5i'
Supplement No. 19
INFORMATION ON SOME
LOUDNESS LOSS RELATED RATINGS
(Melbourne, 1988)
(quoted in Recommendations P.79 and G.III)
Introduction
It is important to determine the electroacoustic performance
of telephone sets in terms of a standard which is universally
recognized. Recommendation P.79 gives the algorithm as agreed by
the CCITT for calculation of loudness ratings (LRs) of telephone
sets. In order to avoid confusion, this algorithm should not be
changed during the 19891992 Study Period. However, it is also
clear from several independent investigations that
Recommendation P.79 represents only with limited accuracy the
speech and hearing characteristics of "ordinary people". This Sup
plement, gives the reader of the PSeries Recommendations a possi
bility to study the background to the problems and provides infor
mation on some other loudness rating systems which have been used.
In particular, SS 1,7 and 8 give examples of algorithms found
useful by some Administrations for their own national planning.
To avoid confusion when dealing with loudness ratings, the
reader should also consider the information given in the Prelim
inary Notes
 to this Volume.
The results of (CCITT) LRs and R25Es calculated by the Chinese
algorithm described in S 3 have been found to be in good agreement
with the subjectively determined values obtained by the CCITT
Laboratory in the past. This algorithm will be used by the CCITT
Laboratory for the objective determination of R25Es for Administra
tions and other organizations.
1 The IEEE algorithm for calculating "objective loudness rat
ings" (Contribution from BNR, Canada)
Abstract
An algorithm for calculating loudness ratings is described.
The agorithm is based on objective measurements and computations
performed in such a manner that the numerical results obtained
reflect the subjective attribute of loudness , but it employs cer
tain simplifying assumptions, to combine simplicity and reasonably
close agreement between objectively determined responses and sub
jective responses.
1.1 Introduction
The algorithm described below is based on a method [1] which
has been in widespread use in North America for several years. The
method has proved very adequate for use both in the planning of
telephone networks and the characterization of individual com
ponents.
The method described may be used for determining the loudness
rating of partial or complete connections. For complete connec
tions, comprising overall or sidetone transmission paths, the pro
cedure involves measurement of acoustic input and output pressures.
For partial telephone connections comprising transmitting, receiv
ing or electrical connection paths, the procedures involve measure
ment of acoustic pressure and electrical voltages. A particular
advantage of this method for planning purposes, is that the sum of
the loudness losses determined for individual parts of a connection
closely approximates the loudness of the overall connection.
1.2 Definitions
1.2.1 loudness rating
The amount of frequencyindependent gain that must be inserted
into a system under test so that speech sounds from the system
under test and a reference system are equal in loudness
(see S 1.3.2).
1.2.2 reference system
A system that provides 0 dB acoustic gain between a mouth
reference point at 25 mm in front of a talker's lips and an ear
reference point at the entrance to the ear canal of a listener,
when the listener is using an earphone. This system is assigned a
loudness rating of 0 dB. The frequency characteristic of the system
must be flat over the range 30033000 Hz and show infinite
attenuation outside of this range.
1.2.3 objective loudness rating (OLR)
The rating of a connection or its components when measured
according to the methodology described within S 1.
1.2.4 overall objective loudness rating (OOLR)
OOLR = 20 log10fISfIM
_______
(11)
where
SM is the sound pressure at the mouth reference point
(in pascals)
SE is the pressure at the ear reference point
(in pascals).
1.2.5 transmitting objective loudness rating (TOLR)
TOLR = 20 log10fISfIM
_______
(12)
where
SM is the sound pressure at the mouth reference point
(in pascals)
VT is the output voltage of the transmitting component
(in millivolts).
1.2.6 receiving objective loudness rating (ROLR)
ROLR = 20 log10(12VfIW
________
(13)
where
VW is the opencircuit voltage of the electric source
(in millivolts)
SE is the sound pressure at the ear reference point
(in pasclas).
1.2.7 electrical objective loudness rating (EOLR)
For an electrical network,
EOLR = 20 log10(12VfIW
________
(14)
where
VW is the opencircuit voltage of the electric source
(in millivolts)
VT is the output voltage of the network
(in millivolts).
1.2.8 Loudness equation
Loudness voltages (in millivolts) and pressures (in pascals)
are determined in accordance with Equation (15).
where
xj is the signal level of SE, SM, VW  or VT  (in dBPa
or dBmV) at frequency fj
X the value SE, SM, VW  or VT  in accordance with the
value used for Xj
fj are specific frequencies of the N  frequencies
selected for analysis.
Loudness voltages and pressures are expressed in decibellike
form using Equation (16).
S `
E, S `
M,
V `
W  r V `
T= 20 log1\d0X
(16)
1.3 Practical considerations
1.3.1 Voltage and pressure levels
Voltage and pressure levels (VJ, VW, SM  and SE) as used in
the definitions above may be measured using exactly the same pro
cedure as used in measuring the corresponding levels (i.e. VJ, EJ,
PM, PE) in Recommendation P.64.
1.3.2 Analysis bandwith
The loudness equation given above in Equation 15 is broadly
applicable to any arbitrary bandwidth. However, for most transmis
sion planning purposes the bandwidth generally selected is
3003300 Hz. This is because the use of partial connection ratings
as engineering tools implicitly requires that for any given connec
tion, the sum of the partial ratings (for example, transmitting
plus receiving) should approximately equal the overall rating.
Thus the bandwidth used to obtain these ratings should approximate
the
bandwidth of the most restrictive element(s) in order to avoid
cumulating bandwidth penalties when summing partial ratings. The
specific limits of 300 Hz and 3300 Hz were selected largely on the
basis of bandwidth capabilities of broadband carrier systems with
a 4 kHz channel spacing. In some cases, for example evaluation of a
telephone sidetone path, a wider analysis band (e.g. 1005000 Hz)
may permit better estimation of the loudness loss. The method
described above may still be used in such cases.
It should be noted that if an actual reference system is con
structed for subjective comparison purposes, the system response
at 300 and 3300 Hz shall be down 3 _ 1 dB relative to the midband
response. The gain of the system shall be adjusted to compensate
for the finite slope of the filter skirts (i.e. in comparison to
the infinite slope inherent in the definition of S 1.2.2) and devi
ation from flatness of the passband. The amount of this adjustment
can be determined by first calculating the OLR (S 1.2.3) over a
frequency range that includes at least the 50 dB points of the
real response, and next calculating the OLR of the ideal response
over the same frequency range. The difference between the OLRs is
the required gain adjustment.
1.3.3 Number of frequency points
As a practical matter, measurement frequencies from which a
loudness computation is made may be evenly spaced on either a
linear frequency scale (1) or logarithmic frequency scale (2).
For (1), no fewer than 31 frequencies should be used. For (2), no
fewer than 12 frequencies should be used, but there is no signifi
cant improvement in accuracy if more than 20 frequencies are used.
1.3.4 Conversion factors between IEEE and Rec. P.79 loud
ness ratings
The following empirical conversion factors have been found
useful among North American Administrations for converting between
loudness ratings derived according to the IEEE method described
above and loudness ratings derived according to Rec. P.79,
for 500type (or equivalent) telephones using the Ghandset.
Send : SLR (P.79) = TOLR (IEEE) + 56 dB
Receive : RLR (P.79) = ROLR (IEEE)  50 dB
Overall : OLR (P.79) = OOLR (IEEE) + 6 dB
Sidetone : STMR (P.79) = SOLR + 8 dB
For send, receive and overall, these relationships give agree
ment between the different ratings with a tolerance of about _ 
dB; for sidetone the tolerance is about _  dB.
1.4 Conclusions
An alternative algorithm for calculating loudness ratings has
been described. This algorithm has been in widespread use in North
America for several years and has been found very satisfactory both
for transmission planning purposes and characterization of indivi
dual network components. One of the main advantages is its relative
simplicity.
2 Algorithms for calculation of loudness ratings (Contribu
tion from the Australian Administration)
2.1 Introduction
There is growing evidence (see S 4) that the algorithm defined
in Recommendation P.79 for the calculation of loudness ratings
(LRs) is nonoptimum, giving undue weight to the lower frequencies.
This prompted a study within Telecom Australia to seek a better
algorithm. The approach involved determining the loudness rating of
many telephone paths and then optimizing the parameters in the
algorithm for best agreement between subjective and computed
values.
An insert earphone type headset and a (pseudo) loudspeaking
telephone were also included in the programme of work. In view of
the physical differences from handset telephones, particularly on
receiving, it was expected that different algorithms would be
required.
2.2 Basic algorithm
A method for the computation of loudness ratings (LRs) is
derived in Recommendation P.79 and results in a formula of the
form:
LR = 10/m log ~" 10
(Si  Wo\di)m /10
where:
m is the loudness growth coefficient
Si is the overall acousticacoustic sensitivity in dB of
the unknown telephone path (completed by the IRS, if necessary)
Wo\di is the (negative) weighting function of frequency,
in dB
i is the 1/3 octave (strictly 1/10 decade) frequency
step number.
In the derivation, Si  efers to real mouth and real ear sen
sitivities, but if the correction factors for using artificial
equivalents are included in the definition of Wo\di, then Sican be
redefined to be the measured sensitivity with artificial mouth and
ear. Wo\dialso includes other components such as the spectral den
sity of human speech, the frequency sensitivity of the human ear,
and normalization so that computed loudness rating of the IRS + IRS
connection is 0 dB.
2.3 Determination of parameters
The weighting function in Rec. P.79 was derived by determining
each of the above components and then combining them. In the
present work, the weighting function was derived directly. This
direct approach leads naturally to consideration of nonhandset
telephones, such as headsets which may have insert type receivers
and handsfree loudspeaking telephones. In the latter case the
weighting function must also take into account the diffraction of
sound around the human head, the effect of listening with two ears
instead of one, and the use of an open rather than occluded ear.
The method involved the insertion of a series of five lowpass
and five highpass filters into various telephone connections,
measuring the LR of each subjectively, and then optimizing the
parameters to give best agreement (in a leastsquares error sense)
with the computed values. The overall acousticacoustic sensitivi
ties of each connection were first measured using an artificial
mouth (B&K type 4219) and an artificial ear (IEC type 318 by B&K)
for handsets, and IEC type 711 (B&K type 4157) for the insert
receiver.
2.4 Telephone paths
The telephone paths involved several different telephone types
which are in use in Australia, and are listed in the first column
of Tables 21 to 25. If necessary, the connection was completed
using the appropriate IRS end. Since the 802 type was fitted with a
carbon transmitter, the send and receive sensitivities were meas
ured using a speech weighted random noise signal. The pseudo
loudspeaking telephone (LST) paths were similarly measured to
reduce the effect of standing waves in the test room. All other
telephones were measured using sine waves. The equalized IRS con
nections were
obtained by first equalizing to give a reasonably flat overall
sensitivity (measured objectively) and then adding further equaliz
ers to give either a falling response or a rising response
(about 6 dB/octave in both cases). The Featherset headset has an
insert type receiver and a noise cancelling electret microphone
which is held near the side of the mouth by a boom.
The pseudo loudspeaking telephone for send measurements con
sisted of a 1/2 inch condenser microphone plus measuring amplifier
with a sound level meter A  weighting function. The microphone was
mounted on a gooseneck extension piece which held the microphone
just above the surface of the table. For receive measurements, the
equipment consisted of a power amplifier and a small loudspeaker
lying on the long side of its enclosure, with the axis horizontal
and pointing to the listener. A real loudspeaking telephone was not
used to avoid complications associated with voice switching.
2.5 Form of weighting function
Various parametric forms of the weighting function were tried,
but a parabola gave almost as good a result as more complicated
forms, including higher order polynomials. A parabola can be
described in terms of the coordinates of its minimum (in this case)
and a coefficient controlling its breadth, by a procedure known as
"completing the square", viz.
Wo\di= A + C (i  B )2
In order to compare the weighting functions derived using dif
ferent values of loudness growth coefficients m , it is more mean
ingful to consider the product Wo. This quantity may be interpreted
as being proportional to the negative of the decibel equivalent of
the weighting function which multiplies the band loudness (as dis
tinct from band power) in each of the 1/3 octave (1/10 decade) fre
quency bands.
The value of i  ranges from 0 to 17 for frequencies
from 100 Hz to 5012 Hz.
2.6 Optimum parameters
The optimum values of m , Am , B , C  nd Cm
 are given in Table 21 for the various telephone paths con
sidered. Also included in the table are the subjectiveobjetive
error standard deviations (means = 0 dB) and the computed LR of
the IRS + IRS connection (which ideally should be 0 dB).
The standard deviations range from 0.1 to 0.4 dB, showing good
fit of the model whem optimized for the particular path. Examina
tion of the distribution of the individual errors showed no trends
with filter cutoff frequencies. The values
of B , Cm  and m  are fairly consistent with different
paths, the biggest differences in B  occurring with the different
equalizer responses used with the IRS. Note that although the
Featherset and loudspeaking telephone have quite different receive
characteristics, B , Cm  and m  are within the range of those
for conventional handset telephones. Am  is significantly dif
ferent, however, and this is reflected in the error of the computed
LR of the IRS + IRS connection. This suggests that a single fre
quency weighting shape may be satisfactory for all telephones,
whether handset, headset or handsfree, provided that a constant
correction factor is applied in certain cases.
Note that the value A  (and hence Am ) for the loudspeaking
telephone on receive is now believed to be in error. This is dis
cussed later in S 2.11.
H.T. [T1.19]
TABLE 21
Optimum parameters for each path and error statistics
____________________________________________________________________________________________________
Parameters Errors
Path m A Am B C Cm Std. dev. IRS
____________________________________________________________________________________________________
802 send 0.255 39.67 10.12 9.64 1.225 0.312 0.2 0.6
802 receive 0.249 42.72 10.63 9.04 0.889 0.221 0.3  0.1
Flipphone send 0.308 34.72 10.69 9.35 0.732 0.226 0.3 1.6
Flipphone receive 0.286 40.69 11.63 8.85 0.513 0.147 0.4  0.5
807 send 0.315 36.38 11.46 9.66 0.648 0.204 0.2 0.3
807 receive 0.263 43.37 11.41 8.95 0.533 0.140 0.1  0.1
Commander T210 send 0.312 33.84 10.56 9.45 0.934 0.291 0.5 0.9
Commander T210 receive 0.279 38.28 10.68 8.72 0.704 0.196 0.4 0.9
Siemens Trans. Cour. send 0.290 35.83 10.39 9.50 1.119 0.325 0.4 0.2
Siemens Trans. Cour. receive 0.337 35.69 12.03 9.33 0.751 0.253 0.3  2.3
Equalized, IRS flat 0.270 42.47 11.47 9.64 0.581 0.157 0.3 0.1
Equalized, IRS falling 0.299 40.21 12.02 10.31 0.398 0.119 0.2 0.6
Equalized, IRS rising 0.300 35.07 10.52 6.66 0.496 0.149 0.3  0.3
Featherset send 0.285 36.48 10.40 9.55 0.684 0.195 0.3 3.2
Featherset receive 0.330 42.63 14.07 9.28 0.525 0.173 0.3  6.9
Pseudo LST send 0.244 40.29 9.83 8.89 0.776 0.189 0.4 3.8
Pseudo LST receive 0.232 27.36 6.35 9.40 0.352 0.082 0.3 23.4
____________________________________________________________________________________________________












































































































































































































Table 21 [T1.19], p.1
2.7 Global optimization
Parameters A , B  and C  are partly dependent on loudness
rating specifics, but m  is a pure psychoacoustic phenomenon.
The average value of m  in Table 21 is 0.2855 (median = 0.29).
The optimization process was therefore repeated with m  held
at 0.2855, with results given in Table 22. The standard deviations
increased only slightly (about 0.1 dB) verifying that a single
value of m  is practicable.
H.T. [T2.19]
TABLE 22
Optimum parameters and error statistics
with m = 0.2855
____________________________________________________________________________________________
Parameters Errors
Path A Am B C Cm Std. dev. IRS
____________________________________________________________________________________________
802 send 39.91 10.25 9.64 1.208 0.345 0.3 0.4
802 receive 37.68 10.76 9.08 0.901 0.257 0.4  0.3
Flipphone send 37.31 10.65 9.29 0.712 0.203 0.3 1.7
Flipphone receive 40.70 11.62 8.85 0.513 0.147 0.4  0.5
807 send 39.76 11.35 9.57 0.603 0.172 0.3 0.5
807 receive 40.17 11.47 9.03 0.569 0.162 0.2  0.3
Commander T210 send 36.55 10.44 9.38 0.919 0.262 0.5 1.1
Commander T210 receive 37.46 10.69 8.74 0.711 0.203 0.4 0.8
Siemens Trans. Cour. send 36.42 10.40 9.49 1.111 0.317 0.4 0.2
Siemens Trans. Cour. receive 41.05 11.72 9.21 0.691 0.197 0.5  1.9
Equalized, IRS flat 40.16 11.47 9.63 0.606 0.173 0.3 0.1
Equalized, IRS falling 42.03 12.00 10.43 0.373 0.107 0.3 0.6
Equalized, IRS rising 36.92 10.54 6.56 0.476 0.136 0.3  0.3
Featherset send 36.42 10.40 9.55 0.685 0.196 0.3 3.1
Featherset receive 47.08 13.44 9.06 0.490 0.140 0.4  6.4
Pseudo LST send 34.42 9.83 9.02 0.817 0.233 0.5 3.4
Pseudo LST receive 18.75 5.35 9.28 0.390 0.111 0.4 23.1
____________________________________________________________________________________________
























































































































































































Table 22 [T2.19], p.2
Next, parameters m , B  and C were optimized globally, but
individual values of A  were permitted, to investigate the feasi
bility of using the same shape for the weighting function, for all
telephone types (handset, headset and handsfree), but permitting a
correction constant if necessary. Optimization gave m = 0.2855, B
= 9.19 and C = 0.7723, with A  and errors as shown in
Table 23. The standard deviations have now increased signifi
cantly, the worst being for the IRS with rising frequency response.
The errors for this path also show a clear trend with filter
cutoff frequency, indicating a lack of fit of the model. Note how
ever that the standard deviations for the headset and handsfree
telephone are still comparable with handset telephones in general.
The value of A  necessary to give a computed LR of 0 dB for the
IRS is 38.45.
Table 24 gives the errors for a new algorithm (denoted D4 for
convenience) based on the above data. The most significant mean
errors are 22.4 dB (but see S 2.11) for the loudspeaking telephone
receive, 6.9 dB for the headset receive, 3.6 dB for loudspeaking
telephone send and 3.0 dB for headset send. There are obvious
reasons why the mean errors on receive would not be zero, but the
main reason for the errors on send are thought to be due to
incorrect pressure distribution as a function of distance of the
artificial mouth (B&K type 4219). Another reason might be due to
the handset mouth cap affecting the pressure of the feedback micro
phone in the artificial mouth, while no significant effect occurred
with the headset and loudspeaking telephone. Errors for handset
telephones are smaller but unfortunately not negligible. These are
thought to be mainly due to limitations of the artificial mouth and
ear, including the effect of earcap leakage which is not modelled
at all, and has to be included in the weighting function.
H.T. [T3.19]
TABLE 23
Optimum A and errors for the case of other parameters
globally optimized
__________________________________________________________________________________________________________________________________
High pass (Hz) Low pass (Hz) Errors
Path A 158 225 380 630 1020 630 780 1260 2040 3120 Std. dev. IRS
__________________________________________________________________________________________________________________________________
802 send 38.03 0.3 0.6 0.2 0.2 0.6 2.2 0.7 0.2 0.4 0.5 0.9 0.4
802 receive 38.62 0.6 0.1 0.3 0.6 0.9 0.0 0.8 0.1 0.5 0.1 0.5 0.2
Flipphone send 36.82 0.0 0.0 0.2 0.0 0.8 0.3 0.2 0.4 0.3 0.5 0.4 1.6
Flipphone receive 38.95 0.3 0.0 0.3 0.9 0.3 0.3 0.5 0.5 0.5 0.7 0.5 0.5
807 send 38.42 0.4 0.2 0.3 0.5 1.9 1.2 1.1 0.6 0.5 0.1 0.9 0.1
807 receive 38.84 0.2 0.2 0.1 0.2 0.2 0.3 0.1 0.5 0.1 0.5 0.3 0.4
Commander T210 send 37.27 0.3 0.2 0.5 0.7 1.0 0.8 0.1 0.3 0.2 0.3 0.6 1.2
Commander T210 receive 37.30 0.0 0.4 0.2 1.3 0.9 1.0 1.4 0.1 0.3 0.1 0.8 1.2
Siemens Trans. Cour. send 38.25 0.1 0.3 0.1 0.4 0.9 1.5 0.4 0.6 0.1 0.0 0.7 0.2
Siemens Trans. Cour. receive 40.53 1.0 0.0 0.1 0.2 0.8 0.3 0.0 0.1 0.2 0.2 0.5 2.1
Equalized, IRS flat 38.89 0.5 0.0 0.9 0.7 1.1 1.0 0.9 1.2 0.5 0.4 0.8 0.4
Equalized, IRS falling 39.37 0.1 0.1 0.1 0.1 2.2 0.1 1.0 0.8 0.6 0.3 0.9 0.9
Equalized, IRS rising 37.39 0.8 1.2 2.1 3.8 5.4 7.1 4.2 1.8 0.0 0.2 3.7 1.1
Featherset send 35.47 0.3 0.3 0.0 0.4 1.4 1.6 0.8 0.6 0.4 1.0 0.9 3.0
Featherset receive 45.31 0.2 0.0 0.2 0.3 1.5 0.1 0.8 0.3 0.2 0.3 0.6 6.9
Pseudo LST send 34.81 0.2 0.3 0.3 0.9 0.9 0.1 0.7 0.2 0.8 0.0 0.6 3.6
Pseudo LST receive 16.07 0.4 0.5 0.6 0.3 0.0 0.4 0.9 0.9 0.4 0.4 0.6 22.3 
__________________________________________________________________________________________________________________________________


















































































































































































































































































































Tableau 23 [T3.19], p.3
H.T. [T4.19]
TABLE 24
Errors for algorithm D4
_____________________________________________________________________________________________________________________________________________
High pass (Hz) Low pass (Hz) Errors
Path 158 225 380 630 1020 630 780 1260 2040 3120 Mean Std. dev.
_____________________________________________________________________________________________________________________________________________
802 send 0.7 1.0 0.6 0.7 1.0 1.8 0.3 0.7 0.8 0.9 0.4 0.9
802 receive 0.4 0.2 0.5 0.7 1.1 0.2 0.6 0.1 0.3 0.1 0.2 0.5
Flipphone send 1.6 1.7 1.9 1.6 2.4 1.4 1.4 1.2 1.9 1.1 1.6 0.4
Flipphone receive 0.8 0.5 0.2 1.4 0.2 0.2 0.0 1.0 1.0 0.2 0.5 0.5
807 send 0.5 0.3 0.4 0.5 2.0 1.1 1.0 0.5 0.4 0.1 0.1 0.9
807 receive 0.2 0.2 0.5 0.6 0.6 0.1 0.5 0.9 0.3 0.1 0.4 0.3
Commander T210 send 1.4 0.9 0.7 0.5 2.1 0.4 1.3 1.5 1.4 1.4 1.2 0.5
Commander T210 receive 1.1 0.8 1.0 0.2 0.2 2.1 2.5 1.2 1.5 1.2 1.1 0.8
Siemens Trans. Cour. send 0.3 0.5 0.3 0.2 1.1 1.3 0.2 0.8 0.3 0.2 0.2 0.7
Siemens Trans. Cour. receive 3.1 2.1 2.1 1.9 1.2 1.8 2.1 2.2 1.9 2.3 2.1 0.5
Equalized, IRS flat 0.0 0.4 0.4 0.2 0.6 1.4 1.4 1.7 1.0 0.0 0.5 0.8
Equalized, IRS falling 1.0 1.0 0.8 0.8 1.3 1.1 1.9 1.7 1.6 0.6 0.9 0.9
Equalized, IRS rising 0.3 0.1 1.1 2.7 4.3 8.2 5.2 2.8 1.1 1.3 1.1 3.7
Featherset send 3.2 3.3 3.0 3.4 4.4 1.4 2.1 2.4 2.6 4.0 3.0 0.9
Featherset receive 7.0 6.9 7.0 7.1 5.4 6.7 7.7 7.1 7.1 6.6 6.9 0.6
Pseudo LST send 3.9 3.9 3.3 2.8 2.7 3.8 4.4 3.4 4.4 3.7 3.6 0.6
Pseudo LST receive 22.8 22.9 23.0 22.1 22.3 22.8 21.5 21.5 22.0 22.8 22.4 0.6
_____________________________________________________________________________________________________________________________________________





























































































































































































































































































Tableau 24 [T4.19], p.4
2.8 Comparison with other algorithms
Table 25 compares algorithm D4 with other algorithms. D2 is
an algorithm based on preliminary work in which m = 0.2976 and the
weighting function is defined by A = 40.50, B = 9.867 and C
= 0.423. P.79 is the current Recommendation while P.XXE is the
draft upon which it is based. Note that draft Rec. P.XXE as pub
lished in [2] is in error. On page 178 it states that the mean
LR\dM\dEis 4.72 dB, but in fact it should be 0.1 dB. Thus 4.6 dB
should be subtracted if the tabulated data are used. Zw is a com
plicated algorithm based on the work of E. Zwicker and published as
ISO Rec. R532B.
The paths are as previously discussed except that the first
item is the set of 14 sidetone responses reported in an earlier
work [3].
The group mean errors for the handset telephones are fairly
small for all algorithms, but the group standard deviation for P.79
is about twice that of the others. Note that the complicated Zw
method does not seem to offer any significant advantage, and still
gives rather large errors for the IRS + IRS + equalizer connec
tions. Naturally D4 gives a reasonably good fit because it was
optimized for these conditions.
The values of Wo  as a function of frequency for
algorithms P.XXE, P.79, D2 and D4 are shown in Figure 21. Note
that P.XXE and D4 are very similar, and that P.79 shows much
smaller (negative) weight at low frequencies.
H.T. [T5.19]
TABLE 25
Comparison of errors for five algorithms
_______________________________________________________________________________________________________________________________________
D2 D4 P.79 P.XXE Zw
Path Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
_______________________________________________________________________________________________________________________________________
Sidetone [3] 1.2 0.5 1.2 0.7 1.9 1.4 0.8 0.6 0.6 0.7
_______________________________________________________________________________________________________________________________________
802 send 0.7 0.9 0.4 0.8 1.4 3.4 1.0 1.2 0.4 1.8
802 receive 0.4 1.5 0.2 0.5 0.1 2.2 0.1 0.5 0.6 0.9
Flipphone send 2.2 1.0 1.6 0.4 1.9 3.0 1.9 0.7 1.1 1.7
Flipphone receive 0.1 1.6 0.5 0.5 0.5 2.6 0.2 0.5 1.3 1.3
807 send 0.6 0.5 0.0 0.9 0.5 3.6 0.3 1.1 0.7 2.1
807 receive 0.2 1.2 0.4 0.3 0.3 2.6 0.0 0.3 1.0 1.4
Commander T210 send 1.6 1.4 1.2 0.5 2.7 2.5 1.6 0.8 0.6 1.7
Commander T210 receive 2.1 1.7 1.2 0.8 1.3 1.8 1.5 0.6 0.7 1.0
Siemens Trans. Cour. send 0.9 1.0 0.2 0.6 0.6 3.1 0.6 1.0 0.0 2.1
Siemens Trans. Cour. receive 1.1 1.0 2.1 0.4 2.1 2.7 1.8 0.8 2.8 1.3
Equalized, IRS flat 0.0 0.4 0.4 0.8 3.2 3.6 0.5 0.9 1.4 2.1
Equalized, IRS falling 0.1 0.5 0.9 0.8 4.7 3.8 0.9 1.2 2.1 2.2
Equalized, IRS rising 1.1 4.7 1.1 3.5 0.3 1.5 1.0 3.4 0.5 2.5
Featherset send 3.3 0.8 3.0 0.8 4.1 3.1 3.3 1.0 2.4 2.0
Featherset receive 6.2 1.1 6.9 0.6 5.0 2.5 6.3 0.8 7.2 1.6
Pseudo LST send 4.4 1.6 3.6 0.6 4.3 1.8 4.0 0.5 3.3 0.9
Pseudo LST receive 23.4 0.7 22.4 0.5 22.0 2.5 22.7 0.6 21.9 1.0
_______________________________________________________________________________________________________________________________________
Handset 0.61 0.95 0.09 1.02 0.09 2.09 0.35 1.06 0.51 1.19
_______________________________________________________________________________________________________________________________________







































































































































































































































































































Table 25 [T5.19], p.
Figure 21, p.
2.9 Validation of new algorithm
Table 26 gives the subjectiveobjective errors for test
results which are not used in deriving D4. Two samples of 802 tele
phone (local designations 82/YZ and 82/IA) were each fitted with
one of four 20E noncarbon transmitters (designated 101, 165, 310
and 313) for send measurements. Only one receive measurement was
made for each telephone. Three lines were used, viz. zero, 1.6 km
and 4.2 km of 0.4 mm cable.
One consistent trend is that the errors become more positive
with increasing line length, and range from 0.6 dB for D2 through
0.9 dB for D4, P.XXE and Zw, to 1.2 dB for P.79. A possible reason
for this trend is the progressive high frequency loss which occurs
with line length, and inadequacies in the loudness models to cope
with this. This is also consistent with the errors associated with
the equalized IRS results in Table 25, where the falling response
gives the most positive error and the rising response the most
negative error of the set of three.
H.T. [T6.19]
TABLE 26
Errors for 802 telephones (20E noncarbon transmitters)
plus lines for five algorithms
Unable to convert table Table 26 [T6.19], p.
2.10 Attempts to reduce errors
In order to explore whether another weighting function would
simultaneously give small errors for the 802 telephone only, with
both filters and lines, the 802 + lines data described above was
combined with the 802 + filter data described earlier. A new
weighting function was then optimized, with A  constrained to
give 0 dB error for the LR of the IRS. It was found however that
the optimum parameters were not greatly different from those in D4
and that the range of errors with line length was only reduced
by 0.1 dB to 0.8 dB.
It was thought possible that forcing a polynomial fit to the
weighting function may be partly responsible for this poor agree
ment, so a piecewise linear weighting function was tried, with
break frequencies at i = 4, 7, 10 and 13 (f = 250, 500, 1000 and
2000 Hz respectively). It was found that the range of errors with
line length was unchanged at 0.8 dB. Thus the weighting function
shape does not seem to be at fault.
A simplification inherent in all algorithms from P.XXE to D4
is that the weighting function does not cause any frequency band to
be masked, whereas it is assumed in the derivation of these models
that it is only the band loudness above threshold which contributes
to loudness. The basic formula was therefore changed to include a
threshold rather than a weighting function. Summation is only over
those bands which are above threshold. A disadvantage of this algo
rithm is that it is now not possible to make loudness rating the
subject of the formula, and an iterative approach is necessary. A
parabolic threshold function was assumed, and it was found that the
range of errors with line length was only reduced a further 0.1 dB
to 0.7. The marginal improvement does not justify the extra compli
cation of this method.
Finally, the effect of frequency masking was included by
investigating whether a better way of using Zwicker's loudness
algorithm could be found. In addition to the sensitivity of hearing
which is inherent in Zwicker's algorithm, a LR algorithm must also
include the spectral density and level of the speech signal, the
ear cap leakage loss and the junction loss to give the same loud
ness through the IRS + IRS path as the NOSFER system with 25 dB in
its junction. These may be combined to form an auxiliary function
analogous to an input signal to the telephone path, where the out
put is fed to Zwicker's loudness algorithm. Assuming a parabolic
shape to this auxiliary function, it was found that the range of
mean errors with line length was 0.8 dB and thus comparable to that
of previous algorithms, such as D4.
A possible reason why none of the methods was successful in
reducing errors to a low and random value (i.e. no trend with line
length) may be that the subjects changed their bases of listening
to the speech from one filter condition to the next. They may not
listen to the signal as a whole, but base their comparison on a
smaller band or bands where the main energy lies (formants). The
location of the band or bands could vary depending on the cutoff
frequencies of the filters. Zwicker based his method on subjective
data gathered on nonspeech signals, but it is known that people
listen to speech in a different way to other sounds, and this may
affect the judgement of loudness. Other possible sources of
discrepancy are possible, including the effect of changes in the
voiceear team membership during the course of the investigation.
2.11 Postscript on the correction factor for loudspeaking
telephone receive
The receive correction factor found initially for the
loudspeaker and amplifier combination was about 22.4 dB, but in
subsequent work a drift in this value was observed. Whether this
was due to setup errors, hardware faults or to changing bases of
rating loudness by the voiceear team has not been resolved. Subse
quent tests repeating those reported above and others have yielded
a correction factor of about 14.0 dB, and this is now believed to
be more correct. (The D4 loudness algorithm continued to give good
consistency in the repeat tests, with a standard deviation
of 0.7 dB over the range of filters.)
2.12 Conclusion
A revised algorithm has been found which is remarkably similar
to the draft Recommendation upon which the present
Recommendation P.79 was based. Using either of these methods gives
about half the standard deviation of the difference between subjec
tive and objective measurements which would be obtained with
Rec. P.79. A general accuracy of about _  dB can be expected,
which is about the order of accuracy of subjective tests, but with
better repeatability and lower cost.
Although it was expected that a different weighting function
would be required for headsets and loudspeaking telephones, in fact
it was found that a constant correction for each path type proved
to be all that was necessay for practical purposes. In particular,
the following corrections should be added to the calculated LRs:
Headset
Send: 3.0 dB
Receive: 6.9 dB ( insert receiver only)
Loudspeaking telephone
Send:  3.6 dB
Receive: 14.0 dB
As far as the revision of Rec. P.79 is concerned, two courses
of action seem possible. Preferably,
i) pool all the data available worldwide and derive
a global average using the principles described above,
or alternatively
ii) return to the algorithm weights of
draft Rec. P.XXE.
3 Uniform algorithms for the calculation of R25 equivalents
and loudness ratings (from the Ministry of Post and Telecommunica
tions of the People's Republic of China)
3.1 Introduction
The subjective test team of the CCITT Laboratory has been
changed since 1985. From the periodic stability check reports of
the CCITT Laboratory, it can be ascertained that the recent subjec
tively determined value x2(see Recommendation P.78) is about 18 dB
which is close to the value determined at other laboratories, and
different from the previously determined value of 12 dB. In addi
tion, the SR25E and RR25E values of telephone systems determined
recently by the CCITT Laboratory are several decibels lower than
the results previously obtained, and close to those measured by
other laboratories.
In this connection, it is possible to use a uniform algorithm,
similar to the simple algorithm in Recommendation P.79, for the
calculation of R25 equivalents and loudness ratings, with values
for the slope parameter m and the G functions different from those
given in Recommendation P.79.
In order to obtain a suitable algorithm and appropriate
parameters, four different algorithms were used in order to calcu
late the values of R25E and LR, and the results were compared.
Three of them are similar to that used for the calculation of loud
ness ratings described in Recommendation P.79, except that dif
ferent values of the slope parameter m and the G functions are
used.
These values:
 are taken from draft Recommendation P.XXE [2];
 correspond to the Chinese test team;
 correspond to the old test team of the CCITT
Laboratory, but with LEcorrected in the NOSFER receiving system.
The fourth algorithm used was the ISO532B (Zwicker) algo
rithm.
3.2 Comparison of various algorithms
The four algorithms used here are labelled as the P.XXE, the
Chinese, the P.79 Cor. and the ISO532B algorithms.
3.2.1 SFC of the reference system
3.2.1.1 The sensitivity/frequency characteristic (SFC) data of
the sending system and the receiving system (without leakage) of
the NOSFER are taken from Recommendation P.42 (Red Book). The cou
pling loss at the receiving part of the NOSFER is included in the
receiving SFC in the calculation.
Several years ago the Chinese Administration pointed out that
the SFC data of the NOSFER receiving system measured by the IEC 318
artificial ear with the flat plate differed considerably from those
measured with the operator's ear, and measured the values of
LEcorresponding to the earphone type DR701 used by the Chinese
test crew in the receiving system of the NOSFER.
This point of view has been verified by many Administrations
and has been generally accepted by CCITT Study Group XII. The
values of LEused here are those corresponding to the CCITT Labora
tory test team, as given by the French Administration (Contribution
COM XII111, 19851988) (see Table 31).
3.2.1.2 The SFC data of IRS are taken from Recommendation P.48
and the SFC values of the receiving system are corrected using the
LEgiven in Recommendation P.79.
3.2.2 Slope parameter m and Gfunctions
Methods for estimating m and G are described in Contribution
COM XII3 and COM XII10 (19811984).
3.2.2.1 P.79 Cor. algorithm (m = 0.175)
The values of the slope parameter m and the G functions in
Recommendation P.79 are derived from the results of the filter
loudness loss test of the old CCITT Laboratory test team; the leak
age between the ear of the operator and the earphone of NOSFER is
not included. The values of the G functions given in
Recommendation P.79 must therefore be corrected. The results of the
G functions with correction of LEare listed in Table 32.
H.T. [T7.19]
TABLE 31
Acoustic coupling loss LvE
used in calculation
_________________________________
Frequency L NOSFER L P.79
_________________________________
100 0.9 20.0
125 0.2 16.5
160 0.6 12.5
200 1.6 8.4
250 2.9 4.9
315 4.2 1.0
400 5.3 0.7
500 5.4 2.2
630 4.9 2.6
800 4.6 3.2
1000 4.5 2.3
1250 3.9 1.2
1600 4.6 0.1
2000 3.3 3.6
2500 3.2 7.4
3150 3.3 6.7
4000 3.7 8.8
5000 2.9 10.0
6300 0.8 12.5
8000 0.8 15.0
_________________________________




























































































Tableau 31 [T7.19], p.8
H.T. [T8.19]
TABLE 32
10 logv1v0 G of various algorithms
____________________________________________
Frequency P.79 Cor. P.XXE Chinese
____________________________________________
100 31.86 35.90 30.67
125 28.58 34.11 30.63
160 27.14 32.94 30.68
200 28.13 31.50 30.81
250 28.48 30.96 31.02
315 31.22 31.21 31.35
400 30.10 31.15 31.79
500 33.02 30.97 32.33
630 33.46 32.13 33.00
800 34.34 33.05 33.83
1000 35.51 34.50 34.74
1250 37.97 35.91 35.78
1600 38.60 37.14 37.10
2000 41.22 38.50 38.46
2500 41.66 39.66 39.96
3150 45.77 41.11 41.70
4000 43.54 43.45 43.68
5000 47.03 45.37 45.71
6300 48.03 48.01
8000 46.32 50.60
____________________________________________



















































































































Tableau 32 [T8.19], p.9
3.2.2.2 P.XXE algorithm (m = 0.225)
The values for the slope parameter m and the G functions are
taken from Table 1, page 185 of COM XII1 [2]. Also see Table 32
of this Supplement.
3.2.2.3 Chinese algorithm (m = 0.2)
Results of smoothed G functions are used [see Contribution
COM XII233 (19811984)]. Values are also given in Table 32.
The coupling loss of the NOSFER earphone is not included in
the estimation of the G functions but this has little effect on
the smoothed result of the G functions.
3.2.3 Wweights for the calculation of R25E
Methods for deriving W weights are described in Contributions
COM XII3 and COM XII10 (19811984).
3.2.3.1 P.79 Cor. algorithm
Weights are derived from the SFC data of NOSFER described in
S 3.2.1.1 and the data for m and G functions given in S 3.2.2.1.
3.2.3.2 P.XXE algorithm
W weights are derived from the SFC data of NOSFER described
in S 3.2.1.1 and the data for m and G functions given in
S 3.2.2.2. In the absence of a complete set of data for the G
functions at high and low frequencies, a number of arbitrary
values have had to be chosen in this contribution.
3.2.3.3 Chinese algorithm
W weights are derived from the SFC data of NOSFER described
in S 3.2.1.1 and the data for m and the G functions given in
S 3.2.2.3.
The derived W weights of the three algorithms discussed above
for the calculation of R25E are listed in Table 33.
3.2.4 Wweights for the calculation of LR
The methods for the derivation of W weights for the three
algorithms are similar to those described in S 3.2.3, except that
the SFC data of IRS (with the LEof P.79) are used instead of the
SFC data of NOSFER.
The derived W weights of the three algorithms discussed above
for the calculation of LR are listed in Table 34.
3.2.5 Source of data for the SFC of telephone systems and
the subjectively determined values of R25E and LR
In making comparisons between the subjectively determined
results and the calculated results, use can only be made of the
data relating to telephone sets with subjectively determined values
established by the new CCITT test team and the corresponding SFC
values.
3.2.5.1 For SR25E
There are only six sets of sending SFC data provided by three
linear telephone sets under 0/L line conditions (i.e. with or
without lines). These data are taken from CCITT Laboratory
Technical Report 808 (Temporary Document 84, Working Party XII/1,
April 1987); the other set of subjectively determined values is
taken from CCITT Laboratory Technical Report 797 (Temporary Docu
ment 78, Working Party XII/1, April 1987).
3.2.5.2 For RR25E
The subjectively determined values of RR25E of some telephone
systems are taken from CCITT Laboratory Technical Report 797 and
the corresponding SFC data were given by the Head of the CCITT
Laboratory in October 1986.
H.T. [T9.19]
TABLE 33
Wweights for the calculation of R25E
___________________________________________
P.79 Cor. P.XXE Chinese
___________________________________________










Frequency W W W W W W
______________________________________________________________________
100 109.6 116.6 106.9 113.9 92.2 99.2
125 82.4 89.8 92.1 99.5 84.5 91.9
160 67.2 75.2 81.2 89.2 78.5 86.5
200 66.6 76.3 69.5 79.2 73.4 83.1
250 60.6 72.7 60.4 72.5 67.2 79.3
315 67.6 82.2 54.3 68.9 61.0 75.6
400 53.8 70.2 47.9 64.3 56.6 73.0
500 63.5 80.8 41.3 58.6 52.9 70.2
630 60.5 75.5 42.4 57.4 51.5 66.5
800 60.8 74.3 42.9 56.4 51.6 65.1
1000 62.6 75.1 45.5 58.0 51.9 64.4
1250 70.5 81.9 47.2 58.6 51.9 63.3
1600 67.3 79.9 47.1 59.7 52.3 64.9
2000 78.4 90.1 50.3 62.0 55.8 67.5
2500 74.8 86.6 50.6 62.4 57.9 69.7
3150 93.2 102.1 53.5 62.4 62.3 71.2
4000 76.7 84.6 61.3 69.2 69.2 77.1
5000 88.8 103.3 63.2 77.7 72.2 86.7
6300 84.9 110.0 92.2 117.3 74.9 100.0
8000 80.4 99.1 





















102.7 





















121.4 





















93.8 





















112.5
______________________________________________________________________
m = 0.175 m = 0.225 m = 0.2
















































Tableau 33 [T9.19], p.10
3.2.5.3 For SLR and RLR
The subjectively determined values are taken from CCITT
Laboratory Technical Report 771 (Temporary Document 42, Working
Party XII/1, May 1986) and the corresponding SFC data [the sending
data measured at LRGP (loudness rating guardring position)] were
also provided by the CCITT Laboratory.
3.2.6 Method of calculation
3.2.6.1 For the P.79 Cor., P.XXE and the Chinese algorithms,
the equations used for the calculation of SR25E, RR25E, SLR and RLR
are as follows:
H.T. [T10.19]
TABLE 34
Wweights for the calculation of LR
___________________________________________
P.79 Cor. P.XXE Chinese
___________________________________________










___________________________________________________________
Frequency W W W W W W
___________________________________________________________
100 149.3 147.6 150.0 150.0 135.7 134.0
125 111.4 112.2 150.0 150.0 117.2 118.0
160 85.3 87.6 150.0 150.0 100.3 102.5
200 74.4 82.5 82.5 90.6 85.0 93.1
250 61.5 73.6 66.7 78.8 71.9 84.0
315 62.2 79.2 54.0 71.0 59.3 76.3
400 46.0 65.0 45.1 64.1 52.4 71.4
500 54.6 75.0 37.4 57.8 47.6 68.0
630 49.4 70.0 36.1 56.7 44.2 64.8
800 48.2 68.6 35.5 55.9 42.6 63.0
1000 50.6 69.2 38.8 57.4 43.6 62.2
1250 59.0 75.0 41.0 57.0 44.2 60.2
1600 57.3 71.0 42.3 56.0 45.8 59.5
2000 71.5 80.7 48.6 57.8 52.5 61.7
2500 71.8 75.7 52.7 56.6 58.8 62.7
3150 88.5 92.6 54.0 58.1 61.5 65.6
4000 116.7 113.5 106.5 103.3 112.7 109.5
5000 155.9 143.2 150.0 150.0 143.0 130.3
6300 170.0 163.6 150.0 150.0 163.8 157.4
8000 180.0 165.0 150.0 150.0 196.7 181.7
___________________________________________________________
























































































































































































_____________________________________
m = 0.175 m = 0.225 m = 0.2
_____________________________________








Tableau 34 [T10.19], p.11
It should be noted that:
 different values of m and W weights are used for
the three different algorithms discussed above;
 the values of Wsfor the calculation of SR25E are
different from the values used for the calculation of SLR within
the same algorithm. The same rule applies to WRfor the calculation
of RR25E and RLR.
 SU\dM\dJare the sending SFC values of telephone
systems measured at RESP (reference equivalent speaking position)
and the RGP for the calculation of SR25E and SLR respectively; and
 the LEvalues listed in Recommendation P.79 are
used to correct the telephone receiving systems for the calculation
of RR25E and RLR.
3.2.6.2 Method using the ISO532B algorithm
a) Input signal: the longterm speech spectrum
given in Recommendation P.50 and 1/3 octave data are used.
b) Reference system: NOSFER sending + 25 dB +
NOSFER receiving for R25E, and IRS sending + 18 dB + IRS receving
for LR.
c) Tested system: varies depending on the items.
For example, for SR25E, sending tested + variable attenuator (X ) +
NOSFER receiving.
Proceed as follows:
1) Calculate the loudness of the reference system
using the ISO532B algorithm on the basis of the output levels in
1/3 octave bands of the reference system.
2) Calculate the loudness of the tested system
using the ISO532B algorithm on the basis of the output levels in
1/3 octave bands of the tested system, change the attenuation
value X of the variable attenuator until the calculated loudness is
the same as that in the reference system.
3) Then:
R 25 E = 25  X
LR = 18  X
In calculating R25E and LR with the ISO532B algorithm, the
SFC data of the reference systems and the telephone systems are the
same as those used in the other algorithms discussed above.
3.2.7 Calculated results
The subjectively determined values of SR25E, RR25E, SLR and
RLR, the result calculated by using various algorithms, and the
differences between the subjectively determined values and the cal
culated results are given in Tables 35 a) to 35 d).
For the sake of comparison, the mean results calculated by the
four algorithms are summarized in Table 36.
3.3 Discusion
Before analyzing the calculated results, it is necessary to
bear in mind the effect of the diffraction by the human head and
the reverberation of the test room on the sending SFC and NOSFER.
As a result of this effect, the difference in the sending SFC
between the mouth reference point of the NOSFER system and a point
140 mm in front of the operator's lips is less than 13.46 dB under
ideal conditions, i.e. with the virtual sound source 6 mm behind
the lips being taken to be the actual human sound source and assum
ing the sound to be transmitted in a free field. In the Chinese
subjective test room, this difference has been mesured with an
average correction of 1 to 1.5 dB for each frequency (see contribu
tion COM XII209 (19851988)). This effect has not been included in
any of the four algorithms discussed above.
3.3.1 The calculated results of SR25E and RR25E using the P.79
Cor. algorithm are about 1.5 to 2 dB higher than the subjectively
determined values. This is understandable because the values of
slope parameter m and the G functions were estimated on the basis
of the filter test results of the old CCITT test team.
3.3.2 Both the P.XXE and the Chinese algorithms can be used as
the uniform algorithm for the calculation of R25E and LR. The SR25E
calculated by the P.XXE algorithm is about 1 dB lower than the sub
jective result, but after correction for diffraction by the human
head and the reverberation of the test chamber as discussed in
S 3.3, there may be fairly good agreement between the subjective
and objective values of SR25. In view of the fact that some values,
at high and low frequencies, of the G functions and W weights
used in the P.XXE algorithm are chosen arbitrarily and that a
correction has to be made to the sending SFC of NOSFER, the Chinese
algorithm may be better than the P.XXE algorithm in use.
3.3.3 The results calculated using ISO532B agree with the
corresponding subjective test results. It has been noticed, how
ever, that the standard deviation for the mean values of the
differences of SR25, SLR and RLR is larger than that for the other
algorithms. Furthermore, the ISO532B algorithm is much more com
plicated than the other algorithms. This algorithm would not there
fore be the best choice.
3.3.4 The difference of the results of SLR and of RLR
calculated with the P.79 Cor. algorithm and with the original P.79
algorithm, respectively, is generally less than 0.1 dB.
3.3.5 It is not advisable to use the P.79 Cor. algorithm to
calculate R25E values because of the considerable difference
between the calculated values and the subjectively determined
values.
The difference between the subjectively determined values of
SR25E for a telephone set obtained by the old and by the new test
team is about 4 to 6 dB, respectively, while the difference calcu
lated by the P.79 Cor. algorithm and by the Chinese algorithm is
about 2 dB.
The value of R25E calculated by the P.79 Cor. algorithm does
not agree with the subjectively determined value of the old test
team either.
3.4 Conclusion
A simpler algorithm such as the Chinese algorithm can be used
as the standard algorithm for the calculation of R25E and LR. There
is good agreement between the calculated results and the results
subjectively determined by the new test team of the CCITT Labora
tory.
The statement appearing in some Recommendations to the effect
that a simple algorithm cannot be used for the comparison of the
loudness of wideband systems should be revised.
4 Loudness rating coefficients derived from subjective meas
urements on highpass (HP) and lowpass (LP) filtered speech (Con
tribution from ELLEMTEL, Sweden)
4.1 Introduction
The exact shape of the frequencyweighting of the loudness
rating (LR) algorithm is not very critical when computing LR values
for routine planning evaluations. However, reasonable realistic
values of the coefficients are needed for a more detailed analysis
of, for instance, attenuation distortion and bandwidth restriction
effects.
Loudness rating parameters may be derived from known statis
tics of the "average" speech power spectrum and the "average" hear
ing threshold frequency response curves.
An alternative direct way is to make use of subjective listen
ing tests of the influence of variable lowpass and highpass
filters in a NOSFER type circuit. Such measurements have been made
many times in the past. This section uses four sets of data of
which three originate from STL [4] and one from the People's Repub
lic of China [5].
As is well known, subjective evaluation of loudness has its
difficulties. A prime requirement is that the test team must
represent "ordinary people" with regard to speech and hearing.
Also, the team must be instructed to judge specifically "loudness
impression" and not "quality impression" of bandwidth limitation.
The CCITT test team seems not to have fulfilled these criterions
when performing the measurements for the P.79 algorithm.
4.2 Derivation of loudness rating coefficients
For complex noise spectra of timeconstant nature the masking
effects between frequency bands have to be considered, i.e. the
Zwicker algorithm should be used for evaluating loudness. However,
it is rather doubtful whether this complex method is really neces
sary (or even correct) for speech signals. Instead, the simpler
conventional " physiological loudness impression " model for
interpreting the subjective results will be tried.
The expression for the loudness loss A caused by a filter
introduced in the electric part of the transmission path of speech
sounds from mouth to ear will be given. To facilitate the mathemat
ical treatment, the usual series summation over the thirdoctave
bands is replaced by a continuous integration over a logarithmic
frequency scale.
H.T. [1T11.19]
_________________________________________________________________________________________
TABLE 35
{
a)
Comparison of subjective and calculated results using the P.79 Cor.
algorithm
}
Unable to convert table
_________________________________________________________________________________________























Tableau 35a) [1T11.19], p.12 a l'italienne
H.T. [2T11.19]
_____________________________________________________________________________________
{
TABLE 35 (continued)
}
{
b)
Comparison of subjective and calculated results using the P.XXE
algorithm
}
Unable to convert table
_____________________________________________________________________________________



























Tableau 35b) [2T11.19], p.13 a l'italienne
H.T. [3T11.19]
_______________________________________________________________________________________
{
TABLE 35 (continued)
}
{
c)
Comparison of subjective and calculated results using the Chinese
algorithm
}
Unable to convert table
_______________________________________________________________________________________



























Tableau 35c) [3T11.19], p.14 a l'italienne
H.T. [4T11.19]
___________________________________________________________________________
TABLE 35 (end)
{
d)
Comparison of subjective and calculated results using
the ISO532B algorithm
}
Unable to convert table
___________________________________________________________________________





















Tableau 35d) [4T11.19], p.15 a l'italienne
H.T. [T12.19]
TABLE 36
Summarized results showing the mean differences and standard
deviations between the subjective
and calculated R25Es
and LRs using various algorithms
______________________________________________________________________________________________________
  

SR25E RR25E SLR RLR
  Mean  ua) Std. dev.  ua) Mean Std. dev. Mean Std. dev. Mean Std. dev.
______________________________________________________________________________________________________
Chinese 0.57 0.22 0.68 0.48 +0.67 0.54 +0.12 0.83 +0.63 1.21
______________________________________________________________________________________________________
P.XXE 1.41 1.06 0.64 0.45 0.33 0.60 +0.12 0.82 +0.80 1.27
______________________________________________________________________________________________________
P.79 Cor +1.56 +1.91 0.87 0.57 +1.83 0.79 0.39 0.85 +0.01 1.11
______________________________________________________________________________________________________
ISO532B 1.15 0.06 0.92 1.05 +0.23 0.52 +0.10 0.93 +0.67 1.42
______________________________________________________________________________________________________






















































































































a) Two values in each block correspond to subjective test results
quoted from different Technical Reports of the CCITT Laboratory.
Tableau 36 [T12.19], p.16
Thus, the loudness loss becomes: Formula [F3.19], p.
where
m is the loudness growth factor
X log1\d0 {  fIF /F0 } F0= 1 kHz
K (X ) is the loudness weighting factors 4.2
L (X ) is the attenuation of the filter
For K (X ) it is stipulated that Formula [F4.19],
p.
Otherwise, K (X ) remains to be determined as well as the
value of m . [In Equation (41) however, the exact value of m has
only a secondorder effect as has been discussed in other contribu
tions.]
For a highpass filter with negligible loss in the pass band
and sharp cutoff at F = Fc(X = Xc) we get:
Similarly, for a lowpass filter
Using Equation (43) we get
For a chosen value of m , we may now plot as a function of Xc
S shaped curves are obtained as in Figure 41 a). If the two
curves more or less coincide as in Figure 41 b) the "best" value
of m has been found. Then a mathematical expression for a curve
Y0which fits the coincidence curve reasonably well is sought. The
derivative of Y0thus gives K (X ).
Figure 41, p.
Of course S shaped curves can be described by an infinite
number of mathematical functions. However, the normal error
integral turns out to be a suitable choice. Plotting Y (A ) on a
"normal distribution diagram" paper gives, in essence, straight
lines.
Figures 42, 43 and 44 present the results from data given
in [4]. (The measurements were made at STL in January 1986,
May 1975 and February 1975.) It is interesting to note how well the
points cluster around straight lines, especially in Figure 42. The
corresponding K (X )curves are plotted in Figure 45 together with
a curve derived from [5] as presented in [6].
4.3 Discussion and conclusions
It is remarkable that the weighting curves depicted in Fig
ure 45 coincide so closely although they were made by very dif
ferent test teams.
Curve 4 in Figure 45 has been used as a kind of reference in
the further development of the "simplified" algorithm P.79A. The
STL HPLP measurements seem to confirm that this "weighting refer
ence" is quite suitable. Thus, the P.79A algorithm will give a rea
sonable estimation of attenuation distortion and bandwidth limita
tion effects.
Another conclusion is that the loudness loss caused by
attenuation distortion and bandwidth limitation can be explained by
the simple loudness rating model without resorting to the Zwicker
algorithm
Curve 4 in Figure 45 was used to compute the corresponding
20weights for the 1/3octave frequencies in the series summation
for the 0.18 kHz band, see [6]. These are shown in Figure 46
together with the equivalent Kivalues for P.79. As can be seen,
the P.79 curve has some absurd peaks and gives more emphasis to
lower frequencies and less to higher frequencies. Thus, P.79 can
be expected to underestimate the effect of how attenuation slope as
a function of frequency influences the loudness loss of a connec
tion. This seems to be verified experimentally, as reported in [7].
FIGURE 42, p.18
FIGURE 43, p.19
FIGURE 44, p.20
FIGURE 45, p.21
FIGURE 46, p.22
5 Loudness ratings and bandwidth in transmission planning (Contri
bution from ELLEMTEL, Sweden)
It is shown below that loudness ratings can be specified as
"basic" parameters in the "common" band 0.33.4 kHz complemented
with an E factor for the band edges down to 0.2 kHz and up to
4 kHz. The E factor can be determined numerically from attenuation
values or by some simple network rules. The advantage of the method
is a simplification for the transmission planner.
5.1 Introduction
Many Administrations seek to maintain good transmission pro
perties in a telephone channel with a band of 200 to 4000 Hz, at
least in the subscriber network. Under those circumstances it may
seem natural to compute loudness ratings (LRs) using parameters
specified for this band [8]. Because in this case the loss distor
tion is limited within the band, the additivity properties of the
LRs will be satisfactory, i.e.:
OLR = SLR + JLR + RLR .
However, a connection may often contain links with an appreci
able band edge attenuation distortion, virtually limiting the band
to 3003400 Hz. (This will be true for many interntional calls.)
Such hard bandlimiting corresponds to an increase of several deci
bels in LR. If LRs are computed for the band 0.18 kHz or even
0.24 kHz they can no longer be added without noticeable errors [9]
which may cause confusion in the transmission planning.
In principle there are several ways of resolving the dilemma.
The first is simply to ignore the improvement of a few dBs which
some "wideband" local connections may possess. Thus, the American
IEEE practice for objective loudness ratings is to use the band
0.334 kHz when computing (or measuring) the LRs.
The second method is to apply bandwidth correction factors to
the LRs. One may compare with the CCITT concept of "corrected
reference equivalents" which is tailored to the subscriber's actual
loudness impression. A 2004000 Hz circuit will have a lower CRE
value than a 3003400 circuit having the same midband loss. The
effect of bandlimiting is taken care of by correcting the wideband
values by adding the socalled D factors according to certain
rules. [The CCITT D factors may not be quite correct, however, as
they were derived from measurements using SRAEN filters. These are
not truly representative of modern transmission circuits [10].]
Considering modern trends of trying to improve the telephone
channel's lowfrequency response, it seems appropriate for the LR
calculations to use a "wideband" method with corrections.
Such a methodology will be described below and this can be
applied to transmission planning.
The LRs are basically calculated for the narrow "common band"
0.33.4 kHz. These LRs can be added without loss of accuracy. A
correction, the E factor, is computed for the transmission at the
band edges. The E factor is subtracted from the "common band" OLR
to obtain the "wideband" OLR (W ).
5.2 The Efactor as a band edge correction of LR
In general, a loudness rating can be thought of as a
"frequencyweighted average" of an electroacoustical attenuation.
According to Recommendation P.79 the electroacoustical pro
perties should be evaluated in the band 0.18 kHz. For practical
reasons the computations are often limited to the band 0.24 kHz.
(The Wiweights are
then diminished by 0.3 dB). However, only in the band 0.33.4
is one assured of a real transfer of signals under all cir
cumstances. At the band edges, 0.2 to 0.3 and 3.4 to 4 kHz, the
attenuation of a specific link in a connection may be so high as
virtually to stop transmission. This could result in a reduction of
several decibels in a subjectively measured loudness impression of
a voice signal.
To handle this properly it is convenient to characterize the
electroacoustical attenuations separately for the "common band"
0.33.4 kHz and for the band edges.
In the common band each link is characterized by the weighted
average of the electroacoustical loss, i.e. SLR, RLR or JLR, and
the LRs can be added. For example, for the circuit as shown in Fig
ure 51, consisting of two telephone sets interconnected via a
number of transmssion links, the following relation should hold at
any interfacte P between the links:
(Any mismatch attenuation effects at the interfaces can be treated
as special forms of JLRs).
Figure 51, p.
At the band edges the connection is characterized by its abil
ity to transmit voice signals, i.e. the E factor. Zero band edge
losses means E = 2.5 dB. (Details will be given later).
For a complete connection as shown in Figure 51, the overall
loudness rating is:
In the common band 0.33.4 kHz OLR = SLR + RLR
(5.2)
In the full band 0.24 kHz OLR (W ) = OLR  E
(53)
In the following, the general mathematical expressions for the
LRs and the E factor are given. It is shown how to apply them to
telephone sets and various transmission links.
The E factor may be designated the "loudness improvement".
5.3 General mathematical expressions
In the common band 0.33.4 kHz the general LR algorithm can be
written as:
LR = L0+
L
(54)
(the summation being made for fi= 0.315 .   3.15, the
1/3octave ISO frequencies)
when
Li are the values of electroacoustic loss for the LR in
question
L0, Ki, m are constants to be specified below.
Note  Equations 54 and 55 are mathematically equivalent to
the "Wialgorithm" as explained in [11] but are more convenient to
use in the following.
When the spread between minimum and maximum values of the Li's
is moderate, the following expression can be used for L .
The full band 0.24 kHz overall loudness rating was given by
Equation 53, i.e.:
OLR (W ) = OLR  E
The expression for the E factor is:
E = C1x
10
0.1m (L 01

L )
+ C2x
10
0.1m (L 02

L )
+ C3x 10
0.1m (L 03

L )
(57)
where
L0\d1, L0\d2, L0\d3 are the band edge losses at 0,2,
0,25 et 4 kHz respectively.
C1, C2 and C3 are constants to be specified below (for
the derivation, see Annex A).
The constants of Equations 54 and 55 may in principle be
derived from any defined LR algorithms. However,
Recommendation P.79 is less suitable for use in the context of
transmission planning because of a lack of accuracy as discussed in
[11] and [7]. The very simple algorithm, designated by "C " in
[11], seems to be just as accurate as any other investigated so far
and is therefore chosen here. (It also has the advantage of closely
resembling the IEEE objective loudness rating.)
The constants used in Equations (54), (55), (56) and (57)
are given in Tables 51 and 52.
H.T. [T13.19]
TABLE 51
Constants used in equations (54), (55) and (56)
______________________________
{
K
= 0.05 for f
=
0.315 and 3.15 kHz
}
{
K
= 0.1 for f
=
0.4, 0.5 .   2, 2.5 kHz
}
m = 0.2
______________________________




























_________________________________
LR SLR RLR OLR JLR
_________________________________
L 0 3 12 9 0
_________________________________
























Table 51 [T13.19], p.24
H.T. [T14.19]
TABLE 52
Constants used in equation (57)
__________________________________
C 1 = 0.5 C 2 = 1 C 3 = 1
__________________________________








Table 52 [T14.19], p.25
A perfectly flat mouthtoear acoustic frequency response in
the band 0.24 kHz will thus result in E = 0.5 + 1 + 1 = 2.5 dB,
i.e. OLR (W ) is 2.5 dB lower than OLR for a "flat" channel limited
to 0.33.4 kHz.
In the following the E factor is computed for a number of
cases including "broadband" and "narrowband" telephone sets in com
bination with different types of transmission links. It turns out
that some rather simple rules can be set up for the approximate
determination of the E factor.
5.4 Telephone sets
Suppose the transmission channel between the sending and
receiving telephone sets is flat within the band 0.24 kHz. Then
the E factor, the loudness improvement, characterizes the
bandwidth performance of the sets. Let this be designated ET.
Table 53 gives some examples. It is worth noticing that the spread
in E around the average value 1.3 is quite moderate.
H.T. [T15.19]
TABLE 53
Examples of Efactors for some telephone sets
_______________________________________
Type of set E
_______________________________________
{
1)
Oldtype carbon microphone
} 1.9
{
2)
Oldtype carbon microphone
} 1.5
{
3)
Oldtype carbon microphone
} 1.1
4) W.E. type 500 1.3
5) Electret microphone 1.8
{
6)
Digital set; new specification
} 0.8
{
7)
Average of 90 types of sets
} 1.3
_______________________________________











































































Table 53 [T15.19], p.
5.5 Transmission links
To characterize the loudness improvement performance of the
transmission links as such, it is convenient to compute the E fac
tor under the assumption that the telephone sets have a flat fre
quency response in the band 0.24 kHz. Let this transmission chan
nel E factor be designated EC. In S 5.6 the resulting E factor
will be given for various typical combinations of ETand EC.
When connecting several transmission links in tandem, mismatch
may occur. These effects can be diminished, however, by using com
plex nominal impedances in the subscriber networks, as many
Administrations already do.
In general, mismatch losses can be considered by computing
their JLRs .
The "common band" performance of the links are characterized
by JLR = L according to Equations (54), (55) and Table 51. As
large attenuation distortions within this band are not allowed, the
very simple Equation (56) can be used for computing L . (It is
interesting to note that this corresponds in effect to averaging
the loss over a log (f )scale, a method which has been verified
empirically a long time ago).
When several links are connected in tandem, ECcan of course be
computed from the total band edge losses. However, some simple
approximate rules can be used for the combination of individual
ECfactors for JRL .
5.5.1 Subscriber cables
Surprisingly, the typical loss curve of a nonloaded sub
scriber cable produces the same loudness improvement as a
fullbandwidth channel, i.e. EC = 2.5 dB. This is due to the fact
that at the lower band edge the loss is lower than the average L
which compensates for the higher loss at the upper band edge.
When a subscriber cable is connected in tandem with a narrow
band device, it turns out that the ECfactor for that device
applies.
5.5.2 Bandlimiting equipment
Bandlimiting in a telephone connection can be caused by
heavily loaded subscriber cables, FDM and PCM equipment. Figure 52
shows some idealized attenuation curves for which the ECfactors
have been computed.
Figure 52, p.
For the attenuation curves in Figure 52 the following
ECvalues are obtained:
EC
Heavily loaded subscriber cable 1.2 dB
1 FDM link 1.4 dB
1 PCM link 1.9 dB
When several PCM and FDM links are connected in tandem the
ECvalues according to Table 54 are obtained.
5.6 Complete connections
The loudness improvement, the E factor, has been computed for
a number of combinations of telephone sets and transmission links.
For each telephone characteristic the "total" E factor has been
plotted against the "channel" ECfactor. The results are presented
in Figures 53 and 54.
H.T. [T16.19]
TABLE 54
EvC for PCM and FDM links in tandem
________________________________________________________
No. of PCM links 0 1 2 3 4 5
No. of FDM links
________________________________________________________
0 2.5 1.9 1.6 1.4 1.3 1.2
1 1.4 1.3 1.2 1.1 1.0 1.0
2 1.1 1.0 1.0 0.9 0.9 0.8
3 0.9 0.9 0.8 0.8 0.7 0.7
4 0.8 0.7 0.7 0.7 0.6 0.6
5 0.7 0.6 0.6 0.6 0.5 0.5
________________________________________________________
















































































Tableau 54 [T16.19], p.28
Figure 53 shows the E factor for the "average" analog and
the "digital" telephone set. (The "average" was taken as the mean
of 90 different types of commercial sets. The "digital" corresponds
to the new CCITT specification for digital sets).
Figure 54 illustrates the spread in the E factor for a
number of widely varying analog telephone characteristics. (It is
worth noticing that the spread, after all, is fairly moderate).
Figure 53, p.29
FIGURE 54, p.30
Considering the general requirements of transmission planning,
there is hardly a need to specify the loudness improvement, the
Efactor, more accurately than within steps of 0.5 dB. Therefore,
instead of calculating the Evalue in each application, one can
follow the rules given in Tables 55 and 56 for analog and digi
tal sets respectively.
H.T. [T17.19]
TABLE 55
Efactor, analog sets
____________________________________________________
E Links in tandem
____________________________________________________
1.5 {
Subscriber cable, nonloaded
}
1.0 {
1 x PCM, .   , 3 x PCM
1 x FDM
}
0.5 {
Subscriber cable, heavy coding
}
4 x PCM 2 x FDM
0.0 {
5 x PCM + 5 x FDM
}
____________________________________________________



















































Note  Nonloaded subscriber cable sections do not affect the E
factor.
Tableau 55 [T17.19], p.31
H.T. [T18.19]
TABLE 56
Efactor, digital sets
_____________________________________________________________
E Links in tandem
_____________________________________________________________
1.0 All digital connection
0.5 {
1 D/AA/D to 6 D/AA/D connections
}
0.0 7 D/AA/D connections
_____________________________________________________________
























Tableau 56 [T18.19], p.32
5.7 Conclusions
The transmission planner can obtain loudness ratings by quite
simple numerical methods: computing and adding individual LRs for
the "common band" 0.33.4 kHz and correcting for the band edge
transmission by subtracting the E factor. The E factor can be
determined by some uncomplicated rules.
The results can be expected to be more accurate than calcula
tions based on Recommendation P.79.
MONTAGE: S 6 SUR LE RESTE DE CETTE PAGE