I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Published Online November 2008 in SciRes (http://www.SciRP.org/journal/ijcns/).
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Towards High Quality VoIP in 3G Networks
an Empirical Approach
Andres ARJONA¹, Cedric WESTPHAL², Antti YLÄ-JÄÄSKI³,
¹ Nokia Siemens Networks, Finland, ² DoCoMo Labs, USA, ³ Helsinki University of Technology, Finland
Email: {andres.arjona, martin.kristensson}@nsn.com, cwestphal@docomolabs-usa.com,
{antti.yla-jaaski, jukka.manner}@hut.fi
Received August 22, 2008; revised October 10, 2008; accepted October 12, 2008
Third generation (3G) packet switched WCDMA networks with high-speed downlink packet access (HSPDA)
are currently being deployed worldwide to provide wireless broadband connectivity. When introducing
HSDPA in 3G networks the end user experience and system capacity with voice over IP applications improve
considerably. When later on adding also high-speed packet uplink access (HSUPA), the system capacity and
end user experience will improve even further. This paper analyzes with measurements the VoIP quality over
current Release 5 HSDPA networks. VoIP is expected to be a widely used application over 3G data services.
The results show that even though the introduction of HSDPA significantly reduces the user-to-user voice
delay, the performance is satisfactory only for selected devices. Overall, the end user experience is still
significantly worse than with circuit switched solutions and is not acceptable. The current limitations with
VoIP in HSDPA networks with a too large delay can be improved by using the RLC UNACK mode,
potentially decreasing the jitter buffer size and reducing the terminal processing delay. In the longer term,
HSUPA and several features in 3GPP Release 7 standards will bring further performance improvements in
both user plan latency and system capacity.
Keywords: HSDPA, VoIP, WCDMA, Voice Quality, MOS
1. Introduction
Voice over IP (VoIP) is becoming a widely deployed
service in data networks, and it will penetrate from the
fixed network domain into wireless network domain. The
characteristics of fixed networks and wireless networks
are fundamentally different, which will impact the
performance of services. In this article we analyze the
VoIP service performance in wireless HSDPA and
WCDMA networks.
High Speed Download Packet Access (HSDPA) [1]
networks are being intensively deployed to provide
broadband connectivity to mobile devices, such as
handheld terminals and laptops. This broadband wireless
access is able to support voice applications over a packet
data connection instead of traditional circuit switched calls.
With the introduction of multi-radio devices with HSDPA,
WCDMA, and WiFi capabilities as well as integrated
VoIP clients, ubiquitous connectivity across any of these
networks is possible using the same mobile terminal.
However, while the mobile terminal and client are the
same, performance differs depending on the wireless
access in use.
Most of the studies of VoIP over 3G network focus on
simulation works. However, there is little data on the
performance in actual networks. Since VoIP is expected to
become a widely used application, and comes pre-
configured in many current handsets, it is of great
importance to better understand the performance of such
application over 3G networks. We set to answer the
following question: is VoIP over 3G network commercially
viable with the current state of the arts networks?
This paper studies the quality of VoIP in wireless
networks with multi-radio mobile devices both in the lab
and in live network environment setups by conducting a
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
methodic performance analysis based on the E-Model
[2,3](we will describe the E-Model in more details in
Section 2). Likewise, our study will encompass the
signaling performance required for VoIP applications.
The key contribution of the paper is to characterize the
performance of VoIP over 3G network, and to identify the
main differences between HSDPA and WCDMA. We
perform a thorough empirical evaluation of VoIP quality
and signaling performance with HSDPA and WCDMA.
From our evaluation, we will observe that:
VoIP performance is acceptable in HSDPA networks
only for VoIP clients on devices with enough
processing power, such as laptops;
VoIP performance is rarely acceptable in WCDMA
networks, even for those high performance clients;
WCDMA performance can be significantly improved
by having retransmissions only at the BTS, not the
The delay introduce by the end-user terminal is a
critical factor in the performance.
Our study takes into consideration both the
performance of the network and also the performance of
real embedded VoIP clients. In addition, we validate the
results of our study by comparing them to the actual
performance in a densely deployed HSDPA network in
Finland. Based on the results, we analyze the primary
differences in performance between simulations found in
the literature, our lab experiences and a live network case
study. Finally, we discuss possible features that can
improve the performance enough in current and future
releases to support VoIP in all handheld devices.
The remainder of the paper is organized as follows.
Section 2 describes our research approach, Section 3, 4
and 5 present results from a laboratory setup, a live
network scenario, and for VoIP signaling performance
respectively. Subsequently, in Section 6 we describe some
standardization improvements. In Section 7 we discuss the
available related works and finally in Section 8 we draw
2. Methodology and Test Environments
Our experiments are composed of measurements in a
HSDPA and WCDMA testbed, as well as a live HSDPA
network of a Finnish operator. We are interested in
measuring both the VoIP service audio quality in both
laboratory and live setups and SIP signaling latencies for
registering users and setting up calls.
2.1. VoIP Quality Methodology
The evaluation methodology consisted of multiple VoIP
tests carried out in a radio interference free environment.
These conditions were achieved in a laboratory setup by
using an RF room for the BTS and clients [4]. The tests
included different wireless accesses technologies and
variable combinations of codecs, signal conditions,
number of clients and fading profiles among others.
The main evaluation was carried out with two similar
tools based on the E-Model [2], which is a ITU-T
recommendation for VoIP evaluation. Firstly, with a NSN
proprietary tool, which is an implementation similar to the
one described in [3], and secondly, with IxChariot, which
is a widely used voice evaluation tool [5]. Finally, a third
tool based on the PESQ evaluation model was used to
determine the average end-to-end delay with real
embedded VoIP clients. With such setup, we can evaluate
the performance of the different wireless access
technologies based on the following test objectives:
VoIP quality performance with the E-Model;
Voice quality characterization for different wireless
accesses, signal conditions, configurations and fading
Benchmark of two voice quality evaluation tools based
on the E-Model;
Estimation of the average end-to-end delay when a real
embedded VoIP client is used;
Effect of simultaneous background traffic during a
VoIP call;
Characterization of delay sources and possible
The E-Model is a voice quality evaluation model that is
based on network performance metrics. It is based on a
mathematical algorithm and provides an “R” performance
value based on the sum of four “impairment factors”
considered to be cumulative. The algorithm is depicted in
Equation (1) where, “Is” is Signal to Noise Ratio, “Id” is
delay (ms), “Ief” is packet loss (%), and “A” is expectation
RIs IdIefA
=− −−+
In practice, ITU-T proposes to use a simplified version
of this algorithm. The simplified algorithm considers that
noise cancellation is encountered in the network and also
dismisses the expectation factor. The expectation variable
is supposed to be used to provide a balance for some
environments in which the user expects a degraded quality,
such as satellite connections. However, since this variable
is merely subjective it is recommended to ignore it. The
simplified algorithm is depicted in Equation (2).
RId Ief
=− −
The R value can be associated with the Mean Opinion
Score (MOS) values, which is a subjective grade for voice
quality based on studies carried out by ITU-T. However,
even though the R-value can match a MOS value, it cannot
predict the absolute opinion of an individual user.
In this paper we calculate the MOS scores with two
tools based on the E-Model: a Nokia proprietary tool and
IxChariot, which is a widely used tool. These tools send
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
dummy packets that resemble VoIP packets. The packet
size and transmission intervals are tied to the modeled
codec. Based on the received packets, network perfor-
mance values are calculated and the E-Model algorithm is
applied to determine a MOS score. Figure 1 shows an
overview of the environment and the E-Model based tools.
In this paper we emphasize the performance of the
G.729 codec, which is the only codec supported by all the
measuring tools used in this research. G.729 is also similar
to AMR-NB. AMR codec is the main building block for a
future codec for 3GPP based networks. ITU-T has set out
standards for maximum voice quality for several codecs,
including G.729 and G.711. However, there is not yet
agreement on a standard AMR codec maximum quality
definition in relation to the E-Model. Therefore, we can
make the assumption that the performance values
measured with G.729 codec are representative and are a
useful basis for our analysis. In addition, G.711 is not an
appropriate codec for wireless networks such as HSDPA
due to its high bitrate. However, G.711 is one of the most
largely supported codecs, and it is widely used in the
Internet. Also, due to legacy equipment it is used in many
cases, even though AMR and other lower bitrate codecs
(e.g. iLBC) are encouraged. For this reason we study both
G.729 and G.711.
2.2. VoIP Signaling Performance Methodology
The evaluation methodology consisted of a variety of
VoIP calls using Nokia N95 terminals. We chose this
terminal due to its widespread penetration in the market
and because it includes an embedded VoIP client by
default. This client can also be configured to work with
other SIP systems (e.g. Gizmo project). We did not use a
3rd party implementation with Skype because there were
no suitable clients for the N95 at the time of the study. We
captured SIP packet traces directly from the mobile
terminal wireless interface [6]. With such variables we
evaluated the different wireless networks available from
the following test objectives:
SIP registration delays
VoIP call signaling delays (post-dial, answer-signal,
and call-release delays)
The two main activities in VoIP calls are: first, a
registration to the VoIP server which is required to make
and receive calls, and second, the voice call setup itself.
The packet captures were carried out with a NSN proprie-
Figure 1. VoIP quality test environment.
Figure 2. VoIP signaling test environment.
tary tool with a function similar to TCPdump, and analyzed
with Wireshark Protocol Analyzer [7].
All the calls were carried out with two identical
terminals with exactly the same setups, registered to the
same VoIP server in the NSN IP Multimedia Subsystem
(IMS) and via the same wireless access in an interference
free environment. The measured scenarios were
to-WiFi calls. The maximum transfer bitrates were set in
the RNC and HLR configurations to model different
wireless access scenarios. For WCDMA, maximum uplink
and downlink transfer rates were fixed at 64/64 kbps and
128/128 kbps. For HSDPA downlink was 3.6Mbps and
the uplink was fixed at 128 kbps. In the case of WiFi,
transfer rates were left with default configuration (802.11g
and maximum transfer rate). Figure 2 shows the test
environment setup.
The core network and IMS system were privately
owned and under very low load. The wireless access
systems were based on NSN Release 5 equipment for
HSDPA and WCDMA tests with default settings. For
WiFi tests, we used a Belkin Pre-N Router with default
configuration. The core network and IMS system were
based on Nokia equipment. The tests executed consisted
of multiple iterations of each of the voice call scenarios
and registration to the VoIP server. We provide average
result values from the measurements. The measurements
took place during February-March 2007.
2.3. Live Network Case Study Methodology
The final stage of our study consisted of evaluating voice
quality in a live HSDPA network. The network evaluation
took place in the Helsinki metropolitan area, and the live
network in use was provided by Elisa, Finland’s largest 3G
operator. The HSDPA coverage in the Helsinki metro-
politan area is densely deployed and assumed to be based
on NSN equipment similar to the one used in our lab
measurements (Release 5 equipment). Therefore, its per-
formance is directly comparable to our previous results.
The test objectives for this phase are as follows:
Characterize the base performance of the network
(throughput and round trip time) under different signal
Evaluate the VoIP quality in different signal conditions
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
(excellent, medium and poor)
Evaluate the average VoIP quality in a mobile scenario
and determine the signal quality distribution for the test
Our approach to the live network measurements was
modeled in the following way. First, we made a basic
network performance evaluation in different radio
environments based on signal-to-noise ratio Ec/N0 levels
Ec/N0 values are an objective figure for quality
conditions because they take into account both signal
strength and the current interference level encountered in
the cell. Based on these basic network performance figures,
we can evaluate the average performance in terms of
maximum downlink and uplink throughput, as well as
average round trip time for a particular Ec/N0 range. As a
result, we are able to define three signal conditions ranges:
1) Good signal 2) Medium signal, and 3) Poor signal.
Second, we evaluate the VoIP quality with the same NSN
Proprietary tool used in previous tests under the three
different signal conditions. This allows us to get a good
metric of what is the quality in a static scenario under
specific signal conditions. Third, we evaluate the average
VoIP quality under a mobile scenario. The test route
chosen crossed a major part of the Helsinki metropolitan
area from West to East. The tests were carried out along
the route in both directions twice. In addition, we
measured the signal levels (Ec/N0) along the whole
driving route and carry out statistical distributions for the
An obvious limitation of our study is the fact that due to
the nature of a live network, we are not able to know or
control the other user traffic that could be taking place at
the same time. Therefore, we are not able to pinpoint the
sources of e.g. a sudden quality drop or reduced bitrate.
However, since we carried out multiple tests, our study
provides a realistic view of what is the actual performance
that could potentially be achieved in the field. The
measurements for the live network study took place during
July and August 2007.
3. VoIP Quality Analysis and Results
3.1. HSDPA/WCDMA VoIP Performance
The tests to evaluate VoIP quality involved the following
variables: signal conditions, wireless access, and fading
profiles. Signal conditions were modeled to provide
different Ec/N0 levels by using attenuators. However, the
results in this paper show that this variable does not make
any sustainable difference and therefore, average result
values are given instead. The wireless access technologies
used were restricted to HSDPA/128, WCDMA 128/128
and WCDMA 64/64. There was no reason to use higher
bitrates in this study since VoIP packets require a low
bandwidth. Therefore, we emphasize the limits in which
VoIP can actually be used with an adequate quality level.
The bitrates were fixed and therefore, features that adapt
bitrate (by increasing or decreasing) during packet
switched connections were not used during the tests.
Fading was applied with Propsim C8 fading simulator
using Pedestrian-A 3km and Vehicular-A 30km fading
profiles. The jitter buffer had a depth of 200ms and first
packet play delay of 120ms. That is, all packets are
delayed at least 120ms to provide a cushion for possible
jitter. These are common settings in VoIP clients for
wireless cellular systems. According to Wang et al. [8], a
conservative jitter buffer playout delay is about 150ms.
Our results are consistent and show that the achieved
quality in the HSDPA system is competitive. Based on
ITU-T G.107 [2] quality was in average medium for
HSDPA with both measurement tools (NSN Proprietary
Tool and IxChariot). The average MOS was roughly 3.7
(see Figure 3 and Figure 4). This is a good figure
especially considering that typical PSTN systems provide
MOS values around 3.5. In the case of WCDMA, quality
differed depending on the bitrate used. WCDMA 128/128
provided low quality and WCDMA 64/64 gave a low/poor
quality level. The results also show a difference between
the measurement tools. Our proprietary tool was able to
differentiate more clearly the quality levels between
WCDMA 128/128 and 64/64. However, IxChariot does
not recognize much difference between these two bitrates.
In any case, both tools show that quality in WCDMA is not
optimal and is around MOS 3.0 at its best. WCDMA 64/64
MOS varied between 2.25 and 2.7. ITU states that MOS
below 2.5 is not recommended for voice services and that
nearly all users will be dissatisfied with such a service.
Therefore, we can expect that the end user experience with
VoIP WCDMA is not stable and will vary.
Table 1 presents the average end-to-end delays in the
experiments (including jitter buffer playout delay). The
results also show very similar performance regardless of
the signal conditions modeled or the fading profile applied.
The reason probably relates to fast power control
mechanisms which are able to handle such changes in
signal conditions in HSDPA and WCDMA. We
recommend that further studies would be performed using
noise or traffic generators instead of only modeling signal
scenarios with attenuators.
Figure 3. VoIP performance evaluation with proprietary tool.
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Figure 4. VoIP performance evaluation with IxChariot.
Table 1. Average VoIP end-to-end delays (including jitter buffer).
Tool IxChariot
Access PedA
3km VehA
3km VehA
HSDPA/128 215ms
WCDMA 128/128
WCDMA 64/64 315ms
Figure 5. Effect of background traffic.
Figure 6. Jitter average.
Figure 7. Packet loss percentage.
Finally, we point out that both measurement tools yield
quite similar results, with exception of WCDMA 64/64.
However, in this case we can observe that our proprietary
tool is actually more accurate than IxChariot, especially
since IxChariot does not seem to recognize any
performance difference between WCDMA 128/128 and
64/64 accesses.
3.2. VoIP Performance with Simultaneous FTP
Background Traffic
We also conducted some experiments where we added
background traffic. The tests included a small number of
simultaneous users running FTP downloads in order to
evaluate if they had any effect on the VoIP performance.
As we expected, a limited number of users cannot affect
VoIP quality (see Figure 5). The reason is tied to the
Round Robin Scheduling used in the system, which
divides bandwidth equally among users. With only 4 simul-
taneous users, each user will be given enough bandwidth
on a timely basis (every few milliseconds). In order to
measure the effect of background traffic we encourage
tests with a much larger number of users, e.g. 15-20 would
be required. This is out of the scope of this document.
Likewise, testing different scheduling techniques such as
Weighted Proportional Fair is of interest. However, there
are several simulation based studies [9,10] that study VoIP
capacity gains for different scheduling schemes including
mixed traffic scenarios. However, note that [11]
analytically showed that QoS constraints on VoIP reduce
the benefit from the Proportional Fair algorithm over
Round Robin scheduling.
3.3. Effect of Jitter and Packet Loss
The next test included experiments with jitter and packet
loss. Jitter and packet loss are presented in Figure 6 and
Figure 7. From the results we can see the average jitter and
packet loss measures for different access networks.
The results show an increased jitter and packet loss for
WCDMA 128/128 and 64/64. Further delay analysis
shows that this increase is most likely caused by constant
RLC retransmissions. RLC retransmissions have an effect
on both jitter and packet loss. Every time a RLC
retransmission takes place, it will cause a ~200ms delay
peak. This peak can potentially fill the jitter buffer causing
an overflow, which results in packet loss. Packet loss also
affects voice quality. The frequency of RLC retrans-
missions is dependent of the access in use. Figure 8 shows
an example of the RLC retransmissions (200ms delay
peaks) for different wireless access technologies.
The performance of these wireless accesses would
improve if RLC retransmissions are avoided as much as
possible. One possibility is to use the unacknowledged
mode (UNACK) feature in the RNC. The principle of
operation in HSDPA [1] is such, that the BTS estimates
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Figure 8. RLC retransmissions.
the channel quality of each user based on the physical layer
feedback on the uplink. Subsequently, link adaptation and
scheduling takes place at a fast pace. When the packets are
first received at the BTS, they are buffered. Then, the BTS
transmits the packet; however, it will still keep it in the
buffer. The reason being that in case of a failure in the
transmission (e.g. decoding failure), a retransmission will
take place directly from the BTS without requiring any
action from the RNC. This is a powerful advantage since
the retransmissions are combined at the terminal. However,
if there is a physical layer failure, such as a signaling error,
then an RLC retransmission is required, and packets are
retransmitted from the RNC (see Figure 9). This obviously
results in an increase in delay, which is not beneficial for
services like VoIP. While RLC retransmissions are not a
very frequent event in HSDPA in static scenarios, they are
more likely in mobility scenarios. In contrast, in WCDMA,
all retransmissions are RLC retransmissions requiring
RNC involvement. In the RLC unacknowledged mode,
packets are not retransmitted even if some are lost, for
example due to cell change operation [1].
3.4. Codec Performance Evaluation
Even though our study focus was on low bit rate codecs
(e.g. AMR or G.729), we also evaluated the performance
of the G.711 codec. Using G.711 codec in wireless
environments is not encouraged due to its higher bit rate.
However, since it is one of the most widely supported
codecs, there are cases in which it will be used due to other
codec incompatibilities. The performance was measured
with a proprietary tool. Tests with WCDMA 64/64 using
G.711 failed most of the time or resulted in very long
delays of several seconds and are therefore excluded.
Figure 9. BTS retransmissions handling.
Figure 10. Codec performance evaluation (G.729 and
Table 2. G.711 codec jitter and packet loss (PedA 3km).
Access Jitter Average Packet Loss %
HSDPA/128 13ms 0.34
WCDMA 128/128
19ms 2.78
Figure 10 shows the VoIP quality comparison for both
G.729 and G.711 codecs. Table 2 summarizes the jitter
average and packet loss encountered when using the
G.711 codec.
3.5. Embedded VoIP Client Evaluation
These tests aimed at determining the additional delay
resulting from real embedded VoIP clients, such as the one
included with the N95. The test setup consisted of
establishing a VoIP call using an IMS system with the
G.729 codec. Subsequently, we measured the offset delay,
that is, the delay between the moment when the original
audio sample occurs to the moment the audio sample is
reproduced in the other calling end. The tool used for
offset measurements was Malden DSLA [12]. Figure 11
shows the measurement environment. The results show
that the total offset delay including the VoIP client
processing delay is rather high (see Table 3). ITU-T
recommends 400ms as the maximum delay for voice
services with a reasonable quality. With delay above this
limit, conversations are not interactive anymore and result
in talker overlaps. Therefore, a voice service with very
high delays results in a situation in which most, if not all
users are dissatisfied. As a comparison, current circuit
switched voice services have a delay of roughly
With the results we can estimate the client processing
delay by subtracting the average end-to-end delay from
our tests based on the E-Model, 215ms, 295ms, and 350ms
respectively. The result is roughly 210ms additional
processing delay when using a real embedded VoIP client.
This value differs considerably from the more optimistic
processing delay estimations of 50-75ms available in
research from [13,14].
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Figure 11. Offset delay measurement environment.
Table 3. Sources of delay (G.729 codec).
/128 WCDMA
128/128 WCDMA
RTT Delay 85ms 170ms 225ms
Jitter Buffer
(100-200ms) 130ms 125ms 125ms
Total E2E Delay 215ms 295ms 350ms
Total E2E Delay,
including embedded
client delay 425ms 505ms 560ms
3.6. HSDPA/WCDMA Overall Effect on VoIP
End-to-End delay is the main reason for low voice quality.
With the total end-to-end total delay average values we
can extend the analysis by dividing the sources of delay
(Table 3).
With this estimation it is quite clear to understand why
VoIP does not perform well in current systems with
handheld terminals, and particularly live networks, even
when the round trip time (RTT) is low. The final
end-to-end delay is just too high. We finalize our VoIP
quality analysis by modeling the resulting VoIP quality
MOS with the additional embedded VoIP client
processing delay based on the E-Model (see formula 2).
Figure 12 shows this estimation. The results represent a
case of a laptop client versus using an embedded client in a
handheld device such as the N95 VoIP client. The figure
considers both delay and packet loss impairment factors. It
must be noted though, that in a laptop client there will also
be an additional processing delay. However, such delay is
considerably lower, ~50ms in a worst case scenario [15].
Thus, still ~160ms lower than with the mobile device
Future features such as HSUPA in further 3GPP
releases will slightly improve performance. For example,
the expected average RTT for HSUPA networks is roughly
65ms (a reduction of 20ms compared with HSDPA). This
reduction however does not improve the VoIP quality
when using a laptop. That is, the average MOS with a
laptop will still be the same. Contrastingly, the expected
quality improvement for an embedded client is about 0.2
Figure 12. Overall VoIP quality with laptop and embedded
handheld clients.
points in the MOS score. The main reason for the limited
quality improvement is that the major sources of delay,
and therefore, main impairment factors reducing VoIP
quality are not directly related only to the wireless access,
but to the VoIP client implementation. However, as we
described previously, if some HSUPA features like
UNACK mode are enabled in the wireless network, it will
be possible to reduce the size of the jitter buffer
implementation without compromising the VoIP quality.
Furthermore, a reduction in the client processing delay is
extremely important in order to seriously improve the
VoIP quality in the mobile environment.
4. VoIP Signaling Analysis and Results
In this section we analyze the latencies for VoIP using the
Session Initiation Protocol (SIP). This is an important
metric because long delays in the call setup seriously harm
the overall VoIP experience; people have certain
expectations based on the current circuit switched services,
and it is crucial to meet those.
4.1. SIP Registration Setup
The signaling [16] and delay measurements for SIP
Registration to the VoIP server in the IMS system are
depicted in Figure 13. The measurements show that the
registration times with HSDPA and WCDMA are about
30% and 50% higher than with WiFi. While this might not
seem much, we should remember that SIP registration
requires a very limited number of messages. Therefore, as
more messages are required, such as with 3GPP based
registration, delays will increase.
4.2. VoIP Call Signaling
ITU E.721 [17] recommends values for call setup delays
in circuit switched calls. The recommended values for call
setup (post-dial delay) are 3s for local, 5s for toll and 8s
for international connections, with 6s, 8s, and 11s as 95%
values. The “call answer” (answer-signal) delay reflects
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Figure 13. SIP registration signaling delays.
Figure 14. VoIP call signaling delays.
the time it takes from the moment the receiving end
accepts the call until the call is actually established. E.721
recommendation is 0.75s for local, 1.5s for toll, and 2.0s
for international connections, with 1.5s, 3.0s, and 5.0s as
95% values. Finally, “call end” (call-release delay) is the
time it takes for the call to be terminated [18,19]. The
signaling [16] and delay measurements for voice call
setups when a PDP context is active and the terminal is
registered to the IMS system are depicted in Figure 14.
The results show an expected increment in the call
signaling delays depending on the access used. Since all of
the network elements were located in a private network,
the environment could be though of as providing local
calls. Our results also show that an embedded mobile VoIP
client experiences an increased delay compared to a PC
client, such as the one measured by Curcio and Lunden [18]
with a WCDMA network.
The setup delays for VoIP calls might be impacted with
additional delays in a cellular system in cases were there is
no active PDP context, and also due to a required regis-
tration to the IMS. The PDP context activation delay was
~3 seconds in our tests. Simulations by Pous et al. [20]
propose 2.24 seconds. Based on these values, the always-
on enabled calls can be in line with E.721 recommen-
dations. However, when the PDP context is not active, the
delay with WCDMA can vary between 11 to 17 seconds,
and thus, exceed the recommended values. HSDPA delay
in this case is around 8 seconds, which is similar to the
recommendation for international calls. However, additional
delays from e.g. traversed networks, gateways, and proxies
could result in larger total delays than those recommended.
Figure 15. Average throughput in Elisa HSDPA network.
Figure 16. Average round trip time in Elisa HSDPA
5. Live Network Case Study
5.1. Generic HSDPA Performance
In this section we describe the generic evaluation of the
live HSDPA (Release 5 equipment) network performance
in Helsinki. The results for throughput and round trip time
are depicted in Figure 15 and Figure 16. The measurement
results show an increase in round trip time delay when
compared to the average values measured in the lab
environment (85ms). This means that the VoIP quality
(MOS) will be worse than our results in Section 3, and
therefore VoIP support will be even more difficult.
Throughput was measured via multiple file downloads and
uploads from a local server in Finland; while RTT was
measured with 32Byte ICMP Echo Request and Reply
(ping) packets to the same server.
5.2. VoIP Quality
The VoIP quality in the Elisa HSDPA network is likewise
slightly lower than our lab measurements (see Table 4).
The mean opinion score was 3.5, 3.5 and 3.3 for good,
medium and poor signal conditions. However, we have to
consider that once again, the VoIP quality was measured
for laptop based VoIP communication. That is, it does not
account for the additional processing delay for the
terminal VoIP client implementation previously described.
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
The results in the live case still indicate that VoIP support
for a handheld embedded client will be poor. However,
these values do take into account the jitter buffer play out
delay. The most noticeable difference between the three
scenarios is the packet loss ratio, which increases as the
signal quality decreases.
5.3. VoIP Quality in Mobility Scenarios
The mobile environment tests were measured from a van
driving through a test route at average speeds of 60-80
km/h without stopping. The selected test route crosses the
Helsinki metropolitan area from East to West and it is
entirely covered by Elisa HSDPA network according to
their publicly available coverage map. The test route was
about 18.5km and it took approximately 15min to travel.
The test route was driven several times to validate the
The results show that the average performance is lower
than in static scenarios. A mobile scenario obviously
brings several additional challenges due to the different
cell changes along the test route. The number of cell
changes along the route was 28 and were characterized via
the changes in scrambling codes used. Table 4 summarizes
the VoIP quality results.
Furthermore, in mobility scenarios the amount of RLC
retransmissions required is very noticeable. To
characterize the retransmissions, we conducted an
additional test along the test route in which we sent
continuous ping packets of 32B (see Figure 17). The
results show a large amount of delay peaks resulting from
these retransmissions. Therefore, it further supports our
lab measurements and emphasizes the importance of the
unacknowledged mode feature. We expect that this mode
would potentially take the majority of large delay peaks,
and thus, improve VoIP quality. However, if this mode is
used, there is a possibility that the packet loss ratio will
increase, and for that reason, it is very important to
validate future results as well even if the feature is enabled.
In addition, during the mobile tests, we recorded the
signal conditions to characterize the signal quality distribution
Table 4. VoIP quality in Elisa HSDPA network (including
jitter buffer).
Scenario Delay
Avg. Jitter
Avg. Packet
Loss %
Good Signal
(Ec/N0 -3 to -5) 288ms
0.4 3.5
Medium Signal
(Ec/N0 -7 to -9) 283ms
1.0 3.5
Poor Signal
(Ec/N0 -11 to -13)
2.6 3.3
Environment 331ms
1.9 3.2
Figure 17. Round trip time during mobility tests.
Figure 18. Signal quality distribution.
Figure 19. Detailed signal quality distribution.
along the test route. The measurements show that in
general, it is highly probable to get a good signal level and
that the coverage is well deployed (see Figure 18 and
Figure 19).
6. Future Directions
At the current moment, the performance of VoIP in 3G
networks is far from optimal. However, with some of the
features and improvements in further 3GPP releases, the
performance will improve. For instance, Release 6
equipment reduces RTT to roughly 65ms, and even lower
with Release 7. Likewise, with Release 7 operators have
other choices for deployment prior to full VoIP rollouts.
For instance, advances such as Circuit Switched voice
over HSPA (CS over HSPA) can improve capacity to
similar levels as with VoIP. In this case, traditional voice
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
is carried over packet data. Hence, since VoIP does not
provide any significantly better capacity figures over CS
over HSPA, operators can delay VoIP deployment can be
delayed until adequately performing terminals and
networks are available. This however, is only possible if
several features are upgraded in several network elements.
These improvements occurred while this manuscript was
under review. CS over HSPA is expected to be included in
3GPP Release 7 [21].
7. Related Work
Although, there is prior work investigating the VoIP
performance in WCDMA and HSDPA systems, it is not
very extensive and mostly based on simulations. For
instance, some papers [22,23,24] study VoIP performance
in WCDMA and provide some baseline results. In addition,
other works [25,26] provide some estimated values for
processing delays. In these studies, the assumption for the
estimations is based on whether the call is towards a
landline or a mobile end. Some performance simulations
are also available [8,10,13,14,27
29]. However, the
simulations only provide a delay budget rather than a
description of the end user experience. Contrastingly, our
study focuses on end user experience and VoIP quality
rather than delay budgets alone. The delay budget values
used in simulations vary from 80-150ms for studies
ignoring encoding/processing delays and jitter buffer
implementations [9,27,28,30,31], to 250-300ms for
studies that assume such delays to some extent [8,10,
13,14,29]. In addition, the estimations used in simulations
are in general overly optimistic in regards to, e.g. client
processing delay. Kim [14] considers the processing/
encoding delay to be 50ms, while Ericson [13] assumes
around 75ms. These delay values include the jitter buffer
playout delay as well. Therefore, it is noticeable they are
too optimistic, especially when compared to our
experiment results with actual handsets and VoIP jitter
buffer client implementations.
Even though, it is understandable that the exact
encoding/processing delays and jitter buffer playout
delays are client specific, unless they are modeled
accordingly, or at least, to some extent, the differences in
performance between simulations and actual deployments
will remain very visible. Therefore, simulations results are
only comparable to laptop based performance at its best
and not to actual handheld performance, which in the end
is the primary use case for VoIP services. Other simulation
study [28], notices the importance of reducing RLC
retransmissions to improve performance in FTP and
HTTP browsing. However, the study does not address its
importance for VoIP services. Finally, Wager and
Sandlund [32] conduct simulations to determine the
amount of possible lost frames of VoIP speech in HSDPA
mobility scenarios.
In regards to VoIP signaling, SIP call setup delays and
signaling performance have been studied previously
mostly for Internet scenarios. ITU E.721 [17]
recommendation and an IETF Internet Draft [33], provide
call setup delays recommendations for circuit switched
and Internet Telephony systems respectively. Additionally,
Eyers and Schulzrinne [19] provide guidelines for Internet
Telephony call setup and signaling transfer delays. In
regards to 3GPP based wireless accesses, Kist and Harris
[34] provide simulations for transfer delays with 3GPP
signaling, while Fathi et al. [35] and Pous et al. [20]
modeled signaling performance. Further, Curcio and
Lunden [9] provide measurements for a WCDMA setup
using laptop clients for local, international and overseas
calls. Most of the mentioned research focuses on
simulations, and does not consider some end user cases
such as calls in wireless environments starting from
different states. Additionally, performance with different
wireless radio accesses and configurations under the same
conditions is not available. Also, the available works do
not use an embedded VoIP client in a handheld mobile
terminal, which yields different delay values compared to
a PC. HSDPA signaling performance has not been
evaluated either. Our research aims at covering these items.
The importance of evaluating a mobile terminal relies in
the fact that the eventual substitution of circuit switched
calls in 3GPP networks (HSDPA and WCDMA) for VoIP
calls will take place with a handheld mobile device and not
with a PC or laptop. Likewise, multi-radio devices can
provide ubiquitous access via different wireless access
technologies with distinct performance characteristics.
The lack of actual measurement performance values in
literature could be mainly due to the unavailability of
integrated VoIP clients in the terminals and available
HSDPA networks. However, with the introduction of
some multi-radio devices with VoIP capabilities (e.g.
Nokia N95, Nokia 6110), it is possible to use VoIP
applications without a PC.
8. Conclusions
Multiple measurements were carried out to evaluate and
characterize the VoIP quality and VoIP signaling
performance in HSDPA and WCMDA wireless accesses.
The results show that HSDPA access is capable of
providing a competitive VoIP quality compared to circuit
switched voice. However, WCDMA in 128/128 and 64/64
bitrate configurations can only provide low and poor
qualities, the main issues are long delays and packet losses,
which occur often due to RLC retransmissions that
overflow the jitter buffer capacity. However, the main
issue with HSDPA is not only tied directly to the wireless
access performance, but to the mobile device capabilities.
Our results show that embedded mobile VoIP clients can
introduce an increased delay due to processing when
compared to laptop performance. This processing includes
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
e.g. encoding/decoding, and other operating system tasks.
The additional delay has a considerable voice quality
reduction effect. Further, the results from the test cases
experimented in a live network resulted in lower
performance when compared with similar laboratory
measurements. Also, the effect of mobility in regards to
VoIP quality degradation is quite noticeable. The
degradation is due to handovers during the test route that
increase the ratio of RLC retransmissions.
Therefore, the main aspects that can potentially
improve VoIP quality performance with the current
systems are mainly to reduce the number of RLC
retransmissions by using unacknowledged mode,
potentially use smaller jitter buffer sizes, and reduce the
embedded VoIP client processing delays. High quality
VoIP in 3G networks will be possible. However, it is tied
to improvements in several areas such as wireless network
delay, client implementation, and client processing delay.
Finally, a main improvement developed while this
manuscript was in process is CS over HSPA, which
improves capacity and thus, can allow operators to delay
VoIP deployment projects until networks and terminals
have better performance.
9. References
[1] H. Holma, and A. Toskala, “HSDPA/HSUPA for UMTS,”
John Wiley, 2006.
[2] ITU-T, Recommendation G.107 “The E-model, a
computational model for use in transmission planning,”
[3] R. Cole and J. Rosenbluth, “Voice over IP performance
monitoring,” ACM SIGCOMM’01.
[4] A. Arjona, C. Westphal, A. Ylä-Jääski, and M.
Kristensson, “Towards high quality VoIP in 3G networks:
An empirical study,” In Proceedings IEEE AICT’08,
Athens Greece, June 8–13, 2008.
[5] IxChariot, http://www.ixiacom.com.
[6] A. Arjona and A. Ylä-Jääski, “VoIP call signaling
performance and always-on battery consumption in
HSDPA, WCDMA and WiFi,” in Proceedings IEEE
WiCOM’07, Shanghai China, September 21–23, 2007.
[7] Wireshark Protocol Analyzer,
[8] B. Wang, K. Pedersen, T. Kolding, and P. Morgensen,
“Performance of VoIP on HSDPA,” IEEE Vehicular
Technology Conference VTC’05 Spring, Vol. 4, pp.
2335–2339, May 30–June 1, 2005.
[9] P. Lundén, and M. Kuusela, “Enhancing Performance of
VoIP over HSDPA”, In Proceedings IEEE VTC’07 Spring,
April 21-24, 2007.
[10] A. Braga, E. Rodriguez, and F. Cavalcanti, “Packet
scheduling for VoIP over HSDPA in mixed traffic
scenarios,” in Proceedings 17th IEEE PIMRC’06,
September 2006.
[11] H. Kim, “Loosing Opportunism: Evaluating Service
Integration in an Opportunistic Wireless System,” IEEE
INFOCOM’07, May, 2007.
[12] Malden DSLA,
[13] M. Ericson and S. Wänstedt, “Mixed traffic HSDPA
scheduling–impact on VoIP capacity,” in Proceedings
IEEE VTC’07 Spring, April 21–24, 2007.
[14] Y. Kim, “VoIP Service on HSDPA in Mixed Traffic
Scenarios,” in Proceedings IEEE CIT’06, September 2006.
[15] Cisco, “Understanding delay in packet networks,”
Document ID: 5125, March 2007.
[16] J. Rosenberg, H. Schulzrinne, G. Camarillo, A. Johnston, J.
Peterson, R. Sparks, M. Handley, and E. Schooler, “SIP:
Session initiation protocol,” RFC 3261, IETF June 2002.
[17] ITU, Recommendation E.721, “Network grade service
parameters and target values for circuit switched services
in the evolving ISDN,” 1999.
[18] I. Curcio and M. Lundan, “SIP call setup delay in 3G
networks,” ISCC’02, Taormina Italy, July 1–4, 2002.
[19] T. Eyers and H. Schulzrinne, “Predicting Internet
telephony call setup delay,” IPTel2000, Berlin, April 2000.
[20] M. Pous, D. Pesch, G. Foster, and A. Sesmun,
“Performance evaluation of a SIP based presence and
instant messaging service,” 3G 2003, June 25–27, 2003.
[21] Nokia Siemens Networks and Nokia, “Supporting CS over
HSPA,” 3GPP R2-073487, August 2007.
[22] F. Poppe, D. Vleeschauwer, and G. Petit, “Choosing the
UMTS air interface parameters, the voice packet size and
the de-jittering delay for a voice-over-IP call between a
UMTS and a PSTN party,” IEEE INFOCOM, Vol. 2, pp.
805–814, April 2001.
[23] F. Poppe, D. Vleeschauwer, and G. Petit, “Guaranteeing
quality of service to packetized voice over the UMTS air
interface,” 8th International Workshop on Quality of
Service, pp. 85–91, June 2000.
[24] R. Cuny and A. Lakaniemi, “VoIP in 3G networks: An
end-to-end quality of service analysis,” in Proceedings
IEEE VTC’03 Spring, April 2003.
[25] ITU-T, Recommendation G.114, “One-way transmission
time,” 2003.
[26] TIATR-41.1.2, “VoIP end to end delay budget planning
for private networks,” Cisco 2000.
[27] G. Rittenhouse and H. Zheng, “Providing VoIP service in
UMTS-HSDPA with frame aggregation,” in Proceedings
IEEE ICASSP’05, March 18–23, 2005.
[28] L. Bajzik, L. Korössy, K. Veijalainen, and C. Vulkán,
“Cross-layer backpressure to improve HSDPA
performance,” in Proceedings IEEE PIMRC’06, Helsinki
Finland, June 2006.
[29] P. Hosein, “Scheduling of VoIP traffic over a time-shared
wireless packet data channel,” in Proceedings IEEE
ICPWC’05, January 23–25, 2005.
[30] Y. Seo and D. Sung, “Performance of VoIP in HSDPA
based on an adaptive power allocation scheme,” in
Proceedings IEEE WCNC’06.
[31] P. Hosein, “Capacity of packetized voice services over
time-shared wireless packet data channels,” in
Proceedings IEEE INFOCOM’05, March 13–17, 2005.
[32] S. Wager and K. Sandlund, “Performance evaluation of
HSDPA mobility for voice over IP,” in Proceedings IEEE
Copyright © 2008 SciRes. I. J. Communications, Network and System Sciences, 2008, 4, 285-385
Vehicular Technology Conference VTC’07 Spring, April
22–25, 2007.
[33] H. Lin, T. Seth, A. Broscius and C. Huitema, “VoIP
Signaling performance Requirements and Expectations,”
IETF Draft, June 1999.
[34] A. Kist and R. Harris, “SIP Signaling Delay in 3GPP”,
IFIP Interworking’02, October 13–16, 2002.
[35] H. Fahti, S. Chakraborty, and R. Prasad, “Optimization of
SIP session setup delay for VoIP in 3G wireless
networks,” IEEE Trans. on Mobile Computing, Vol. 5, No.
9, September 2006.