Prof. Milan Milosavljevic
Belgrade University
Faculty of Electrical Engineering
Kralja Aleksandar 73
11 000 Belgrade

Yugoslavia
http://control.etf.bg.ac.yu/Nastavnici/Nastavnici.htm  

School of Medicine

...........................

phone:   381 11 324 84 64
fax:   381 11 324 86 81
mmilan@etf.bg.ac.yu
bisop@ptt.yu

Belgrade University

BUY is leading faculty in former and present Yugoslavia in the field of signal and speech processing. It has been involved in almost all theoretical and practical projects from this field during last 50 years. Its scientists and professors achieved remarkable results in the field of speech modeling, coding and recognition. For instance, they developed first domestic channel, predictive and CELP vocoders; first speech isolated spoken digit recognition system, first system for recognition of Serbian phonemes etc. Former fellows of these two institutions are leading researchers in many important scientific centers all over the world.  

Mathematical Institute SANU, Belgrade, is a leading institution in the former and nowadays Yugoslavia from the field of Applied Mathematics and Computer Sciences. Especially, during last two decades, Mathematical Institute has been engaged in the several theoretical and technological projects, funded by Serbian Ministry of Sciences. The most important scientific results are achieved in the fields of pattern recognition, computer sciences, signal processing, speech processing, image processing, linguistics, cryptography, theoretical aspects of applied mathematics, pure mathemtics, etc. Today, Mathematical Institute SANU is the Serbian institution with probably the largest number of high-quality papers published in the most significant international journals and international conferences. In the field of speech processing, the main results are achieved in robust speech analysis and coding based on pattern recognition approach, Huber's M-estimation theory, T-distribution based methods, neural network based methods, etc. Scientific project proposals obtained from Mathematical Institute SANU almost always have the highest score of scientific points which resulted from summation of scientific points of individual researchers.

Faculty of Electrical Engineering, Belgrade University, Belgrade, Yugoslavia,  is leading faculty in former and present Yugoslavia in the field of signal and speech processing. It has been involved in almost all theoretical and practical projects from this field during last 50 years. Particularly, collaborating with Institute of Applied Mathematics and Electronics, Belgrade, Yugoslavia, its scientists and professors achieved remarkable results in the field of speech modeling, coding and recognition. For instance, they developed first domestic channel, predictive and CELP vocoders; first speech isolated spoken digit recognition system, first system for recognition of Serbian phonemes etc. Former fellows of these two institutions are leading researchers in many important scientific centers all over the world.

Professor Milan M. Milosavljevic and Branko Kovacevic are leaders of this speech and signal-processing group. They published during last 20 years about 500 scientific papers, and 10 monographs. Their approach to robust speech modeling based on Huber’s theory, as well as active learning principle in adaptive filtering is recognized in the international scientific community as remarkable results. His practical and theoretical works was served as a basis for many professional communication and signal processing systems.

Our research group could be engaged in the following activity of development of new high accuracy phoneme recognition system (RecoPhone):

- Feature extraction subsystem – Application of robust estimation techniques for extracting reliable, discriminative features describing phonemes on the whole phoneme and sub phoneme levels.

- Automatic segmentation – reliable automatic extraction of broad range of change point events in time domain, from slow to abrupt, as well as associated values of predefined class membership functions of these events.

- Redefinition and recalculation of starting features in accordance with segmentation results. Final features possesses important information regarding time domain clues in the vicinity of change points of signal characteristics.

- Engineering of adaptive structures (such as Neural Networks, Support Vector Machines, Bayesian Dependence Networks) for transforming primary signal description in the feature space to variety of σ – sub fields induced  by intended category description.

- Design and analysis of adaptation and learning procedures associated with adaptive transformation structures. Resampling and active learning principles will be introduce in order to improve intermediate and finale recognition accuracy.

- Estimation of the lower bound of recognition accuracy of a given feature set. This could be useful for automating feature selection and transformation procedure, before designing any classification procedure.

- Investigation of  influence of glottal signal to the accuracy of phoneme recognition.

- Investigation of accuracy of developed RecoPhone system in the case of different acoustical, lexical and lingual settings.