In this paper, we propose a new method to extract noise-robust features and reduce the number of them for developing a small-sized speech recognition system. Although it is assumed that no correlation between features occurs in typical recognition systems, high correlations occur in practical cases. In consideration of this point, we apply principal component analysis to typical features and reduce the correlations between them. Furthermore, we introduce multi-condition training to improve recognition performance under noisy environments. From a large amount of experiments, when the number of features was reduce by two-thirds, the proposed method allowed us to maintain high performance at assumed SNR levels. Finally, we consider the relation between recognition performance and the amount of information when we use the proposed features.