Speech recognition using dual-stage hierarchical structure: Gross and fine phonetic features

A speech recognition system using a dual-stage feature matching scheme, one on spectral slope and the other on slope-normalized all-pole model representation, is investigated. The speaker normalization process is accomplished by the combination of a nonlinear spectral transformation and an adaptive spectral slope compensation. The spectral slope function has been derived from concatenation and damping of the all-pole model. Normalization of the slope is equivalent to inverse filtering and controlling the damping factor. The effectiveness of this recognition scheme and its signal processing techniques for speaker-independent speech recognition are evaluated and discussed in this report.