my interest focus on integrating linguistic and para-linguistic information into speech recognizers: I am convinced this is the most promising way to improve speech recognition and to go beyond recognition, towards speech understanding.

Contact

cerisara@loria.fr
03 54 95 86 25
LORIA, tranche C
Campus Scientifique - BP 239 - 54506 Vandoeuvre-les-Nancy Cedex