Original idea is that ring tone is generated by same exchange that provides the ringing voltage to analog local loop that rings. Idea there is that it tests that the whole connection is built correctly and can pass this ring tone. On the other hand this is horribly inefficient and thus often ring tone is generated on another places. For this cell phon behavior it's entirely imaginable that (G)MSC (lets just use GSM terminology) starts generating ring tone when it had found correct MS in location registry and sent paging request to it, obviously there can be significant latency between that and actual user observable ring on the terminating MS (on the order of 200ms in best case, few seconds or call being dropped in worst).