elsnet |
Central and Eastern European Survey |
phoneme based speech recognition,neural nets,hidden Markov model
Characteristics of SRL KTU
I. Scientific qualification and stuff of SRL KTU - SRL KTU head dr. Algimantas Rudzionis in the field of speech processing is active since 1969 ( PhD thesis in speech vocoding in 1972); - till 1990 he carried out fundamental researches according to projects of the Academy of Sciences of the former USSR (since 1986 a member of the Speech processing council ); - 1992-94 senior researcher for the IPI (Oregon, USA); - participant of the European Union programs: COST 232 "Speech Recognition over Telephone Line" COST 249 "Continuous Speech Recognition over the Telephone" COST 250 "Speaker Recognition in Telephony". - currently SRL stuff consists of 4 persons: 2 researches and 2 engineers
II. Previous activities of SRL KTU We consider that our outstanding achievements are as follows: - time resolution effect on phoneme perception accuracy (we observed strong dependences on vowel context); - speaker independent phoneme classifying (training, dichotomy-based classifier, coarticulation evaluation) - very close coincidence was achieved between the results of perception and classification; - speaker dependent phoneme-like isolated word recognition ( projection method was proposed for evaluating of every phonetical event in separate way); - fully phonetical speaker dependent recognition of 200 isolated words (less than 1% error for 200 words recognition by 32 bit/word reference).
III. Facilities IBM PC computers , CD ROM player, device for telephone speech input, SoundBlaster and VoiceMaster sound cards. Proramming language C/C++. Operating systems MS-DOS and Windows.
[Survey] [Organisation] [General Info] [Training] [Resources] [Research] [Staff] [Publications] |