Speaker-independent speech recognition using nonlinear predictor codebooks

A neural spectrum-prediction mechanism is implemented in the predictor codebook for speaker-independent speech recognition. The nonlinear predictor codebook consists of neural predictors generated through LBG (Linde-Buzo-Gray) based predictor quantization procedures. Nonlinear prediction functions insulate each predictor code from the other codes, and accomplish high phoneme separation without decreasing the robustness to speaker variation. The structure of each predictor is equivalent to a three-layer neural network, but it is not trained by error backpropagation. The predictor is first optimized as a linear prediction function. Then, the nonlinear (sigmoid) function is implemented in it. A set of nonlinear predictors is totally optimized by the predictor quantization algorithm.<>