0% found this document useful (0 votes)
30 views

Calculating Reverberation Time From Impulse Responses - A Comparison of Software Implementations

Uploaded by

Paulo Gonçalves
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
30 views

Calculating Reverberation Time From Impulse Responses - A Comparison of Software Implementations

Uploaded by

Paulo Gonçalves
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

Acoust Aust (2016) 44:369–378

DOI 10.1007/s40857-016-0055-6

TECHNICAL NOTE

Calculating Reverberation Time from Impulse Responses:


A Comparison of Software Implementations
Densil Cabrera1 · Jianyang Xun1 · Martin Guski2

Received: 1 February 2016 / Accepted: 16 April 2016 / Published online: 4 May 2016
© Australian Acoustical Society 2016

Abstract In room acoustics measurement, calculating reverberation time from room impulse responses is often done, aided
by software. This paper compares the performance of nine software implementations for calculating octave band reverberation
time, including two written by the authors. Synthetic impulse responses are used to test decays without and with a steady
noise floor, and an impulse response from a real measurement is also used for comparison. Results indicate no significant
reverberation time calculation problems for noise-free exponential decays, and for exponential decays leading to a steady noise
floor. Frequency selectivity is identified as an area for potential improvement in filter-bank design, and a highly selective octave
band filter-bank is shown to be effective without introducing errors. Testing with a real measured impulse response, which
had been used in a 2004 study comparing reverberation time analysis implementations, showed greater agreement between
software than was found previously. This might reflect an improvement in software performance in the years between the
two studies. However, it also might reflect the smaller scale of the present study. Nevertheless, the results can contribute to
confidence in current software implementations.

Keywords Room acoustics · Reverberation time · Software

1 Introduction is quite common for researchers and professionals in archi-


tectural acoustics to use software available for purchase or
Reverberation time is the most used and best-known parame- distributed freely from external sources. Hence, with a range
ter in room acoustics and has a long history of development of software available, measured values of reverberation time
in terms of theory, measurement and application. It is estab- might be affected by software choice.
lished in key standards, such as the ISO3382 series, and The majority of published work comparing reverbera-
almost all projects concerned with room acoustics use rever- tion time calculation from an impulse response has focussed
beration time in some way. Measurement of reverberation on the efficacy of algorithms, often proposing refinements
time can be done in various ways, and is commonly done or novel approaches to the calculation method (e.g. [1–3]).
by recording a room impulse response, often indirectly (e.g. Related to this, research has examined the errors associated
from a swept sinusoid signal). Deriving reverberation time with certain types of input, most often examining how to
from a room impulse response relies on software, and while deal with a steady-state noise floor, and proposing required
some researchers implement their own analysis software, it impulse response criteria for results within acceptable error
margins (e.g. [2,4–6]). Another less common type of compar-
B Densil Cabrera ison study is between software, without particular knowledge
[email protected] of the algorithm, and there is some value in this considering
1 Faculty of Architecture, Design and Planning, The University
that many acousticians rely on closed source software. The
of Sydney, Sydney, NSW 2006, Australia present study is primarily a software comparison, but it also
2 Institute of Technical Acoustics, RWTH Aachen University,
considers some algorithm-related matters as they arise from
Aachen, Germany the comparison.

123
370 Acoust Aust (2016) 44:369–378

Lundeby et al. [7] and Bradley [8] conducted studies 1.1 Issues in the Calculation of Reverberation Time
in the 1990s comparing systems for reverberation time
measurement and analysis. In Lundeby’s study the whole In general, the procedure for calculating reverberation time
measurement and analysis system (including physical equip- from a single measured room impulse response will involve
ment) was included, whereas Bradley used an electronic the following steps:
artificial reverberator to remove physical variability. How-
ever, both studies were comparisons of measurement and 1. Truncation of the impulse response at the start, although
analysis systems, rather than isolating just the analysis part. this should have little effect on reverberation time calcu-
Subsequent studies by Ikegami and Uchida [9] and Katz lation and is mainly important for energy ratio parameters
[10] focussed only on analysis by using an impulse response such as clarity index (ISO3382-1 suggests starting where
as the test input for software. Katz’s impulse response (a the energy first rises significantly above the background
recording of a balloon pop in a large room) is of partic- noise, but is more than 20 dB below the peak);
ular interest because of the large scale of the study and 2. Applying a filter-bank (e.g. octave band);
because the impulse response has a significant noise floor. 3. Truncation of the impulse response at the end, usually at
The impulse response was sent to researchers and practition- a different point in each band;
ers to analyse, using software of their choice. Results showed 4. Squaring the impulse response;
a wide range of reverberation time values in the lowest octave 5. Optionally compensating for the decay energy lost from
band (125 Hz), for which signal-to-noise ratio was poorest. end-truncation and/or subtraction of the excess energy
Katz’s study is also of particular interest because some of due to a steady-state noise floor;
the software that was used for the tests is still in use today, 6. Integrating the impulse response in reverse time, and con-
albeit in updated versions. Since Katz’s study, some smaller verting to decibels; and
scale software comparison studies have been published, for 7. Finding reverberation time parameters by linear regres-
example, by Tronchin et al. [11] (four software implemen- sion over the appropriate evaluation ranges (−5 to
tations), Ţopa et al. [12] (three implementations), Machín −25 dB for T20 ; −5 to −35 dB for T30 ).
et al. [13] (three implementations, two external and one by
the authors) and Mansilla et al. [14] (four implementations, Steps 2 and 3 and the optional step 5 have some scope
three external and one by the authors). The conclusions of for variations in approach, which could affect the resulting
these studies vary: Tronchin et al., using impulse responses values.
measured in an opera theatre, found average T30 (reverbera- Filter-banks for reverberation time measurement can
tion time) deviations of up to 8.3 % between software, with a influence the result for two reasons. Firstly, the time-response
maximum deviation of almost 25 %; Ţopa et al. tested three of the filter-bank can have its own decay, which could artifi-
rooms (a school hall, a church and a university auditorium) cially increase the measured reverberation time [16]. This is
and found that the software did not affect the results; the only likely to be noticeable when reverberation time is very
main aim of the study by Machín et al. was validate soft- short, and especially in the lower octave or 1/3-octave bands.
ware developed by the authors through a set of measurements However, this can be avoided by using reverse-time filter-
in one auditorium, and they found only minor deviations in ing [3,17]. Secondly, the selectivity of the filter-bank is only
reverberation time between the tested software; and Mansilla loosely defined by standards [18–20], and cases where there
et al. tested 59 diverse impulse responses, finding average is significant variation in reverberation as a function of fre-
octave band deviations of about 0.3 s in T30 between the quency will be better represented by the results of a highly
external software tested. Concurrent with the present study, selective filter-bank. Venturi et al. [21] suggested the use
Álvarez-Morales et al. [15] conducted a detailed software of highly selective filters (144 dB/octave skirts, equivalent to
comparison based on a large set of in-situ measurements 24th order) for octave band reverberation time measurement,
in a university auditorium, using four software implemen- finding that high-order filtering always improved the accu-
tations, with extensive statistical analysis. They found that racy reverberation time results compared to sixth-order filters
the implementations mostly agreed when used as recom- (both of which comply with IEC Class 0 response limits from
mended by the software documentation. Hence the extent the standard at the time [19]). However, they speculated that
to which software agrees or disagrees appears to depend on lower selectivity might be needed for very short reverbera-
the type of testing done, and the present paper focuses on test- tion times and for short evaluation ranges such as early decay
ing specific potential vulnerabilities of software, along with time and T10 .
a reassessment using Katz’s impulse response with present Much has been written about end-truncation, because it
day software. has a large effect on the results in practical measurements

123
Acoust Aust (2016) 44:369–378 371

[4,6,7,22]. If end-truncation is omitted, then the background value due to lack of validity. Where available, noise com-
noise will accumulate in the reverse integration, potentially pensation is applied (including in cases where the impulse
yielding artificially high reverberation time values. Follow- response is essentially noise-free), and where reverse-time
ing ISO3382-1, the end-truncation point should be at the filtering is chosen if it is a user setting. For the sake of
intersection between a horizontal line representing the back- succinctness, only T30 values are reported (i.e. reverbera-
ground noise and a sloping line representing the level decay tion time calculated over the −5 to −35 dB evaluation range
(prior to reverse integration) [18], but the method for finding in the reverse-integrated decay), as these are more vulner-
this could vary between implementations, and it is likely that able than T20 to background noise, truncation and filter
variation in results between software for impulse responses response. Indeed, Hak et al. [5] show that T30 is the most
that include noise will be partly due to the way in which sensitive of the ISO 3382-1 parameters to signal-to-noise
end-truncation is done. ratio.
Compensation for the energy lost from end-truncation is Seven software implementations were tested, together
an optional step in ISO3382-1, which helps straighten the with two of the authors’ own implementations. The seven
reverse-integrated decay. Without doing this, there may be external software instances in this study include prominent
an underestimation of reverberation time. The specifics of reverberation time calculation software that tends to be used
how this is done are open to interpretation, but ISO3382-1 in professional and research work: in alphabetic order, the
recommends a value derived from the decay rate in the final software is ARTA, Aurora, Dirac, EASERA, Odeon, Smaart
10 dB. A variation of this, by Lundeby et al. [7], provides and WinMLS. Software that required a superseded operat-
a 5–10 dB safety margin in finding this late decay rate, to ing system or particular hardware was excluded. All of this
reduce the influence of the noise floor on the late decay rate software has been included in previous software compar-
calculation. isons by: Katz (Aurora, Dirac, Smaart and WinMLS) [10],
Noise power subtraction is another technique that helps Tronchin et al. and Ţopa et al. (Aurora, Dirac and Win-
straighten the reverse-integrated decay [2]. Although this MLS) [11,12], Machín (Odeon, WinMLS) [13], Mansilla
approach is not compliant with ISO3382, noise compensation et al. (Aurora, Dirac, EASERA) [14] and Álvarez-Morales
is implemented (although not necessarily described) in com- (ARTA, Dirac, EASERA, WinMLS) [15]. One of the con-
monly used software. Guski and Vorländer [6] recommended cerns with software selection was that the software be
that this, together with truncation energy compensation, recognised as a serious analysis tool, and inclusion in pre-
be included in future modification of ISO3382. Where vious software comparisons done by acoustics researchers
available, it was included in the software tests reported provides one indication of this. Two other software imple-
here. mentations, which were not in these previous comparison
There are many non-standard alternative approaches to studies, were tested in preparing the present paper, but were
calculate reverberation time from an impulse response. excluded from this study due to the combination of deficien-
For example, Xiang [23] proposed non-linear regression cies in performance (yielding outlying values in the tests)
accounting for the reverse-integrated noise floor. Morgan and a lack of external evidence that the software was recog-
[4] proposed using an evaluation range that depended upon nised as a serious analysis tool (such as multiple citations
the noise floor (finishing 5 dB above the noise, in conjunc- in room acoustics literature or authorship associated with
tion with end-truncation at the elbow) instead of using fixed acoustics research). Although Smaart tends to be used more
evaluation ranges. However, mostly reverberation calcula- for professional audio system tuning than for room acoustics,
tion software aims to follow a standard approach (often it was included in the study because of its inclusion in
with the non-standard option of noise power subtraction), Katz’s. Every software version tested was current in Decem-
and it is common to see references to ISO3382 in such ber 2015. Like some previous software comparison studies
software. in room acoustics, including Katz’s, the external software is
de-identified, and is referred to arbitrarily using the letters
A to G.
2 Software and Test Concept The two implementations that were developed by the
authors and their colleagues are the ITA Toolbox (developed
The test concept of the present study is that the software in the Institute of Technical Acoustics, RWTH University
settings are fixed (the same settings used for every impulse Aachen) [24] and AARAE (developed at the University of
responses) so that human decision making is removed. Sydney) [25]. The code for these was written independently.
Hence, some results may be inferior if there are cases for Both of these projects are freely available as open-source
which an expert might have optimised the analysis by chang- Matlab-hosted projects, aiming to support teaching and
ing the settings, by editing the impulse response, by choosing research in room acoustics and related areas. These projects
the most appropriate value to report or by not reporting a have a wide range of possible settings for reverberation time

123
372 Acoust Aust (2016) 44:369–378

calculation, but the results reported in this paper are from responses with uniform octave band reverberation times in
the respective authors’ preferred settings. In both cases, this the 125–8000 Hz range.
included automatic end-truncation, compensating for decay
energy lost from truncation and background noise power sub- 3.2 Exponential Decay with Steady Noise Floor
traction. However, the implementations differ in details, such
as the filter-bank design, and the methods to find the trun- Apart from the relatively trivial task of calculating reverber-
cation point and to do the abovementioned adjustments. For ation time from a noise-free impulse response, probably the
the ITA Toolbox, the settings are exactly those described by most telling simple test is to introduce a steady-state noise
Guski and Vorländer [6] as ‘Method E’. The most distinc- floor. This means that automatic end-truncation must be done
tive feature of the chosen AARAE settings was that they by the analysis software, as well as associated adjustments.
applied the suggestion by Venturi et al. [21] for highly selec- An exponentially decaying envelope with a perfectly steady
tive octave band filtering (implemented as a reverse-time, or noise floor can be generated by adding a tonal noise term to
maximum phase, filter-bank in the frequency domain, with Eq. 1, as expressed by Eq. 2, where PNR is the peak-to-noise
144 dB/octave skirts). ratio expressed in decibels. Note that the tonal noise is in
quadrature with the tonal decay, so that they add similarly
to incoherent noises in terms of the resulting envelope, but
3 Tests and Results without random fluctuations. The noise term is multiplied by
the square root of 2 because PNR is a ratio of maximum to
3.1 Noise-Free Artificial Impulse Responses root-mean-square amplitude.
7  
A simple and trivial test of reverberation time software y (t) = sin 2k π f 0 t e−3ln(10)t/Tk
k=1
performance is to calculate the reverberation time of an expo- √  
nentially decaying waveform. To avoid random fluctuations + 2 × 10 P N R/20 cos 2k π f 0 t . (2)
in the envelope, we use exponentially decaying sinusoidal
waves. Equation 1 expresses the waveform, y, as a function Results from testing this are more interesting than without
of time, t (starting at 0 s and finishing at a duration suffi- noise, and they indicate that the tested software is generally
ciently long for reverberation time calculation). The lowest well behaved in dealing with steady noise. Figure 1 shows
tone frequency in the present study is f 0 = 125 Hz, and results for T30 with a PNR of 30, 35, 40 and 45 dB. In inter-
seven octaves spanning 125 Hz to 8 kHz are used (k = 1–7). preting this, it should be borne in mind that a signal-to-noise
In Eq. 1, Tk is the octave band reverberation time, and so the ratio of 45 dB or greater is required for standard calculation
error in the reverberation time calculation software can be of T30 , so the lower PNR values do not meet this require-
found by comparison to this. ment. Admittedly there is some subtlety in signal-to-noise
ratio metrics for impulse responses as explained by Hak et
7   al. [5], but in this paper we use PNR because of its simplic-
y (t) = sin 2k π f 0 t e−3ln(10)t/Tk . (1)
k=1

The synthetic impulse responses tested for this paper were


generated in Matlab using a sampling rate of 48 kHz, and
saved as 16-bit wav files, which were then analysed by the
reverberation time software. Using a reverberation time of
1 s in all seven bands and a waveform duration of 2 s, all
software implementations tested yield negligible errors. Sim-
ilarly, negligible errors are found with longer and shorter
uniform reverberation times. For a very short reverberation
time of 0.1 s, results mostly show negligible errors, although
software implementation A only returned results in the two
lowest octave bands. Implementation C returned an increased
result in the 125 Hz band (of 0.112 s), which is the expected Fig. 1 Reverberation time (T30 ) calculated from synthetic impulse
effect of a filter’s own reverberation contaminating the result responses consisting of octave-spaced exponentially decaying sine
(e.g. for a minimum phase filter). Considering that 0.1 s is a tones mixed with an octave-spaced cosine tone noise floor, for peak-to-
noise ratios from 30 to 45 dB. The synthesised reverberation time in all
shorter reverberation time than we would normally encounter
bands for all impulse responses was exactly 1 s and the waveform dura-
in non-anechoic rooms, the conclusion is that no substantial tion was 2 s. Calculated reverberation time values shown are averaged
issues were revealed by testing noise-free artificial impulse over the seven octave bands 125–8000 Hz

123
Acoust Aust (2016) 44:369–378 373

ity as a parameter for synthesis and analysis. In two cases I


(implementation E and the ITA Toolbox), the software did
not output a result for PNR of 30 and 35 dB. In other cases
(including AARAE), although a result is returned at a PNR of
30 dB, its validity is counter-indicated by other output para-
meters such as measured PNR, correlation coefficient of the
linear regression, and/or the ratio of T20 to T30 . Nevertheless,
some implementations (B, C and F) are remarkably robust in
this test, yielding errors of only about 1 % in the worst case
PNR.
A 5 % error margin is generally considered acceptable in
reverberation time measurement, and almost all of the results
in Fig. 1 are within this margin. At the required PNR of 45 dB, II
all but one result (implementation A) have an error of 1 %
or less. The results do not reveal any significant problems
with any implementations, and provide confidence that end-
truncation and associated adjustments by the software are
generally functioning well, at least for simple exponential
decay with added steady-state noise.

3.3 Frequency Selectivity

To examine the effect of octave band filter-bank selectivity,


decaying octave-spaced sinusoids without added noise floor
were generated, with reverberation time varying markedly III
between adjacent octave bands. The two examples presented
here have reverberation time alternating between 0.1 and
1 s over the seven octave bands from 125 to 8000 Hz, or
between 0.5 and 1 s over the seven bands. The first of these
is an extreme case, with a 10:1 ratio that is highly unlikely
to be encountered in octave band measurements of any real
room. The second has a large but not implausible ratio of 2:1
between adjacent bands. The results in Fig. 2 show that in
many cases the implementations return a value much higher
than the short reverberation time (0.1 or 0.5 s) in the 125,
500, 2000 and 8000 Hz bands, due to leakage from the adja-
cent band or bands. On the other hand, the long reverberation Fig. 2 Reverberation time (T30 ) calculated from a synthetic impulse
time (1 s) in the 250, 1000 and 2000 Hz bands is accurately response consisting of octave-spaced exponentially decaying sine tones
analysed in all cases. The filter-bank for the AARAE imple- with a synthesised reverberation time alternating between 0.1 and 1 s
over the 125–8000 Hz bands (top chart, labelled I) or 0.5 s and 1 s over
mentation was designed to be highly selective (following the 125–8000 Hz bands (middle and bottom chart, labelled II and III).
the suggestion of Venturi et al. [21]), and returns results For visual clarity, the bottom chart shows the same data as the middle
with negligible error in both examples. The ITA Toolbox chart, but omits the bands that have a reverberation time of 1 s
implementation was not configured for extreme selectivity
for this test (and used filters equivalent to 10th order), but
nevertheless is one of the most selective. Some implemen- ratio of reverberation time between adjacent bands, three of
tations, especially A, B and G, yield results indicating low the nine implementations return errors of less than 5 % in
selectivity. Cabrera et al. [21] previously compared software all bands. If the 24th-order filter-bank is considered to be
for speech transmission index measurement, and observed unnecessarily selective, a 12th-order filter-bank (tested using
that the modulation transfer functions of A and B were con- AARAE) yields a maximum octave band T30 error of 0.3 %
sistent with that expected for sixth-order filters (A and B (or T = 0.5017 s in the 0.5 s bands) in the 2:1 ratio case (not
in the present study are labelled F and C, respectively, in shown in the figure).
the previous study), which comply with IEC Class 1 octave The waveforms tested here have unrealistically high con-
band filter criteria [20]. Considering the case with the 2:1 trast between bands, and so the effect of frequency selectivity

123
374 Acoust Aust (2016) 44:369–378

will be less important in most real cases. However, the inter- because it results in a shift of energy backwards in time.
esting point here is that very high selectivity is achievable Hence zero-phase filtering is conceptually ideal for centre
without causing problems with the measurement of short time, but not ideal for reverberation time. The theoretical
reverberation times. To examine this further, a synthetic tonal centre time, Ts , of a simple exponential decay is proportional
impulse response with octave band reverberation times of to its reverberation time (Eq. 4).
0.01 s was tested with the AARAE 24th-order filter-bank
implementation. The measured result for T30 is 0.01 s in all T
Ts = . (4)
seven bands (using ‘double’ precision format for the input 6ln (10)
waveform, instead of the 16-bit signed integer wav file format
that was used for the other tests reported here). This confirms Tests using AARAE with ideal exponential decay of
the proposition by Venturi et al. [16] that a highly selective T = 1 s (for which Ts = 72.38 ms) using a time-reverse
filter-bank provides robust analysis, and it extends this by filter of order 6 to 24 yielded octave band centre time values
showing that the filter-bank does not introduce significant with errors no more than 0.1 ms, which can be neglected. A
time-related errors for the shortest conceivable reverberation shorter reverberation time is more vulnerable to filter effects,
time, at least for T30 calculation from an exponential decay. and a decay of T = 0.1 s (for which Ts = 7.238 ms) was
also tested with filter orders 6 to 24. In that case, the errors
3.3.1 Further Implications of Filter-Bank Design: Other were higher in absolute terms and much higher in relative
Parameters terms, with a maximum error of 0.27 ms. This error can still
probably be regarded as acceptable, especially since Ts is
In the present work we focus on reverberation time, but a unlikely to provide any insights to situations with very short
filter-bank ideal for that might not be ideal for another para- reverberation times. These maximum errors occurred in the
meter. For example, for speech transmission index (STI) 125 Hz band, and errors were an order of magnitude smaller
calculation, zero or linear phase filters are ideal because in high frequency bands. Errors did not increase as the fil-
phase distortion within each octave band must be min- ter order increased. Considering this, not using a zero-phase
imised to avoid changing the octave band envelopes [26]. filter-bank for octave band centre time appears to be unprob-
Furthermore, excessive filter order artificially reduces the lematic from a practical standpoint, at least for the bands
modulation transfer function [27]. Therefore, Cabrera et al. centred on 125 Hz and higher.
[26] selected a linear phase 12th-order octave band filter-bank Although the aim of this study is to examine T30 results of
for STI analysis. On the other hand, with regard to the clar- the surveyed instances of software, the considerations in this
ity and definition energy ratio parameters in ISO3382-1, the section of the paper naturally raise the question of software
filter phase response and order do not fundamentally affect performance for clarity index and centre time. Figure 3 shows
results (apart from controlling spectral leakage) because the the C50 and Ts errors from a noise-free synthetic impulse
relevant time periods are extracted prior to filtering, so a high- response with T = 1 s in all seven bands. For the most part,
order reverse-time filter can be used. However to avoid error, the errors in C50 are negligible, although one software imple-
the filters’ reverse-time decay should be captured, which may mentation (F) deviates by more than 1 dB in the 125 Hz band.
require zero-padding the relevant time periods extracted from On the other hand, Ts errors in excess of 5 % are common in
the waveform prior to filtering. The theoretical clarity index the lower two octave bands, which could be due to forward
of a simple exponential decay is known from Eq. 3, where temporal smearing by the octave band filters.
t E is the limiting time between early and late (0.05 s for C50 ,
0.08 s for C80 ), and T is reverberation time. 3.4 Realistic Measured Impulse Response
 
1 − e−t E 6ln10/T Real measured impulse responses are much more compli-
C1000t E = 10log10 . (3)
e−t E 6ln10/T cated than the synthetic waveforms tested here. While they
may raise multiple diverse challenges for analysis, they do
This was tested using AARAE, with a reverse-time not have known reverberation time values a priori, and so
octave band filter-bank spanning 125–8000 Hz. Exponen- there is not an exactly correct result for error analysis. We
tially decaying octave-spaced tones over a range of rever- use the same impulse response as Katz [10] in the present
beration times (0.1–2 s) yielded negligible errors (<0.2 dB) paper, more than 1 decade after Katz’s study. It should be
in the seven octave bands for C50 and C80 , and increasing noted that Katz allowed expert analysts to contribute to the
the filter order from 6 to 24 did not increase the error (in fact analysis—for example, they might manually choose the start
it slightly reduced it). and end times to prepare for the analysis. By contrast, the
Conceptually, centre time may be affected by reverse-time present study excluded pre-manipulation by a human expert,
filtering, especially at high orders in the low frequency bands, and relied entirely on the automatic processes in the software

123
Acoust Aust (2016) 44:369–378 375

for the results. Visualising the impulse response in octave


bands provides some insight into the challenges faced by
the software. Figure 4 shows the squared impulse response
(expressed in decibels) in each of the octave bands. In all
bands, the noise floor is essentially steady over the long term,
and the decay appears to be linear. Greater short-term fluc-
tuations are seen in the lower bands, both in the decay and in
the noise floor. The 500–4000 Hz bands have a distinct peak
from the direct sound, and have a visually smoother decay
than the lower two bands.
Figure 5 shows the T30 results for the software implemen-
tations, together with deviations from the median in each
octave band. The ITA Toolbox and AARAE results are nearly
identical, and implementations E and F also return near-
identical results. Like the results reported by Katz, the largest
variability between implementations is in the lowest octave
band. Implementations A and C have the most distinctive
results. In the 125 Hz band, implementation C returns a value
substantially lower than the others, but is in general agree-
ment in the remaining five bands. In the 500–2000 Hz bands,
the results are in close agreement, except that implementation
A returns outlying values at 500 and 1000 Hz. Implementa-
Fig. 3 Octave band error in clarity index (C50 , upper chart) and centre
time (Ts , lower chart) for a synthetic impulse response consisting of tions B, C and F, which had very similar results with minimal
octave-spaced exponentially decaying sinusoids from 125 Hz to 8 kHz errors for the synthetic test of PNR, have a wide spread of
with a uniform reverberation time of exactly 1 s. The theoretical C50 results in the 125 Hz band for the real impulse response, sug-
value is −0.02 dB, and the theoretical Ts value is 72.38 ms. The AARAE
gesting that there is more behind the divergence of results
values were calculated using a 24th order reverse-time octave band filter-
bank than just the PNR in each band. Unlike the ideal decay lead-

Fig. 4 Octave band


instantaneous level of the
impulse response published by
Katz [10]. Values, in decibels,
are with reference to full-scale
amplitude of the wav file format

123
376 Acoust Aust (2016) 44:369–378

Apart from the 4 kHz band, the ITA Toolbox and AARAE
implementations return results that are essentially at the
median. Although AARAE returns an outlier at 4 kHz, its
deviation in that band returns only 0.01 s from the median.
Table 1 provides a comparison between the results of the
present study and those of Katz. The median values are sim-
ilar, but the standard deviations and interquartile ranges in
the present study are much smaller. The interquartile range
excludes outliers, such as implementations C and A in the
lower bands. This reduced spread of results in the present
study could be from a combination of factors, including
improvements in software implementations, the selection of
software, the human contribution to analysis in Katz’s study
(which was not in the present study) and statistical issues
relating to sample size. The end-truncation method in the
current standard (ISO3382-1-2009) provides less flexibility
than in the version at the time of Katz’s study (ISO3382-
1997) [28], which may have contributed to greater agreement
between implementations.
To put the statistical dispersion of results in context, the
last row of Table 1 shows the greatest absolute value devi-
Fig. 5 The upper chart shows octave band reverberation time (T30 ) ation from the median in the present study, expressed as a
calculated from the impulse response published by Katz [10]. The lower percentage. In other words, this represents the furthest out-
chart shows the deviation from median (i.e. the calculated value minus lier (either above or below the median), and provides a simple
the median value) in each octave band
way of comparing with the 5 % acceptable deviation rule of
thumb. This is not to say that the median is correct—it is
merely used as a reference for comparison. The PNR in the
125 Hz band is almost 45 dB, and is greater than this in the
ing to steady noise that was used for the synthetic test of PNR, other bands (see Fig. 4), and so the signal-to-noise ratio is
the real impulse response has fluctuations in both the decay essentially that required for T30 calculation. Of the nine soft-
and the noise floor. Hence the differences exhibited here by ware implementations, five had deviations from the median
B, C and F may be due to the sensitivities of their particular greater than 5 % in the 125 Hz band, three in the 250 Hz band,
algorithm used to identify the truncation point, and perhaps and one in the 500 Hz band.
the way in which the late decay rate or the noise floor values In comparing dispersion of T30 with the PNR in each
are calculated. The lower frequency selectivity of B’s filter- octave band, PNR appears to be a plausible predictor of vari-
bank might also contribute to its divergence from C and F, ation between the implementations’ results. Figure 6 shows
which is supported by the similarity of its results with those this comparison for the present study and for Katz’s study,
of G (which has similarly low selectivity) in the lower two with dispersion expressed in relative terms, as interquartile
bands. ratio (interquartile range divided by median). This concurs

Table 1 Octave band median


125 (Hz) 250 (Hz) 500 (Hz) 1 (kHz) 2 (kHz) 4 (kHz)
reverberation time (T30 ) and
indicators of statistical Median 1.98 s 1.61 s 1.22 s 1.17 s 1.15 s 1.09 s
dispersion calculated from the
impulse response published by Median from Katz 1.92 s 1.62 s 1.22 s 1.18 s 1.16 s 1.12 s
Katz [10], with results from this STD 0.16 s 0.08 s 0.02 s 0.01 s 0.00 s 0.02 s
study and the values reported in STD from Katz 0.35 s 0.17 s 0.08 s 0.07 s 0.09 s –
Katz’s study
IQR 0.22 s 0.11 s 0.01 s 0.00 s 0.00 s 0.01 s
IQR from Katz 0.49 s 0.16 s 0.03 s 0.02 s 0.02 s –
Greatest deviation from median 20.1 % 7.6 % 5.5 % 2.6 % 0.6 % 3.4 %
Standard deviations (STD) and interquartile ranges (IQR) from both studies are shown, along with the
greatest absolute value deviation from median (expressed as a percentage) in results of the present study.
Katz did not report STD or IQR at 4 kHz, and 8 kHz was not included at all in his study. In the present study
not all software returned a result at 8 kHz, and so it is omitted

123
Acoust Aust (2016) 44:369–378 377

bands, and substantially reduced variability compared to a


software comparison conducted more than one decade prior
to this one. Nevertheless, the range of results still exceeded
±5 % from the median in the low frequency bands (includ-
ing bands for which the signal-to-noise ratio requirement
for T30 was met), and this indicates that improvements in
the software would be useful. Now that indirect methods
for measuring impulse responses are well developed and
widely implemented, it is usually possible to achieve high
PNR values in room acoustics measurements (with appro-
priate equipment and expertise), certainly higher than those
in a balloon pop recording. Considering that PNR is clearly
Fig. 6 Interquartile ratio of octave band reverberation time (T30 ) cal-
culated from the impulse response published by Katz [10], shown in one of the sources of error and deviation between software,
relation to PNR. Interquartile ratio is the interquartile range divided by current indirect methods should result in reduced variabil-
the median, and so is a measure of dispersion relative to the median. ity in reverberation time calculated from well-made impulse
Results are shown for the present study and for Katz’s study. PNR values responses.
were measured using a 12th-order octave band zero-phase filter-bank,
with the period from 2.5 s to the end of the waveform taken as the noise There is evidently some room for improvement, arguably
floor in all of the software implementations tested, including those
of the authors. Mostly these improvements would yield
minor increases in robustness, or reductions of errors that
with the pattern of results for the synthetic impulse responses are already well within acceptable margins. This paper has
tested in the present study, which also exhibited increased focussed especially on filter selectivity, and results support
dispersion at lower PNR (except that the synthetic impulse the suggestion by Venturi et al. [21] that highly selective
response’s dispersion of results is smaller, e.g. IQR of 0.01 s octave band filters improve results when reverberation time
for PNR of 45 dB). However, it must be borne in mind that for varies between adjacent bands. Implemented as reverse-time
Katz’s impulse response, PNR is correlated with octave band filters, a highly selective octave band filter-bank did not have
centre frequency, which can plausibly contribute to standard any unwanted side effects in the tests performed. That is not to
deviation of T30 in itself, due to the increased irregularity of say that a 24th-order octave band filter-bank is necessary—
the reverse-integrated decay at lower frequencies (this irreg- results indicate that a significantly lower order filter-bank
ularity was avoided in the synthetic impulse responses by will return negligible errors for a 2:1 reverberation time ratio
using decaying tones instead of decaying noise). Apart from between adjacent bands (with energy at the centre of each
the interquartile ratio result for the 4 kHz band (which is not band). Filter artefacts seem to be evident in the centre time
available from Katz’s paper), there is nothing to disentangle values returned by some of the software implementations
the influence of PNR from the influence of frequency in this tested, including the ones with relatively low frequency selec-
real impulse response. tivity. This suggests that there is room for improvement in
filter-bank design in some of the external software imple-
mentations tested, not only with regard to selectivity, but
4 Conclusions also with regard to time-response—and that improving one
will not necessarily damage the other.
The main conclusion of these tests is to confirm that the tested This study is complemented by other room acoustics mea-
software mostly provides good reverberation time results. surement software comparison studies that use more realistic
This is confirmed for noise-free impulse responses, includ- measured impulse response sets—for example full sets of
ing very short reverberation times—the results for noise-free measurements within an auditorium, allowing spatial statis-
impulse responses were essentially in perfect agreement with tics to be examined. Both types of studies are limited, but
the synthesised decay. Good performance is also confirmed together they provide guidance on software vulnerabilities
for when a steady noise floor is added, even at peak-to- and the degree of uncertainty.
noise ratio (PNR) values significantly lower than required for
standard analysis. With noise present, results deviated notice-
ably from the ideal, but were all well within the generally
accepted 5 % error margin at the required 45 dB signal-to- References
noise ratio for standard T30 analysis, and mostly within 1 %.
A further positive indicator is that the results for the real 1. Schroeder, M.: New method of measuring reverberation time. J.
recorded impulse response show strong agreement in most Acoust. Soc. Am. 37, 409–412 (1965)

123
378 Acoust Aust (2016) 44:369–378

2. Chu, W.: Comparison of reverberation measurements using 16. Jacobsen, F.: A note on acoustic decay measurements. J. Sound
Schroeder’s impulse method and decay-curve averaging method. Vib. 115, 163–170 (1987)
J. Acoust. Soc. Am. 63, 1444–1450 (1978) 17. Kob, M., Vorländer, M.: Band filters and short reverberation times.
3. Jacobsen, F., Rindel, J.: Time reversed decay measurements. J. Acta Acust. United Acust. 86, 350–357 (2000)
Sound Vib. 117, 87–190 (1987) 18. International Organization for Standardization: ISO 3382–1
4. Morgan, D.R.: A parametric error analysis of the backward inte- Acoustics-Measurement of Room Acoustic Parameters Part 1.
gration method for reverberation time estimation. J. Acoust. Soc. Geneva (2009)
Am. 101, 2686–2693 (1997) 19. International Electrotechnical Commission: IEC 61260 Ed. 1,
5. Hak, C., Wenmaekers, R., van Luxemburg, L.: Measuring room Electroacoustics-Octave-band and fractional-octave-band filters,
impulse responses: impact of the decay range on derived room International Electrotechnical Commission, Geneva (1995)
acoustic parameters. Acta Acust. United Acust. 98, 907–915 (2012) 20. International Electrotechnical Commission: IEC 61260-1 Ed. 1,
6. Guski, M., Vorländer, M.: Comparison of noise compensation Electroacoustics-Octave-band and fractional-octave-band filters-
methods for room acoustic impulse response evaluations. Acta Part 1: Specifications, International Electrotechnical Commission,
Acust. United Acust. 100, 320–327 (2014) Geneva (2014)
7. Lundeby, A., Vigran, T., Bietz, H., Vorländer, M.: Uncertainties of 21. Venturi, A., Farina, A., Tronchin, L.: On the effects of pre-
measurements in room acoustics. Acustica 81, 344–355 (1995) processing of impulse responses in the evaluation of acoustic
8. Bradley, J.S.: An international comparison of room acoustic mea- parameters on room acoustics. J. Acoust. Soc. Am. 133, 3224
surement systems. Institute for Research in Construction, IRC (2013). (POMA 19, 015006)
Internal Report IRC-IR-714 (1996) 22. Faiget, L., Legros, C., Ruiz, R.: Optimization of the impulse
9. Ikegame, M., Uchida, K.: Comparison of various types of acoustic response length: application to noisy and highly reverberant rooms.
indexes calculated from the same impulse response (in Japanese). J. Audio Eng. Soc. 46, 741–750 (1998)
In: Proceedings of the Symposium on Impulse Response Measure- 23. Xiang, N.: Evaluation of reverberation times using a nonlinear
ments and Evaluation of Acoustic Indexes, pp. 7–13, Architectural regression approach. J. Acoust. Soc. Am. 98, 2112–2121 (1995)
Institute of Japan (2001) 24. ITA-Toolbox: A MATLAB Toolbox for the Needs of Acousti-
10. Katz, B.F.G.: International round robin on room acoustical impulse cians. Institute of Technical Acoustics, RWTH Aachen University,
response analysis software 2004. Acoust. Res. Lett. Onl. 5, 158– Aachen (2013) http://www.ita-toolbox.org/
164 (2004) 25. Cabrera, D., Jimenez, D., Martens, W.L.: Audio and Acoustical
11. Tronchin, L., Tarabusi, V. and Bortilitti, A.: On the uncertainties of Response Analysis Environment (AARAE): a tool to support edu-
calculations of acoustical parameters in room acoustics measure- cation and research in acoustics. In: Proceedings of Internoise,
ments. In: Proceeding of 6th AMTA/WSEAS Conference, Buenos Melbourne (2015)
Aires (2005) 26. Cabrera, D., Lee, D., Leembruggen, G., Jimenez, D.: Increasing
12. Ţopa, M.D., Toma, N., Kirei, B.S., Homana, I., Neag, M., De Mey, robustness in the calculation of the speech transmission index from
G.: Comparison of different experimental methods for the assess- impulse responses. Build. Acoust. 21, 181–198 (2014)
ment of the room’s acoustics. Acoust. Phys. 57(2), 199–207 (2011) 27. International Electrotechnical Commission: IEC 50268-16 Ed. 4,
13. Machín, J., Arregui, A., San Martín, R., Arana, M.: Areto 1.0: Sound System Equipment-Part 16: objective Rating of Speech
Software para la caracterización acústica de recintos (Areto 1.0: Intelligibility by Speech Transmission Index. International Elec-
Software for the characterization of room acoustics). In: Tech- trotechnical Commission, Geneva (2011)
niacustica 2013: 44◦ Congresso Español de Acústica, Valladolid 28. International Organization for Standardization: ISO 3382
(2013) Acoustics-Measurement of the reverberation time of rooms with
14. Mansilla, J., Sato, S.I., Rodiño, L., Arias, A., Masson, F., De reference to other acoustical parameters. Geneva (1997)
Palma, I.C.: Comparison of room acoustic parameters calculated
by different types of software. In INTER-NOISE and NOISE-CON
Congress and Conference Proceedings. Institute of Noise Control
Engineering vol. 250(3), pp. 3333–3343 (2015)
15. Álvarez-Morales, L., Galindo, M., Girón, S., Zamarreño, T., Cib-
rián, R.M.: Acoustic characterisation by using different room
acoustics software tools: a comparative study. Acta Acust. United
Acust. 102 (2016) (in press)

123

You might also like