An adaptive predictor for media playout buffering

Receiver playout buffers are required to smooth network delay variations for multimedia streams. Playout buffer algorithms such as those commonly used in the Internet, autoregressively measure the network delay and variation and adjust the buffer delay accordingly, to avoid packets arriving too late. In this work, we attempt to adjust the buffer delay based on a prediction of the network delay and a similar measure of variation. The philosophy here is that the use of an accurate prediction will adjust the buffer delay more effectively by tracking rapid fluctuations more accurately. Proper buffer delay can lead to either (or both) a lower total end-to-end delay for a fixed packet lateness percentage or fewer late packets for a fixed total end-to-end delay which are both important metrics for applications such as IP telephony. We present a playout algorithm based on a simple normalized least-mean-square (NLMS) adaptive predictor and demonstrate using Internet packet traces that it can yield reductions in average total end-to-end delays.