Bernard F Berry, Berry Environmental Ltd. Nicole Porter, Independent Consultant.

Transcription

1 A review of methods for quantifying tonal and impulsive features in environmental noise Bernard F Berry, Berry Environmental Ltd. bernard@bel-acoustics.co.uk Nicole Porter, Independent Consultant. nicolebenton@btopenworld.com Abstract. The UK Government s Department of Trade and Industry, through its National Measurement System Directorate, has been funding a major research project on Environmental Noise. The aims were to improve the quality and face validity of environmental noise measurements, and to contribute to the improvement of British, and International Standards on environmental noise. The project had three related aspects, with a separate work package being organized for each aspect: A critical review and inter-comparison of methods of quantifying acoustic features in environmental noise, such as tones and impulses Work Package 1. Identifying and quantifying sources of uncertainty in environmental noise measurements. Work Package 2. Publishing practical guidance on minimizing such uncertainties - Work Package 3. The authors are part of the consortium of consultants undertaking the project and have been responsible for the first work package, on acoustic features. In a paper at Internoise 2004 in Prague [1], we outlined the background to the project, explained the approach and methods used in the review and summarized the status of the project. Progress in Work Packages 2 and 3 was reported in another paper at Internoise 2004 [2]. In the present paper we describe the later phases of our specific part of the project, including; a detailed review, and synthesis of a large number of previous research publications in which intercomparisons of the performance of methods of quantifying acoustic features have been documented a survey of the state of the art on practical implementations of various methods, in commercial instrumentation, and in systems used in research, some new small-scale intercomparisons involving a selected number of methods. our overall conclusions and recommendations.

2 1. INTRODUCTION- THE NEED FOR THIS REVIEW There exist a plethora of objective methods for detecting features in noise and/or their character correction, the latter being related to the supposed subjective annoyance likely to result from the feature. See Porter, Flindell and Berry [3]. Indeed, there has been no robust inter-comparison of the various objective methods, either in terms of their relative effectiveness or in terms of their ease of implementation. A previous UK Department of the Environment-funded project, carried out at the National Physical Laboratory, NPL, involved subjective listening tests on the judged annoyance of specific types of industrial noise. This explored the effect of two acoustic features - impulsivity and tonality - on subjective annoyance, and the performance of objective rating and assessment methods [4,5]. An outcome was the confirmation of the uncertainties relating to the identification and assessment of acoustic features. The conclusions recommended that future work should develop better descriptors for the identification and quantification of the physical magnitudes of various features. That work was concluded over 10 years ago, and there has been more development of the methods since. The ultimate aim of identifying features is to provide an accurate and comprehensive description of the noise that is meaningful in terms of the subjective characteristics it represents. However, there are not a great number of published examples where the results of subjective experiments have been used to validate methods of quantifying features. A better approach, which we have adopted, is to concentrate first on deriving objective methods to detect the presence of audible features, rather than attempting to determine a formal relationship between the level of a feature and increased subjective response. However during later phases of the project, we did review intercomparisons of methods which involved data in terms of annoyance, subjective penalties etc. 2. STAGES IN THE REVIEW The work was originally set up in the following main phases; Phase 1: Familiarisation Phase 2: Identify state of the art methods o Literature review o Consultation with key contacts Phase 3: Compare methods and make recommendations o Examine principles of each method o Evaluate extent of real world testing o Evaluate practicality of implementation o Objective comparison of methods 3. CURRENT STATUS Progress in the project, on Phase 1, Phase 2, and parts of Phase 3 were described in detail in our Internoise 2004 paper [1]. However, for completeness, we have summarised it here. We have conducted a systematic literature review, and an extensive structured consultation with key contacts, on an international basis. Arising from this, we identified for further study, 29 methods used to describe tonal features and 16 methods used to describe impulsive

3 features. Detailed tables have been assembled, organized in chronological order, listing the methods and the associated key references describing the principles of each method. Using each of the tables outlined above, a further set of tables has been developed in which each of these methods was described and documented in terms of its development and application, together with the basic principles of the feature classification method Each method has been examined carefully, and documented in terms of: the level of sophistication/technical complexity e.g. subjective versus objective, filter bandwidth analysis requirements, the extent to which the method represents the capabilities of the ear, and whether the method estimates the sensory level of the feature and/or the extent of an adverse response. We have examined the key stages in the evolution and development of the various methods, to show how one method may be linked to, or may have led to the development of another method, and how the methods have developed chronologically. The findings were summarized in two sets of flowcharts. This information, together with the information from earlier work stages, was used to categorize the methods into groups for analysis. For the tonal methods, the main method of categorization seems to fall most easily into groups based on the type of the analysis i.e. octave, 1/3-octave band and narrow band analysis (with the narrow band analysis being broken down further). For impulse methods, we formed 5 distinct groups; sound-level meter methods, time-series methods, spectral methods, loudness methods, and subjective methods. In this last group, the classification of a noise as impulsive is based on the judgment of the listener. From these main groupings, we selected the initial methods suitable for examination in the next stage of work. For the tonal methods, two current 1/3-octave band methods and a selection of narrow band methods were chosen. For impulse methods, we concentrated on time-series methods and spectral methods. The methods selected for subsequent examination were summarized in tables giving a brief description of the method, an explanation of the choice of descriptor, together with some brief information on the extent of real world testing and initial comments on the practicality of their implementation. We will now outline progress in the final stages of the project. 4. FINAL STAGES OF THE PROJECT 4.1 Evaluating the extent of real world testing We have reviewed previous research on the intercomparison or testing of tonal and impulsive classification methods and developed tables which sum up the main points of previous

4 research, either comparing methods and/or testing the objective methods against subjective response. The tables include details on the tester, methods tested, number of subjects and general comments on the tests and key outcomes. This previous research is described in considerable detail in reports which will be available on the DTI website. The main results are outlined below which include a summary of the shortcomings of a group of methods. Tonal character The most tested methods are within the group of methods based on tone-to-noise ratio, prominence ratio (TNR, PR and DIN). These methods have been tested on both artificial and real environmental test sounds with between 190 and 215 test subjects. Shortcomings of these methods have been related to the perceived tonality of two tone complexes, and the overestimation of low frequency tonal sounds. The DIN standard 2002 appears to perform better than the TNR and PR methods, thought to be due to the introduction of the masking index function in the DIN 2002 standard. Research findings also suggested that trackable non-stationary behaviour of frequency modulated tones leads to difficulties when using tonal metrics that are derived from estimated spectra. The methods are found not to be adequate for time varying tones for which an alternative model has been proposed, the Pitch Model, which is suggested to perform better for these types of sounds. The need for a frequency correction, modifications for the criteria of prominence, the handling of tonal harmonics for PR and TNR have been recognised and, in 2002, modifications have been proposed to overcome these. However, it is notable that despite much within group testing and external review, no single method has been recommended at this time for rating the prominence of tonal character of sounds. The Joint Nordic Method presents a method for assessing the audibility of tones in noise. It has been widely tested on both artificial and environmental sounds. It appears to be the most adopted method in Europe for assessing the audibility of tones in noise. In particular variations of this method have been used for wind turbine noise. It is included in the draft ISO standard ISO 1996 Description and measurement of environmental noise, Part 2, Annex C. It has identified shortcomings in relation to non-stationary (time varying) signals. Additionally, research findings also suggested that trackable non-stationary behaviour of frequency modulated tones leads to difficulties when using tonal metrics that are derived from estimated spectra. The ISVR method is still really in a research phase at present, and has only had very limited testing in the laboratory using artificial sounds. There are a number of uncertainties, including dealing with two or more tone complexes, non-stationary or non-broadband background noise and the effects of modulation. Although the method shows real promise, more development may be required before it can be recommended for implementation today. The PR, TNR, JNM and ISVR methods all relate to prominence of tones in noise rather than perceived tonality, such as the Aures method. The Aures method has been tested on over 20 artificial sounds. Although these experiments were specifically aimed at finding solutions to overcome recognised problems with effects of roll-off rates on the perception of the tonalness of bandpass filtered and how frequency modulation affects tonalness. There therefore appears to be some within group testing, but no inter-group testing. All the methods are being developed and refined, and this is an ongoing evolutionary process. This

5 work can only provide a snapshot at the time of writing. However, at this time there appears to be a main theme of refinement for all methods focussed around frequency corrections, time varying issues, and multiple tones. More research testing would be needed to overcome the identified shortcomings. Impulses Early work at NPL involved intercomparisons of a range of methods. These included several methods based on processing of the time-series of L Aeq [10ms]. The Increment method which arose from this NPL work was independently supported by the work of Brambilla in Rome. The ISVR Masked Threshold method, while conceptually interesting has limited testing in terms of the number and type of noises used. It should also be noted that the methods relies on calculation of the difference between FFT spectra for the impulse and the spectrum for a background noise. Whilst this is an important part of the research undertaken to develop the method it becomes a major shortcoming for more general application. It would be necessary to analyse both the impulse and the background noise, or for the analysis system to hold some kind of reference noise spectrum. Although the method shows real promise, more development may be required before it can be recommended for implementation today, and therefore this method was not considered in the remainder of this work. It can be seen in respect of the Prominence method that a reasonably large range of noises has been used in the testing, and that a range of candidate methods has been compared. However one possible limitation is that a small number of listening subject have been used. 4.2 Evaluating the practicality of implementation We have investigated commercial or research implementation of the methods. It was decided that we needed a systematic enquiry into whether any of the set of methods we were investigating were already fully implemented in existing commercial instrumentation. It was also felt necessary to ascertain if such instrumentation might already provide the basic frontend processing of acoustic signals, as a basis for implementing the methods by later postprocessing. We also noted that, during various stages in the course of the review of methods, information was being acquired, and more could be obtained, on the specific details of how certain methods had been implemented within the research studies being reviewed. We therefore documented the information obtained on Commercial and also on Research Implementation. Although information was available on the websites of instrumentation companies, and could be searched, a more direct enquiry was preferred. Essentially, having set out the context by summarising the project, and supplying a copy of the relevant Internoise 2004 paper, we asked recipients to look at the descriptions of the methods and advise us; if any of the methods are already implemented in current commercially available instrumentation, or if any instrumentation you are familiar with could provide the basic "front-end" processing, such as frequency analysis, or very short-term L Aeq, from which the methods could be implemented with additional processing.

6 We received four detailed replies. These were summarised in tables which presented the information gathered on commercial and research implementation, and indicated the extent of implementation for each of the selected methods, bringing together both the information on the commercial implementation, inquiries and the responses on research implementations. For tones, it seemed that examples of practical implementation are available for nearly all the listed methods. However, many of these are research implementations, noted from the various publications reviewed, rather than commercially available implementations. For impulses, full implementations are less available, in comparison to the situation for tones. For both tonal and impulsive methods, there are many cases where methods can be implemented by post-processing and current instruments are designed in such a way that post-processing can be used to implement almost any method. However such a situation means that there is a lack of standardisation in exact details of how methods are implemented, leading to greater variability in measured values. 4.3 Objective comparison of methods It became clear during the later stages of this project that the topic was in a constant state of evolution, and that we could provide only a snapshot in time. When we examined, in detail, the nature and scale of the research effort which had gone into the testing of various methods, it was clear that we could not expect, with our resources, to apply further meaningful testing, nor to refine any methods but merely document and comment on progress. There are several levels of intercomparison that could be envisaged. The most complex of these would require a wide range of noise samples to be evaluated on a subjective basis by a large panel of listeners and for the same noise samples to be assessed by various objective methods. It was felt that further work on intercomparisons should be focussed on a smallscale intercomparison to get some information on practicality and the most suitable purposes / applications. The final stage of the project therefore adopted this approach. We decided to choose noise samples, which had already been analysed by other researchers, both subjectively and objectively. We then aimed to carry out additional objective analysis of the same noises and to comment on the findings. Tones Although we intended to include a new intercomparison using the Joint Nordic Method for tonal noise using specifically tailored instrumentation options, this initially proved problematic due to the unavailability of suitable equipment at the time of the work. Due to time constraints, it was only possible to carry this work at a late stage. In terms of implementation, we have found that tonal analysis using instrumentation with specifically included tonal options may in practice be more difficult than one would expect, as one would have to have a dedicated instrument readily available. These instruments can be fairly costly and do not appear to be held by the average environmental officer or consultant. Post-processing may be the more common option available to all but is more time and labour intensive. Further conclusions may have been possible if we had managed to complete this analysis earlier.

7 For the limited analysis that was completed, we found that the Draft ISO Annex C method was not superior to two of the objective methods tested by Daniel et al, namely E DIN , ANSI S1.13, but was similar to (and even slightly better than) ANSI S1.13 PR, based on the regression fit. An example from these new analyses is shown in Figure 1 below. New Analysis, R=0.691 Subjective penalty (db) Objective penalty ISO 1996 [Nor-121] Figure 1: Relationship between Pompetzki s subjective data and the objective analysis of the tonal signal according to DIS ISO 1996 Annex C, R 2 =0.48 These data relate to the extent of a subjective response and not the detectability of the actual tonal feature itself. Therefore, although our work has suggested that for these noises this method may not predict the subjective response as well as for two of the other objective methods, we cannot comment here on its performance at detecting whether a feature is present or not. If more time and resources had been available, it may have been possible to compare the ability of each of these methods to predict the prominence of the tonal character. Data on the performance of methods at predicting the detectability of tonal features was included in earlier stages of the work and we found that no intercomparison of a number of methods had been performed. Principally the findings of these previous studies have led to the further refinement of the methods. Notably no single method has been recommended from these tests for detecting the prominence of tones. Impulses We established that the NPL Increment method, originally developed more than 10 years ago with special instrumentation, can be readily implemented today with a combination of modern instrumentation and standard PC calculation software. The Prominence method is considerably more complex to implement than Increment, but the comparison may not be entirely fair, as the Prominence method has been the subject of considerable development in recent years, with the associated requirement for detailed description, so that it could satisfy the requirements of the Nordtest system in Denmark, and also be considered for incorporation into ISO1996.

8 An example from our new analyses is given below in Figures 2 and 3. Details of the analyses will be available in Reports on the DTI website. Our analyses indicate that, where the subjective parameter is either Prominence, or the Subjective Penalty, on a scale of 0 to 5dB, the objective Prominence method gives higher correlations than the Increment method. SUBJECTIVE IMPULSE PENALTY PROMINENCE Figure 2. Subjective Impulse Penalty, on a scale of 0 to 5dB, against objective Prominence: R 2 = SUBJECTIVE IMPULSE PENALTY INCREM ENT Figure 3. Subjective Impulse Penalty and Increment R 2 = This reflects the fact, recently indicated by Torben Holm Pedersen of Delta, in a Personal Communication by , that the NPL Increment method formed the basis for development of the Prominence method, but the method was extended to include the use of the standardised F time-weighting, in order to provide a better psychoacoustic element than L Aeq 10ms. Towards the end of the project, we were awaiting the outcome of more recent work at Delta in Denmark, in which new software for their noise analysis system Noiselab, [see has been developed to include automatic classification of impulses by the Prominence method and has then be used by them to re-analyse the same set of signals used in previous studies. It would be interesting to see if this particular new implementation of the Prominence method produces significant numerical differences in the values associated with the noises discussed above. However it is unlikely that such differences would be sufficient to have any impact on the comparison of the objective values with subjective data and hence our conclusions on this topic.

9 5. SUMMARY AND RECOMMENDATIONS The project s overall aim is to improve the quality and face validity of environmental noise measurements in the UK. The project had three related aspects, with a separate work package being organized for each aspect. We have met the original aims of the project in critically reviewing and comparing methods of quantifying acoustic features in environmental noise. This topic is in constant state of evolution, particularly for tonal methods, and we are only able to provide a snapshot of the current state of the art. In particular, rapid developments in instrumentation mean that the availability of implementations of specific methods may have moved on from the time when our survey was conducted. We recommend that websites of relevant manufacturers be checked for detailed specifications of current instruments. We have noted that there were many cases where methods can be implemented by postprocessing and current instruments are designed in such a way that post-processing can be used to implement almost any method. We have drawn attention to the fact that there is a lack of standardisation in exact details of how methods are implemented, leading to greater variability in measured values. We recommend that users take note of this in interpreting any results. There are ongoing developments in some of the relevant Standards organisations, e.g. DIN and ISO. We recommend that users of specific methods covered in this review should note whether the Standards documents they are using match the versions referred to in this review and/or indeed the most up to date available. Users of methods covered in this review should also note the various shortcomings and limitations of methods which have been brought to light. We recommend that users make themselves aware of the shortcomings of a specific method and judge its suitability and reliability for their specific purpose, given its shortcomings/limitations. Because of constraints on resources, our own inter-comparisons, in which we implemented a number of tonal and impulse methods on noises for which there existed subjective and objective data, were necessarily limited in scale. They did however provide valuable new information, both on the issue of practicality of implementation, and on relative performance of methods. Taken together with the other information, we believe we have provided information to enable a potential user of methods to decide if a specific method is fit for the purpose, bearing in mind the context in which the method is to be applied, its complexity, ease of implementation. Therefore we further recommend that a method is selected on the basis of fit for purpose i.e. the most suited to the task in hand, rather than the selection of the most complex or accurate available. In view of the above limitations, and even with the very detailed review of previous research on inter-comparison or testing of methods, our review does not itself constitute a Consumer Report in which we can recommend specific methods over other candidate methods. However, from the considerable amount of information given in the individual reports, we believe that the interested reader, whether a user of methods, a researcher, or one involved in the process of standardisation in this field, will be able to make informed decisions.

10 We would also recommend that a number of loose ends in our review could be followed up, subject to additional resources being available. One example would be to take note of the ongoing work on the Impulse Prominence method at the Delta organisation in Denmark. This involves new software for their noise analysis system Noiselab, to enable automatic classification of impulses by the Prominence method. We recommend that full intercomparisons be made which test between group methods, this extending the within group testing. This would allow for improved information to facilitate the selection of a method best on fit for purpose. It was outside the scope of this work to fully explore implications to international standards such as ISO However, the published ISO1996-1, 2003 says that there is no suitable objective method for assessing impulses. This review confirms the usefulness of the Impulse Prominence approach. We therefore recommend that the development of future standards should take note of our findings. 6. ACKNOWLEDGEMENTS The authors gratefully acknowledge the contribution of the other members of the consortium involved in this DTI NMS project Dr Andrew Bullmore and Justin Adcock of Hoare Lea Acoustics, Dr Ian Flindell of Ian Flindell Associates, and Colin English of the English Cogger Partnership. We also wish to acknowledge the invaluable help provided by all those who responded to our requests for information during the consultation process, and those who contributed to the survey of commercial implementations of methods. We would particularly like to thank Torben Holm Pedersen of DELTA, Doug Manvell of B&K Denmark, Prof. Dr-Ing. Hugo Fastl of the Technical University of Munich, Ian Campbell of Campbell Associates, UK, Dr Peter Daniel previously of Cortex (now at B&K), Dr Giovanni Brambilla of the Istituto di Acustica IDAC in Rome, and Norman Bolton of NMSD, DTI. The financial support of the UK s Department of Trade and Industry is also acknowledged. 7. REFERENCES NOTE- Additional technical references will be found in detailed Reports on this project on the DTI website [1] B F Berry and N D Porter A critical review and inter-comparison of methods for quantifying tonal and impulsive features in environmental noise. Proceedings of Internoise Prague.Paper 389. CDROM. [2] A J Bullmore, I H Flindell, N D Porter and J Adcock Recent developments in the UK approach to environmental noise management. Proceedings of Internoise Prague. Paper 843. CDROM. [3] N D Porter, I H Flindell, and B F Berry An acoustic feature model for the assessment of environmental noise. Proc. Inst. Acoustics, 1993, 15, Part 3, [4] N D Porter and B F Berry Subjective effects and objective assessment of combined tonal and impulsive noise. Proc. 6th International Congress of Noise as a Public Health Problem, Nice, 1993, 2, [5] B. F. Berry and N.D. Porter, The evaluation of acoustic features in industrial noise. Proceedings of Internoise 94, Yokohama, Japan, 1994, pp