Speech processing systems are actively considered for voice recognition, control and communication systems due to tremendous improvement in the digital signal processor and digital programmable logic devices. Preprocessing, feature extraction and classification are the fundamental steps involved in speech processing systems. Mostly transform domain and soft computing techniques are employed for feature extraction and classification respectively. Low computational complexity is desired to make speech processing systems as commercially viable solution. In this paper, speech recognition system for Marathi digit recognition using discrete wavelet transform (DWT), reduced order LPC coefficient and ANN is proposed. The main objective was to estimate the performance of the speech processing system at minimum feature vector length without significantly sacrificing on recognition rate. Better representation of speech frame was obtained using LPC coefficients derived from discrete wavelet-decomposed subbands instead of modeling the frame directly. Artificial neural network was used in the implementation of speech recognition system as classifier. An experimental result on Marathi digits from one to nine has demonstrated 78 % of the recognition rate at the minimum feature vector length of 24.