Incorporation of time varying AR modeling in speech recognition system based on dynamic programming

Two isolated-word speaker-independent recognition systems are based on non-stationary representations of the speech signal. The first system is based on a time-varying linear prediction model. This recognition system is distinguishable by incorporation of the model and appropriate extension of the log likelihood ratio (LLR) distortion measure. The second system is based on the lattice model represented by time-varying reflection coefficients In this System, the distortion measure used is the Euclidean distance. In order to support this system properly, the extension of the LLR was modified. >