A new block-based noise reduction system is proposed which focuses on the preservation of transient sounds like stops or speech onsets. The power level of consonants has been shown to be important for speech intelligibility. In single-channel noise reduction systems, however, these sounds are frequently severely attenuated. The main reasons for this are an insufficient temporal resolution of transient sounds and a delayed tracking of important control parameters. The key idea of the proposed system is the detection of non-stationary input data. Depending on that decision, a pair of spectral analysis-synthesis windows is selected which either provides high temporal or high spectral resolution. Furthermore, the decision-directed approach for the estimation of the a priori SNR is modified so that speech onsets are tracked more quickly without sacrificing performance in stationary signal regions. The proposed solution shows significant improvements in the preservation of stops with an overall system delay (input-output, excluding group delay of noise reduction filter) of only 10 milliseconds.