This paper uses a speech recognition method which combines a continuous density hidden Markov model (CDHMM)-based preprocessor with window-based neural network (WNN) architectures. The method also employs modularity of NNs. It removes fixed-sized input constraint of NN and improves recognition performance. The CDHMM preprocessor performs a priori Gaussian shaping and normalization using statistically modelled state vectors in contrast to simple distance metric between acoustic vectors in dynamic time warping (DTW) preprocessor. Then, normalized and a priori Gaussian shaped speech features are applied as input to WNN and modular WNN architectures. NIST TI-46 E-set experiments are performed and the results are compared with a baseline CDHMM results. The proposed system improves the recognition performance. Modular WNN provides further significant improvement on the performance.