This paper studies a new semiparametric estimation procedure of the memory parameter ($d$) in models of fractional integration. The procedure is called modified log periodogram regression and it is especially useful for nonstationary time series where $d\geq 1/2$. The modified estimator is shown to be consistent and asymptotically normally distributed with variance $\pi^2/24$ under mild regularity conditions. Simulations reveal that the estimator performs well for all $d\geq 1/2$.