SECOND QUARTER 2004, VOLUME 6, No.

_
IEEE
COMMUNICA TlONS

IS U R V E Y S'-----
The Electronic Magazine of
Original Peer-Reviewed Survey Articles

www.comsoc.org/pubs/surveys

VOICE OVER THE INTERNET:

A TUTORIAL DISCUSSING
PROBLEMS AND SOLUTIONS
ASSOCIATED WITH ALTERNATIVE TRANSPORT
GEORGE SCHEETS, OKLAHOMA STATE UNIVERSITY
MARIOS PARPERIS, WILTEL COMMUNICATIONS
RITU SINGH, NDRS TECHNOLOGIES

ABSTRACT
This article provides a tutorial overview of voice over the Internet, examining the
effects of moving voice traffic over the packet switched Internet and comparing this
with the effects of moving voice over the more traditional circuit-switched telephone
system. The emphasis of this document is on areas of concern to a backbone service
provider implementing Voice over IP (VoIP). We begin by providing overviews of
the Plain Old Telephone Service (POTS) and VoIP. We then discuss techniques
service providers can use to help preserve service quality on their VoIP networks.
Next, we briefly discuss Voice over ATM (VoATM) as an alternative to VoIP.
Finally, we offer some conclusions.

T
his article provides a tutorial overview of one of the POTS
Internet traffic types likely to see significant future
growth, Voice over the Internet (a.k.a. Voice over IP or POTS today consists of a mix of the very old and the very
VoIP). The effects of moving voice traffic over a packet- new. Figure 1 presents a simplified view of POTS connectivity.
switched statistically multiplexed network such as the Internet, End-user telephones typically use twisted pair copper cabling
which was never really designed to handle this type of source, to deliver analog voice signals to the central office (CO). The
are examined and compared with the effects of moving voice vast majority of inter-office communications are now carried
over the more traditional circuit-switched time division multi­ over fiber optic trunks.
plexed telephone system, more affectionately known as POTS. Today, most long-haul voice traffic is digitized. At the CO
The emphasis of this article is on areas of concern to a back­ switch, the analog signals are passed through a narrow-band­
bone service provider implementing VoIP. width bandpass filter, sampled at a rate of 8,000 samples/sec­
This article is organized as follows. We first provide an ond and quantized (rounded off) to one of 256 unequally
overview of POTS. We then examine a non-traditional, pack­ spaced voltage values. A technique known as Pulse Code
et-based, voice transport, VoIP. We next survey techniques Modulation (PCM) is then used to assign an 8-bit code word
service providers can use to help preserve quality on their to each voltage, resulting in 64 kb requiring transport every
VoIP networks, while the following section considers the second. This 8-bit code word is the basic transmission entity of
effects that design choices and failings in the network have on POTS. POTS backbones use Time Division Multiplexing
this quality. We discuss the trade-offs involved in terms of (TDM) and circuit switching to efficiently transmit this entity.
maximizing the number of calls supported while simultaneous­ Circuit switching is used to dedicate sufficient bandwidth to
ly preserving end-to-end delivery bounds and meeting mini­ support this bit rate for the duration of the phone call. An
mum quality standards. Comments regarding another end-to-end path is set up, and resources are dedicated, prior
alternative voice transport system, ATM, are offered. Our to the initiation of the actual voice conversation. As PCM out­
conclusions are then summarized. puts 8 bits every l/8000th of a second in a predictable, deter-

22 IEEE Communications Surveys & Tutorials· Second Quarter 2004

ministic manner, for every trunk traversed, regardless
of the trunk speed used, TDM will set aside 8 bits
every l/8000th of a second for each simplex call. The
resulting traffic is moved to the destination central
office switch, and converted back to an analog signal
for transmission over the local loop [1].
Transmission protocols originally developed for the Analog Analog
voice network, including state of the art Synchronous Digital T DM
64 kb/s
Optical Network (SONET) and the now mostly obso-
lete T Carrier systems, are all built around frames, of
• FIGURE 1. Typical POTS connectivity.
which there are 8,000 each second. Frames typically
consist of some control bits or bytes, and time slots
into which 8-bit voice code words can be inserted.
Few central office switches are directly connected as
Source CO
shown in Fig. 1. More typically one or more digital
=
TDM telephone switches are in the end-to-end path a
POTS call traverses. To maximize network use, inter­ Q Local loop
PCM
coder
TOM trunk
POTS
TSI
mediate voice switches (known as tandem switches)
may move voice code words to an output line time slot
that differs from the input line's time slot. This may be Trunk resources are dedicated
I •
Intermed iate
necessary, for example, when voice traffic occupying to each voice ca l l via TOM •
digital
and circuit switching. • vOice
the same time slot is received on two different input
switches
lines, but both need to be placed on the same output
line. This time slot interchange (TSI) may be accom­
I
=
plished by writing voice code words into the tandem
Q
PCM POTS
switch's memory in the order they are received, and coder TSI
Local loop TOM trunk
then reading them back out of memory at the appro-
Destination CO
priate time but in a different order [1]. Time slot inter­
change may cause a worst-case delay of l/8000th of a
second, which can occur if a voice code word occupying • FIGURE 2. Sources of POTS delay
time slot k of a frame must be delayed to time slot k-l
of the following frame.
VOICE CODER
Voice delays through a POTS network are smaller than
those seen through a VoIP network. Essentially, the end-to­ POTS networks almost universally use the lTV 0.7 1 1 64 kb/s
end delivery delay on a POTS network consists of the propa­ PCM standard. As previously noted, this type of coder outputs
gation delay of the electromagnetic energy over the physical 8 bits every l/8000th of a second, that is, the frame of this
cabling, the PCM encoding delay at the source central office coder is 1/8000th of a second, and the frame size is 8 bits.
(ideally l/8000th of a second), TSI delays on intermediate tan­ Other coders exist that can reduce the generated bit rate,
dem switches (:0:: l/8000th of a second per switch), and the but usually with a slightly reduced perceived quality. Table 1
PCM decoding delay at the destination central office that is compares selected lTV coders, including their associated bit
also ideally l/8000th of a second. Practically speaking, coding rates [2, 3].
and decoding delays may be approximately 10-20 times larger, Of these codes, 0.729 has evoked considerable interest for
depending on the demands placed on the source and destina­ VoIP providers as it has comparable quality to 0.711, but at a
tion digital signal processors. Figure 2 shows a simplified dia­ greatly reduced bit rate. 0.729 has a frame length of 10 msec,
gram of these delays. and its compression algorithm outputs 80 bits every 1/ 100th of
a second, yielding an 8 kb/s bit rate.
POTS TDM backbones are based around fixed-rate coders.
NON-TRADITIONAL TELEPHONY: VolP For example, a 0.7 1 1 coder outputs 64 kb/s at all times,
regardless of whether or not the voice source is talking or lis­
Probably the most significant difference with VoIP, as com­ tening. The statistical as-needed allocation of backbone band­
pared to POTS, is that backbone trunking resources are not width on VoIP networks offers the opportunity to deploy
assigned in a dedicated methodical manner to support a voice variable-rate coders, which output traffic at one bit rate when
call. Instead, trunk bandwidth is assigned by routers on an the voice source is talking, and a lower bit rate (possibly zero)
seemingly random, as needed basis via statistical multiplexing when the voice source is quiet. For example, 0.729B with
and packet switching. The result is that while statistics can be silence suppression examines each 10 msec voice frame and
generated for the average packet delay and variation, the makes a voice/no voice decision. If the coder detects voice
delivery times for each individual packet are not predictable. energy, it will output the standard 80 bits of compressed, digi­
With POTS, voice traffic arrives at the destination PCM tized voice for that frame. If the coder decides this frame does
decoder in a constant 8-bits-every-l/8000th-of-a-second man­ not contain voice energy, the coder will output a reduced
ner. With VoIP, packetized voice traffic arrives at the destina­ block of bits containing comfort noise, information that the
tion decoder at unpredictable intervals and ordering. This receiver will use to generate background noise so that the
randomness necessitates a more complex environment in user does not think the connection has been lost [3]. Alterna­
order to smooth the flow of voice bits to the destination voice tively, the coder could output nothing at all and the receiver
decoder, which typically requires a steady bit arrival rate. could generate comfort noise.
Figure 3 shows a simplified diagram of the delays associat­ Experiments have shown that in a typical two-way interac­
ed with this type of network. Below we highlight the key func­ tive voice conversation, voice sources are only active 40 per­
tions of each block. cent of the time [4]. The 60 percent idle time includes pauses

IEEE Communications Surveys & Tutorials· Second Quarter 2004 23

 =
Q -
Voice
coder
Packet
assembler
T ransm iss ion
buffer
f--
Packet
switch with the irregular packet arrivals and provide a
smooth traffic flow to the voice decoder. As a

Trunk resources are stochastically

I
• I n termed iate
result, the receiver buffer is also known as a de-jit­
ter buffer. Generally, this buffer is allowed to par­
assigned to each voice call via statistical • packet
• tially fill before any play-out to the voice decoder
mu ltiplexing and packet switching. switches
commences.

=
I An appropriately sized receiver buffer is critical
to the proper operation of VoIP systems. By sizing,
Q -
Voice
decoder
Receiver
buffer
Packet
switch
we are referring to two parameters: the fill delay
StatMux mentioned immediately above, and the buffer stor­
trunks age capacity. If the fill delay is chosen too small
and a group of packets arrives later than expected,
• FIGURE 3. Sources of VolP delay it is possible that the buffer will empty, which will
result in a loss of voice output. If the fill delay is
chosen too large, then excessive time is lost in the
while listening to the other party talking, as well as pauses receiver buffer waiting to be played out, which reduces the
between sentences, and even pauses between some words. A time allotted to other devices in order that end-to-end deliv­
G.729B coder with silence suppression will output 8 kb/s dur­ ery goals be met. If the buffer storage capacity is too small
ing talk spurts (40 percent of the time), and nothing or a then a burst of packets received may result in inadequate stor­
reduced bit rate during intervening silence intervals (60 per­ age space being available and dropped packets.
cent of the time). The use of silence suppression potentially With the use of time lines, and assuming no packets are
allows a G.729 coder to reduce its average output from 8 kb/s dropped by intermediate packet switches, it is possible to
to 3.2 kb/s (alternating 8 kb/s and 0 kb/s bursts). show that to be 100 percent certain that the receiver buffer
does not empty, packet playback should commence no earli­
PACKET ASSEMBLER er than WCDeliv seconds after assembly, where WCDeliv is
equal to the worst case packet end-to-end delivery time. A
One decision faced by every operator of a VoIP telephony more complicated statistical analysis would be necessary to
voice switch is how many frames from the voice coder to determine tolerable receiver buffer delay if packets are
include in each transmitted packet. A typical VoIP packet dropped by the network, or if a possibility of the receiver
requires about 47 bytes of overhead: 8 bytes for the User buffer emptying is acceptable. If VoIP traffic is routed over
Datagram Protocol (UDP), 12 bytes for Real Time Transport an Internet backbone augmented with voice gateways han­
Protocol (RTP), 20 bytes for the Internet Protocol (IP), and 7 dling connection admission control to limit the number of
bytes for the Point-to-Point Protocol. Header compression of voice calls, and these gateways are knowledgeable of the
the UDP and IP headers can reduce the amount of overhead end-to-end voice paths through the network and worst-case
on low-speed links, but is not a standardized option on high­ delays at transited routers, the above technique is, at least in
speed carrier backbones. It would not be cost effective to take theory, possible.
the one byte frame output of a G.71 1 coder, packetize it, and However, if the traffic is routed over the commodity Inter­
immediately transport this 48 byte packet, as 98 percent (47 net, where the exact path through the network and the worst­
out of 48) of the bytes transmitted would be overhead. In this case delay through routers is not known, determination of
case, it is better to place the output of multiple frames into WCDeliv may be difficult, if not impossible, to obtain. In this
one packet. The disadvantage of this is that time is lost at the case the packet playback might be delayed a value large
transmitter site waiting for the desired number of voice enough such that it is (hopefully) rarely exceeded by the actu­
frames to be generated. al packet end-to-end transit time. An alternative option, and
The choice of the number of frames in a packet can seri­ one that tends to be favored today, would be to adaptively
ously impact the number of phone calls a VoIP network can adjust the de-jitter buffer delay to account for the time vary­
support. Too few frames/packet results in a considerable per­ ing changes of the voice packet delivery delays.
centage of bandwidth being allocated to the movement of To summarize, in comparing Fig. 2 with Fig. 3, it should be
overhead bytes. Too many frames in a packet results in an noted that a VoIP system is more complex than POTS. VoIP
excessive amount of time being lost assembling a packet at the has more sources of delay, and the delay through these ele-
transmitter, reducing the time remaining to meet the end­
to-end delivery time bound. In the latter case, trunk loads
may have to be decreased in order to maintain switch Codec G.711 G.723.1 G.726-32 G.729

queuing delays at tolerable levels.

Once assembled, the transmission buffer provides a Coding rate (kb/s) 64 5.3-6.3 32 8
place for the packet to be stored while waiting for net­
Frame si ze (ms) 0 . 1 25 30 0. 250 10
work bandwidth to become available.
Algorithm PCM MP -MLQ ADPCM CS-ACELP
RECEIVER BUFFER
Processing delay (ms) 0 . 1 25 30 0. 250 10

Unlike the POTS network where voice traffic arrives at

Look ahead delay (ms) 0 7.5 0 5
the destination in byte sized pieces at regular time inter­
vals, in a VoIP network traffic arrives irregularly in packet Payload (Bytes) 1 20-24 1 10
sized chunks. Voice decoders are typically expecting a
steady feed rate. For example, a G.71 1 decoder expects to Qual ity Good Good/fair Good Good

receive 8 bits every 1/8000th of a second. A G.729 decoder

Comp lexity Lowest H ighest Low H igh
expects 80 bits every l/100th of a second. The receiver
buffer's primary function is to remove the jitter associated • Table 1. Selected ITU voice coders and key characteristics.

24 IEEE Communications Surveys & Tutorials • Second Quarter 2004

 Proba b i l i ty of
dropped packets
first out routers can be deployed on the backbone. Figure 4
shows a classic delay versus load plot for a statistically multi­
I plexed switch, modified to account for finite-length buffers.
I
I As the trunk load increases, a point will eventually be
Average delay for I
del ivered packets I
reached where the switch buffers fill up, the probability of a
, packet being dropped becomes unacceptably high, and the
,
-------- � --
---
- � delay seen traversing a switch by packets that are not
0%
dropped is limited by the buffer size. At low loads, the aver­
Trunk load 1 00%
age delays through a router will be low, the probability of
dropped packets will be small, and quality will be acceptable
All traffic has low delay, j itter, and probab i l ity of d iscard. for all traffic. Note, however, that this option may not be a
cost effective approach as it does not scale well if VoIP min­
utes increase significantly.
• FIGURE 4. Light trunk loading.
At heavier loads Quality of Service (QoS)-enabled routers
should be deployed, and priorities used to give preferential
treatment to time-sensitive traffic such as interactive VoIP.
ments is generally considerably larger than that experienced Figure 5 shows an example where the prioritized voice traffic
crossing corresponding POTS elements. Note also that indi­ sees similar performance from a QoS-enabled router (dashed
vidual packets will see different transit times across the net­ orange line) that the unprioritized voice would see in a more
work. VoIP designers have a more difficult job insuring that lightly loaded best effort router (orange line). The widespread
end-to-end delivery delays and overall quality remain at toler­ deployment of Differential Services (DiffServ) or Internet
able values. Protocol Version 6 will be essential in order to offer standard­
Examining Fig. 3, note that to meet the required end-to­ ized priorities.
end delivery delay, the sum of the voice coding delay (VCDe­
lay), packet assembly delay (PADelay), service time and CONTROLLING PACKET Loss AT PACKET SWITCHES
queuing delays at the voice source and all intermediate packet
switches (QuDelays), receiver de-jitter buffer delay (IitDelay), Related to the problem of controlling delay at intervening
end-to-end propagation delay (PropDelay), and voice decoding switches is the problem of controlling packet loss. Both are a
delay (VDDelay), all must be less than or equal to the target function of the amount of buffer space in the switch. In the
end-to-end voice delivery delay (Target). Mathematically, delay-limited case, there is plenty of buffer space such that the
probability of packet discards can be ignored. The allowable
VCDelay + PADelay + QuDelays (1) delay through the switch sets the load that the output trunks
+ IitDelay + PropDelay + VDDelay� Target.
can support. In the buffer-limited case, the finite amount of
buffer memory in a switch is the limiting factor that sets the
tolerable load that the switch can handle. This finite storage
PRESERVING QUALITY ON VolP NETWORKS
for queued traffic results in packet discard probabilities being
Preserving the quality on VoIP networks requires careful the limiting factor, not the tolerable delay.
engineering by the system engineer [5]. In this section we dis­ Choosing an upper bound on the allowable packet loss
cuss available design options [6-14]. on a VoIP network is not a trivial task. The subjective
voice degradation resulting at the receiver is a complex
CONTROLLING THE NUMBER OF function of the amount of coder compression (generally,
INTERMEDIATE PACKET SWITCHES the greater the amount of compression the greater the
impact of a loss of information), the number of coder
It is not unusual for traffic traversing the commodity Internet frames placed in each packet (the loss of a packet carrying
to traverse 10-20 packet switches (routers), even on short-dis­ multiple frames is more serious than the loss of a packet
tance trips. As examples, a Traceroute launched from the lead carrying one voice frame), and the decoder's ability to
author's house in Stillwater, Oklahoma to Cisco System's Web mask any lost traffic [ 1 5]. This issue is discussed in more
site in California traversed routers in Oklahoma City, Dallas, detail in the next section.
Fort Worth, Anaheim, and San Jose, 13 in all. A Traceroute
launched from the same location to Oklahoma State Universi­
ty's Web server located two miles away was first shipped to

�
Oklahoma City, then back to Stillwater, traversing a total of
10 routers. �0'"
If)

�
�
Most of the variability in the inter-arrival time of VoIP 4- Q)
>,-"
co u Low pcioc", do'o,
packets at the destination is due to varying queue times at the -co
Q)o..
"0"0
intermediate packet switches. A large amount of resulting jit­ i H igh priority delay
ter will require a large receiver de-jitter buffer delay to
Q) Q)
Ol�
co Q)
>
� :'=
Overa l l delay
: /;
smooth things out. Keeping the number of intervening packet Q) :
> Q) "
switches low can make a significant difference here. This may
be difficult to accomplish on the commodity Internet.
«"0
::::::::::::;;:.�
_______ ___
__ - ... ...
.......
i� ,

CONTROLLING THE DELAYS AT PACKET SWITCHES D Offered

trunk load

Queuing delays in packet switches can be controlled by either Preferential treatment reduces the delay seen by high·priority
controlling the load on interconnecting trunk links, prioritiz­ traffic and increases delays seen by lower·priority traffic.
ing the voice traffic, or enabling some combination of the two.
lf the trunks are kept lightly loaded, best-effort first in, • FIGURE 5. Heavy trunk loading

IEEE Communications Surveys & Tutorials· Second Quarter 2004 25

 SEGREGATE OR INTEGRATE CODER DISTORTION

Another choice faced by the VoIP system designer is whether Coder distortion is frequently assessed via the mean opinion
or not to segregate the time-sensitive voice traffic from the score (MaS), which is a measure that specifies the "perceptu­
data traffic, where the latter refers to man-computer or com­ al similarity between an uncoded source signal and its coded
puter-computer communications. Potentially, carriers will gain version" [1 6]. This score is subjective in nature, and is assigned
the most economy by integrating all types of traffic over a by a panel of trained listeners awarding values ranging from 1
common core. However, given the current state of the com­ to 5, with a "5" being no degradation noted and a "4" being
mercial Internet, which by and large remains a best-effort net­ perceptible degradation noted, but not annoying. An MaS
work, and traffic increases that strain a carrier's ability to score of "1" is awarded if the degradation is very annoying. Of
deploy bandwidth fast enough to remain ahead of the growth the coders listed in Table 1, the POTS mainstay, G.7 1 1 , has
curve, integrating time-sensitive voice traffic with other traffic the highest MaS score of 4.3, G.723.1 has the lowest with a
on Internet backbones is not necessarily a good idea if the 3.9, and the other two lie in between [1]. While the caliber of
carrier's goal is to offer a VoIP service with quality approach­ these coders is very similar, a VoIP system using G.729 does
ing that of POTS. begin its journey from the source coder with a slight quality
Given the state of the art, some carriers have chosen to disadvantage compared to POTS.
segregate their VoIP traffic and move it via a dedicated VoIP
network separate from the network moving data. PACKET Loss

MAINTAINING THE NETWORK: TRAFFIC ENGINEERING Packet loss can severely affect playback quality. One packet
can contain one or more speech frames depending on the
One problem with the Internet today is that, by and large, the implementation. While the actual impact of a lost packet will
system still transmits traffic as datagrams wherein switching depend on the amount of compression delivered by the voice
devices forward each packet independently of all other pack­ coder and the number of voice frames per packet, a typical
ets. As a result there is no guarantee that traffic that is part of requirement is that the loss rate be one percent or less [ 1 5,
the same information transfer will follow the same path. As 17].
routers typically update their routing tables several times an Error masking can be used to help overcome the deteriora­
hour it is possible that the end-to-end path taken by the voice tion of quality of the decoded speech due to packet loss.
traffic may shift, possibly to a path with more hops, a larger Information in the last correct frame can be used by the
delay, or more delay variation. Should this happen, the care­ receiver decoder to mask missing information. Unfortunately,
fully engineered VoIP system could be thrown into disarray packet losses encountered in digital transmission over the
and fail to meet specifications. commodity Internet generally appear in bursts. This tends to
One advantage of Multi-Protocol Label Switching (MPLS) reduce the effectiveness of voice decoder error concealment,
is that it enables virtual circuits, which allow the path through as the decoder may need to mask the loss of several consecu­
the system to be designated in advance and fixed in place [9]. tive packets, each carrying multiple frames [1 8]. Good CAC
Designating the path in advance allows the option of setting procedures at gateways to a carrier VoIP backbone can help
aside and reserving switch resources, such as buffer space and reduce the probability of lost packets.
bandwidth, for specific traffic flows [10]. Having predictable
paths through the system makes traffic engineering simpler END-To-END DELIVERY DELAY
and more feasible, increasing the probability that the system
can be configured to perform reliably. Tests have shown that the perceived quality of an interactive
Installing voice gateways at access points to the VoIP back­ voice conversation depends heavily on the time that elapses
bone can allow the implementation of connection admission between voice energy hitting a microphone and playing out on
control (CAC), such that the number of allowed voice calls the destination speaker. As this time increases, the quality
can be monitored and calls blocked if network resources are steadily degrades in that users become more likely to acciden­
insufficient to support a new request. tally talk over each other. This time need not be very large
before problems commence. The lead author has personally
experienced this difficulty during satellite-based phone calls
DEALING WITH DEGRADATION: between Korea and the United States back in the early 1980s.
The one-way delivery time of approximately 0.3 to 0.4 sec­
THE IMPACT OF PACKET Loss,
onds, coupled with brain processing time at the far end while
CODER DISTORTION, AND a reply was formulated, and then the return trip of the audio,
was perceived as an unnaturally long time interval that
END-TO-END DELAY ON VOICE QUALITY
required some practice to adjust to. There are other effects as
The previous section discussed techniques that carriers can well. For example, a pause due to a lengthy round-trip time
use to help insure that a backbone carrying VoIP traffic is might be mistaken as hesitation on the part of the party with
able to provide reliable and timely delivery of voice packets. whom you are talking. International Telecommunications
However, despite careful engineering, end-to-end delivery Union (ITU) standard G. 1 14 recommends 1 50 msec as the
delay is going to be greater than that on POTS, and packets maximum allowable end-to-end delivery value for VoIP sys­
will be lost. In this section, we discuss how these factors, cou­ tems, and that will be considered the worst case target value
pled with distortion in the voice coding algorithms, can effect for this article.
the perceived quality of the reconstructed voice at the receiv­
er. We also note one important tool that can be used to esti­ ECHO CONTROL [19]
mate the overall voice quality.
The POTS local loop typically uses two wires which carry both
inbound and outbound analog voice traffic, whereas the long
distance telephone network uses what are called four wire sys­
tems with the inbound audio (ideally) separated from the out-

26 IEEE Communications Surveys & Tutorials • Second Quarter 2004

 User
satisfaction
100
bound audio and each carried on its own path. Connec­ - Tel R=65 dB Very
tions between two wire and four wire systems are known Tel R=60dB satisfactory
" - Tel R=55dB
as hybrids. Imperfections in these interconnections allow 90 ,
, - Tel R=50dB
some of the inbound long distance audio to couple into \
._- TeIR=45dB Satisfactory
,
the outbound audio path and return to the sender. Echo \
80 \
in the network results from this coupling between the \ Some
\ users
transmit and the receive path. C>: \
\ dissatisfied
While an end-to-end VoIP network will have no ,
70 , Many
hybrids, and hence would have no internal echo paths, it ,
,
,
may still be subject to echo from other sources. Coupling ,
60 ,
between a speaker and a microphone, either acoustically ,
,
,
or via mechanical vibrations, is one potential return path ,
,
back to the original source. Electrical pick up between cir­ 50
cuits or wires (known as cross-talk) is another. These echo 0 100 200 300 400 500
sources are usually weaker than hybrid echo. ( s)
One·waydelay m
The severity of an echo depends on both the strength
of the echoed signal and the time it takes to return to the
• FIGURE 6. The effect of echo loudness and end-to-end delay on the
talker. Given a constant amplitude echo, the longer the
perceived quality of a voice call [20). (Reproduced under written per­
round trip time, the louder the echo will sound. As men­
mission from Telecommunications Industryr Association.)
tioned previously, VoIP networks will have a longer end­
to-end delay than POTS. As a result, echo that is
inaudible on POTS may become noticeable in a packet
based VoIP system. and varies between 0 and 100. Given the R rating, the quality
Two key techniques used to control echo in the Public is somewhat akin to the standard school grade scale where :2:
Switched Telephone Network (PSTN) and on VoIP networks 90 is an "A," :2: 80 is a "B," etc.
are echo cancellers and loss planning. Depending on the The R rating is defined as
strength of the expected echo and its associated time delay,
R Ro - Is - Id - Ie + A,
= (2)
these methods could be used alone or may be combined.
An echo canceller is a device that detects and eliminates where Ro accounts for the effects of noise, Is the effects of
echo. Adaptive filters are usually used here. These filters speech transmission impairments such as too much side tone
search for delayed and attenuated replicas of the inbound or a too loud signal, Id the impairments associated with delay
audio in the outbound return path. The adaptive filter will (including echo), Ie equipment imperfections, and A is an
then generate an estimate of this echo signal and subtract that advantage factor that accounts for the willingness of users to
estimate from the return path, reducing the strength of the accept substandard performance on a new service.
echo by up to 2 5-30 dB. Any residual echo can then be For example, the talker echo loudness rating (TELR) is the
removed by a processor that eliminates all weak signals below loudness loss between the talker's mouth and the ear via the
a certain threshold. echo path. It is the sum of the losses around the loop, from
Echo control on calls with short end-to-end delays can be one set's transmitter back to the receiver on the same set. The
managed effectively by deliberately introducing loss in the family of curves in Fig. 6 shows the effect of echo on the quali­
path, reducing the echoes to a point below the typical individ­ ty of the voice signal, as predicted by the E-Model. Note that
ual's audio threshold. Loss planning (also known as padding) as TELR is reduced and an echo is louder, the amount of end­
in the Public Switched Telephone Network is largely intended to-end delay available to the connection for a given perfor­
to keep all sources of echo inaudible. Loss planning can only mance quality objective on the R scale is also reduced.
remove echo where the mouth-to-ear delay is up to 20-25 ms. Given some target voice quality, the system designer can
H the delays become greater than this, echo begins to again use the E-Model to gain an understanding as to how backbone
become audible due to the previously mentioned phenomena and access network impairments might affect the caliber of the
of echo subjected to longer round trip times sounding louder transmitted voice signal over the VoIP network, and insight
than the same echo signal with a shorter round trip delay. regarding which steps might be taken to bring a substandard
Even with short delays, it is generally necessary to control sig­ result up to the quality level desired. We shall use this model
nal levels to ensure that echo control devices function proper­ shortly when we estimate the expected quality of the signals in
ly. Any echo must be low enough that an echo canceller will the test network described in the following section.
not mistake it for a direct signal. Level control becomes of For a more detailed discussion of the E-Model, the inter­
great importance in VoIP networks because of the increase of ested reader is referred to [15] or [20].
the end-to-end delay. This requires the deployment of tech­
nologies such as automatic level control and Noise Reduction
(as in lTV Recommendation G.1 69) to insure natural, com­ ENGINEERING ON-TIME DELIVERY AND
fortable speech.
MAXIMIZING THE NUMBER OF
THE E-MoDEL SUPPORTED VOICE CALLS

lTV standard G.107, known as the E-Model, attempts to pre­ To maximize resulting revenues, the system engineer will want
dict how a typical user would rate the overall quality of a to maximize the number of calls the network will support,
phone call. Based on a large set of subjective experiments, the while meeting or exceeding the quality level indicated above.
model can generate a numerical score that accounts for fac­ To better understand the impact certain design choices have
tors such as coder distortion, packet loss, end-to-end delay, on the ability of a VoIP system to support calls, it is construc­
and echo, as well as other parameters such as noise. This tive to look in some detail at a specific example. Fig. 7 shows
resulting score is known as a the transmission rating factor R, the network used in this example, where phones are connect-

IEEE Communications Surveys & Tutorials· Second Quarter 2004 27

 Gateway B

• End-to-end packet flows are connection orient­

VolP backbone
ed, and therefore packets will arrive at the desti­
nation in order of transmission.
• Using the information provided in R TP, the
receiver de-jitter buffer delays playback until
WCDeliv seconds after packet assembly in order
to insure the receiver buffer never empties. All
follow-on packets will therefore also commence
playing WCDeliv seconds after assembly, spend­
ing a total of WCDeliv seconds in transit, stored
in the queues of intermediate switches, or stored
in the receiver buffer. Note that in this situation,
QuDelays + IitDelay + PropDelay = WCDeliv
for all packets.
As mentioned earlier, current deployments of
VoIP are likely to use fixed or adaptive de-jitter
• FIGURE 7. Example carrier VoIP network. buffer delays. A large fixed de-jitter buffer delay
would have the same effect as above. An adaptive de­
jitter buffer would likely result in a reduced de-jitter
ed to PBX telephone switches which, in turn, direct external buffer delay (since the worst case delay may not be that com­
traffic to a VoIP gateway and a VoIP network. This configura­ mon), but also occasional glitches due to an emptied playback
tion is similar to that used by carriers that have segregated buffer, adversely affecting the R rating.
their interactive IP voice traffic from the data. For the numer­ Considering this scenario and Eq. 1, we have
ical example to follow, we make the following assumptions:
VCDelay + PADelay + QuDelays + IitDelay
• The voice signal is analog until it hits the PBX, where 64
kbls G.7 1 1 coding is used. The coding delay here is + PropDelay + VDDelay =
l/8000th of a second. POTS technology then transports .015 + .010* N + QuDelays + IitDelay
this digital signal to the gateway. (3)
• At the gateway, an analog signal is reconstructed and fed + PropDelay + .010 =

to a G.729 voice coder that outputs one frame of 1 0 .015 + .010* N + WCDeliv + .010 :::; Target, or
octets every 1 0 msec (8 kb/s). T o help prevent audible
.010* N + WCDeliv:::; Target - .025
glitches from frame-to-frame, G.729 includes 5 msec of
look-ahead information which overlaps the next frame. To maximize profits, a VoIP carrier desires to move as
Hence 15 msec are required to acquire the necessary many voice calls as feasible over their VoIP backbone. Equa­
voice information to code a frame. tion 3 illustrates two of the key trade-offs that can affect this:
If silence suppression is being used and the voice source is the number of frames in a packet and the tolerable delays at
silent, 0 kbls is assumed output with any necessary comfort intervening switches. As one increases, the other must
noise generated at the receiver. decrease. As the number of frames in a packet decreases,
• At the gateway, N G.729 Frames are acquired for place­ more time can be allocated to queuing delays, and trunk loads
ment in a packet for transmission. This yields a 47 + can be increased in order to carry more traffic, but a larger
N*10-byte packet for transmission (7 bytes Point-to-Point percentage of that traffic is packet overhead, not more phone
Protocol, 20 bytes IP, 8 bytes UDP, 12 bytes RTP, and calls. Conversely, as the number of frames in a packet increas­
10* Nbytes of traffic). es, less time is available for packets to spend in intervening
• It requires 10* N msec + 5 msec to acquire sufficient switches. Trunk loads must be decreased, reducing the num­
voice to generate NG.729 frames for emplacement in a ber of phone calls that can be conveyed. Using a technique
packet. The voice coder will then have up to 10 msec to outlined in [21], Figs. 8a and 8b show plots of the number of
complete compression before the next following frame is voice calls supportable over a backbone OC-12 trunk for both
collected and ready for processing. Considering this, we the fixed-rate (Fig. 8a) and variable-rate (Fig. 8b) G.729
use a voice coding and packet assembly delay of 15 msec coders, for 100 msec and 150 msec target end-to-end delay
+ 10* Nmsec below. bounds, using the network of Fig. 7. A target delivery time of
• The PBXs are connected to a gateway with a T -1 line. 50 msec is seen to be impractical for this example network,
Not shown are other PBXs or CO switches connected to even with one frame per packet, as from Equation 3 it is seen
the gateways. that WCDeliv must be less than 15 msec, an infeasible value
• The gateways are connected to the VoIP backbone by when the propagation delay alone is 20 msec. A similar situa­
OC-3s. Packet switching and statistical multiplexing are tion occurs at 1 1 frames per packet with a target delivery time
used on these connections. Not shown are other gate­ of 150 msec, and at six frames per packet and a target delivery
ways that are also attached to the backbone routers. time of 100 msec.
• The VoIP routers are connected by OC-12s. These con­ Figure 8 indicates that multiple frames per packet are nec­
nections are also packet switched and statistically multi­ essary to maximize the number of voice calls the system can
plexed. support, and that given some network configuration the
• Propagation delays between the PBX and the gateways choice of the number of voice frames to carry in a packet can
are inconsequential and are ignored. The propagation significantly impact the number of calls the system is able to
delays between the routers are assumed to be 10 msec. support. As a comparison, note that an OC-12 can carry 8,192
• The network is engineered such that packet discards due POTS phone calls. The combination of compression and
to queue overflows or bit errors are negligible and can be silence suppression allows the VoIP system to potentially ser­
ignored. Hence CAC procedures at gateways are in vice a significantly larger number of customers.
effect, and fiber optic systems are widely deployed. The choice of receiver de-jitter buffer delay also has a con-

28 IEEE Communications Surveys & Tutorials • Second Quarter 2004

 Fixed rate G.729 Variable rate G.729

_ 1
J � � �g ��:� l
40000 i---------:=------;===='ll
: 100000 i------------r=====il

I
100msec '"
'" :0
D 150msec
:0
'" .e 80000
________________

'5 30000
------------------------- ----- ----- --------------------------------------------
a
0.
0. 0.
0.
::J i.il 60000 ._-----------------------------------------

'"
!!!.
� 20000 -- ----_ . ._---- ----- ._----------------------------------
(ij
u
u
� 40000 ---- ----- ---- ---- ----------------------------------

n�
'"
u ·0
>
.� 10000
---------------------------

"'" � 20000
----- ----- ---- ---------

n
c
2
2
I
>-
>-
o������·� �����u-� n o +-L4���r-L4�
�LL����LL��-�
2 3 4 5 6 7 8 9 10 11 2 3 4 5 6 7 8 9 10 11
Number of frames per packet Number of frames per packet

• FIGURE 8_ a) Fixed rate G. 729 VolP calls supportable over OC-12 trunks for 100 and 150 msec delivery delays; b) Variable rate G. 729
VoIP calls supportable over OC-12 trunks for 100 and 150 msec delivery delays.

siderable impact on the number of calls the network can sup­ backbones can be used_ An analysis of this type of network
port. Equation 3 is based on the conservative choice of delay­ would differ from the network of Fig_ 7 in several significant
ing the initial packet of a call or talk spurt such that it plays ways_ First, the mobile phone voice coder would take longer
back WCDeliv seconds after construction at the transmitter. than a 0_7 1 1 coder to generate voice bits for movement over
This choice insures the buffer will not empty, but it delays the mobile phone radio frequency (RF) link_ This would
play back an amount of time that is only necessary in the reduce the time available for the VoIP backbone to move the
worst case, an event that may be unlikely_ If the de-jitter traffic to the destination_ If this destination were another
buffer delay is decreased, time can be freed up and trans­ mobile phone, the voice decoder would also require more
ferred to another entity of Equation 1, such as the packet decoding time than POTS' 0_7 1 1 . Second, the probability of
assembly delay (allowing a reduction in the percentage of voice frames being lost over the mobile phone's RF link will
overhead) or delays spent at packet switches (allowing an generally be higher than that of a wired link, adversely effect­
increased trunk load)_ The interested reader is referred to ing the R rating_ Third, were trans-coding necessary to con­
[22] and [23] for further information on these issues_ vert the mobile phone-compressed voice to a different
What kind of quality does the E-Model predict for the net­ protocol being used on the VoIP backbone, some additional
work of Fig_ 7? We use the following values in this example voice degradation would result compared to the example
[20]: above, further impacting the R rating_ In short, if a mobile
• A base Ro value of 94, corresponding to the value associ­ telephone is the source or destination of a call being carried
ated with a typical PSTN voice connection_ by a VoIP backbone, additional engineering attention must be
• A delay impairment value Id of 2, 4, and 6 for echo focused on maintaining quality_ A reduction in the allowed
delays of 50, 100, and 150 msec, respectively_ This corre­ end-to-end delay (to reduce echo delay impairments) may be
sponds to a system using echo cancelers sufficient to gen­ necessary to offset quality degradation caused by frame losses
erate a TELR of 60 dB in Fig_ 6_ Perfect echo over the RF link and trans-coding degradation_ A reduction in
cancellation requires a TELR of 65 dB. the number of supportable calls over the backbone would
• An equipment impairment value Ie of 10 for a 0_729 result.
fixed-rate coder, and 11 for a variable-rate coder. These Another type of system likely to be connected to a carrier
values reflect the voice quality degradation of these VoIP backbone has one (or both) of the POTS PBX systems
reduced bit rate coders as compared to the standard of Fig_ 7 replaced with LAN-based VoIP systems_ Traffic from
0_7 1 1 coder. the LAN VoIP phone will be mixed in with data traffic on the
• A and Is are both set equal to zero, indicating the user corporate LAN prior to being segregated and routed to the
expects similar quality on a VoIP phone as on a regular VoIP network gateway_ It will be much more difficult for a
phone, and that there are no speech impairments other carrier to offer and engineer high QoS when a significant part
than those associated with the voice coder, as reflected in of the end-to-end network is on a LAN, especially if in the
Ie_ LAN the VoIP traffic is not prioritized_ This situation will
Substituting these into Eq_ 2 yields R values ranging from have some of the same characteristics as that of mobile phone
78 to 82 for the fixed-rate coder, and 77 to 81 for the variable­ voice sources and sinks, specifically higher delays at the LAN
rate coder (with the lower values associated with the higher voice coders and decoders, possible trans-coding degradation
end-to-end delivery delays)_ Note that these R values would at the gateways, and higher packet losses on the LAN_ Engi­
be reduced further were packet losses of any significance_ neering steps to offset these would be similar to that men­
If the minimum acceptable R rating is a 70, these all pass tioned in the paragraph above, but would also include an
muster in terms of quality_ But the values calculated do corre­ increase required in the receiver de-jitter buffer delay to off­
spond to a perceived voice quality on the borderline between set the jitter inherent with LAN traffic_ Again, a reduced
where most users are satisfied and where some of the more number of supportable calls over the backbone would likely
critical users will be unsatisfied_ be required in order to meet QoS goals_

ALTERNATE TRAFFIC SOURCES FOR A VolP BACKBONE

WHAT ABOUT ATM?
Two other types of networks are likely to be connected to a
carrier VoIP backbone and are mentioned briefly below_ Designed in the late 1980s for a mixed traffic environment,
Instead of using standard PSTN techniques to interconnect ATM already has many of the QoS protocols tested and in
mobile telephone cell sites to the rest of the world, VoIP place that IP networks are now seeking to deploy_

IEEE Communications Surveys & Tutorials· Second Quarter 2004 29

 ... •
Line speed
� Cell switch
u
.

co TOM/StatMux ... .... •

0..
co
u
Active I d le
Packet switch
OJ
c StatMux .. ... •
.� Traffic Overhead
ro Circuit switch
u
TOM
Carrying capac ity = traffic(b/s)/Iine speed (b/s)

0% bu rsty Offered traffic m ix 1 00% bursty

1 00% fixed rate 0% fixed rate • FIGURE 10. Carrying capacity. Accounting for overhead and
idle bandwidth.
• FIGURE 9. Switched network carrying capacities for high-speed
trunks.
developed in the late 1980s and early 1990s, and was exten­
sively adopted by providers as they realized that a cell­
Figure 9 shows a plot of the carrying capacity for the three switched architecture allowing a combination of TDM and
main digital technologies that have been deployed by telecom­ statistical multiplexing would be more cost effective in an
munications carriers over the last 40 years: circuit switching environment with a mixture of fixed-rate and bursty traffic. As
and TDM; cell switching and a combination of TDM and sta­ the traffic mix continues to become even more heavily data
tistical multiplexing; and packet switching and statistical multi­ oriented at the beginning of the 2 1st century, Internet tech­
plexing. Carrying capacity is defined here as the ratio of nologies are now justifiably receiving considerable attention.
application traffic moved, divided by the line speed required In terms of its ability to support voice calls, Voice over
to carry it. From Fig. 10 it can be seen that this parameter ATM (VoATM) is superior to any other technology previous­
accounts for both the inability of certain multiplexing tech­ ly discussed in this article. If the routers of Fig. 7 are replaced
niques to fully load a line and maintain acceptable perfor­ by ATM switches, and fixed-rate 8 kb/s G.729 voice coders
mance, and the overhead associated with moving the traffic. supported by CBR virtual circuits are used over the backbone,
Referring back to Fig. 9, the carrying capacity is plotted the OC-12s would be able to support approximately 55,000
against offered traffic mixes ranging from 100 percent fixed­ phone calls. If, instead, variable-rate G.729B coders using
rate and no bursty traffic on the left hand side of the plot, to silence suppression are used, the ATM system using variable
100 percent bursty and no fixed-rate traffic on the right side. bit rate/real-time virtual circuits would again be able to sup­
In the middle of the horizontal axis, the offered load is there­ port a greater number of phone calls than the variable-rate
fore half fixed-rate traffic and half bursty traffic such as data. VoIP system, at a comparable quality level.
Endpoints on this graph were plotted based on the following However, the viability of ATM likely lies not in its ability
logic. If fixed-rate traffic is fed to a circuit-switched TDM net­ to move voice, but in its ability to adapt to the changes in the
work, trunk lines can be fully loaded. Likewise, an ATM net­ offered traffic mixes faced by carriers today, specifically the
work could map the fixed-rate traffic onto constant bit rate increasing percentage of bursty data. The cost effectiveness of
virtual circuits whose cells could be moved by TDM, also fully ATM is not so strong in this type of an environment.
loading the trunk, but in this case slightly more than 10 per­
cent of the bandwidth would be lost to ATM overhead. Fixed­
rate traffic fed to a packet network would have to be CONCLUSIONS
packetized and statistically multiplexed onto the trunk band­
width. Due to a mix of variable-length packets, traffic sources This article has provided a tutorial overview of some of the
randomly entering and leaving, and the need to try to hold jit­ problems and possible solutions identified with packet­
ter down to negligible levels in order to mimic "fixed-rate" switched voice transport. The focus was on carrier issues asso­
quality, we assumed for this plot that approximately 30 to 40 ciated with preserving the quality of VoIP, but traditional
percent of the bandwidth will be lost due to overhead and an POTS technology as well as non-traditional VoATM were
inability to fully load the trunk. Endpoints on the right hand also discussed. For papers that focus on other aspects of
side of the graph for the packet-switched and cell-switched Internet telephony, especially the signaling process, the reader
points were assigned based on average packet sizes of 300 is referred to [26] and [27].
bytes requiring movement (including overhead) [24], and 70 VoIP technology is clearly viable, but deploying a high­
to 80 percent statistically multiplexed trunk loads for the quality system requires careful system engineering, which is
packet and cell networks. The cell network maps the packets not a trivial task. Quality comparable to POTS can be
into ATM cells. The circuit-switched TDM network's right obtained. Couple this with VoIP's potential ability to support
hand end point is assigned under the assumption that circuits a greater number of phone calls than a POTS network of
(leased lines) do not have their bursty traffic aggregated. It is comparable bandwidth, and the recent shift to a traffic mix
somewhat arbitrarily specified at 25 percent based on average dominated by bursty computer-to-computer data that is best
traffic statistics observed on corporate Frame Relay access carried by packet switching, and the conclusion is that voice
lines. Points in between the endpoints are weighted averages. over the Internet is a technology with a promising future ... at
A more rigorous analysis would clearly be required to pin­ least for the immediate future.
point the exact traffic mixes where the different technologies However, the communications industry does not stand still.
cross over in terms of their efficacy in moving the offered As MPLS penetration increases and operators become more
load. In [25] it is discussed in detail how this can be accom­ familiar with its use, voice over MPL S may see increased
plished. Nevertheless, Fig. 9 does offer some insight into why usage as, like VoIP, this technique can easily take advantage
different technologies have come to the forefront over the of voice compression, and it has considerably less overhead
years. In the 1970s and 1980s the traffic mix offered to the than that associated with a typical VoIP packet [28].
typical telecommunications carrier was almost wholly fixed­ Additionally, there may be at least one more major
rate voice. Figure 9 shows that particular mix is most effec­ paradigm shift in the coming years, that being the emergence
tively carried by a circuit-switched TDM network. ATM was of high-quality, high-bit-rate video that will require reliable

30 IEEE Communications Surveys & Tutorials • Second Quarter 2004

and timely delivery. Internet technologies may not be the best [20] TSB1 1 6, "Voice Oual ity Recommendations for IP Telephony, "
choice when video dominates the traffic mix offered to carri­ Te/ecommun. Industry Association, Mar. 2001 .
[2 1 ] G. Scheets, R. S i ngh, and M. Parperis, "Analyzing End-to-End
ers [29]. If either of these proves to be the case, the transport
De l ivery Delay in P u re Vo l P Networks, " 45th IEEE MWSCAS,
techniques used to move voice will have to evolve once again.
Aug. 2002.
[22] D . Vleeschauwer et al., "An Accurate Closed-Form Form ula to
ACKNOWLEDGMENTS Calcu late the Dej ittering Delay in Packetized Voice Transport. "
Proc. IFlP- TC6/European Commission Int 'l. Conf. Ne t . , May
The authors wish to acknowledge the benevolent prodding of 2000.
the editor and reviewers, which resulted in a much improved [ 2 3 ] D . Vleeschauwer, J. Janssen, and G. Petit, " Delay Bou n d s for
article than that originally submitted. Low- B i t- Rate Voice T r a n s p o rt over IP Networks , " Proc. SPIE
Conf. Perf. a n d Control of Network Systems III, v o l . 3 8 4 1 ,
Sept. 1 999, pp. 40-48.
REFERENCES
[24] K. Thompson et al., "Wide Area I nternet Traffic Patterns and
[ 1 ] J . B e l l a m y , Digital Telephony, 3 r d Ed . , J o h n W i ley & S o n s ,
Characteristics," IEEE Network, Nov.!Dec. 1 99 7 , p . 1 0-22.
2000.
[25] G . Scheets and M . Allen, "Switched Network Carrying Capaci­
[ 2 ] R. C o x , " T h ree N e w S p e e c h C o d e r s f r o m t h e I T U C o v e r a
ties, " chapter in CRC Handbook of Communica tions Technolo­
Range of Appl ications, " IEEE Commun. Mag., Sept. 1 99 7 .
gies: The Next Decade, 2000, CRC Press.
[ 3 ] A . Benyassine e t al., " I T U · T Recommendation G . 7 2 9 Annex B :
[ 2 6 ] B. Goode, "Voice Over I nternet Protocol (Vo I P) , " IEEE Proc.,
A S i lence Compression Scheme for U s e with G . 7 2 9 Opt i m i zed
Sept. 2002.
for V . 7 0 D ig ita l S i m u l taneous Voice a n d Data Appl ications, "
[ 2 7 ] W. Jiang et al., " I ntegrating I n ternet Telephony Service, " IEEE
IEEE Commun. Mag., Sept. 1 99 7 .
Internet Computing, June 2002.
[ 4 ] P . Brady, " A M o d e l f o r O n · Off Speech Patterns i n Two·Way
[ 2 8 ] D. Wright, "Voice over MPLS Compared to Voice over Other
Conversation, " Bell System Tech. Journal, Sept. 1 969.
Packet T r a n s port Tec h n o l o g i es, " IEEE Commun. Mag . , Nov.
[ 5 ] P . P . M ishra a n d H . Sar an, " Capacity Management and Rout·
2002.
in g Polic ies for Voice over IP Traffic , " IEEE Network, Mar.!Apr.
[ 2 9 ] J. L u b a c z , " Th e IP S y n d r o m e , " IEEE Commun. Mag. , F e b .
2000.
2000.
[6] S. Sh e n ker, " F un d amenta l Design Issues for the Future I nter·
net," IEEE JSAC, Sept. 1 9 95.
[ 7 ] C. Metz, "IP O oS : Trave l i n g in Fi r st Class on the Internet," IEEE BIOGRAPHIES
Internet Computing, Apr. 1 999, pp. 84-88.
[8] X. Xiao and L. Ni, " I n ternet OoS: A Big Picture," IEEE Network, G E ORGE SCHE ETS (scheets@okstate . e d u ) received a B . S . i n general
Mar'!Apr. 1 99 9 . e n g i n eer i n g from the U n ited States M i l itary Aca d e my at West
[ 9 ] T . Li, " M P L S a n d the Evo l v i n g I n ternet Arch itectu r e , " IEEE Point i n 1 97 5 , and following five years i n the U . S. Army as a Sig­
Commun. Mag., Dec. 1 999. n a l Corps officer, he received M.S. and P h . D . degrees i n electrical
[1 0] C. Metz, "RSVP: General·Pu rpose Signaling for I P," IEEE Inter­ e n g i n e e r i n g from K a n sas State U n ivers i ty in 1 9 8 4 a n d 1 9 8 7 ,
net Comp., June 1 999. respectively. H e then joi ned the faculty a t Oklahoma State U n iver­
[1 1 ] M . Perkins et a l . , "Speech Transmission Performance Planning sity. H is research i n terests i n c l u d e telecom m u n ications network
in Hybrid IP/SCN Networks, " IEEE Commun. Mag., J u ly 1 99 9 . analysis and design, signal processing for intercept receivers, and
[ 1 2 ] F . Sch n e i d e r e t al., " B u i l d i n g Trustworthy Systems: Lessons teac h i ng i n a virtual environment.
from the PTN and Internet. " IEEE Internet Comp., Dec . 1 999.
[1 3] B . L i et al., "OoS-Enab led Voice Support i n the Next-Genera­ MARiaS PARPERIS (marios. parperis@wcg.com) received the H igher
tion I nternet: Issues, Existing Approaches, and C h a l l en ges, " National D i p loma in electrical engineering from the Higher Tech n i ­
IEEE Commun. Mag., Apr. 2000. c a l I n stitute i n N icosia, Cyprus i n 1 9 9 1 , a n d the B . E n g . and
[1 4] L. Mathy et al., "The I n ternet: A G lobal Teleco m m u n ications M . En g . degrees i n electrical engineering from Carleton U n iversity
Solution?, " IEEE Network, JU ly/Aug . 2000. in Ottawa, Canada in 1 9 9 7 a n d 1 9 9 9 , respective ly. In 1 9 9 9 h e
[ 1 5 ] J. J a n ssen et al., " Assess i n g Voice O u a l ity in Packet-Based jo ined Newbridge Networks/Alcatel as a strategic network e n g i ­
Telephony," IEEE Internet Computing, June 2002. n e e r . He is currently a network arc h i tect with W i l T e l Commun ica­
[ 1 6 ] P. N o l l , " Au d i o Cod i n g , " c h apter in The Communications t i o n s . H i s research i n terests a r e in v o i c e q u a l ity i ssu es in Vo l P
Handbook, 1 99 7 , CRC & I E E E Press. networks. He has been a n I E EII E E E member since 1 990.
[ 1 7 ] P. Goyal e t al., " I ntegration of C a l l S i g n a l i n g and Resource
Management for I P Telephony, " IEEE Network, May/June 1 99 9 . RITU SINGH (rcsingh @hotm a i l .com) received the B. A. degree with
[ 1 8 ] M . Bore l la et al., " I n ternet Packet L o s s : Measurement a n d honors in Engl ish l i terature from Kurukshetra U n iversity, Ind ia, in
I m p l i c a t i o n s for E n d - t o - E n d Oos, " Proc. 1998 ICPP Wksps. , 1 9 80, B.S. (magna cum laude) and M . S . degrees in electrical engi­
Aug. 1 99 8 . neering from the U n iversity of Tulsa, Tulsa, OK, in 1 990 and 1 996,
[1 9] Nortel Networks White Paper, "Voice over Packet: A n Assess­ respectively, and the P h . D . degree i n electrical engineering from
m e n t of V o i c e P e r fo r m a n c e o n P a c k e t N e t wo r k s " , Oklahoma State U n iversity in 2002. Cu rrently she is president and
http ://www . n o rte l n etwo rks . c o m/ p r o d u cts/I i br a ry/ c o I I ater­ CEO of N D R S Tec h n o logies and engaged i n the deve l o pm ent of
al/7 4007 .25-09-01 . pdf software for medical and electromagnetic applications.

IEEE Communications Surveys & Tutorials· Second Quarter 2004 31