On NLMS Estimation for VoIP Playout Delay Algorithms - Improving Delay Spike Detection

Voice over IP (VoIP) applications are now very popular and widely used on the Internet. Such applications use receiver playout buffers to smooth delay variations so as to reconstruct the periodic form of the transmitted packets. Packets arriving after their scheduled playout time are considered late and are not played out. Playout delay control algorithms often operate by updating the playout delay between periods of silence. A recent class of playout control algorithms has received particular attention; this class of algorithms uses autoregressive measures on the network delay so as to estimate future packet delay values and adjust the playout delay accordingly. In this work, we compare two algorithms previously proposed that use such autoregressive approach; both playout algorithms use a normalized least-mean square (NLMS) adaptive predictor. The difference between both algoritms is that the second one is an extension of the first that adds delay spike detection. We demonstrate, by using Internet audio packet traces that, contrary on what was claimed, the algorithm that uses spike detection does not overperfom the first one. Finally, we propose an algorithm based on the original NLMS algorithm with delay spike detection that overperforms the previous two NLMS playout algorithms.