A robust speech analysis in speech recognition

This report presents an adaptive speech analysis method specially used for a speech recognition system. The designed speech recognition system consists of an adaptive speech analysis, a self-organized clustering/pseudo-labeling method and a DTW. All methods are redesigned in fully parallel and pipelined mechanism. In the speech analysis method, an adaptive ARMA lattice modelling is introduced for the reduction of distortion, noise and disturbance. In addition, the speech analysis keeps robust condition where an adaptive method is usually considered to be sensitive as to the convergence property and the parameter estimation. By using the recognition system including the robust adaptive speech analyzer, speech recognition results are shown