Research Interests

My current research interests include functional data analysis, nonlinear regression, kernel methods and sequential learning, with applications in prediction tasks. Functional data arise in a number of scientific fields associated with continuous-time monitoring process whose final outputs are curves (fonctions). Hilbert spaces theory is a well-motivated way to model nonlinear regression for functional responses. In this work, I am interested in using kernel methods to deal with efficient estimation for various nonparametric models. Further, I am working on using sequential analysis tools for estimating regression function in the general case of multiple functional predictors and responses.

My past research interests aim to investigate and develop efficient and robust methods to automatically answer questions about audio content like:
        • Did an audio file contain speech, music or other audio entities?
        • How many speakers are contained in a speech segment?
        • What gender they are?
        • Which persons are speaking?
Answering such questions has become recently more necessary because of the amassing of large volumes of audio, including broadcasts, voice mails, meetings and other spoken documents. I am interested in developing and applying signal processing and machine learning techniques across the wide range of audio signals commonly encountered in daily life. This includes extracting many kinds of information from speech signals, music recordings, and environmental sounds.

Sequential Functional Regression

Within the many Machine Learning problems, that of regression consists in learning a function from data. More precisely, a regression algorithm outputs a function f such that f(x) ≈ y when x and y are respectively covariates and responses. Many algorithms have been proposed for cases where x is a numeric space and y is a numeric a one-dimensional space. Some approaches have benn proposed for cases where y hs dimension 2 or 3. Here, we propose to adress the very general case where x is any metric space, possibly infinte(in the latter case, it is a space of functions). Moreover, we consider the case where data set is made step by step: it is thus necessary to learn a new f for each new observed pair (x,y). Of course, this must be done sequentially without training from scratch at each iteration. Sequential functional regression has many applications, ranging from estimating and prediction of workload curves, growth curves, or ever 2D functions such as rainfall maps.

Audio Speaker Segmentation

Speaker-based segmentation can be defined as splitting and labeling a spoken audio stream associated with an unknown number of unknown speakers into homogeneous regions according to speaker identity. Speaker segmentation is highly associated with the traditional speaker recognition but it is more difficult. In general, in a speaker recognition system, speaker models are usually well trained. But in real-time speaker segmentation system, there is no prior knowledge on speakers (speaker identities and the number of speakers). Thus, no data can be achieved to train appropriate models for speakers a priori. My thesis work looked at unsupervised methods for the detection of potential speaker change points in a speech stream in real time and to segment the stream into homogeneous speaker clips. I am interesting in realizing this task using the hybrid approach which is based on two-pass algorithms where in a first pass many change points are suggested (more than there actually are) and in a second pass such changes are reevaluated and some are discarded. I am also interested in the use of optimized One-Class Support Vector Machine (1-SVM) for novelty detection to increase robustness in speaker segmentation (simultaneous speeches, short speaker changes, environmental noise, etc).

Impulsive Sound Classification

I am interested in the segmentation and the classification of various general sounds (such as glass breaks, explosions, phone rings, machines, etc), with immediate application to surveillance and security. In general, the purpose of sounds (events) recognition is to understand whether a particular sound belongs to a certain class. This is a recognition problem, similar to voice, speaker or speech recognition. Sound recognition systems can be partitioned into two main modules. First, a detection stage isolates relevant sound segments from the background by detection abrupt changes in the audio stream. Then, a classifier try to assign the detected sound to a category.I am interested in using and illustrating the potential of SVMs on detecting and recognizing impulsive audio signals belonging to a complex real-world dataset.