Archives of Acoustics, 15, 3-4, pp. 287-300, 1990

Simplified system for isolated word recognition

R. Gubrynowicz
Institute of Fundamental Technological Research, Polish Academy of Sciences, 00-049 Warszawa, ul. Świętokrzyska 21
Poland

K. Marasek
Institute of Fundamental Technological Research, Polish Academy of Sciences, 00-049 Warszawa, ul. Świętokrzyska 21
Poland

W. Mikiel
Institute of Fundamental Technological Research, Polish Academy of Sciences, 00-049 Warszawa, ul. Świętokrzyska 21
Poland

W Więźlak
Institute of Fundamental Technological Research, Polish Academy of Sciences, 00-049 Warszawa, ul. Świętokrzyska 21
Poland

This paper presents a general-purpose system for recognition of a limited set of words uttered in isolation. Such a system is intended for voice control of robot's movements. In order to minimize the number of operations performed during the recognition process and to limit the memory requirements frequency analysis of the signal was performed in adeq uately selected bands. Output signals from filters undergo detection and through an ¬A/D converter are introduced into a computer where they undergo further processing logarithmic conversiop and linear time standarization, among others. This leads to a reduction of the number range in further calculations. The DTW algorithm was used in the recognition process, while templates of individual words are introduced once, in principle separately for individual operators. The developed system speaker-dependent, in principle was verified experimentally for various vocabularies (containing 20 to 60 words) uttered by 11 voices (including 1 female voice). The average recognition accuracy for a 60 word wocabulary exceeded 98% for individual voices, while in a case of recognition whithout system accomodation to given voice the average error of recognition increased by about 10%.
Full Text: PDF


Copyright © Polish Academy of Sciences & Institute of Fundamental Technological Research (IPPT PAN)