Characterizing packet audio streams from Internet multimedia applications

We analyzed 70 voice traces collected from IP-telephony applications, multicast lectures, and multimedia conferencing sessions which involve multiple speakers and different dynamics of interaction beyond two-way conversations. Results show that application differences have significant impact on the traffic characteristics. The conventional exponential model, established for telephone conversations, fails to capture accurately the packet level activity observed in these traces, e.g., the heavy-tail distributions of the talk-spurt and silence periods. We classify the traces into four types based on their audio contents: audience, lecture, multi-party conferencing and conversation. Further analysis shows that Weibull is a better matching statistical model and achieves lower mean-square-error than the exponential model (by 1 to 2 orders of magnitude) in approximating the audio streams for all four cases.