The performance of most standard Automatic Speech Recognition (ASR) systems degrades severely under noisy environments, because they are usually trained with clean speech data. There exist several speech enhancement or compensation schemes that can improve the robustness of ASR to different levels. In this paper, an effective feature compensation method called In-phase Feature Induction (IFI) is proposed. It makes use of the phase relationship between the spectra of noisy input and the corresponding noise signal to accurately obtain the clean speech spectrum. Experimental results show that recognition systems with IFI compensation yield much higher accuracy rates under various noisy conditions compared with the Spectral Subtraction based method.