Vahid is a graduated PhD student who has received his PhD degree in computer science on 16th January 2013. He has conducted his doctoral studies at the French national institute of research on computer sciences (INRIA), inside GEOSTAT team under the helpful supervision of Dr. K. Daoudi. He holds an Electrical Engineering M.Sc. degree majored in communication systems and digital signal processing (Iran Univ. of Science and Tech., 2008). He also has two years of experience as research engineer working on research, development and implementation of hardware efficient signal processing algorithms.
Major contributions and research
The thematic of his PhD thesis, titled as "Speech analysis using advanced non-linear multi-scale methods", concerns the investigation of novel non-linear multi-scale methods (the MMF) for the analysis of speech signal. The idea is to relate the non-linear character of speech to the concept of predictability of complex systems. By evaluating the limiting power-law scaling behaviour of a multiscale functional and estimation of the associated scaling exponent, a geometric local notion of predictability is defined (regardless of traditional Linear Predictive analysis), which allows geometric characterization of information content and dynamical transitions of the signal in terms of scales. The definition of the multi-scale function is of highest importance in this context. It can be a wavelet, or several other gradient based measures, whose appropriate definition for speech analysis has been the heart of discussions during this thesis. Once the exponent is defined and precisely estimated, the notion of the most informative (the less predictive) subset of points can be used to classify speech samples according to their information content. Initially, extensive experiments over large databases were performed to evaluate the validity of the aforementioned measurements. Consequently the methodology found to show outstanding performances in several applications such as phonetic segmentation, Irregularly spaced sampling (waveform coder), and multi-pulse source approximation (inside a parametric AbS coder). Recently, Vahid has developped an efficient algorithm for Glottal Closure Instant (GCI) detection and has applied it to the problem of sparse linear prediction analysis. A closed-form formula is derived for estimation of Linear Prediction coefficients such that the resulting residuals are as sparse as possible. The proposed method is much more efficient than the usual approach of recasting the norm-1 minimization problem into a linear program and using convex optimization methods. Currently, he is working on the development of a unified coding framework using the MMF.
His M.Sc thesis dealt with Blind Source Separation (BSS) of speech mixtures in reverberant environments. He developed a novel Convolutive BSS technique, which has the byproduct of speaker localization in reverberant environment. For making the latter possible, he proposed a method for solving the global permutation ambiguity of the CBSS by using several speaker specific features of the separated sounds such as pitch period, MFCC, PLPCC, formants and a "summary measure" representing the relative periodic/Aperiodic energy of each frame of the speech signal.
He was among the finalists by ISCA for Best Student Paper at the Interspeech 2010 conference.
He obtained a CORDIS doctoral grant for pursuing a PhD thesis at INRIA Bordeaux Sud-Ouest since November 2009
He is a reviewer for Elsevier’s DSP Journal.
Admitted for IUST's master program through a national entrance exam held by by the Iran Ministry of Higher Education in 2006 (ranked. 158 among 12000 participants).
Admitted for IUST's B.Sc. program through a national entrance exam held by by the Iran Ministry of Higher Education in 2006 (ranked. 732 among 300K participants).
