This paper proposes a weighted group delay distance measure (WGD) for noisy speech recognition. The WGD distance measure approximates an absolute group delay spectral distance weighted by the power spectra, which is sensitive to lower formant peaks and insensitive to the global spectral slope variation. The recognition performance of WGD was evaluated through speaker dependent and independent word recognition tests in additive white Gaussian and low-pass filtered noise. The results showed that WGD not only attains significant gains on robustness to additive noise over WLR and RPS, but also maintains the same robustness to interspeaker variation as WLR.