Speech recognition is an important part of human-machine interaction which represents a hot area of researches in the field of computer systems, electronic engineering, communications, and artificial intelligence. While speech signal is very complex and contains huge number of sampling points, the extraction of features from its time and frequency domain is very complex by analytical methods. The neural network capabilities to estimate the complex functions make it very reliable in such applications. This paper presents speech recognizer based on feed forward neural network with multi-layer perceptron structure. The speech is preprocessed by two methods; discrete wavelet transformation (DWT) and principal component analysis (PCA). The results and structure are presented and comparison is making over them.