644
IEEE COMMUNICATIONS LETTERS, VOL. 8, NO. 10, OCTOBER 2004
A New Approach for VoIP Traffic Characterization
A. Estepa, R. Estepa, and J. Vozmediano
Abstract—This letter proposes a new frame generation model
for those audio codecs which handle SID frames and deduces an
analytical expression for the mean bit rate at the input of the IP
network as a function of the number of frames per packet. The
new model is experimentally validated for the G.729B, G.723.1, and
GSM AMR codecs. Results show that the error of the estimation
of the mean bit rate can be reduced significantly compared to the
traditional ON–OFF model.
Index Terms—Audio codecs, IP telephony, voice communications, voice over IP (VoIP), voice traffic modeling.
I. INTRODUCTION AND RELATED WORK
T
HE current audio codecs can improve the speech quality
by reproducing the speaker’s background noise. This feature is supported by a special frame type called SID, generated
at the speaker’s side, which describe the main characteristics
of the background noise. The SID frames are generated during
voice inactivity periods, and their coding scheme differs from
that of the voice frames (ACT frames). SID frames are generated by the codec’s DTX algorithm according to changes in the
background noise energy and to some specific rules depending
on the codec’s implementation [3].
Human speech has been traditionally modeled as a sequence
of alternate talk and silence periods whose duration is exponentially distributed in the so-called ON-OFF model [1], [2]. The traditional ON-OFF model does not consider the effect of the SID
frames in the traffic pattern generated by voice sources. This
have been commonly accepted since the effect of short SID
frames in the codec bit rate is reduced (typically less than 2%)
[3].
consecutive codec
However, in VoIP, a sequence of
frames is sent in a single IP packet. The packet overhead significantly increases the mean bit rate, specially for low values
. Thus, the presence of SID frames can severely affect
of
the traditional ON-OFF traffic mean bit rate estimation [4], [5].
We propose a new model for the frame generation pattern of
one voice source in which the SID frames are included. An analytical expression for the mean bit rate at the input of the IP
network is deduced for this model accounting the packetiza. Experimental validation will let us compare
tion factor
the mean bit rate obtained from both models with trace measurements for the G.729B, G.723.1, and AMR codecs.
Manuscript received February 18, 2004. The associate editor coordinating the
review of this letter and approving it for publication was Dr. P. Cotae. This work
was supported in part by the Spanish Science and Technology Ministry under
the code TIC2003-04784-C02-02.
The authors are with the Department of Ingenier a de Sistemas y Automaetica,
University of Sevilla, Seville 41092, Spain (e-mail: [email protected];
[email protected]; [email protected]).
Digital Object Identifier 10.1109/LCOMM.2004.835318
II. THE ON-SID MODEL
The new model (from now on named ON-SID) is based on the
following frame generation pattern: during the ON periods, the
ACT frames are generated with deterministic interarrival time
(as in the ON-OFF model), but during the voice inactivity periods,
SID frames are generated randomly.
For mathematical tractability, we assume that, in the discrete
, the codecs continuously generate frames
time space
that can be either type: ACT, SID, or NoTXN. The latter corresponds to a zero-length frame used to model the moments when
no frames are generated.
as the number of consecWe define the random variable
utive NoTXN frames between the generation of two consecutive SID frames during a voice inactivity period. Its probability
can be experimentally
density function
determined for different background noise environments. Thus,
the interarrival time of the SID frames during inactive periods
.
is
, ,
This new model is described by the parameters ,
from the ON–OFF model [1], and the additional parameter .
III. MEAN BIT RATE OF THE ON-SID MODEL
The overall mean rate can be decomposed into the sum of the
contribution of the packets sent during the active voice periods
and voice inactivity periods
separately. Thus
(1)
where is the conversation activity factor, which is also the
probability of being in an ON period.
A. Rate During the ON Periods
We assume that during the ON periods the packets payload is
ACT frames and packets are generated every
. This
is the formula traditionally considered in the ON–OFF model [4]
(2)
where and
are the packet overhead and the size of the
ACT frames, respectively.
B. Rate During the Voice Inactivity Periods
We can add separately the contribution of the packet overhead
and the SID frames carried into the packet. Thus the rate during
inactive periods is
(3)
In order to calculate the terms
and
, let us first
define as the time fraction during the SID periods, where the
. Then
interarrival time of the SID frames is
1089-7798/04$20.00 © 2004 IEEE
ESTEPA et al.: A NEW APPROACH FOR VoIP TRAFFIC CHARACTERIZATION
Fig. 1.
SID frames generation process for different values of N
645
and .
The contribution of the SID frames does not depend on
but just on the frame generation rate. Using the previous definition of we can stand
Fig. 2. Experimental probability density function P .
where
,
and
are the contributions of the overhead,
the voice frames and the SID frames respectively, to the overall
mean bit rate.
IV. VALIDATION
A. Experimental Measurements
(4)
is the SID frame size.
where
The contribution of the packet overhead can be derived inductively1 from Fig. 1 as
(5)
C. Overall Mean Bit Rate
Replacing (3) and (2) in (1) and grouping we have
Following a methodology similar to [6], both ends of a set of
ten conversations of 15 minutes in length were recorded from an
dB .
ISDN line in a low-noise office environment
The raw sample audio files were encoded with the G.729B,
G723.1, and GSM AMR codecs. The output of the codecs was
processed to obtain the sequences of generated frame types.
These sequences (ftype files) were processed to determine the
as well as other model parameexperimental distribution of
, ). Fig. 2 shows the
distribution found for the
ters ( ,
codecs2 under study. The percentage of SID frames obtained for
the codecs G.729B, GSM AMR, and G.723.1 was 7.69%, 7.58%
and 3.59%, respectively. The activity rate was 0.4559, 0.4717
and 0.4697, respectively. The mean duration of the ON periods
was 336, 1026 and 1490 ms, while the OFF periods
mean duration was 420, 1171, and 1722 ms, respectively. To
measure the mean bit rate at the network edge, the packetization process over the ftype audio files was repeated for different
, accounting for a header size which included the
values of
IP, UDP, and RTP protocols.
B. Results
(6)
1The method followed was to develop the rate expressions for different values
and i.
of N
In Table I, the measured mean bit rate is compared to the analytical expression given by both, the ON-OFF and the ON-SID
2According to the respective DTX algorithm specifications, the G.729B codec
should keep a hangover period of 2T after generating a SID frame. The AMR
codec generates a SID frame every 8T and at the third T out of every SID period.
All codecs generate a SID frame at the beginning of a voice inactivity period.
646
IEEE COMMUNICATIONS LETTERS, VOL. 8, NO. 10, OCTOBER 2004
TABLE I
OVERALL MEAN BIT RATE TRACE AND MODELS (kb/s)
V. CONCLUSIONS AND FUTURE WORK
Fig. 3. Models’ error compared to traces.
models for several values3 of
. The first column shows the
measured mean bit rate and the second and third corresponds
with the rate obtained using the ON-SID model proposed and
the ON-OFF model, respectively. The error due to the underestimation of SID frames in ON-OFF model ranges from the 20% for
G.729B codec, to 35% for the GSM AMR4 codec. However, the
difference between the measured rate and the ON-SID model
mean rate is always less than a 3%. This improves the traditional ON-OFF model mean bit rate estimation up to a 32%. The
G.723.1 mean rate estimation error remains small (around 5%
in ON-OFF and 2% in ON-SID) in both models due to the longer
ms , and the small number
frame generation period
of SID frames generated. Fig. 3 summarizes the errors obtained
from both models.
3We have selected a range from 1 to 10 for all codecs, which we find useful
for comparison purposes. However, in a QoS environment high values of N
forces a packetization delay which would result unacceptable specially for the
G.723.1 codec.
4The lower rate mode (MR4.75) was selected for AMR codec because it offers
=L
is the smallest.
the worst rate estimation as the ratio L
The traditionally accepted ON-OFF model ignores the SID
frames, which causes significant errors in the estimated bandwidth requirements of a voice source in an IP network. The
proposed ON-SID model allows to use an analytical expression
for the mean bit rate at the input of the IP network. The experimental validation of this new model shows an improvement of
the mean bit rate estimation obtained with the ON-OFF model
up to a 32%. This new model allows dimensioning with more
precision the network resources in VoIP environments.
with well
Further work is in progress to approximate
known distributions, as well as having into account different
background noise types and levels.
REFERENCES
[1] R. Guerin, H. Ahmadi, and M. Naghshineh, “Equivalent capacity and
its application to bandwidth allocation in high-speed networks,” IEEE J.
Select. Areas Commun., vol. 9, pp. 968–981, Sept. 1991.
[2] K. Sriram and W. Whitt, “Characterizing superposition arrival processes in packet multiplexers for voice and data,” IEEE J. Select. Areas
Commun., vol. SAC-4, pp. 833–846, Sept. 1986.
[3] A. Estepa, R. Estepa, and J. Vozmediano, “Paquetization and silence
influence on VoIP traffic profile,” in Lecture Notes on Computer Science 2899—Multimedia Interactive Protocols and Systems MIPS 2003,
Naples, Italy, Nov. 2003.
[4] B. Goode, “Voice over Internet protocol (VoIP),” Proc. IEEE, vol. 90,
pp. 1495–1517, Sept. 2002.
[5] H. P. Sze, S. C. Liew, J. Y. B. Lee, and D. C. S. Yip, “A multiplexing
scheme for H.323 voice-over-IP applications,” IEEE J. Select. Areas
Commun., vol. 20, pp. 1360–1368, Sept. 2002.
[6] W. Jiang and H. Schulzrinne, “Analysis of On-Off patterns in VoIP and
their effect on voice traffic aggregation,” in Proc. 9th IEEE Int. Conf.
on Computer Communication Networks, Las Vegas, NV, Oct. 2000, pp.
82–87.