travauxderecherche2

From 2003 to 2008, my research work has been in the topic of automatic speech recognition.

During my Final training course of Engineer studies in the spring 2003, in the LORIA, I worked on automatic speaker segmentation and clustering. It consists in detecting how many speakers are talking in an audio recording, and when each of them is speaking. This is a preprocessing for speech recognition that allows specific speaker adaptation, and thus leads to better results.

My master degree training course has also been spent in the LORIA, and had the same topic as the upper. I implemented and tested a method that has been developed by Sylvain Meigner in his PhD thesis.

For my PhD thesis, it is hosted by the LORIA and is partially funded by the European project HIWIRE. The main topic of my PhD is the automatic recognition of non-native speech. That is, speech that is uttered in a target language that is different from the mother tong of the speakers. The main aim is to enhance the performances of ASR systems against such non-native accents. For that matter, we have developed a new approach that takes advantage of the knowledge about the origins of speakers. We have also developed a method for the detection of the mother tong of non-native speakers. The latter could be a preprocessing in order to select a well suited non-native enhanced ASR, according to the detected mother tong.

Lately, I have been collaborating with Mr Michael Cai on "fast likelihood computation". We have tested several methods such as "Feature Component Reordering (FCR)", "Best Mixture Prediction (BMP)", "Gaussian Selection", "Single Instruction Multiple Data (SIMD)". We have developed a new technique called "Dynamic Gaussian Selection (DGS, or Extended BMP)". The results of our research will be published soon.