Bergen, Norway

BNAM 2010
May 10-12

Engineering principles and techniques in room acoustics prediction

Bengt-Inge Dalenbck
CATT, Mariagatan 16A, SE-41471 Gothenburg, Sweden, bid@catt.se

Room acoustic prediction software are now often considered as standard tools but basic engineering principles and techniques remain as important. These programs are tools and will not solve an acousticians problem. The acoustician still has to come up with a sound design idea, based on knowledge and experience, and the prediction program can then, within its limitations, offer an independent check of how well a given design may work. A good description of these software is "a qualified discussion partner". This paper will list some engineering techniques that sometimes seem forgotten once a prediction software is used. One example, that relates to that the standard measurement of absorption coefficients is discussed again, is that in some types of rooms it is not the absorption coefficients that are most crucial to estimate (once a basic estimate has been made) but the scattering coefficients. On the other hand there are also special cases where nearly everything hangs on an accurate estimate of the absorption coefficients. Further, there is an increasing need to emphasize and teach the limitations of geometrical acoustics when these software have become common since many times the expectations have become unrealistically high. One misunderstanding is that if a sufficiently detailed model is made, and if a sufficiently high number of rays are used, the results will always be accurate. Computers and algorithms may get faster, and more complex models can be handled, but the limitations of geometrical acoustics stay the same and including a few wave-related phenomena will have limited impact on the overall results. Examples will be shown of how a too detailed model will lead to bad results. Unfortunately, marketing statements such as "highly accurate" can be found also regarding geometrical acoustics-based software and that further increases the risk for this misunderstanding, especially for new users. Software like this can, for many reasons, never be "highly accurate" but applied correctly they can be sufficiently accurate to be very useful tools in room acoustics. The accuracy of room acoustic measurements is naturally also a part of the discussion, since they are used to estimate the prediction accuracy, and common measurement and evaluation problems will be exemplified.

Introduction

Geometrical acoustics (GA) based software were 15-20 years ago almost exclusively used for predicting room acoustics in concert halls, operas, and auditoria while sound system prediction was almost exclusively based on direct sound coverage, possibly with a classical approach based on a Sabine RT to estimate some intelligibility measures. Since then the two types of prediction fields have essentially merged and concert halls, operas and auditoria now represent only a small part of what these software are used for and predictions e.g. for big voice alarm systems have become common. Many types of models are now extremely challenging, many times probably too challenging for GA. Examples are very big and flat models with pillars, airport halls with many pillars and coupled side volumes, multi-room, multi-floor, or both. In concert halls and auditoria the complexity is mainly concentrated on the walls and in the ceiling and has been possible to simulate well by a simplified geometry and frequency dependent scattering coefficients but with many pillars or doorways in between the source and the receiver diffraction effects will be much more important. At the same time these types of projects seldom have good input data available so it is uncertain if a wave-related treatment would be meaningful or even feasible due to the big sizes. Unfortunately, the limitations of GA are not always recognized by users and seem to be less and less emphasized by software developers.

Geometrical acoustics

2.1 Basics
Once upon a time it was taught that GA was applicable if the wavelength was much smaller than the smallest dimension of a surface (d) i.e. << d, which soon changed to < d in practical modelling and now seems to have arrived at >> d since even the 63 Hz octave is claimed to be predicted using GA methods ( = 7.7 m at the lower band limit). It is interesting to go back to a central reference in room acoustics [1] and chapter IV of the 1973 edition, The Limiting Case of Very High Sound Frequencies: Geometrical Room Acoustics. The latest 2009 edition uses similar formulations and adds that the limiting case is around 1 kHz ( 0.34 m) but below that it is useful qualitatively. The increase in computer speed and improvements in CAD modelling tools allows handling of more and more complex and detailed models but unfortunately the limitations of GA stay right where they are. Figure 1 shows schematic curves indicating how the ability to handle complex models has constantly grown while the limits of GA in principle were reached around the time of the 1st international round-robin [2], and where a few wave-related later additions have had limited impact on the overall accuracy.
model complexity/detail

accuracy using GA

1995 (1st round-robin)

time

Figure 1: Schematic curve showing how GA method results have not improved with model complexity. Frequency dependent scattering (FDS) can be used to allow prediction for a wider frequency range but that assumes that the model is simplified and details are replaced by an FDS coefficient, if a model is made with too small details it can only work well at high frequencies for where << d actually holds. After frequent and constant contact with software users during more than 15 years it does seem that the limitations of GA are now not as emphasized in education as they used to and should be. Cases are frequently seen where there would be an extremely slim chance to achieve meaningful predictions. One example is where the source is placed in a far corner on the 1st floor of a five-floor library and the receiver is placed in a far corner on the 5th floor. The only communication between the source and the receiver two is via a central staircase and around many corners and bookshelves. Only a few rays, if any, will ever reach the 5th floor receiver, any sound reaching that position will have to be via multiple diffraction. It has been claimed that regarding modelling one either i) omit the details and model a flat surface with an FDS coefficient, or ii) model the details and they will give the diffusing effect of that surface. Within the limitations of GA the i) approach can be successful but the ii) approach will only work for very high frequencies. Figure 2 shows an example where ii) fails completely. It indicates direct sound, 1st and 2nd order reflection paths where a diffuser is placed on one wall at the location where a 1st order specular reflection would occur. As can be seen from d) when using the actual diffuser geometry as in a) there is no specular reflection at all from the diffuser when modelled, there should have been for the detailed model to act as a diffuser, while in e) where the diffuser is modelled flat with an FDS coefficient the principle behaviour of the diffuser is achieved, i.e. an attenuated specular reflection followed by a diffuse tail. Predicted details in d) differ from the measured case in e) but a big reason for the visual difference is the use of a dodecahedron "omni" source. This will be shown in section 4 about measurement comparisons. One solution to this problem, within GA, would be to create say three models, one simplified for low frequencies, one less simplified for mid frequencies and a more detailed one for high frequencies. This approach has been discussed many times but is impractical and even if automatic gradual "geometrical low-pass filtering" of the room geometry is possible it is likely to be risky and require human intervention.

a)

b)

c)

diffuse tail

d)

e)

f)

Figure 2: A room with a diffuser model in two ways vs. measured, all graphs show wideband 10log10(h2) where h is the impulse response (IR). a) modelled detailed showing up to 2nd order specular reflection paths, b) as a) but modelled flat with an FDS coefficient, c) room photo (here showing a box rather than the diffuser, one of the test cases), d) predicted IR when modelled detailed, e) predicted IR when modelled flat with an FDS coefficient, f) measured with the diffuser installed. Picture and measurement courtesy SynAudCon. Another solution is to allow parts of the geometry to have three versions (similar to a level of detail, LOD, algorithm in computer graphics) but while elegant it is probably easier to simply make three models. Both these approaches require prediction algorithms with true frequency independence between bands so that the three models can be run independently. Such algorithms exist since long so the procedure can be tested. What is a detail and is a detail always a detail? This has much to do with the room size and shape. If a table in a corner of a big concert hall is omitted it will not affect the results, but if the tables in a classroom are omitted it will lead to the wrong results since there they have a clear mixing and obscuring effect. The same with classroom chairs but since they are smaller in relation to the room, and if they have no absorbing seats or backs, they are often better modelled by an increased FDS coefficient for the tables.

2.2 Algorithm properties

During many years software, commercial as well as in research, did not include FDS [3] but today it is finally found in just about all commonly used software. The first round-robin in 1995 [2] indicated the need for handling diffuse reflection but since it only covered the 1 kHz band it was not seen if scattering was implemented with frequency dependence. However, even when FDS is implemented it may be done so in several ways, algorithms and scattering functions may differ. All application of FDS has to be based on the physics of the case at hand (surface detail in relation to wavelength) but sometimes FDS coefficients seem to be treated as if they were magic numbers that are there only to fix a problem in GA prediction but they are directly related to reality, even if it is to a simplified version of reality. One clear example of a not physically reasonable application was when all scattering was assigned to one arbitrary surface in a model and nothing to the rest of the surfaces. No reference will be given to this, and it was 10 years ago. Suffice to say that it was in a peer-reviewed journal and the reviewer did not object to that procedure indicating that the magic number assumption was common. Do the various FDS functions applied always behave in a reasonably physical way? The ubiquitous Lambert function has been accused of not being appropriate for acoustics but there is about 50 years experience from its use [4] and it can be implemented so that it mimics the behaviour of actual diffusing surfaces, in principle if not in detail, as can be seen from Figure 2e and more clearly in Figure 4. Other scattering functions are in use that instead of attenuating the specular part and cause a diffuse reflection moves the specular sector, has low back-scattering, and thus gives a forward directional bias. To the authors knowledge the first example of a function like this can be seen in [5] and it can lead to unexpected results vs. the mean free path and prediction of the RT. For low and high values of the scattering coefficient there is not much difference from Lambert, but in the range in between (where most practical cases end up) the differences can be big. There may well be dedicated diffusers designed to move the specular sector, and where a function like this may be applicable, but if applied for general FDS RT predictions may suffer. Apart from the FDS function itself, in ray-tracing-like algorithms scattering can be applied in two ways. Either via random scattering where the ray exit angle is determined by the scattering function, or deterministically by spawning

many new rays where the strength of each diffuse ray is determined by the scattering function. Random scattering has the benefit of being fast and simple to implement and has been in use for more than 40 years together with the Lambert function [6] but will lead to some run-to-run variation and will not predict flutter echoes well. A deterministic diffuse ray split-up method will have no such random effects and can also give the diffuse tail after arrival of an attenuated specular reflection found with real diffusers and can predict flutter echoes. However, a brute force split-up algorithm will lead to an enormous calculation time since every ray is again split up at the next reflection etc. For this reason it is most often used for early reflections but there are ways to implement it without the long calculation times also for fulllength echograms. An early example is seen in [7] and a new more general implementation can be found in the TUCT software used for the predictions in this paper. A different technique with similar properties is the Room Acoustics Rendering Equation [8]. An interesting question is if a very high degree of repeatability (i.e. avoiding random scattering) always is beneficial, may it not give a false sense of accuracy? If results are without random variation it does not mean that they are more correct. A common example of the difference is of two rifles where one has a random spread, but hits the target most of the time and is centred on the target, and the other rifle instead hits all its shots very close but all of them are beside the target. Deterministic results may result in a focus on too small details that can never be accurately predicted or even controlled in practice, and if the overall method (GA with a few extensions) is anyway an approximation and input data is uncertain, why demand a very high degree of repeatability? The run to run variation actually helps to concentrate on the broader picture. A good example of the opposite is a specular-only algorithm that is 100% repeatable but that does not predict well in most practical cases.

2.3 Algorithm benchmarks

A good way to check prediction software in general, as well as their implementation of FDS, is to run benchmark "end" cases where the results can be known in advance assuming reasonably shaped single-volume rooms, and see if an algorithm actually predicts those results: in a model with 100% scattering on all surfaces, will the algorithm predict the Eyring RT? Will it, as it ought to, do so independently of the room shape and how the absorption is distributed? will the algorithm predict a mean free path of <l> = 4V/S ? <l> should not be dependent on the scattering functions (if they are uniform) or if there is any scattering at all since <l> is a room shape property, but if the scattering function has a directional bias it can happen that it will not predict the expected <l>.

in a room with no explicit scattering (and a non-mixing geometry such as a rectangular shape), will the algorithm predict a T30 much longer than the Eyring RT if the absorption is unevenly distributed or the width, height and length dimensions differ considerably? If not, the algorithm has a hidden property that gives somewhat diffuse reflections even if the scattering coefficients are set to zero.

2.4 Adding some wave effects

It may be tempting to try to get away from the limits of GA by adding a limited number of wave effects to an otherwise energy-based GA model. One example of a wave-effect is diffraction which is likely to improve prediction in special cases such as office screens and for very early reflections if the model is detailed and accurate (which is a problem since GA works best without a detailed model which again leads to more than one model) and if diffraction algorithms exist for the cases at hand, e.g. soft edges appears to have no practical solution similar to the BTM [9] for hard edges. It has been seen that diffraction algorithms like [10], for infinite length barriers, has been implemented for use with limited sized surfaces. Is that an improvement? Another example is to estimate the reflection phase, which often has to be done crudely since all that is know in the absorption coefficient in 1/3- or 1/1-octave bands. It can be useful to think about if these advanced-looking additions actually have a significant positive effect vs. accuracy. For example, consider if crude reflection phase is included under some assumptions (such as that surfaces are locally reacting) and calculated from the random incidence absorption coefficients, but one surface is either so small that diffraction would really be needed or that the phase simply can't be well determined. What point is there then to include reflection phase at all? Figure 3 illustrates the summation with phase for three consecutive reflections where the phase of just one is unknown and the phasor angle can be anything between 0 and 2, and how the resulting magnitude with phase summing will vary.

a)

Imag

b)

Imag

c)

Imag

Real

Real

Real

Figure 3: An illustration of how complex vectors may add up if not all phase contributions are known. For simplicity in drawing, only the 3rd reflection has an unknown phase (dashed arrow) resulting in a different magnitude and phase of the sum phasor (thick arrow) in a), b) and c). In auralization using GA methods, phase has to be added to predicted 1/n-octave band reflections so that an impulse responses can be synthesized, and that is one of the fundamental limitations of auralization using GA since below 500 Hz, it will affect the energy and give different energy measures than those originally predicted. i.e. what you predict is not exactly what you will hear. This is a known limitation common to all GA-based auralization software, and has to be considered, but it is different from when adding just a bit of reflection phase and claiming that it improves the overall prediction. Another way to illustrate how it may work with a varying level of accuracy is to see the various parts as numbers in a sum or product. Let us assume that there are five numbers corresponding to input data or phenomena (it could be absorption, diffusion, phase, diffraction, input data uncertainty etc.) each with a varying degree of accuracy or input data uncertainty, here indicated by a varying number of decimals: sum of effects: 1 + 2 + 1.1234 + 3.2 + 3.2 = 10.5234 If that is correct, it looks very accurate but let us assume that all values would take on their possible end values with four decimals: min sum: max sum: 0.5551 + 1.5551 + 1.1234 + 3.1551 + 3.1551 = 9.5438 1.4999 + 2.4999 + 1.1234 + 3.2499 + 3.2499 = 11.6230

showing that the sum can only be said to be somewhere between 9.5438 and 11.6230. Now let us instead sum the rounded values (corresponding to applying a method with a more even accuracy): sum of rounded: 1 + 2 + 1 + 3 + 3 = 10 which ends up in the middle of the 9.5438 to 11.6230 span. This was only a simplistic game of numbers to illustrate a point but it is important to remember this old proverb, here with an extra twist: A chain is no stronger than its weakest link; but if you show how admirably the last few are united, half the world will forget to test the security of the parts which are kept out of sight. [11] Modes is another interesting topic since adding phase to reflections calculated from GA will create modes, unless a random phase is used which may give no modes at all, another type of error. A Fourier transform of the synthesized IR will show peaks and dips at low frequencies that will look like modes, the problem is that they are not the correct modes, that can happen with GA only in a few cases like rectangular rooms with hard boundaries and no objects.

Input data

Independent of the prediction method practical prediction suffers from unknown, inaccurate or incomplete input data, often it even has to be estimated. However, the GA prediction methods are actually at their best when applied at an early stage when not all details are known, i.e. to investigate main effects of room size and shape and the spatial distribution of absorbing and diffusing surfaces, all of which will be harder or impossible to change at a later stage. It has happened that users spend days searching for an as accurate absorption coefficient as possible for a certain material not realizing that no matter the values used it would have had a very limited effect on the RT, simply because the relative surface area of that material was small, but it is really crucial if big areas use the same material. An example of the difference is when a small patch of concrete is found in a hall, it does no matter if a coefficient of 0.01 or 0.02 (or even 0.05) is used but if the room is all concrete (e.g. a reverberation chamber), the values chosen are critical since going from 0.01 to 0.02 causes a factor of two difference in reverberation time (at frequencies where the air absorption does not dominate).

It is safe to say that in most practical cases the input data, such as absorption coefficients and even more so scattering coefficients, is uncertain. How can that be handled, should a long time be spent searching for (perhaps) accurate values? No, typically not, it is then time to bring out an old engineering technique when data is uncertain, i.e. to make several calculations using high/low input data combinations. If a model geometry is mixing, or has many dedicated diffusing surfaces modelled with FDS, then calculating two cases using the highest and lowest reasonable absorption coefficients is sufficient, but if also scattering coefficients are crucial and uncertain (and they often are) then four cases are required (min/max absorption x min/max scattering) and from that measures such as T30 can be bound within reasonable limits. It must be stressed that problems like these are not due to the prediction software but are purely input data related.

Measurement comparisons

GA-based prediction can never be very accurate, but differences to measurements are often blamed only on the prediction software but the fact is that measurements often show big deviations between various measurement hard/software. A sensitive case is T30 estimates with background noise, examples can be found in a round-robin [12], but it is not only in such extreme background noise cases that there can be a problem. Some measures, notably EDT, suffers from sensitivity regarding the direct sound arrival time estimate and also regarding the shape of the 0 to -10 dB decay where a linear regression is used according to ISO 3382 [13] in spite of that the early decay often is far from linear, especially close to a source or if a source is directive. A measure with ever increasing importance is STI and with the added simplified STIPa it is clear that the results so far are not consistent depending on which actual STIPa meter is used [14], the complex nature of STI together with standard changes also makes it hard to know if comparisons are made with exactly the same STI calculation procedure. A further problem that has been seen in especially T30 comparisons is that the echogram in prediction has been too short. If an actual T30 of say 2 s is to be measured or predicted an IR of at least a 1.5 s has to be used [13], preferably 2.0 s. It is beneficial to see a prediction as a virtual measurement and while few would try to measure a 2 s T30 using an IR length of only say 1 s it can frequently bee seen in prediction. A perhaps unexpected problem is the use of an source omni-directional (omni) source. An physical source such as a dodecahedron (dodec) is typically far from omni-directional from 2 kHz and up, the variation over angle is of the order of +/-5 dB, which is within the allowed deviation according to [13]. If the microphone is far from the source it has little impact but close up a -5 to +5 dB difference will affect most measures strongly depending on the rotation of the dodec as compared to when modelling using an ideal omni source. In the measurement in Figure 2f at 8 kHz (not shown) the first reflection was actually stronger than the direct sound. To illustrate the uneven directivity Figure 4c shows a simulated 4 kHz directivity balloon and further related information can be found in [15,16]. An additional issue with a dodec is that its impulse response often is quite messy and long. Figure 4 shows a comparison between predicted/measured from Figure 2e (flat diffuser model with an FDS coefficient) but now with the IR predicted with an ideal omni having been convolved with the on-axis IR of the dodec (the gap between 6 and 7.5 ms is because the dodec impulse response could only be estimated up to when the first reflection occurred and is therefore shorter than the actual).

a)

c)

b)

Figure 4: a) predicted IR when modelled flat with a FDS coefficient and convolved with the dodec IR, b) measured with the diffuser installed (same as Figure 2f), c) a modelled dodec in the 4 kHz octave-band.

Summary

Do we really need very accurate prediction methods? Can we ever get the input data to utilize them? Many times these software are best used in the early stages of a project when room shape and major distribution of absorbing and diffusing surfaces are decided. To perform very detailed calculations at that stage may result in a too myopic view (the proverbial not seeing the woods for the trees). It is especially dangerous if users of these software think they actually can be very accurate. To exemplify that developers do not always indicate the limitations of GA, a number of current accuracy-related marketing statements are here given without further comment or reference (they can all be found on various web sites): have developed over the last decades an highly accurate prediction , fast and accurate, the method also yields accurate results..., ...has been developed, which allows a highly accurate and realistic ..., allows calculations of acoustic parameters with exceptional accuracy and speed, making X the most accurate, fastest and easiest to use program, one method is utilized for highly accurate auralisations, they help designers learn and grow by graphically displaying accurate predictions of real-world acoustics, over the last years Y has become one of the most powerful and accurate acoustic prediction programs in the world. GA based programs can not be highly accurate, but they can be sufficiently accurate to be very useful tools in consulting. They can give say 80% of the answer with a very reasonable amount of work using readily available, or even estimated, input data. The alternative is to use very much more complex wave-based methods such as FEM where say 90 or 95% of the answer may be found (never 100%) but with a very much bigger effort due to the complex input data required (wall impedances for locally reacting surfaces but to catch the true behaviour of vibrating walls or resonant absorbers that is not sufficient). An important distinction to make is between how advanced an algorithm is, mathematically and/or geometrically, and its potential for prediction. Again a good example is that some very advanced algorithms have been developed for specular reflections but that does not mean that the prediction results using only specular reflections are good. It can be argued that, for consulting use, it is typically more useful with an even accuracy than if some parts are calculated with high claimed accuracy while other parts are more or less neglected.

References
[1] H. Kutttruff, Room acoustics, Applied Science Publishers Ltd, London, 1973. [2] M. Vorlnder, International Round-robin on Room Acoustical Computer Simulation, Proc. 15th ICA, 689-692, Trondheim, 1995. [3] B.-I. Dalenbck, The Rediscovery of Diffuse Reflection in Room Acoustics Prediction, ASA Cancoun, 2002. [4] J. Allred and A. Newhouse, Applications of the Monte Carlo Method to Architectural Acoustics, J. Acoust. Soc. Am, 30(1), 1958, 1-2. [5] M. van der Voorden, L. Nijs1, G. Vermeir and G. Jansens, The Applicability of Ray-tracing Based Simulation Tools to Predict Sound Pressure Levels and Reverberation Times in Coupled Spaces. IBPSA Conference, 1999. [6] M. R. Schroeder, Digital Simulation of Sound Transmission in Reverberant Spaces, J. Acoust. Soc. Am, 46(2), 1970, 424-31. [7] B.-I. Dalenbck, Room Acoustic Prediction Based on a Unified Treatment of Diffuse and Specular reflection, J. Acoust. Soc. Am, 100(2), 1996, 899-909. [8] S. Siltanen, T. Lokki, S. Kiminki and L. Savioja, The Room Acoustics Rendering Equation, J. Acoust. Soc. Am 122(3), 2007, 1624-35. [9] U. P. Svensson, R. I. Fred and J. Vanderkooy, An Analytic Secondary Source Model of Edge Diffraction Impulse Responses, J. Acoust. Soc. Am. 106, 1999, 2331-2344. [10] U. J. Kurze and G. S. Anderson, Sound Attenuation by Barriers, Applied Acoustics, 4, 1971, 35-53. [11] L. Stephen, Cornhill Mag. XVII., 1868, 295. [12] B. F. G. Katz, International Round Robin on Room Acoustical Impulse Response Analysis Software 2004, Acoustics Research Letters Online, 2004.

[13] Acoustics Measurements of the reverberation time of rooms with reference to other parameters, International Standard ISO 3382:1997 [14] P. Mapp, New Techniques for Measuring Speech Privacy & Efficiency of Sound Masking Systems, 122nd AES Convention, paper 7126, Vienna, Austria, 2007. [15] R. San Martn, I.B. Witew, M. Arana and M. Vorlnder, Influence of the Source Orientation on the Measurement of Acoustic Parameters, Acta Acustica united with Acustica, 93(3), 2007, 387-397. [16] P. Dietrich, Modeling Measurement Uncertainty in Room Acoustics, Acta Polytechnica, 47(45), 2007, 63-67.