Source Models for Speech Traffic Revisited

In this paper, we analyze packet traces of widely used voice codecs and present analytical source models which describe their output by stochastic processes. Both the G.711 and the G.729.1 codec yield periodic packet streams with a fixed packet size, the G.723.1 as well as the iLBC codec use silence detection leading to an on/off process, and the GSM AMR and the iSAC codec produce periodic packet streams with variable packet sizes. We apply all codecs to a large set of typical speech samples and analyze the output of the codecs statistically. Based on these evaluations we provide quantitative models using standard and modified on/off processes as well as memory Markov chains. Our models are simple and easy to use. They are in good accordance with the original traces as they capture not only the complementary cumulative distribution function (CCDF) of the on/off phase durations and the packet sizes, but also the autocorrelation function (ACF) of consecutive packet sizes as well as the queueing properties of the original traces. In contrast, voice traffic models used in most of today's simulations or analytical studies fail to reproduce the ACF and the queueing properties of original traces. This possibly leads to underestimation of performance measures like the waiting time or loss probabilities. The models proposed in this paper do not suffer from this shortcoming and present an attractive alternative for use in future performance studies.