Acoustic-phonetic module for continuous Slovene speech recognition

The authors present a possible configuration of an acoustic-phonetic module for Slovene speech recognition. They tested several feature sets, deriving the orthogonal transformation of the log energies of the short bandpass spectra components and loudness. Segmentation is done on the basis of the reference values of the phone components and fundamental phonetic knowledge of the statistical properties of phones, phone duration times, and certain other phonetic rules of the Slovene language, which are useful in the segment classification process. The module was tested on speech data consisting of 44 simple Slovene sentences. The results of testing and plans for the future development are presented.