Speaker-independent word recognition for large vocabulary using pre-selection and non-linear spectral matching

This paper proposes a method of speaker-independent isolated word recognition for large vocabulary size using pre-selection and non-linear spectral matching. Pre-selection is used for reducing computation time for word matching. Non-liner spectral matching is used to normalize the individual difference in vocal-tract length of the speakers. The parameters for pre-selection are computed from the low order coefficients of Fourier-cosine expansion of the time pattern of the three gross features. The parameters represent the type and position of phonemes in a word. Speaker normalization of vocal-tract length is carried out by spectral matching of the optimal frequency warping function in logarithmic frequency scale. The experiment of speaker-independent isolated word recognition was carried out for 212 words uttered by 4 males and 4 females. The recognition score was 91.8%.