Perceived Quality of Packet Audio under Bursty Losses

We examine the impact of bursty losses on the perceived quality of packet audio, and investigate the effectiveness of various approaches to improve the quality. Because the degree of burstiness depends on the packet interval, we first derive a formula to re-compute the conditional loss probability of a Gilbert loss model when the packet interval changes. We find that FEC works better at a larger packet interval under bursty losses. In our MOS-based (Mean Opinion Score) listening tests, we did not find a consistent trend in MOS when burstiness increases if FEC is not used. That is, in some occasions MOS can be higher with a higher burstiness. With FEC, our results confirms the analytical results that quality is better with a larger packet interval, but T should not be too large to avoid severe penalty on a single packet loss. We also find that low bit-rate redundancy generally produces lower perceived quality than FEC, if the main codec is already a low bit-rate codec. Finally, we compare our MOS results with objective quality estimation algorithms (PESQ, PSQM/PSQM+, MNB and EMBSD). We find PESQ has the best linear correlation with MOS, but the correlation is still not high enough to be used in isolation to predict MOS. Keywords—packet audio, IP telephony, voice over IP, perceived quality, subjective quality, loss model, Gilbert model, forward error correction, low bit-rate redundancy.