In this paper, our work on Marathi digit speech recognition is proposed using comparative analysis of discrete wavelet transform (DWT) and discrete cosine transform (DCT) through soft computing technique. Marathi digit speech signals are obtained from twenty different speakers that include ten males and ten females. The words are selected from one to nine digits pronounced in Marathi language. Transform domain and soft computing approaches are mostly preferred for speech feature extraction and classification respectively. The main objective was to estimate the performance of the DWT and DCT in speech feature extraction process so as to achieve better speech representation, low computational complexity and better recognition rate. Linear predictive coding (LPC) coefficients were derived from DWT decomposed sub bands and artificial neural network (ANN) was used as classifier in the implementation of speech recognition system. An experimental result was obtained to measure the performance of DWT and DCT on the speech recognition rate. Thus experimental results clearly demonstrate that DWT with LPC coefficients achieves better recognition rate at lower feature length as compared to DCT for same dataset.