Information Technology Reference
In-Depth Information
on a link-by-link basis to select the redundant piggybacking degree. When
multiple speakers are talking, we combine their packets destined to the same
client into one (without exceeding the MTU). Although our scheme requires
each node to maintain retransmission buffers for storing recently received
frames, the overhead is manageable when the piggybacking degree is four
or less.
Our POS scheme is different than the two-party counterpart. In the multi-
party case, the order of the speakers is unknown, and the network conditions
among the participants may have large disparities. We address this problem
by equalizing the MSs observed by different listeners in the same turn. For
listeners whose MED affects the efficiency of the whole conversation (bottle-
neck clients), we use stricter MED values that closely hugs the delay curve,
similar to that in Figure 2.10a. For other listeners, we use less strict MEDs. In
comparison to the MSs in Skype in Figure 2.5, Figure 2.12 shows that our sys-
tem has a lower average MS as well as less variations. We are in the process
of developing a general equalization algorithm with parameters that operate
in continuous equalization levels [10]. Using simulated conversations gen-
erated under various control parameters and network and conversational
conditions, we plan to conduct subjective JND tests to determine the per-
ceptual sensitivity of humans in conversational quality as a function of the
level of equalization. We will develop a statistical procedure similar to that
described in Section 2.2 and experiment on some limited cases to within a
prescribed confidence level. Finally, we plan to train a classifier off-line and
apply it at run time to determine the control parameters that will lead to the
most preferred conversational quality. We expect our approach to be general
and lead to better equalization than that in Figure 2.12.
1800
1600
Our VoIP prototype
1400
1200
1000
800
1
2
3
4
5
6
7
8
9
10
11
12
Conversation turn index (between one speech segment and the next)
Person A
Person B
Person C
Person D
Person E
Figure 2.12
MSs experienced in our VoIP prototype.
