Measuring Network Parameters in a Testbed

This Section presents the testbed that we used to identify the most relevant parameters and their ranges for real-time transmission of speech. In addition to the parameters mentioned in the introduction, we noticed that real-time audio in general and speech in particular are sensitive to other factors that, when combined with the network behavior, give a different resulting perceived quality. One of these parameters is the spoken language [79]. A series of tests with different languages showed in fact that they do not equally tolerate the losses. For these reasons, all our MOS experiments were done for three different languages (Arabic, Spanish and French). A survey on IP-telephony users today showed that they generally fall into the following categories. The first one concerns IP-telephony nationwide. For that reason, we did three series of measurements between Rennes (ENST-B) and peers laying respectively in Rennes (Irisa) (2 Km distance), Brest (300 Km) and Sophia Antipolis (1300 Km). The second category is related to international phone calls. For that reason tests were accomplished between Rennes and Mexico City (Mexico). The number of hops, the minimum, the average, and the maximum one-way delay in ms between Rennes (ENST-B) and the other sites are shown in Table 5.1. The third category of communications in a LAN was not considered because it was difficult to find a real practical situation where a 10Kb/s IP flow does not easily find its way through a 100Mb/s or even a 10Mb/s LAN. Each session consisted of sending a 160-byte packet every 20 ms as real-time traffic carried by RTP protocol. The duration of the session was 10,000 packets. The receiver reported the total number of packets, the total loss rate, and percentage rate of each n-consecutively-lost packets. The destination considered any packet that arrived after the playback threshold as lost. This is to avoid the jittering problem. In fact, there are several algorithms to choose the best value of the playback threshold in order to minimize the percentage loss rate and to avoid the jittering problem [37].

Table 5.1: Number of hops and one-way delay statistics between the peers (Site) and Rennes (ENST-B) one.

Site	Hops	Minimum delay	Average delay	Maximum delay
Irisa	6	4.2	11	70
Brest	7	7.1	43.3	117
Sophia	12	24	35	60
Mexico	28	149	159	221

For each site, we repeated the tests 50 times in working-day hours and in different days. Then we selected the results that gave the maximum and the minimum percentage loss rate. By varying the playback time length, the percentage loss and the loss distribution change accordingly. Figure 5.1 depicts the minimum and the maximum percentage loss rate. As expected, the loss rates decrease when we increase the playback buffer length of the receiver and the number of hops. In Figure 5.2 and Figure 5.3 we plot the rates of consecutively lost (CL) packets against the buffer size. The first Figure shows minimum values and the second shows maximum ones. The two Figures show the frequency of $n^{th}$-consecutively lost packets where $n$ varies from one to ten. As it is clear from these two Figures, in national sites, the three consecutive loss pattern is considered the limit. In the international case, five consecutive losses is the limit. The same approach was taken in [17] (with almost the same paths for tests); however, the number of years separating the two experiments clearly shows an improvement in delay and loss bounds. The tests described were only intended to give a realistic figure on the average values taken by the parameters that we used in the neural networks training. Therefore, the databases used to train the neural network are as close as possible to real network situations.

Figure 5.1: Maximum and minimum percentage loss rates as a function of the playback buffer length between Rennes and the other different sites.

Figure 5.2: Minimum rates for one-to-ten consecutively lost packets as a function of the playback buffer length between Rennes and the other different sites.

Figure 5.3: Maximum rates for one-to-ten consecutively lost packets as a function of the playback buffer length between Rennes and the other different sites.