With this paper we report various experiments we conducted on the noisy digits recognition task of AURORA 2. We are aiming not only for optimized feature extraction algorithms, but also for improved strategies throughout the entire ASR system. Starting from a detailed examination of the noise-estimation procedure in spectral subtraction and its influence on ASR performance, we introduce an alternative speech enhancement approach that is based on singular value decomposition. Discriminative techniques during HMM training, namely rival training, are shown to improve system performance significantly, especially for small-footprint recognizer set-ups. In our full resource system we obtain slight additional improvements over previous results by applying LDA.