A very good research, headed by Georgia Tech Institute: PinDr0p.
It works by analysing the various characteristic noise artifacts left in audio by the different types of voice network – cellular, VoIP etc. For instance, packet loss leaves tiny gaps in audio signals, too brief for the human ear to detect, but quite perceptible to the PinDr0p algorithms. Vishers and others wishing to avoid giving away the origin of a call will often route a call through multiple different network types.
The PinDr0p analysis can’t produce an IP address or geographical location for a given caller, but once it has a few calls via a given route, it can subsequently recognise further calls via the same route with a high degree of accuracy: 97.5 per cent following three calls and almost 100 per cent after five.