Playing out of packets at the receiver is simulated offline using stream 1 only, stream 2 only, and using both streams, so that the three different schemes are compared under the same conditions.
We first compare the delay - loss tradeoffs by setting small
and varying
in (1) during playout, which
suggests that we are less concerned about full reconstruction of both
descriptions at the receiver. The results are plotted in Fig. 4. The quantities we are interested in here are average total delay
and loss rate. The average total delay,
, is the average value of
over all the received packets in a playout session. The total
loss rate
is the percentage of lost packets, no matter the
loss is a result of channel erasure or late arrival. We consider a packet to
be lost only if we receive neither description from the two streams in time.
We also define the burst loss rate
here, which is the
percentage of burst loss occurrences in a session. In calculating
, M consecutively lost packets are counted as M-1
occurrences, where
. Burst loss is of our great concern because it
is difficult to conceal and it impairs voice quality severely. Burst loss is
not only a result of consecutive channel erasure, but also caused by delay
jitter and sustained outage period.
Figure 4: Loss - delay tradeoff, Experiment 1
Figure 5: Standard deviation of playout time, Experiment 1
Figure 6: SNR - delay tradeoff, Experiment 1
From Fig. 4, we observe significant reduction of loss rates by using multi-stream transmission. At the same average delay of 70 ms, the total loss rate is reduced from more than16% to less than 2%, compared to using single-stream FEC. More importantly the burst loss rate is reduced from more than 3.5% to 0.5%, which is significant for burst losses. Packet loss in this experiment does not quite result from channel erasure according to Table 2, but from the high delay jitter. The loss rates are reduced by a great amount due to the independent jitter behavior of multiple paths, and the isolated outage periods.
In Fig. 4, the average delay is also observed much lower for multi-stream scheme. At the same total loss rate of 5%, the delay is cut by more than 20 ms. The delay reduction can be explained by the possibility to playout the description with the lower delay if obtaining full voice quality is not a priority. For this same reason, the variation of playout rate is lower for multi-stream at different loss rates, as is seen in Fig. 5. The variation of playout rate is a direct result of the adaptive playout scheme, whose effort is targeted at improved tradeoff between loss and delay [10]. The unusually high playout variation of stream 1 is a result of its high delay jitter.
In the case the user is more concerned about the voice quality instead of
low latency, the multiplier in (1) can be
increased to give more emphasis onto receiving both copies of MDC. In Fig. 6 we have plotted SNR degradation vs. delay as we vary
while keeping
moderate and fixed. The SNR of
received single is compared to that of the quantized (in full resolution)
signal at the sender. It is seen from Fig. 6 that, as the
delay goes up, chances are better that both MDC copies can be played out
successfully, resulting in higher sound quality. In Fig. 6,
the voice quality corresponding to multiple streams are better than that
using single-stream FEC. One reason is that, more noise is introduced when
losing the primary copy of FEC than losing either copy of multi-stream MDC.
The other reason is that, due to the independent characteristics of multiple
paths, if one description of MDC is lost, the chance of receiving the copy
from the other stream is still quite good. However, with single-stream FEC,
whenever the primary copy is lost, no matter due to channel erasure or the
packet's late arrival, the chance of receiving the secondary copy is not
optimistic, since the delay and loss statistics are correlated across
successive packets. Fig. 6 demonstrates that the improved of
tradeoff between delay and loss using MDC is not obtained at the cost of
compromising voice quality.
Fig. 7-9 show the performance results in parallel for Experiment 2. Similar improvements made by multi-stream MDC are observed, except that the delay reduction is less than that in Experiment 1. This can be explained by lower STD of network delay and milder jitter in this experiment (Table 2). However, from Fig. 7, it is obvious that the reduction of total loss rate and burst loss rate is still considerable.
Figure 7: Loss - delay tradeoff, Experiment 2
Figure 8: Standard deviation of playout time, Experiment 2
Figure 9: SNR - delay tradeoff, Experiment 2