Predictive playout delay adaptation for Voice over Internet

Packet switching is an effective technology for integrating voice and data in a single network. An example for such a kind of a network is Internet. New service is emerging, and it is a Voice over Internet Protocol (VoIP). An important aspect of packet-switched voice is the reconstruction of a continuous stream of speech from the set of packets that arrive at the destination DTE (data terminal equipment). Because of the statistical nature of packet switching, each packet can encounter a different amount of delay in traversing a packet network. The problem described in this paper is how to design an appropriate voice reconstruction strategy that reproduces acceptable-quality speech from packets that arrive with varying delay. The basic idea proposed is to measure a delay jitter between several points on the Internet. Our voice traffic generator, implemented on a workstation, sends packets that are 20 ms long. The receiver measures the inter-arrival time and keeps track of mixed sequences and lost packets. The measurement results are given in a form of average delay jitter. The receiver also keeps a record of all relevant packet information in a text file that is used as an input to a delay jitter predictor. A traffic prediction element predicts talk-spurt characteristics and directly influeuces playout delay. For the prediction element, we propose an adaptive neuro-fuzzy inference system (ANFIS).