A combined analysis of SKS, SKKS and P660s seismic phases detected in the records of BCAO/BNG seismic station in central Africa reveals that the waveform and amplitude of the T component of P660s deviate significantly from those predicted from the observations of shear wave splitting in SKS and SKKS. The inverse problem is investigated by using the anisotropic Thomson-Haskell code and a genetic algorithm. The discrepancy between SKS/SKKS and P660S can be explained by P-SH conversion at the boundaries of a weakly anisotropic layer 30--40 km thick lying atop 660 km discontinuity beneath BCAO/BNG. This anisotropy is probably induced by horizontal flow at the bottom of the upper mantle, in the transition zone. ¿ 1998 American Geophysical Union