2013 International Conference on Computing, Networking and Communications, Communication QoS and System Modeling

Symposium

Monitoring VoIP Call Quality Using Improved

Simplified E-model
Haytham Assem, David Malone Jonathan Dunne, Pat O’Sullivan
Hamilton Institute, National University of Ireland, Systems and Performance Engineering, IBM Dublin,
Maynooth Software Lab
Hitham.Salama.2012, [email protected] Jonathan_dunne, [email protected]

Abstract—ITU-T recommendation G.107 introduced the E- of the listeners are considered an important factor in
model, a repeatable way to assess if a network is prepared to estimating accurate scores. Thus, subjective testing using
carry a VoIP call or not. Various studies show that the E- MOS is time consuming, expensive and does not allow real
model is complex with many factors to be used in monitoring time measurement. Consequently, in recent years new
purposes. Consequently, simplified versions of the E-model methods were developed for measuring MOS scores in an
have been proposed to simplify the calculations and focus on objective way (without human perception): PESQ [4], E-
the most important factors required for monitoring the call model [5] and several others.
quality. In this paper, we propose simple correction to a PESQ, Perceptual Evaluation of Speech Quality, is
simplified E-model; we show how to calculate the correction considered an objective method for predicting the speech
coefficients for 4 common codecs (G.711, G.723.1, G.726 and
quality. It is an intrusive testing method which takes into
G.729A) and then we show that its predictions better match
PESQ scores by implementing it in a monitoring application.
account two signals; one is the reference signal while the
other one is the actual degraded signal. Both signals are sent
Keywords-VoIP; E-model; PESQ; Monitoring through the test that uses a PESQ algorithm and the result is
a PESQ score. Consequently, this approach cannot be used to
I. INTRODUCTION monitor real time calls.
Nowadays, a new objective method proposed by TU-T
The evaluation of data networks depends on several G.107 [5] defines the E-model, a mathematical model that
factors. Thus, it is argued that it is not appropriate to use a combines all the impairment factors that affect the voice
single metric to evaluate the quality of data networks. Yet in quality in a single metric called R value that is mapped to
the telephony world, a single number is typically given to MOS scale. The E-model was designed to provide estimated
rate call quality. Such value is used as a basis of monitoring network quality and has shown to be reasonably accurate for
and tuning the network. Voice over Internet Protocol (VoIP) this purpose. It has not been accepted as a valid measurement
is an example of such data network application [1]. tool for live networks. The ITU-T G.107 Recommendation
In previous years, VoIP has become an important [5] states at the beginning of the document that “it is
application and is expected to carry more and more voice considered only estimates for the transmission planning
traffic over TCP/IP networks. In real-time voice applications, purposes and not for actual customer opinion prediction”
the speech quality is impaired by the packet loss, jitter, delay unlike the PESQ [4] which is developed to model subjective
and bandwidth. Consequently, VoIP applications require low tests commonly used in telecommunications to assess the
delay, low packet loss rates, low jitter and sufficient voice quality by human beings.
bandwidth in order not to affect the interaction between call Increasingly and against ITU recommendations, the E-
participants. model is being used nowadays by industry and research as a
VoIP is based on IP network; however IP networks live voice quality measurement tool. Thus, simple versions
frequently provide best effort services, and may not of E-model [1, 6] have been proposed to simplify the
guarantee delay, packet loss, and jitter [2]. So, the prediction complexity of the original E-model [5] and focus on most
of voice quality in different environments and traffic loads important part that affect the VoIP call quality.
may be as important part of network monitoring in order to The objective of our work is to provide a monitoring
measure voice quality and prevent critical problems before system using a simplified version of the E-model corrected
they occur. for 4 common codecs to better predict PESQ MOS scores as
As measuring voice quality is important to the service PESQ is generally considered to provide more accurate
providers and end users, ITU-T provides two test methods predictions of user experience than the E-model.
subjective and objective testing. Subjective testing was This paper is organized as follows: Section 2 describes
considered the earliest attempts on this issue to evaluate the the proposed improved simplified E-model. In Section 3 we
speech quality by giving Mean Opinion Scores (MOS). The show how we derived the correction coefficients used in the
MOS test is one of the widely known accepted tests that give improved simplified E-model. In Section 4 we propose our
a speech quality rating. ITU-T Rec. P.800 [3] presents the results using the derived model by implementing it in a
MOS test procedures as users can rate the speech quality monitoring application. Finally, we conclude and summarize
from 1(Poor) to 5 (Excellent) scale. Of course, the numbers
the paper in Section 5.

978-1-4673-5288-8/13/$31.00
Authorized licensed use limited to: Institut©2013 IEEE
Teknologi Bandung. Downloaded927
on September 14,2021 at 03:57:06 UTC from IEEE Xplore. Restrictions apply.
 II. IMPROVED SIMPLIFIED E-MODEL (4)
In this section, we will first give a brief description of the
simplified E-model [6] and then we will describe our TABLE I. CODECS SPECIAL COEFFICIENTS
proposed improvements to the simplified E-model with the Codec a b c
method of calculation of the various parameters used in the
model in order to be applicable in monitoring purposes. G.711 0.18 -27.90 1126.62
G.723.1 5.3k 0.039 -4.2 166.61
A. Simplified E-Model
G.726 24k 0.046 -4.53 168.09
The original E-model is very complex [7] and involved
with many factors. Moreover, the voice processing is not G.729A 0.063 -8.08 311.72
related significantly to the instantaneous judgment of QoS.
Thus, a simplified version of the E-model [6] has been
introduced to focus on the most important parts and 1.1)
afterwards it was used in a monitoring system [2]. This is the basic signal to noise ratio, including noise
model takes in to account the codec and the present network sources such as circuit and room noise. However, currently it
conditions which are the main two factors that affect the is really difficult to calculate directly. Thus, ITU-T
voice quality. The simplified E-model is expressed by
equation (1) by calculating the evaluation value R. G.113 [8] provides the common value of . Since, the
inherent degradation that occurs when converting actual
R= R0 – Icodec – Ipacketloss – Idelay (1) spoken conversation to a network signal and back reduces
the theoretical maximum R-value (94.2) with no
Where R0 represents the basic signal to noise ratio, Idelay impairments to 93.2 [5]. So, we set the R0 value to 93.2.
represents the delays introduced from end to end, Icodec is 1.2)
the codec factor and the Ipacketloss is the packet loss rate
within a particular time. Finally the R value is mapped to is the equipment impairment (codec quality) factors as
MOS score. defined in [8] and [9]. It represents the codec distortion
which leads to voice distortion and impairments arising
B. Improved simplified E-Model because of signal conversions. Nowadays, its value is
The objective of this model is to determine the voice determined by looking up the codec in the ITU-T
quality MOS rating by a simplified modified version of the Recommendation G.113 literature [8] as Table II is part of it.
previous E-model described above. The computational
model consists of a mathematical function of parameters of TABLE II. SOME CODING INFORMATION
the transmission system. The computation itself can be split Encoder References BitRate Ie value
into several elements and can be expressed by the following Type
equation (2). (Kbit/s)
PCM G.711 64 0
(2)
ACELP G.723.1 5.3 19

Where is a second order function corrected using curves ADPCM G.726 24 25

fitted to PESQ scores which is the standard objective method CS-ACELP G.729A 8 11
defined by ITU-T recommendation P.862 [4], is the
average delay time within specified period and A is the
expectation factor due to the communication system. The 1.3) :
description and method of calculating the previous is the packet loss percentage within a particular
parameters ( , and A) in (2) are as follows: period measured by certain number of packets. The
percentage measured is the loss of packets occurred when the
1) : sender’s packets is not received by the receiver. It can be
as mentioned above is a second order function model expressed by the following formula (5).
corrected with PESQ scores to obtain more accurate results
in our monitoring system. Ry can be expressed by the (5)
following equation (3).
(3) Where DS is the difference between the largest and smallest
sequence number of N packets. Statistics and calculation of
Where is a part of the simplified E-model (1) which is the Real-time Transport Protocol (RTP) packets can be used
corrected with PESQ scores, can be obtained by the to calculate this percentage by the following expression (6).
following expression (4) and a, b, c are codecs coefficients as DS=LS-SS+1 (6)
shown in Table I and derived in section III.

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded928

on September 14,2021 at 03:57:06 UTC from IEEE Xplore. Restrictions apply.
Where LS and SS are the largest and smallest sequence
numbers respectively. They are extracted from the RTP
header of the sequence number field from the packets
received.
2) :
The delay components contributing to provided in
ITU-T G.107 [5] are , the average absolute one way
mouth to ear delay. T, the average one way delay from the
receive side to the point in the end to end path where a
signal coupling occurs as a source of echo. , the average
trip delay in the 4 wire loop. G.107 [5] gives a fully
analytical expression for the function , interms of , T, Figure 1. versus one-way delay
and parameters associated with a general reference
connection describing various circuit switched and packet 3) A:
switch inter-working scenarios. Assuming perfect echo The advantage factor, A represents an “advantage of
access”, introduced into transmission planning for the E-
cancellation, all the factors in can be collapsed in a
model (ITU-T G.107) [5]. This value can be used directly as
single points as shown in (7) and (d) is now function only an input parameter to the E-model. Provisional A values are
of the one way delay d. (d) can be calculated by a series listed in [5] as show in Table III. Assuming our
communication system is conventional then we neglect A
of complex equations in ITU-T G.107 [5] as shown with the value.
plotted curve of vs one way delay in Fig.1 (labeled
“G.107”). TABLE III. EXAMPLES FOR THE ADVATAGE FACTOR A
Communication System Maximum value
(7) of A
Conventional (wire bound) 0
Mobility by cellular networks in a 5
The one way delay (d) is the time it takes to get data building
across the network. The one-way delay measured from one Mobility in a geographical area or 10
end of the network to the other end is mainly composed of moving in a vehicle
four components that can be expressed in equation (8). Access to hard-to-reach locations, e.g., 20
via multi-hop satellite connections
d=t0+t1+t2+t3 (8)
The R value of the E-model is finally transformed to
Where t0 is the propagation delay, t1 is the transport delay, MOS score that will reflect the user level of satisfaction as
t2 is the packetization delay and t3 is the jitter buffer delay. shown in Table IV, theoretical range of transmission
In this paper we approximate these four components by performance rating factor R from 0 to 100. R=0 represents of
measuring the response time (round-trip delay) as in most the worst quality and R=100 represents the best quality. The
modern devices t1, t2 shall be small. Thus, ping should be R factor value for estimated average score of MOS can be
reasonable. expressed by equation (10).

In our model we used a simplified version of (9) as For R<0: MOS=1

provided in [10]. This model shows accuracy for one way For
delay less than 400ms as shown in Figure 1 (labeled “AT&T
simplified model”). We found this model reasonable as ITU- For R>100: MOS=4.5 (10)
T recommend that one-way delay should not be more than
150 ms for good speech quality[11].
TABLE IV.RELATIONSHIP BETWEEN R-VALUE AND USER’S SATISFACTION
R-Value Satisfaction Level MOS
(9) 90-100 Very satisfied 4.3+
80-90 Satisfied 4.0-4.3
H(x) =0,if x<0 70-80 Some users dissatisfied 3.6-4.0
Where 60-70 Many users dissatisfied 3.1-3.6
50-60 Nearly all users dissatisfied 2.6-3.1
H(x) =1,if x>=0 0-50 Not recommended 1.0-2.6

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded929

on September 14,2021 at 03:57:06 UTC from IEEE Xplore. Restrictions apply.
 III. CORRECTION OF THE SIMPLIFIED E-MODEL
In this section we show how we derived the values of a, b
and c (Table I) used in our improved simplified E-model
described in the previous section.
In our experiment shown in Figure 2 we have developed
a java program that stream RTP packets using 4 main audio
codecs (G.711, G.726, G.723.1 and G.729A). We recorded
the voice at both ends and measured the PESQ scores under
different random packet loss rate ranges from 0-20%. For
each packet loss rate, we repeated the experiment 10 times
taking the average MOS PESQ score in order to increase the
accuracy of the results as much as possible.
Figure 4. Relationship between and (G726 and G729A)

IV. MONITORING SYSTEM DESIGN AND RESULTS

The monitoring system could target specific number of
RTP packets to capture and perform an effective MOS value
calculation. The system will use a coefficient database for
the codec used in the call. This monitoring system is
developed for monitoring VoIP quality at the network
terminals, and the environment could be a personal or family
network with voice quality monitoring.
Figure 2. Deriving codecs coefficients a,b and c The whole system works as follows: The system uses
network capturing module to capture a certain number of
The PESQ scores are converted from MOS to R value packets passed to specific IP and port. Non RTP packets will
and this can be conducted by a complicated Candono’s be filtered. When this process completes the packet capture,
Formula as in [12] or by the simplified 3rd-order polynomial the system will analyze the data, delay and packet loss rate as
fitting [13] as shown in (11). described previously in section II. The MOS score is
calculated to assess voice call quality in this period of the
call. We took our results online with introducing random
(11) packet loss rate in the network using Dummynet [14].
We compared our monitoring system using MOS scores
The converted PESQ scores from (11) will represent the based on the codec’s coefficients (see Table I) derived for 4
values shown in the below graphs (Fig.3-Fig.4) on the y main codecs with the simplified version of the E-model that
is used in monitoring purposes [1, 6] and the PESQ scores.
axis. Since PESQ does not take the delay factor in its
The graphs (Figure 5-Figure 8) show our results for the 4
account, so we correct the rest part of the model which we
codecs. It can be observed that the MOS scores of our
name it (see equation 3) represented on x axis. improved simplified E-model based on the coefficient
We found that it is well matched second order relation database (Table I) are very close to the PESQ scores unlike
function and then we derived the coefficients a, b and c as in the simplified E-model which gives an advantage for the
Table I using least-squares fitting method. The graphs below corrected model in monitoring purposes for the VoIP call
(Figure 3-Figure 4) show the correlation between the quality.
converted values from PESQ and the R values from the
simplified E-model for individual codecs in different loss
ranges.

Figure 3. Relationship between and (G723.1 and G711)

Figure 5. Comparative Analysis (G.723.1)

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded930

on September 14,2021 at 03:57:06 UTC from IEEE Xplore. Restrictions apply.
 Consequently, we have proposed an improved simplified
E-model and show how we derived the coefficients used in
the model for 4 common codecs (G.711, G.723.1, G.726 and
G.729A). We demonstrate its results by implementing it in a
monitoring system; our system analyzes the impact of voice
quality encoding factors under various network conditions
and uses our simplified improved E-model to assess voice
quality. The main advantage of our improved simplified
version that, it is less complex than the original E-model
model and it is more accurate than the simplified versions
used.
We stress three benefits of our work. The first as
confirmed by the experiment, the simplified version of E-
model does not provide accurate results compared to PESQ
scores. The second, the correction coefficients derived
Figure 6. Comparative Analysis (G.711) enhance the simplified E-model to monitor/predict the call
quality. The third, proposing a complete design of
monitoring system using our improved simplified E-model
for 4 common codecs. Another output of our work is a java
application that stream RTP packets using number of codecs.
ACKNOWLEDGMENT
The authors were supported by Science Foundation Ireland
(SFI) grants 07/SK/I1216a and 08/SRC/I1403.
REFERENCES
[1] John Q. Walker, “Assessing VoIP Call Quality Using the E-model”,
NetIQ Corporation.
[2] Junsheng Zhang and Xiaohua Sun, “The VoIP phone QoS protection
in the wide-are network”, Computer learning .2006 No.6 .17-18.
[3] ITU-T Recommendation P.800, “Methods for subjective
determination of transmission quality”, Geneva, 08/1996.
[4] ITU-T Recommendation P.862, “Perceptual evaluation of speech
Figure 7. Comparative Analysis (G.726) quality(PESQ): An objective method for end-to-end speech quality
assessment of narrow-band telephone networks and speech codecs”,
Febrauary 2001.
[5] ITU-T Recommendation G.107, “The E-model: A computational
model for use in transmission planning”, Geneva, 04/2009.
[6] Chunlei Jiang and Peng Huang, “Research of Monitoring VoIP Voice
QoS”, International Conference on Internet Computing and
Information Services, 2011.
[7] Pystechnics Limited, “The E-Model, R Factor and MOS”, 23
Museum street Ipswitch, Suffolk United Kingdom. December 2003.
10-41.
[8] ITU-T Recommendation G.113, “Transmission impairments due to
speech processing”, 2001.
[9] ITU-T Recommendation P.833, “Methodology for derivation of
equpment impairment factors from subjective listening-only tests”,
2001.
[10] R.G.Cole and J.Rosenbluth, “Voice over IP performance monitoring”,
ACM comput. Commun. Rev., vol. 31, no. 2,pp. 9-24, April 2001.
Figure 8. Comparative Analysis (G.729A) [11] S.Pracht and D.Hardman: “Voice Quality in Converging Telephony
and IP Networks, Agilent Technologies”, White Paper,
http://literature.agilent.com/litweb/pdf/5980-0989E.pdf
V. DISCUSSION AND CONCLUSION
[12] C. Hoene, H. Karl, and A. Wolisz, “A perceptual quality model for
The E-model brings a new approach to the computation adaptive VoIP applications”, Int. Symp. Performance Evaluation of
of estimated voice quality. The main advantage of using E- Computer and Telecommunication Systems(SPECTS’04), SanJose,
CA.
model that it is classified as an objective non intrusive
[13] L. Sun, “Speech Quality Prediction for voice Over Internet Protocol
method that can be applied in real time. On contrary to the
Networks”, Ph.D dissertation, Univ. Plymouth, UK., Jan 2004.
ITU-Recommendation, simplified versions of E-model have
[14] Marta Carbona and Luigi Rizzo, “Dummynet Revisited”, ACM
been introduced by researchers and industry to be used for SIGCOMM Computer Communication ReviewVolume 40 Issue 2,
monitoring purposes and predicting the VoIP call quality. April 2010, 12-20.

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded931

on September 14,2021 at 03:57:06 UTC from IEEE Xplore. Restrictions apply.