ON NLMS ESTIMATION FOR VOIP PLAYOUT DELAY
ALGORITHMS
Improving Delay Spike Detection
Karen S. Miranda-Campos and V´ıctor M. Ramos R.
Metropolitan Autonomous University, Electrical Engineering Department
Networks & Telecommunications Research Team, Mexico
09340 Iztapalapa, Mexico
Keywords:
Playout control delay algorithms, VoIP, autoregressive estimation.
Abstract:
Voice over IP (VoIP) applications are now very popular and widely used on the Internet. Such applications
use receiver playout buffers to smooth delay variations so as to reconstruct the periodic form of the transmit-
ted packets. Packets arriving after their scheduled playout time are considered late and are not played out.
Playout delay control algorithms often operate by updating the playout delay between periods of silence. A
recent class of playout control algorithms has received particular attention; this class of algorithms uses au-
toregressive measures on the network delay so as to estimate future packet delay values and adjust the playout
delay accordingly. In this work, we compare two algorithms previously proposed that use such autoregressive
approach; both playout algorithms use a normalized least-mean square (NLMS) adaptive predictor. The differ-
ence between both algoritms is that the second one is an extension of the first that adds delay spike detection.
We demonstrate, by using Internet audio packet traces that, contrary on what was claimed, the algorithm that
uses spike detection does not overperfom the first one. Finally, we propose an algorithm based on the original
NLMS algorithm with delay spike detection that overperforms the previous two NLMS playout algorithms.
1 INTRODUCTION
Delay, jitter, and packet loss in packet-switched wide-
area networks, such as the Internet, are the main fac-
tors impacting audio quality of interactive multime-
dia applications. Today’s Internet still operates in a
best-effort basis, and thus, the impact caused by such
phenomena must be alleviated by employing end-to-
end control mechanisms. Audio applications such as
NeVoT (Schulzrinne, 1992), Rat (Sasse and Hardman,
uary), or more recently Skype and Google Talk gen-
erate packets spaced at regular time intervals. The
traffic generated by an audio source is divided into
periods of activity, called talkspurts, and periods of
silence. Silence periods are periods where no audio
packets are transmitted.
Audio packets encounter variable delay while
crossing the Internet, this is mainly due to the vari-
able queueing time in routers. Such delay variability
modifies the periodic form of the transmitted audio
stream. In order to playout the received stream, an
application must reduce or eliminate this delay vari-
ability, by buffering the received packets and playing
them out after a certain deadline. Packets arriving af-
ter their corresponding deadline are considered late
and are not played out. If the playout delay is in-
creased, the probability that a packet will arrive be-
fore its scheduled playout time also increases. This
reduces the number of packets artificially dropped in
the playout buffer. However, very long playout de-
lays have a negative impact on the interactivity of an
audio session. Obviously, there exists a trade-off be-
tween delay and loss due to late packets. For interac-
tive audio, packet delays up to 400 ms and loss rates
up to 5%–10%, depending on the audio codec used,
are considered adequate (Jayant, 1980).
In this work, we compare two algorithms previ-
ously proposed in the literature that use an autore-
gressive approach; both playout algorithms use a nor-
malized least-mean square (NLMS) adaptive predic-
tor. The difference between both algoritms is that the
second one is an extension of the first that adds delay
spike detection. From here, we call this second al-
gorithm E-NLMS. We demonstrate, by using real In-
342
S. Miranda-Campos K. and M. Ramos R. V. (2008).
ON NLMS ESTIMATION FOR VOIP PLAYOUT DELAY ALGORITHMS - Improving Delay Spike Detection.
In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 342-347
DOI: 10.5220/0001940703420347
Copyright
c
 SciTePress

Playout
t
k
i
Talkspurt
k
Silence
k
Talkspurt
k+1
Talkspurt
k
Talkspurt
k+1
Silence
k
∆
Receiver
a
k
i
p
k
i
Sender
Figure 1: The timings between the transmission, reception
and playout of packets.
ternet audio packet traces that, contrary on what was
claimed, the E-NLMS algorithm does not overperfom
the original NLMS algorithm. Finally, we propose an
improvement with spike detection that overperforms
both, the NLMS and the E-NLMS algorithms.
The remainder of the paper is as follows. In Sec-
tion 2, we provide some background about playout
delay algorithms. In Section 3 we describe the related
work on which we base our proposal, and particularly,
on the algorithms using the NLMS playout algorithms
just cited. In Section 4, we present our ongoing work
on an improvement of the original NLMS algorithm
proposed by DeLeon. Finally in Section 5, we con-
clude and present our current and future work.
2 BACKGROUND
Receivers use a playout buffer to smooth the stream
of audio packets. This smoothing is done by delay-
ing the playout of packets to compensate for variable
network delay. The playout delay can be either con-
stant through the whole session, orcan be adjusted be-
tween talkspurts. Moreover, in a recent work (Liang
et al., 2001), it has been shown that by using a tech-
nique called packet scaling, it is possible to change
the playout delay from packet to packet while keep-
ing the resulting distortion within tolerable levels. In
this paper we only focus on the per-talkspurt playout
delay adjusting approach.
Figure 1 shows the different stages incurred in an
audio session. The i-th packet of talkspurt k is sent at
time t
i
k
, it arrives at the receiver at time a
i
k
, and is held
in the smoothing receiver’s playout buffer until time
p
i
k
, when it is played out. Within a talkspurt, packets
are equally spaced at the sender by time intervals of
length ∆ seconds.
By delaying the playout of packets and dropping
those that arrive after their deadline, we are able to
reconstruct the original periodic form of the stream.
This adjusting mechanism results in a regenerated
stream having stretched or compressed silence peri-
ods compared to the original stream. These changes
are not noticeable by the human ear if they are kept
within tolerable small levels.
In Fig. 1, a dropped packet due to a late arrival is
represented by a dashed line. A packet is artificially
dropped if it arrives after its scheduled deadline p
i
k
.
The loss percentage can be reduced by increasing the
amount of time that packets stay in the playout buffer.
An efficient playout control algorithm must take into
account the trade-off between loss and delay in order
to keep both parameters as low as possible.
Throughout the paper, we use the notation de-
scribed in Table 1. For the validation of our algo-
rithm, we consider the packet traces generated with
the NeVoT audio tool that are described in (Moon
et al., 1998). We choose to use those traces since they
are packet traces generated during real audio conver-
sations. In (DeLeon and Sreenan, 1999) and later
in (Shallwani and Kabal, 2003), DeLeon and Shall-
wani use traces generated with the ping program for
the latter, and traces generated between three hosts in
the US and one host in the UK. Those traces are not
available to everyone and thus we cannot verify the
delay behavior in them. On the other side, the au-
dio traces provided by Moon, contain the sender and
receiver timestamps of transmitted packets that are
needed for the implementation of any playout delay
control algorithm. In these traces, one 160 byte audio
packet is generated approximately every 20 ms when
there is speech activity. A description of the traces
(reproduced from (Moon et al., 1998)) is depicted in
Table 2.
A typical sample of packet end-to-end delays is
shown in Fig. 2. A packet is represented by a dia-
mond and talkspurt boundaries by dashed rectangles.
The x-axis represents the time elapsed at the receiver
since the beginning of the audio session. Only the
variable portion of the end-to-end delay (d
i
k
) is rep-
resented on the y-axis of Fig. 2. To this end, the
constant component of the end-to-end delay (mostly
caused by the propagation delay) is removed by sub-
tracting from packet delays their minimum over all
the corresponding trace. By considering the variable
portion of the end-to-end delay, synchronization be-
tween sender and receiver clocks can be avoided.
We observe in Fig. 2 the presence of delay spikes.
This phenomenon in end-to-end delay has been pre-
viously reported in the literature (Ramjee et al., 1994;
Bolot, 1993). Delay spikes represent a serious prob-
lem for audio applications since they affect the per-
formance of playout delay adaptation algorithms. A
delay spike is defined as a sudden large increase in the
end-to-end delay followed by a series of packets arriv-
ing almost simultaneously, leading to the completion
ON NLMS ESTIMATION FOR VOIP PLAYOUT DELAY ALGORITHMS - Improving Delay Spike Detection
343

Table 1: Definition of variables.
Param. Meaning
L The total number of packets arriving at the receiver during a session.
N The total number of talkspurts in a session.
N
k
The number of packets in talkspurt k.
t
i
k
The time at which the i-th packet of talkspurt k is generated at the sender.
a
i
k
The time at which the i-th packet of talkspurt k is received.
d
i
k
The variable portion of the end-to-end delay of the i-th packet in talkspurt k.
d
i
k
= a
i
k
− t
i
k
− min
1≤k≤N
1≤i≤N
k
(a
i
k
− t
i
k
).
p
i
k
The time at which packet i of talkspurt k is played out.
10 11 12 13 14 15 16 17 18
0
0.5
1
1.5
2
2.5
Time [s]
Delay [s]
Packet arrivals and talkspurt boundaries
(a) A delay spike spanning through two consecutive talk-
spurts.
550 551 552 553 554 555 556 557
0
0.5
1
1.5
2
2.5
Time [s]
Delay [s]
Packet arrivals and talkspurt boundaries
(b) A delay spike spanning through three consecutive talk-
spurts.
Figure 2: Delay spikes in end-to-end delay measurements.
of the spike (Ramjee et al., 1994).
Delay spikes can be contained within a single talk-
spurt or can span over several talkspurts. Figure 2(a)
shows a delay spike spanning through two consecu-
tive talkspurts. Figure 2(b) shows a delay spike span-
ning over three talkspurts. Since the playout delay
is generally updated between talkspurts, a playout al-
gorithm behaves better when delay spikes span over
more than one talkspurt. Only in this way, a playout
algorithm can react adequately to the spike by setting
the playout delay according to the experienced delay.
If the spike vanishes before the end of a talkspurt, the
playout algorithm will not have enough time to set the
playout time accordingly.
In the next section, we briefly describe the al-
gorithms proposed by DeLeon. Playout delay is
adapted from talkspurt to talkspurt based on past
statistics of the delay process. The playout delay
of the first packet of each talkspurt is the basetime
of the deadlines for subsequent packets in the same
talkspurt. This principle is the basis for most of the
existing playout adaptation algorithms (Kansal and
Karandikar, 2001; Ramjee et al., 1994; Moon et al.,
1998; Pinto and Christensen, 1999; Ramos et al.,
2003).
3 RELATED WORK
The algorithms that perform better in (Ramjee et al.,
1994) are 1 and 4. Algorithm 1 estimates the delay
using a FIR filter; algorithm 4 operates similarly but
it adds delay spike detection. To calculate
ˆ
d
i
k
and ˆv
i
k
,
the packet’s sender and receiver timestamps, t
i
k
and
a
i
k
, are read from a trace file. Both algorithms differ
only in the way they calculate
ˆ
d
i
k
and ˆv
i
k
. Algorithm 1
computes these statistics as follows:
ˆ
d
i
k
= α
ˆ
d
i−1
k
+ (1− α)d
i
k
and
SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications
344

Table 2: Description of the traces.
Trace Sender Receiver Start time Length [s] Talkspurts Packets
1 UMass GMD Fokus 08:41pm 6/27/95 1348 818 56979
2 UMAss GMD Fokus 09:58am 7/21/95 1323 406 24490
3 UMAss GMD Fokus 11:05am 7/21/95 1040 536 37640
4 INRIA UMass 09:20pm 8/26/93 580 252 27814
5 UCI INRIA 09:00pm 9/18/93 1091 540 52836
6 UMass Osaka University 00:35am 9/24/93 649 299 23293
ˆv
i
k
= αˆv
i−1
k
+ (1− α)|
ˆ
d
i
k
− d
i
k
|,
where d
i
k
= a
i
k
− t
i
k
, and α has the default value of
0.998002. Once
ˆ
d
i
k
and ˆv
i
k
are computed, the play-
out time of the i-th packet of talkspurt k is set by both
algorithms as follows:
p
i
k
=

t
i
k
+
ˆ
d
i
k
+ βˆv
i
k
, for i = 1 .
p
1
k
+ (t
i
k
− t
1
k
), for 1 < i ≤ N
k
.
(1)
This is how most algorithms falling into this class
operate.
From now on, we focus only on the algo-
rithms proposed first by DeLeon (NLMS) in (DeLeon
and Sreenan, 1999) and later by Shallwani (E-
NLMS) (Shallwani and Kabal, 2003). The work
in (Shallwani and Kabal, 2003) aims to improve the
NLMS algorithm proposed by DeLeon by introducing
delay spike detection. Table 3 shows the parameters
of the NLMS algorithm used in both papers:
The algorithm proposed by DeLeon uses a simple
adaptive NLMS approach:
h
i+1
= h
i
+
µ
n
T
i
n
i
+ a
n
i
e
i
(2)
This NLMS algorithm minimizes the mean
squared error (mse) between the current delay sample
and its estimate. Previous samples are passed through
a FIR filter to compute the current estimate. The mse
is then used to update the weights of the adaptive fil-
ter.
Later, Shallwani proposed an extension of the
DeLeon algorithm by adding delay spike detection
which we call E-NLMS. He claims that, by adding
spike detection the performance of the NLMS algo-
rithm for playout delay is improved. We show in next
section that the latter is not true for most of the cases
Table 3: NLMS parameters.
Parameter Value
h
0
[1 0 ...0]
N
nlms
18
µ 0.01
using real audio traces rather than ping traces measur-
ing the RTT.
4 RESULTS
Before we can test the algorithms proposed by
DeLeon (NLMS) and Shallwani (E-NLMS), and next
the improvement we propose to improve the perfor-
mance of the NLMS algorithm, we define the set of
performance measures we use in our paper. To as-
sess the performance of a playout adaptation algo-
rithm, we focus on the total number of packets that
are played out during an audio session, as well as on
the experienced average end-to-end delay. Suppose
we are given a packet audio trace with the sender and
receiver timestamps of audio packets. Let p
i
k
, N, L,
N
k
, t
i
k
, and a
i
k
be defined as in Table 1. As in (Moon
et al., 1998), we define r
i
k
to be a variable indicating
if packet i of talkspurt k is played out or not. So, r
i
k
is
defined as:
r
i
k
=

0, if p
i
k
< a
i
k
.
1, otherwise.
The total number of packets, T, played out in an
audio session is thus given by:
T =
N
∑
k=1
N
k
∑
i=1
r
i
k
. (3)
The average playout delay, D
avg
, is equal to :
D
avg
=
1
T
N
∑
k=1
N
k
∑
i=1
r
i
k
[p
i
k
− t
i
k
]. (4)
Finally, the loss percentage, l, is equal to :
l =
L− T
L
× 100. (5)
With the performance measures defined above, we
choose to replace the delay spike detection in the
Shallwani’s algorithm by a modification of the delay
spike detection proposed by Ramjee. We choose to
adjust the var parameter in line 8 of the Algorithm
ON NLMS ESTIMATION FOR VOIP PLAYOUT DELAY ALGORITHMS - Improving Delay Spike Detection
345

(a) Comparing DeLeon’s, Shallwani’s and NLMS-mod algo-
rithms with Trace 1.
(b) Comparing DeLeon’s, Shallwani’s and NLMS-mod algo-
rithms with Trace 3.
Figure 3: Performance comparison.
shown bellow to an adaptive value as a function of the
previous recent delay spikes. The original NLMS al-
gorithm (DeLeon) is kept during the “ NORMAL ” mode.
Algorithm 1 Algorithm 4 by Ramjee et al.
1: n
i
= Receivertimestamp− Sendertimestamp;
2: if (mode == NORMAL) then
3: if(abs(n
i
− n
i−1
) > abs(ˆv) ∗ 2+ 800);
4: var = 0;
5: mode = IMPULSE;
6: else
7: var = var/2+ abs((2n
i
− n
i−1
− n
i−2
)/8);
8: if (var <= 63); then
9: mode = NORMAL;
10: n
i−2
= n
i−1
;
11: n
i−1
= n
i
;
12: return;
13: end if
14: end if
15: if mode == NORMAL then
16:
ˆ
d
i
= 0.125∗ n
i−1
= 0.875∗
ˆ
d
i−1
;
17: else
18:
ˆ
d
i
=
ˆ
d
i−1
= n
i
= n
i−1
;
19: end if
20: ˆv
i
= 0.125∗ abs(n
i
−
ˆ
d
i
) + 0.875 ∗
ˆ
d
i−1
;
21: n
i−2
= n
i−1
;
22: n
i−1
= n
i
;
23: return;
We call our new NLMS algorithm NLMS-mod.
Then we test our algorithm with the packet audio
traces provided by Sue Moon and compare it with the
DeLeon’s and Shallwani’s algorithms.
Figure 3 shows the corresponding results for
traces 1 and 3. From the figures we note the follow-
ing:
• Our NLMS-mod algorithm overperforms the two
other algorithms. Moreover, it behaves better for
the loss rates of interest (bellow 5%–10% depend-
ing on the audio codec used).
• One very important result is that, contrary on what
was claimed by Shallwani, the E-NLMS algo-
rithm he proposed in (Shallwani and Kabal, 2003)
does not perform better than the DeLeon’s NLMS
algorithm. This is a a pretty important result,
since it demonstrates that the delay process of
ping packets may not model well a delay process
obtained by a real voice conversation. Moreover,
this result might reflect a bug in the Shallwani’s
algorithm.
5 CONCLUSIONS AND FUTURE
WORK
We presented in this paper an improvement of the
original NLMS algorithm for playout delay control
proposed by DeLeon in (DeLeon and Sreenan, 1999).
With trace-based simulation, we compared our algo-
rithm with DeLeon’s and Shallwani’s algorithms and
we proved that our NLMS-mod algorithm overper-
forms them for the loss rates of interest in interactive
applications as VoIP.
An additional result is that we prove that, contrary
on what was claimed, the Shallwani’s E-NLMS algo-
rithm does overperform the DeLeon’s algorithm.
Our future work will focus on adapting the size
SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications
346

of the memory of the NLMS algorithm in a dynamic
fashion, so the memory size of the predictor adapts
to the network state. Other research directions we are
interested are exploring the behavior of different au-
toregressive estimators for playout delay control.
REFERENCES
Bolot, J. (1993). End-to-end packet delay and loss behav-
ior in the Internet. In Proceedings of the ACM SIG-
COMM, pages 289–298.
DeLeon, P. and Sreenan, C. (1999). An adaptive predictor
for media playout buffering. In Proc. of IEEE Inter-
national Conference on Acoustics, Speech, and Signal
Processing (ICASSP), pages 3097–3100.
Jayant, N. (1980). Effects of packet loss on waveform coded
speech. In Proceedings of the International Confer-
ence on Computer Communications, pages 275–280.
Kansal, A. and Karandikar, A. (2001). Jitter-free audio
playout over best effort packet networks. In ATM Fo-
rum International Symposium, New Delhi, India.
Liang, Y. J., Farber, N., and Girod, B. (2001). Adaptive
playout scheduling and loss concealment for voice
communications over IP networks. IEEE Transactions
on Multimedia.
Moon, S. B., Kurose, J., and Towsley, D. (1998). Packet
audio playout delay adjustment: Performance bounds
and algorithms. ACM/Springer Multimedia Systems,
6:17–28.
Pinto, J. and Christensen, K. J. (1999). An algorithm for
playout of packet voice based on adaptive adjustment
of talkspurt silence periods. In Proceedings of the
IEEE Conference on Local Computer Networks, pages
224–231.
Ramjee, R., Kurose, J., Towsley, D., and Schulzrinne, H.
(1994). Adaptive playout mechanisms for packetized
audio applications in wide-area networks. In Proceed-
ings of the IEEE Infocom, pages 680–688.
Ramos, V., Barakat, C., and Altman, E. (2003). A moving
average predictor for playout delay control in VoIP.
In Proceedings of the XI International Workshop on
Quality of Service.
Sasse, A. S. and Hardman, V. (February). Multi-way multi-
cast speech for multimedia conferencing over hetero-
geneous shared packet networks. RAT-robust audio
tool. Technical report, EPSRC Project #GRIK72780.
Schulzrinne, H. (1992). Voice communication across the
Internet: a network voice terminal. Technical report,
University of Massachusetts, Amherst.
Shallwani, A. and Kabal, P. (2003). An adaptive playout al-
gorithm with delay spike detection for real-time VoIP.
In Proceedings of the IEEE Canadian Conference on
Electrical Computer Engineering, pages 997–1000.
ON NLMS ESTIMATION FOR VOIP PLAYOUT DELAY ALGORITHMS - Improving Delay Spike Detection
347