In a speech recognition system, environmental mismatch between speech models and test speech causes serious performance degradation. To solve this environmental mismatch problem, smoothing process is one of the most widely used techniques. In this paper, an adaptive signal limiter (ASL) is developed to smooth speech features so that the undesired spectral variations could be effectively reduced. In addition, we propose the method for combining ASL with other noise-robust techniques in the recognition of noisy speech.