The rapid growth of IP-based packet switched network and the overall bandwidth efficiency of an integrated IP network make it an attractive candidate to transport real time voice connections. However, high quality voice over IP (VoIP) remains a challenge because interactive voice imposes many performance requirements (such as loss rate and latency) on the transport network. This cannot be easily achieved in the current Internet's "best-effort" transport. In this paper, we investigate the aggregate behavior of multiplexed voice flows and try to leverage the statistical multiplexing property to design a better network. With proper network provisioning, it is possible to carry VoIP traffic with excellent voice quality without serious under-utilization of resources in a well-managed IP network. We review VoIP architecture over Differentiated-Service model, and analyze the performance using both analysis and simulation. We show that knowledge about the mean m and variance sigma-squared of the individual voice source is sufficient to estimate the bandwidth usage, Cv, of the aggregate voice traffic under specific performance requirements. We estimate the bandwidth usage (where N = number of users) and investigate how k captures the multiplexing gain as well as the specific loss rate required. The results are useful for making such decisions as choosing appropriate output link bandwidth or proper bandwidth allocation for VoIP traffic in VPNs. Our experiments show that if we leverage the statistical multiplexing property, we can admit more than twice the number of voice sources compared with allocating the peak rate to each flow. We propose a sender-assisted call admission control policy for Voice VPNs to handle voice set up requests based on our results.