Nonlinear ground response at two strong-motion arrays in Taiwan is studied using the spectral ratio technique. At the SMART1 array, we calculate the frequency-dependent soil amplification functions as a ratio of the spectra at alluvium to rock sites, and study their dependence on the excitation level. Horizontal components of shear waves are considered. We compare (1) the average spectral ratios on weak and strong motions, (2) the ratios for the mainshocks and aftershocks, and (3) the ratios for the strong shear waves and their coda. At the SMART1 array, “weak motions” have a peak horizontal acceleration (PHA) less than 30 Gal. “Strong motions” are in the range of 100 to 267 Gal. Comparison of the average weak- and strong-motion spectral ratios shows a significant deamplification of strong motion between 2 and 9 Hz, exceeding the error margin estimated by the standard deviations. The maximum deamplification occurs at approximately 6.5 Hz where the average weak-motion amplification is 2.9 versus 0.40 in the strong motion. A similar pattern is exhibited by the ratios calculated for the mainshocks and the aftershocks, as well as for the shear waves and their coda. The spectral ratio calculated from a single realization of coda is identical to the average ratio obtained from many small earthquakes. At the SMART2, we analyze spectral ratios between the stations on Pleistocene terrace deposits and recent alluvium, which characterize the relative response at these two types of sediments. Weak motion is PHA less than 13 Gal, while strong motion extends from 100 to 295 Gal. Strong-motion spectral ratios between terrace and alluvial sites are consistently reduced in the frequency range from ∼1 to 10 Hz, compared with the weak motion. This effect is insensitive to the variation in distance between stations from 7.9 to 11.4 km, as well as the azimuthal change of up to 80° in the station pair strike. We attribute the observed discrepancies between weak- and strong-motion amplifications to the differential nonlinear response occurring at terrace and alluvial sites. Our results document a significant nonlinear ground response at both arrays.