| Dr.Volker Fischer University of Erlagen |
|
| Dr.Volker Fischer University of Erlagen |
The Chair for Pattern Recognition was established in 1975 as a part of the Department of Computer Science of the Friedrich-Alexander-Universitat and since is headed by Prof. Dr.-Ing. H. Niemann.
General research areas at the chair for pattern recognition include signal processing, continuous speech recognition, spoken dialog systems, natural language processing, image processing, image analysis, knowledge based image interpretation, and robotics. At present, the chairs research staff comprises 17 members, working in these areas, ten of them in cooperation with industrial companies, in joint projects sponsored by the BMBF or the German Research Foundation (DFG), or in the ESPRIT projects.
In image processing prior work deals with low-level image processing as well as with knowledge based image analysis using semantic networks, which results in the development of several image understanding systems for medical applications. Current research deals with 3D object recognition and segmentation, the generation of statistical and geometrical 3D models, and motion detection and description. Work on active vision has resulted in a closed loop real-time object tracking system and the extension of this work to high-level vision will be a forthcoming research theme.
Experience and prior research in speech recognition comprises the design and implementation of an operational (German) speech understanding and dialog system for intercity train inquiries together with companies and other universities in joint BMFT research projects. Now this work is continued in the VERBMOBIL project, which aims on the development of a portable speech translation system. Contributions to word recognition, polyphones, and polygraphs gathered excellent results in several VERBMOBIL word recognition tests.
In the ESPRIT SUNDIAL project, research was performed on word recognition and dialog. An operational version of the developed dialog system presently is available for scientific evaluation on a phone line. A recently started project aims on the extension of the work on spoken dialog systems towards a multi-lingual and multi-functional system. Further work deals with prosodic analysis, the improvement of F0 estimation, integration of prosodic analysis into dialog control, and the prosodic marking of phrase boundaries.
In 1994 the chair became a subcontractor in the Real World Computing Partnership (RWCP). Long term research under the RWC program aims on the development of a unified approach for the understanding of multi-modal information meeting the requirements of real world applications. Members of the RWCP research group are Prof. Niemann (principal investigator), Dr. V. Fischer and Dipl.-Inf. J. Fischer.
In June 1995, we had the opportunity to join the annual RWC Symposium at TEPIA Hall, Tokyo, to present our ideas on a real time speech understanding system. Being a new contributor to the RWC program at that time, the visit was extremely valuable to get an insight into other research projects in the Partnership. In fiscal year 1995 our institute was visited twice by delegations from Japan. In September, we were able to discuss some major points of speech understanding research with Mr. Yoshiaki Itoh and Dr. Ryuichi Oka from the RWCP and Dr. Yoichi Yamashita from Osaka University. A visit from RWCP planning section (Mr. Kazutoshi Hashimoto and Mr. Shoyu Watanabe) for an interim inspection of the project management was also very helpful for a better understanding of RWCP's business procedures.
Since speech is expected to become one of the most important sources of input to communicate with computers up to the end of the decade, recent research under the RWC program concentrates on the development of a real-time speech understanding system. The general framework for our work is provided by a speech understanding system that is able to answer inquiries about the German intercity train time table. The system is connected to the public telephone line, thus allowing the collection of large samples of spontaneous speech for both an improvement of the systems modules (like e.g. word recognition, prosodics) and the development of new components (like e.g. a stochastic dialog model).
Speech understanding requires the mapping of the speech signal into the systems internal representation of the task domain to obtain a reaction intended by the user. Due to both imperfect results from word recognition and ambiguities in the systems linguistic knowledge base, this mapping is far from being unambiguous and therefore the computation of an optimal mapping requires the exploration of large search spaces.
Our basic approach to real time speech understanding employs iterative optimization techniques (like e.g. simulated annealing, the great deluge algorithm, or genetic algorithms) and parallel processing of the knowledge base on a subconceptual level. The use of iterative optimization results in an any-time behaviour of the system, which is important for real world applications where usually no fixed processing time is given. Improved any-time capabilities are obtained from parallel optimization and massively parallel processing of the semantic network based knowledge base. Whereas the former allows the simultaneous exploration of different parts of the search space, the latter decreases the computing time needed for a single iteration.
A first operational version of the system is now under evaluation and implementation of the parallel system in a local area network of high speed work stations has recently started. Results obtained with this prototype system will contribute to a further improvement of optimization algorithms and parallel constraint processing for search space reduction.
Our future research under the RWC program will concentrate on the further improvement of our parallel speech understanding system as well as on our long term research on multi-modal information processing.
The completion of the speech understanding system demands an extension of our work to the dialog level. Moreover, for the achievement of real time capabilities, an approach to incremental knowledge based processing has to be made. For that purpose, we plan to examine the use of prosodic information, which could provide the linguistic analysis with boundaries of meaningful portions of the speech signal.
Since our semantic network formalism and the iterative control algorithm were already tested successfully in an image understanding task, we are convinced that this approach will be feasible for the knowledge based processing of multi-modal information.
More recently we started some preliminary studies on the integration of different sources of information which include a connectionist approach to enhance speech recognition by means of lipreading and a natural language interface to a stereo head.
Further information about our institute is also available in the world wide web (www) under the URL http://www5. informatik.uni-erlangen.de