Hi Thomas, What is the pre-processing of the noise when computing autocorrelations? Is it the same as for the cross-correlations? If yes, would it be better to not fully whiten the data before autocorrelation as the only information comes from the amplitude spectrum (and not the phase)? I suggest to modify a bit the whiten.py function to keep the structure of the whole code similar. Following Gorbatov et al. (2013), instead of putting the amplitude spectrum =1 (with tapper) I would use a normalisation with a water-level in the form A(freq) = A(freq)/max(A(freq),c*max(A(freq))) With A the amplitude spectrum and c the water-level factor (which is typically on the order of 1%). It should not change the results for the cross-correlations and improve the autocorrelations. Does it make sense or did I missed something? Cheers, Aurelien REF: Gorbatov, A., Saygin, E., & Kennett, B. L. N. (2013). Crustal properties from seismic station autocorrelograms. *Geophysical Journal International*, *192*(2), 861-870. -- ========================== Postdoctoral fellow Department of Earth, Atmospheric and Planetary Sciences Massachusetts Institute of Technology 54-526 Cambridge, MA 02139-4307 e-mail: mordret(a)mit.edu web: http://web.mit.edu/mordret/www