The extraction of Green’s functions by cross correlation of continuous seismic records at station pairs is best achieved in a diffuse wave field, where energies radiated by random sources have equal power. To partially satisfy the diffuse wave‐field condition, original seismic records must be normalized as they are highly nonstationary and may have amplitude variations of orders of magnitude within the cross‐correlation time window. We adapt a frequency–time normalization method to obtain seismograms with an even spectrum at all times within the data‐processing unit. Compared with the commonly used one‐bit normalization, the new method improves the signal‐to‐noise ratio of empirical Green’s functions by a factor of ∼2 and increases the effective data‐recording duration by a factor of 4, assuming random local and instrument noise. It yields useful Rayleigh waves at periods up to 300 s for Program for Array Seismic Studies of the Continental Lithosphere (PASSCAL)–type deployments and 600 s for permanent stations with very‐broadband sensors. Thus, the new method makes it possible to extend surface‐wave signals to frequencies beyond those in traditional earthquake‐based surface‐wave tomography at both the high‐ and low‐frequency ends.