Wire Communications Laboratory (WCL), Electrical Engineering Department, University of Patras, GR.
The research activity of the Wire Communications Laboratory is mainly of an applied nature. At present it includes the following topics of research and development which are pursued with funding from various sources (EU, Greek General Secretariat for Research and Technology, industry, etc.): Design and construction up to the level of industrial prototype of advanced telecommunications systems and microcomputer controlled devices, optimization of telecommunication networks with applications in the Greek network, digital processing of speech with applications in automated speech recognition, speaker recognition and speech synthesis of Modern Greek, Natural Language processing, linguistic processing of Modern Greek texts by computer, coding of acoustic signals, processing of noisy acoustic signals, development of artificial intelligence systems, multimedia processing and transmission of medical information. The WCL has research teams working in the above fields and suitable equipment, so as to be in a position to deal with complex problems. The WCL cooperates with Greek and foreign firms and participates in large national and international research projects supported by the European Community, the National Telecommunications Organization, the Greek Telecommunications Industry, the Ministry of Industry and other National and Regional bodies. Patras was a partner in the SPHERE HCM network.
Patras will lead the work on auditory techniques for speech enhancement in noise as a basis for robust speech recognition (task 3.4,). In previous work (see the references below and ?3.3), it has been shown that such methods can reconstruct intelligible speech irrespective of the initial SNR, provided that suitable estimators of the sparse psychoacoustic data can be developed. These data are estimates of the clean speech Auditory Masking Threshold values per critical band (typically for 16 KHz, 22 values per frame must be estimated). The estimators are based on Minimum Mean-Square Error criteria for successive data frame statistical distributions and can be linked to HMM recognisers. The work will include improvement of the existing noise masking model, and the evolution of suitable enhancement filters for reverberant and cocktail-party speech (in collaboration with Bochum).
In task 3.4, Patras will work with Bochum and Ericsson. To extend its expertise as a promising lab in a less-favoured region of the community, Patras will also become involved in the precedence effect and binaural cues tasks lead by Bochum (1.3 and 2.3). Some work at Patras has already been done in this area using psychoacoustically optimised time-frequency analysis (windows), resulting in much-reduced response order and reflection pattern-smoothing. In addition, we will seek to recruit Patras researchers for placements on other tasks, so that on their return they will broaden the Patras research-base.