Conversational Model Based VoIP Traffic Generation

Transcription

1 Conversational Model Based VoIP Traffic Generation Li Ji 1, Xia Yin 1, Xingang Shi 2, Zhiliang Wang 2 1 Department of Computer Science and Technology, Tsinghua University, Beijing, China 2 Network Research Center, Tsinghua University, Beijing, China jili@csnet1.cs.tsinghua.edu.cn; yxia@tsinghua.edu.cn; shixg@cernet.edu.cn; wzl@csnet1.cs.tsinghua.edu.cn Abstract VoIP (Voice Over IP) traffic generation has significant meaning for the measurement of VoIP quality in networks, but existing generation methods are insufficient both in reality and interactivity, and thus affect the accuracy of VoIP quality measurement. This paper proposes a conversational model based VoIP traffic generation algorithm, which simulates the behavior of two users in a VoIP session, and generates VoIP traffic for quality measurement according to specific voice codec and VAD (Voice Active Detection) parameters. Our experiments show that the algorithm illustrated in this paper can effectively demonstrate the influence of conversational users speech characteristics and interactive process to VoIP traffic, and thus provides more realistic and interactive traffic for practical quality measurement activities. Keywords-VoIP traffic; conversational model; measurement 1. Introduction VoIP [18], which means transmitting voice traffic over IP networks, has gradually become an important application on the Internet. Compared with traditional telephone networks, VoIP is much more cost-efficient and flexible, and thus attracts more and more corporation and individual users. VoIP applications like Skype [3] also have achieved great success. However, due to the complexity of the Internet, it is unpractical to calculate VoIP performance metrics only through mathematical modeling, as what was done in telephone networks, so the performance evaluation of VoIP requires actual measurement activities. Before the deployment of VoIP applications, an ISP or a corporation can verify the availability and deployment scale limitation through VoIP measurement on the targeted network, and then plans for the deployment based on the result; after the launch of VoIP applications, VoIP quality can be monitored periodically through measurement among nodes in the network to gather quality variation information, and avoid the ignorance of unacceptable VoIP quality caused by network failure or bandwidth bottleneck. From these we can see that VoIP quality measurement is very critical for the proper deployment and usage of VoIP applications. VoIP quality measurement includes three steps: VoIP traffic generation, VoIP traffic transmission and measurement result analysis. Firstly it is necessary to generate proper VoIP traffic traces, which simulate the traffic of one or multiple VoIP streams. Then the generated traces are distributed to measurement nodes in the network and actual traffic is transmitted according to the traces between pairs of them. Finally, after the transmission, VoIP performance metrics, such as loss rate, end-to-end delay and jitter can be measured and VoIP quality can be analyzed based on these metrics. For VoIP traffic generation, there are two main existing methods. The first method is to use Constant Bit Rate (CBR) traffic as simulated VoIP traffic. This method is used because some popular voice codecs, like G.711 [13] and G.729 [14], generate constant sized frames at a constant interval. However, currently most VoIP products have adopted Voice Active Detection (VAD) technique, so the actual VoIP traffic is not simply CBR, but depends on the voice signal power that VAD technique detects, and codecs only code voice signal when the power is above the threshold which VAD technique recognizes as the lowest human voice level. As a result, VoIP traffic has an onoff pattern according to speaker s behavior, so using CBR to generate VoIP traffic is not realistic. The second method is to code real voice files (usually in.wav format) with a specific kind of voice codec, and then use the coding result as VoIP traffic. Although in this way VAD technique can be adopted to detect voice, the problem is that the generated VoIP traffic is limited by the number and size of voice source files, and more important, that a voice file can only represent monolog and be used to generate single speaker s VoIP traffic, but can not reflect the interactive process between two speakers in a conversation. As it ignores two speakers interactivity in a VoIP stream, coding voice files can not generate representative VoIP traffic either. To conclude, existing VoIP traffic generation methods lack reality and interactivity, so they can not simulate real VoIP products traffic correctly. If we use these methods in measurement activities, it is obvious that the accuracy of measured quality will be damaged, and thus the expected objective of VoIP measurement can not be achieved. Focusing on the current methods deficiency illustrated above, we consider providing a more effective VoIP traffic generation algorithm in this paper. To achieve this goal, we need to analyze the traffic characteristics of VoIP applications. Firstly, traffic of VoIP applications is different from traffic of other traditional applications, as it is affected by user behavior not only when the stream is opened or closed, but also in the whole duration of the stream. Because of the use of VAD technique, traffic sending rate has a positive correlation with user s real talking duration, so it is necessary to consider state change of user s talking and silent duration. In addition, as VoIP is an interactive application between two users, traffic generation should consider the effect of these two speakers interactive relationship. Summing up these two requirements, this paper proposes a conversational model to simulate two

2 speakers conversational state change and interactive process, analyzes how user behavior, voice codec and VAD technique affect VoIP traffic, and generates more realistic and interactive VoIP traffic, which we believe can improve the accuracy and reliability of VoIP quality measurement. The remainder of this paper is organized as follows. Section 2 outlines related work in VoIP measurement and VoIP traffic generation. Section 3 presents our conversational model based VoIP traffic generation algorithm. Experiments on our algorithm s validity and practical usage are discussed in Section 4. Finally, we make concluding remarks in Section Related Work There have been numerous studies on VoIP measurement. A. Markopoulou [16] measured loss and delay characteristics of American backbone networks, and analyzed how these characteristics impact VoIP quality. As he did not actually generate VoIP traffic to transmit on measured networks, the result was indirect to illustrate VoIP quality. I. Marsh [17] generated VoIP traffic by coding per-recorded voice files with G.711 codec, transmitted it between pairs of 9 sites in different continents, and analyzed VoIP quality through Emodel [12]. He claimed that VoIP quality had been improved from 1999 to 2002, and that it showed especially good results in academic networks. Although this work provided some valuable statistics of VoIP quality, the traffic used in the measurement was generated by coding voice files, which could not simulate two speakers interactive process in VoIP applications. As it could not represent real VoIP traffic, the accuracy of the results needs to be investigated. Besides these active measurement activities, there were studies on passive measurement of VoIP quality, in which most work focused on monitoring and analyzing performance of actual applications, like MSN and Skype [7][8]. However, as VoIP applications usually encrypt their payload and choose random UDP port, it is difficult to accurately capture all packets of a VoIP stream. Correspondingly, estimation of characteristics like loss, delay and jitter may be affected and not accurate. Besides, passive measurement is not reproducible and may invade users privacy. For these reasons, we believe that active measurement is more reasonable and useful for VoIP quality assessment. However, existing active measurement work can not generate realistic and interactive VoIP traffic, which damages the accuracy and reliability of measurement results. Because of the importance of traffic generation in VoIP quality measurement, there have also been some specific tools for VoIP traffic generation. D-ITG [2] is a popular traffic generation tool in academic society, which adopts CBR method for the generation of VoIP traffic. In network testing industry, Spirent s Abacus 5000 [1] and Agilent s NetworkTester [20] can use various codecs to code.wav format voice files into packet stream, and Abacus 5000 is able to configure VAD parameters. However, as illustrated in Section 1, both CBR and voice file coding methods have problems in VoIP traffic s reality and interactivity, so they can not accurately simulate actual VoIP applications traffic. The fundamental reason for the problems existing in VoIP traffic generation methods above is that they ignore user behavior and its impact on VoIP traffic. Indeed, in telecommunication area, researchers began to study user behavior s influence on telephone system very early. P. T. Brady analyzed user behavior in trans-atlantic phone calls in 1960s, and designed a six-state user conversational model [4] to describe the transition process of users talking, double talking (two speakers are talking simultaneously) and silent states. ITU-T Recommendation P.59 [21] simplified this model into 4 states to describe the same process. These models efficiently reflect two users interactive process and speech characteristics in a telephone call, but they can not be used directly in VoIP traffic generation. Firstly, models in telephone system are usually based on continuous time, but for generating packets of VoIP traffic in IP networks, discrete-time model is more suitable and flexible to use. Additionally, VAD technique in telephone system usually uses constant parameters, so in models introduced above no consideration has been given to the variation of VAD parameters. However, advanced VAD technique used in VoIP applications can tune parameters to meet requirements in different scenarios, so VoIP traffic generation needs to consider the impact of variation of VAD parameters. For these reasons, generated traffic in [11][19] was not realistic enough as they used P.59 directly in the wireless networks. There were also studies on the modeling and analysis of on-off pattern for single VoIP user end [6][9][10][15]. But as these models are without consideration of conversational interactivity between two users in a VoIP stream, they can not be used to generate two-way interactive VoIP traffic either. As a result, we need a new conversational model to generate VoIP traffic, which is suitable for packetized IP networks as well as can consider the influence of VAD parameters variation. 3. VoIP Traffic Generation Algorithm Voice codec parameters Transition probability (Section 3.1) Start User behavior parameters User behavior simulation (Section 3.1) Time point determination (Section 3.2) Traffic trace generation (Section 3.3) End Conversational model (Section 3.1) VAD technique parameters Figure 1. Framework for VoIP traffic generation algorithm

3 Conversational model based VoIP traffic generation algorithm introduced in this paper synthetically considers the influence of user behavior, VAD technique and voice codec to VoIP traffic. As shown in Fig. 1, after setting up all input parameters, first we construct our conversational model and calculate transition probability to describe two users behavior and their interactive process. Then based on the simulation result of this model, VAD technique is considered and the time points to generate packets are determined. Finally, according to the specific codec s attributes, VoIP packet size at every determined time point is decided and VoIP traffic traces are generated. 3.1 Conversational model construction The conversational model introduced in this paper is shown in Fig. 2. Its has the same 4 states as ITU-T Recommendation P.59, but in P.59, the duration of every state is simulated in continuous time, which is not suitable for VoIP. Additionally, P.59 doesn t provide details on how to generate the equation for every state s duration time from input parameters, so if the parameters are changed in different scenarios (for example, conversations in different languages may have different parameter values), corresponding duration time can not be recalculated. For these reasons, we design a discrete-time statetransition model in this paper. Its 4 states respectively represent user A s talking (state 1), user B s talking (state 2), double talking (state 3), and mutual silence (state 4). Transition of states happens at discrete time points, as VoIP packets should be generated discretely, and P ij represents the probability of transiting from state i to state j. At every discrete time point, a random number between 0 and 1 is generated, and according to the current state and transition probability, this number is used to choose the state at the next discrete time point. Following this process, the model can determine the state at every discrete time point in the measurement duration, and thus two users behavior can be presented as a series of states at discrete time points. P11 P33 State 1: A is talking B is silent State 3: A is silent B is silent P14 P41 P32 P23 State 4: A is talking B is talking P31 P13 P24 P42 State 2: A is silent B is talking Figure 2. Discrete-time conversational model Transition probability is the key factor of the whole conversational model. As in our conversational model, state transition happens at discrete time points based on transition probability, the possibility that state i remains for consecutive k time points before transiting to another state at k+1 time point follows a geometric distribution: k 1 Pk i( ) = pii (1 p ii ) In [5], P. T. Brady modeled user s talking and silent duration as exponential distribution through experiments. Although there P44 P22 have been controversies on this model, it is still commonly used. As a discrete analog of exponential distribution, geometric distribution is then suitable to simulate the duration of user behavior in discrete-time model. In it, the expectation of k is calculated as below: Ei( k) = 1/(1 pii) This value should be proportional to the average duration on the state i from user behavior parameters, which should be gathered by experiments from real human conversations. With these parameters, the loopback transition probability for every state can be calculated by following equations: p11 = p22 = 1 Tinterval / Ttalk avg (1) p33 = 1 Tinterval / Tstop avg (2) p44 = 1 Tinterval / Tdouble avg (3) Within these equations, T interval represents the duration between two consecutive discrete time points, and its configuration should refer to the specific voice codec given in voice codec parameters. For example, if we choose voice codec G.711 in the measurement, T interval can be configured using default value, 20 milliseconds. T talk-avg,t stop-avg,and T double-avg respectively represent the average duration of state 1 (same as state 2), state 3 and state 4. These values are configured in user behavior parameters and can be retrieved by measuring human conversations. For example, P.59 provided second, second and second as the parameters for these three average durations, which were average results from conversations in English, Italian, and Japanese. After the calculation of loopback transition probabilities, it is time to consider the transition probabilities between different states. As the probabilities should satisfy the following equation: p ij = 1 j and because transitions from state 3 to state 1 and state 2 are symmetric, their probabilities satisfy equation: p31 = p32 = (1 p33) / 2 (4) The same situation happens for possibilities from state 4 to state 1 and state 2: p41 = p42 = (1 p44 ) / 2 (5) Because of symmetry, transition probability from state 1 to state 3 equals with the one from state 2 to state 3, and transition probability from state 1 to state 4 equals with the one from state 2 to state 4. Here we use K talk-double to represent the probability that state 1 (or state 2) transmits to state 4 on the condition that state change happens. Then the equations are: p14 = p24 = (1 p11) Ktalk double (6) p13 = p23 = (1 p11) (1 Ktalk double ) (7) Ktalk-double is also from user behavior parameters, gathered from statistics on human conversations. P.59 suggested this parameter should be configured as 0.6. All transition probabilities required in our conversational model can be calculated by equations (1) (7), with which two users behavior in a conversation can be modeled. Fig. 3 shows a segment of user s conversational interactive process simulated by our model, with user behavior parameters configured as P.59 suggested and T interval as 20 milliseconds. It

4 describes the states for all discrete time points from 35th second to 41st second in the modeling period, in which marked node at the time point means that at that specific time the user is talking. Fig. 3 illustrates the interactive process between two users and each user s talking and silent duration. We will show that the generated result is consistent with user behavior in Section 4. Now we add the influence of hangover time to the conversational model introduced in Section 3.1. When transmitting from state 1 to state 3, or from state 4 to state 2, user A should send packets in configured hangover time; when transmitting from state 2 to state 3, or from state 4 to state 1, user B should send packets in configured hangover time. Fig. 5 shows the result after considering VAD hangover time (set 200 milliseconds here) in the conversational model. Here we put time points marked by VAD handover time a little higher than those marked by the conversational model. Compared with Fig. 3, we can see more traffic needs to be generated as more time points are marked, and the actual amount of extended traffic is due to VoIP hangover time configuration. We will give a quantitative analysis for this in Section 4. Figure 3. Simulation result by the discrete-time model 3.2 VAD hangover time operation VAD technique adopted in VoIP is quite different from the one used in traditional telephone system. In telephone system, parameters for VAD technique are constant, while in VoIP parameters such as hangover time and energy threshold can be configured manually[15]. For VoIP traffic generation, configuration of hangover time can have a significant impact. The basic function of hangover time is shown in Fig. 4, X-axis representing time and Y-axis representing voice energy that VAD detects. From T0 to T1, the energy is above the threshold that VAD recognizes as the lowest human voice level, so VoIP traffic is generated. After T1, the energy decreases below human voice level, but during the period from T1 to T2, called hangover time, VoIP traffic is still generated. The reason why VoIP traffic does not stop just at time point T1 is due to the existence of background sound in a VoIP call. If the traffic immediately stopped at T1, the user on the other side would feel the abrupt disappear of background sound, which would make him feel uncomfortable and damage VoIP quality, so it is necessary for the traffic to last some time after a user stops talking. It is obvious that the configuration of hangover time influences bandwidth consumption of VoIP applications and must be considered in VoIP traffic generation. Energy threshold configuration also affects the amount of VoIP traffic, but as it is much less influential compared to hangover time and quite difficult to model, we ignore it in this paper. Power Threshold Hangover Time T0 T1 T2 Figure 4. Usage of VAD hangover time Figure 5. Simulation result with VAD hangover time 3.3 Traffic trace generation Based on the conversational model and hangover time operation introduced above, it can be determined whether a VoIP packet needs to be generated at each time point, so now it is necessary to consider how to construct payload format for each packet. The payload format is composed by payload content and payload size. For payload content, as routers and switches on the network do not change their operation to VoIP packets due to payload content, it can be filled by random data in measurement. For payload size, according to the type of voice codec and interval between time points, it is calculated by the following equation: 3 Rcodec 10 Spayload = Tinterval 8 S payload represents the calculated payload size (in bytes), R codec represents the voice codec sending rate (in Kbps), and T interval represents the interval between time points (in seconds). For example, if we use G.711 codec (64Kbps) and the interval is configured as 20 milliseconds, according to the equation, the payload size should be 160 bytes; if we use G.729 (8Kbps) with the same time interval, the payload size should be 20 bytes. Now, packet sending time (Section 3.1 and Section 3.2) and packet payload format (Section 3.3) of VoIP traffic for two users in a conversation have all been determined. For the convenience of measurement work, we store the traffic information into two trace files, each representing one user in the conversation. Table 1 shows a segment of a trace file. For every record in the file, there is information on packet sequence number, packet sending time and packet payload size, based on which a measurement node can generate proper sized packet at proper time to simulate the traffic of a VoIP application.

5 Table 1 Segment of a traffic trace file Number Sending time (ms) Payload size (byte) Generated trace deployment To be used in the actual measurement activities, traffic generated following above steps needs to be deployed in the measurement environment. Firstly, locations of measurement nodes in the networks are determined and workstations for transmitting VoIP traffic are deployed at these locations; then these workstations are organized into pairs as the two ends of a VoIP stream and generated traffic trace files are distributed to them; after that, two workstations in a pair negotiate UDP ports for transmitting traffic and the start time of the measurement; finally, according to pre-stored traffic trace files, workstations generate VoIP traffic, encapsulate with RTP [22] header, and transmit via UDP ports. Characteristics such as loss, delay and jitter can be gathered after the transmission and VoIP quality can be evaluated based on these characteristics and standards like Emodel. The scale of VoIP traffic used in the measurement is decided by its objective. If the objective is to estimate the maximum number of VoIP streams which can be supported in a specific network before actual deployment, large-scale traffic of VoIP streams needs to be deployed in multiple nodes in the network, and in this case, special network testing equipments can be used instead of normal workstations to improve the traffic handling capacity; if the objective is to monitor VoIP quality variation in the network, single VoIP stream can be transmitted between specifically chosen measurement nodes and normal workstations are capable to transmit the traffic traces. 4. Experiments and Analysis In this section, we first illustrate the validity of our conversational model, then give a quantitative analysis on the influence of VAD hangover time to VoIP traffic, and at last provide actual measurement results using the generated VoIP traffic in our lab and campus networks. It is a critical requirement that the simulation result from our conversational model is consistent with user behavior. Here we use experimental data from ITU-T Recommendation P.59 as user behavior parameters, so the simulation results for every state s duration, which represent user s talking and silence situation, should be consistent with these parameters. Fig. 6 shows the average duration for every state simulated by our conversational model under different measurement durations. It can be seen that when the whole measurement duration is less than 200 seconds, as the amount of simulation data is not enough, the average durations have some difference with user behavior parameters, and have an intense oscillation; when the whole measurement duration is over 200 seconds, the average durations simulated by our model are approximate to user behavior parameters, and the difference percentage is always below 10%, so user behavior in a conversation can be simulated correctly. To conclude, it is necessary to generate enough amount of traffic in the measurement to improve the accuracy of our conversational model. Based on this experiment, the whole measurement duration should be more than 200 seconds to generate enough traffic, which can guarantee that the simulation result for each state is consistent with experimental parameters for user behavior. Figure 6. Average duration for each state in the model As illustrated in Section 3.2, configuration of hangover time can mitigate the effect of abrupt disappear of background sound at the listening side, but at the same time it increases the bandwidth consumption of a VoIP stream. Here we provide some quantitative analysis for its effect. We construct G.711 and G.729 VoIP traffic for 2000 seconds, in which hangover time is configured from 0 second to 5 seconds. Fig. 7 shows the traffic stream rate for varying hangover time. From it we can see that stream rate of G.711 changes from 26.68Kbps to 62Kbps, with an increased percentage of 132.5%, and it is getting close to G.711 s coding rate 64Kbps; stream rate of G.729 changes from 3.5Kbps to 7.77Kbps, with an increased percentage of 121.6%, and it is getting close to G.729 coding rate 8Kbps. These results present hangover time s great impact to the amount of VoIP traffic, so considering its influence in traffic generation is very important. Figure 7. Analysis of VAD hangover time s impact to VoIP traffic Using the generated traffic trace files, we do some preliminary measurements in both our laboratory and campus

6 networks. In the lab environment, we select two computers in the same laboratory building, and in the campus environment, we choose one computer in the dormitory building and the other in the laboratory building. We generate VoIP traffic using our algorithm with the interval for time points configured as 20 milliseconds, user behavior parameters from P.59, VAD hangover time configured as 200 milliseconds, and G.711 and G.729 codecs. Loss rate and jitter are calculated using RTP header s sequence number and timestamp, and RTT delay is calculated using RTCP [22] SR and RR packets. Table 2 shows the results of our measurement. Results for lab environment are better than those for campus, as we expect. But what is surprising is that the two paths for campus measurement have quite different characteristics. For example, traffic from A (dorm node) to B (lab node) has less than 0.01% loss rate, while traffic from B to A has average 0.36% loss rate. Although the reason for this phenomenon has not been clarified, it shows that two users in the same VoIP stream may feel quite different qualities, and thus illustrates the importance to generate interactive VoIP traffic to simultaneously evaluate the qualities for both paths in the measurement. Codec L G.711 a b G.729 C a m p u s G.711 G Conclusion Table 2 Measurement result using generated traffic Speed Loss rate RTT Delay Jitter Path (Kbps) (%) (ms) (ms) A->B B->A < A->B 3.62 < B->A 3.95 < A->B < B->A A->B 3.62 < B->A This paper begins with the current situation of VoIP measurement, analyzes the deficiency of existing VoIP traffic generation methods, and then proposes our conversational model based VoIP traffic generation algorithm. We use a new conversational model to simulate two users behavior in a VoIP application, and generate realistic and interactive VoIP traffic according to specific VAD parameters and voice codec. This generation algorithm is suitable for VoIP measurement and can help improve its accuracy and reliability. From the evolvement of the Internet, we can see that user behavior has had more and more influence on traffic pattern. The traffic of applications including VoIP, video conference and online game has a close correlation with user s activities. Traditional traffic generation methods usually only focus on protocol behavior but ignore user behavior, so that they can not generate effective traffic to measure and evaluate these applications. For this reason, we will continue to analyze the impact of user behavior in these applications, use modeling methods to simulate it and generate representative traffic for measurement and performance evaluation. 6. References [1] Spirent Abacus 5000, [2] S. Avallone, A. Pescapè, and G. Ventre, "Distributed Internet Traffic Generator (D-ITG): analysis and experimentation over heterogeneous networks", in Proc. of ICNP, Atlanta, Georgia (USA), November [3] S. A. Baset, and H. Schulzrinne, "An analysis of the Skype peer-to-peer internet telephony protocol", In Proc. of IEEE INFOCOM, Barcelona, Spain, Apr [4] P. T. Brady, "A model for generating on-off speech patterns in two-way conversation", Bell Syst. Tech. J., Sept , pp [5] P. T. Brady, "A technique for investigating on-off patterns of speech", Bell Syst. Tech. J., Jan. 1965, pp [6] E. Casilari, H. Montes, and F. Sandoval, "Modelling of Voice Traffic Over IP Networks", CSNDSP 2002, Network Communications K1.5, July [7] K. Chen, C. Huang, P. Huang, and C. Lei, "Quantifying Skype user satisfaction", In Proc. of ACM SIGCOMM, Pisa, Italy, [8] W. Chiang, et al., "A performance study of voip applications: Msn vs. skype", In Proc. of MULTICOMM, [9] C. Chuah, and R. Katz, "Characterizing packet audio streams from internet multimedia applications", in Proc. of the International Conference on Communications, [10] T. Dinh, B. Sonkoly, and S. Molnár. "Fractal Analysis and Modeling of VoIP Traffic", In Proc. of International Telecommunication Network Strategy and Planning Symposium (Networks), pvienna, Austria, June [11] Y. Gwon, J. Kempf, R. Dendukuri, and R. Jain, "Experimental Results on IP-layer Enhancement to Capacity of VoIPv6 over IEEE b Wireless LAN", In Proc. of WiNMee, [12] ITU-T Recommendation G.107, "The Emodel: A computational model for use in transmission planning", Dec [13] ITU-T Recommendation G.711, "Pulse code modulation (PCM) of voice frequencies", [14] ITU-T Recommendation G.729, "Coding of Speech At 8 kbit/s Using Conjugate-Structure Algebraic-Code-Excit ed Linear-Prediction(CS- ACELP)", Geneva, Mar [15] W. Jiang, and H. Schulzrinne, "Analysis of on-off patterns in VoIP and their effect on voice traffic aggregation", In Proc. of Computer Communications and Networks, Las Vegas, NV, USA, [16] A. Markopoulou, F. Tobagi, and M. Karam, "Assessment of VoIP quality over Internet backbones", in Proc. of IEEE INFOCOM, New York, NY, Jun [17] I. Marsh, F. Li, and G. Karlsson, "Wide area measurements of VoIP quality", in Pro. of QoFIS, [18] D. Minoli, and E. Minoli, Delivering voice over IP networks. John Wiley & Sons, Inc., New York, NY, [19] S. Shin, and H. Schulzrinne, "Balancing Uplink and Downlink Delay of VoIP Traffic in WLANs using Adaptive Priority Control(APC)", In Proc. of QShine, Waterloo, Canada, Aug [20] Agilent NetworkTester, [21] ITU-T Recommendation P.59, "Artificial conversational speech", [22] H. Schulzrinne, S. Casner, V. Jacobson, and R. Frederick, "RTP: A Transport Protocol for Real-Time Applications", RFC 3550, June 2003.