Using a vertical source array, directive horizontal transmission of 1200 cps sound pulses was studied in the Caribbean north of Bonaire Island. Both the fixed source and a linearly moving receiver were operated at 800 m depth and ranges between 0.35 and 5.6 NM. Close to the source, the pulse consisted of two sets of 12 wavefronts, each in immediate sequence with their phases exactly reversed. The directly propagated pulse exhibited a well-recognized phase-reversal point out to a range of 4 NM, beyond which it gradually merged into the basic pulse. Measuring transmission times to this phase reversal, a total of 4000 were statistically analyzed, using a moving average to eliminate the effect of the linear motion of the receiver. The rms deviation of transmission time from this average remained between 0.3 and 0.5 msec for ranges up to 4 NM, beyond which it climbed to 2 msec. The normalized autocorrelation coefficient calculated with respect to time interval always had the same shape and a first null near 1.5 sec. If we use the velocity inhomogeneity that was measured at the same time and the Chernov theory for an idealized model, we would infer an rms deviation in transmission time 100 times smaller than was measured. It seems that, particularly for horizontal transmission, one must consider the large and rapid variation that can be caused by the motion of inhomogeneous water across the band of multipaths connecting the source and receiver. [Hudson Laboratories, Columbia University Informal Documentation No. 43. Work supported by the U. S. Office of Naval Research.]