Soundscape and Psychoacoustics Using the resources for environmental noise protection Standards in Psychoacoustics Roland Sottek HEAD acoustics GmbH roland.sottek@head-acoustics.de Satellite symposium on August 25 and 26, 2016 1 Introduction The evaluation and design of noise is becoming increasingly important. Different kind of sources contribute to the noise: Broadband and narrowband noises, tonal components, modulated sounds. The perception is often predicted by psychoacoustic parameters in order to reduce time-consuming listening tests. 2 1
L/dB[SPL] L/dB[SPL] L/dB[SPL] Why do we need psychoacoustics? Time data and 3 rd octave spectrum of an electrical motor and a Mozart concert Almost identical levels and 3 rd octave spectra 3 Different time structures Almost identical 3 rd octave spectra and thus identical sound pressure levels and stationary loudness values (calculation from 3 rd octave spectra) Clearly audible difference also with respect to loudness because of different time structures 3 rd octave spectra Noise 20 100 f/hz2000 20k Impulse 20 100 f/hz2000 20k Diesel 20 100 f/hz2000 20k 80 60 50 40 80 60 50 40 80 60 50 40 4 2
Same sound pressure levels different loudness values 5 Modulated signals, tone complexes sound frequency / Hz modulationrate / Hz fluctuation roughness 1000 0 1000 1 1000 4 1000 20 1000 70 1000 & 2000 0 6 3
frequency Perception and physical measurement Optical illusion 7 Acoustical illusion ( signal-estimation ) time time pulsation thresholds 9 4
Acoustics vs. psychoacoustics (1) In acoustics the sound source is in the center of focus: What signals? What are the amplitudes? What vibrations / frequencies? What energy? Psychoacoustics provides the recipient (the people) in the center of focus : What loudness, sharpness, roughness, tonality, annoyance are perceived? What are the expectations, attitudes, experiences of those affected? 10 Acoustics vs. psychoacoustics (2) In acoustics, the entire sound event is often reduced to a simple variable in the form of A-weighted sound pressure level db (A). A-weighting considers in a highly simplified manner, the frequencydependent sensitivity of human hearing; low-frequency tones are perceived softer as high-frequency tones at the same sound pressure level. Psychoacoustics describes the auditory sensation of a human being as a complex function of the signal composition, the temporal patterns, the interaction of different frequencies. Here cognitive and contextual aspects are very important: attitude to noise, the information content and the cause of the noise. 11 5
What is psychoacoustics? Psychoacoustics deals with the sound perception of human hearing (sound recording, analysis in the inner ear, processing and analysis in the brain) and is engaged in addition also with the acoustically correct recording (e.g., with an artificial head measuring system) and the hearing-related analysis of sound events. Taking into account the cognitive aspects of noise psychoacoustics enables one to describe the transformation of a sound event in an auditory event. 12 6
Introduction to loudness 14 Definition of loudness Loudness is a psychological term used to describe the magnitude of an auditory sensation. * (Fletcher, Munson, 1933) Die Empfindungsgröße der zur Schallstärke gehörenden Intensitätsempfindung ist die Lautheit. ** (Zwicker, 1982) (The auditory sensation corresponding to the perceived sound intensity is loudness.) * FLETCHER, H. MUNSON, W.A. Loudness, its definition, measurement, and calculation. J. Acoust. Soc. Amer. 5, p. 82 (1933) ** ZWICKER, E. Psychoakustik. Hochschultext, Heidelberg, New York, Berlin, Springer Verlag, p. 79. (1982) Roland André Fiebig Sottek Internoise 2016, Hamburg Was & ist Berlin Lautheit? - Standards in Psychoacoustics 15 7
Influence of masking on loudness Sound Sound N 5 =10,6 sone GF N 5 =13,5 sone GF 16 Influence of duration on loudness (temporal integration) Tone pulse N max =8,5 sone Tone pulse N max =14,0 sone Tone pulse N max =16,4 sone 17 8
Factors influencing loudness Frequency Tones with the same sound pressure level but different frequency are not perceived as equal loud. Spectral composition Sounds with different spectral composition but the same sound pressure level are not perceived equally loud, e.g. broadband sounds are perceived louder than narrowband sounds at the same level. Sound pressure level Changes in sound pressure level do not lead to the same degree to loudness changes. Simultaneous masking At the same level, loudness varies by different masking effects in the spectral domain. Backward and forward masking The temporal structure is influencing the perceived loudness (test tone is presented before or after the masker). Signal duration Loudness sensation increases with signal duration up to a duration of 1 s. 18 Critical band concept Critical bandwidth can be considered as the bandwidth of auditory filters. Critical bandwidth is constant below 500 Hz and is about 20 % of the center frequency at higher frequencies. Third octave filters have a similar bandwidth, thus they can "be considered as a useful approach to auditory filters * The concept of critical bands is for many auditory sensations of particular importance. db(a) does not consider critical bands and masking! *FASTL, H. Psychoakustische Methoden, in: Kalivoda, M.T. und Steiner (Hrsg.). Taschenbuch der Angewandten Psychoakustik, Wien, New York, Springer (1998) 19 9
Loudness standards 20 Standardization of psychoacoustic parameters (1) Loudness evaluation has become a central focus for assuring better consideration of sound intensity phenomena than frequency-weighted levels like db(a). Different loudness standards available for stationary sounds: ISO 532:1975 section 1 (method A) [Stevens method] ISO 532:1975 section 2 (method B) [Zwicker method] ANSI S3.4:2007 [Moore/Glasberg method] DIN 45631:1991 [Zwicker method] Loudness standard available for time-varying sounds: DIN 45631/A1:2010 [Zwicker method] WG 9 of ISO TC43 (Acoustics) has worked on (available end of 2016) ISO 532-1 Methods for calculating loudness Part 1: Zwicker method for stationary and time-varying sound based on DIN and ISO 532-2 Methods for calculating loudness Part 2: Moore/Glasberg method for stationary sounds based on ANSI S3.4:2007. 21 10
History of loudness standardization (Zwicker method) DIN 45631: 1967 stationary loudness ISO 532B : 1975 stationary loudness DIN 45631: 1991 stationary loudness widely used! corrections to match the ISO equal loudness contours (ISO 226:1987) DIN 45631/A1: 2010 time-varying loudness widely used! ISO 532-1: 2016 stationary and time-varying loudness based on DIN 45631, but with test implementation (source code in appendix), detailed description from time signal to (specific) loudness vs. time function; nearly identical results can be obtained by ISO 532-1 and DIN 45631 or DIN 45631/A1 22 History of loudness standardization (Moore/Glasberg method) ANSI S3.4: 2007 stationary loudness ISO 532-2: 2016 stationary loudness, based on ANSI S3.4 The ISO 532-2 method does not fully describe obtaining a result from time-signals, only from described levels versus frequencies. User interaction is required: description of tones, noise bands,! Moore/Glasberg made the source code of their loudness model available as part of ISO 532-2. A time-varying loudness model is planned for a later update. 23 11
Standardization of psychoacoustic parameters (2) Sharpness standard (the weighted first moment of the critical-band rate distribution of specific loudness, only stationary signals): DIN 45692 (2009) Tonality standards: ECMA-74 (IT products), DIN 45681 Standardization of roughness is currently being pursued in a DIN working group. Following: Overview of loudness calculation procedures Sharpness calculation procedure Tonality calculation procedure Roughness model discussed as one option in the DIN working group Blind source separation 24 12
Loudness calculation procedures 26 Loudness calculation procedures band pass filter bank temporal effects and post-masking (DIN 45631/A1) 3 rd Oct. (DIN) LP N NL g 1 p(t) FF Bark (Zwicker) LP N NL g 2 DF ERB (ANSI) LP N NL g k + LP N(t)? LP N NL g K 27 13
Comparison: ANSI - DIN Choice of sound field Band pass filter bank Envelope formation ANSI S3.4-2007 DIN 45631 40 filters, constant bandwidth on ERB scale Free or diffuse 28 third-octave filters, approx. 24 Bark bands Rectifying and low-pass filtering Frequency weighting Strong attenuation Less attenuation Nonlinearity Square root law between sound pressure and loudness (highly simplified) Opposite effects! 28 Equal-loudness contours (60 phon) Equal-loudness contours (60 phon): IS0 226 new/old, DIN 45631, ANSI S3.4-2007 8 db ISO 226: 2003 ISO 226: 1987 DIN 45631 ANSI S3.4-2007 L/dB[SPL] 120 110 100 90 80 70 20 50 100 200 f/hz 1000 2000 5000 10k Tones are judged much softer by ANSI S3.4-2007 than DIN 45631! GENUIT, K., SOTTEK, R. AND FIEBIG, A., Comparison of Loudness Calculation Procedures in the Context of Different Practical Applications, Internoise 2009, Ottawa, Canada (2009). 29 60 50 40 14
Input signal Loudness of technical sounds sound source electric motor wind noise vehicle noise at constant speed power seat park (New York) electric screwdriver electric saw exhaust system DIN 45631 4.7 sonegf 25.4 sonegf 24.1 sonegf 4.6 sonegf 18.7 sonegf 30.6 sonegf 23.2 sonegf 28.8 sonegf ANSI S3.4-2007 6.8 sone 28.3 sone 25.7 sone 6.2 sone 23.8 sone 37.9 sone 29.8 sone 28.9 sone DIN 45631 and ANSI S3.4-2007 standards provide significantly different loudness values in many cases! 30 Signal processing scheme: time-variant loudness (ISO 532-1) A 25 Hz f LP 1 L T 1 Weighting (table A.1) A 3rd octave f LP i L T i Weighting (table A.1) L CB 1 Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Core loudness N C 1 (eq. A.1.1) NL Mapping to 24 critical bands (Bark) using table A.6 A 80 Hz f LP 6 L T 6 Weighting (table A.1) Start index z = 0.1 Bark. Step size = 0.1 Bark. Approximated core loudness, filter number: x = 1 A A 100 Hz 125 HZ f f LP 7 LP 8 L T 7 L T 8 Weighting (table A.1) Weighting (table A.1) L CB 2 Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Core loudness N C 2 (eq. A.1.1) NL Specific loudness = core loudness x Calculation of slope loudness using table A.7 slope loudness > specific loudness false true Step to next CB: x(z) A A A 160 HZ 200 Hz 250 Hz f f f LP 9 LP 10 LP 11 L T 9 L T 10 L T 11 Weighting (table A.1) Weighting (table A.1) Weighting (table A.1) L CB 3 Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Core loudness N C 3 (eq. A.1.1) NL true false Specific loudness = slope Specific loudness = core loudness loudness x true z < 24 z = z + step size false Summation of specific loudness A A 315 Hz 3rd octave f f LP 12 LP i L T 12 L T i L CB 4 L CB i Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Core loudness N C 4 (eq. A.1.1) Core loudness N C i (eq. A.1.1) NL NL t=3.5 ms LP 1 LP 2 0.47 0.53 t=70 ms A 12500 HZ f LP 28 L T 28 L CB 20 Level corrections a 0,DL DF, DL CB (tables A.2-A.5) Core loudness N C 20 (eq. A.1.1) NL Applicaton to ISO 532-1 method B Total loudness 31 15
ISO 532-1 Reducing uncertainties in loudness calculation of time-variant sounds In addition to the loudness standard for stationary sounds: Specification of the third-octave filter bank Rectification and intensity averaging Non-linear temporal decay of the hearing system Temporal weighting of total loudness 32 Temporal weighting of total loudness t=3.5 ms LP 1 LP 2 t=70 ms 0.47 0.53 Two 1 st order low-pass filters (time constants 3.5 ms and 70 ms) applied to sum of specific loudness values Simulates duration dependent behavior of loudness perception for short impulses: signal with duration of 10 ms is perceived as about half as loud as one with duration of 100 ms Total loudness is weighted sum (factors 0.47 and 0.53) of filtered signal 33 16
Representative value for time-variant loudness? N 5 percentile of loudness vs. time representation Considers peaks -> better than mean value Especially in the case of many events (cognitive effects) Even better: root mean cubed average of loudness vs. time function 39 Representative value for time-variant loudness? N 5 not suited for impulses: For this special case only: maximum value 40 17
Summary and conclusion (loudness calculation procedures) Application of ANSI S3.4-2007 and DIN 45631 standards to technical sounds provides significantly different loudness values 3 factors influence strongly the results of loudness models Frequency weighting (differences in equal-loudness-level contours according to ISO 226:1985, 2003; especially a low frequencies; new listening tests for normal equal-loudness-level contours needed!) Frequency scale (Bark, ERB) Nonlinearity between sound pressure and spec. loudness Presentation of an updated ISO 532 standard Complete description of each signal processing step starting from wave form to specific or total loudness vs. time functions Code example for implementation of all algorithms will be available Reduction of uncertainties in loudness calculation 41 18
Sharpness calculation procedure 43 Sharpness Ratio of high frequency level to overall level (basic description). Center of gravity, on frequency scale of spectral envelope: the higher the c.g., the sharper the sound. Integration of specific loudness multiplied by a weighting function, divided by total loudness (hence, sharpness is levelindependent). Normalized to a reference sound, a narrow band of noise centered at 1 khz at a level of 60 db and a bandwidth of 160 Hz, which has an agreed value of 1 acum. 44 19
Calculation of sharpness (DIN 45692) S k z 24Bark N'( z) g( z) z /Bark dz z 0 z 24Bark N'( z) dz z 0 acum 1 g( z) 0,42 z / Bark 15,8 0,15 e 0,85 z 15,8Bark for z 15,8Bark Solid line: weighting function according to DIN 45692 (see above) Dashed line: weighting function according to v. Bismarck Weighting function according to Aures depends on loudness N! 45 Comparison of sharpness methods: von Bismarck / Aures white noise Sharpness vs. Time white noise, 8 khz amplified Sharpness vs. Time 46 20
Influence of time structure on sharpness stream stream random 47 21
Tonality calculation procedure 49 Introduction to tonality Technical and natural sounds often contain prominent tonal components that can significantly influence the individual perception and evaluation of the sound event, increase annoyance. Tonality of sounds is increasingly important, even at very low levels regarding sound quality and sound design applications. Products may emit tonally-perceived noises due not only to pure tones but also to narrow noise bands, and to same-vicinity combinations of pure tones and narrow elevated noise bands. Established methods for tonality calculation such as (specific) prominence ratio and tone to noise ratio exhibit problems when compared to listening test data. 50 22
Available tools for assessing audible tonality Tone-to-Noise Ratio (TNR: ECMA-74; Information Technology main use). Prominence Ratio (PR: ECMA-74; Information Technology main use). DIN 45681 tonality (German standard, 2006) Psychoacoustic tonality ( Tonality : Aures/Terhardt) 51 Perception of tonality Recent research results show a strong correlation between tonality perception and the partial loudness of tonal sound components. HANSEN, H., VERHEY, JL. AND WEBER, R. The Magnitude of Tonal Content. A Review, Acta Acustica united with Acustica, Vol. 97, pp. 355-363 (2011). HANSEN, H. AND WEBER, R. Zum Verhältnis von Tonhaltigkeit und der partiellen Lautheit der tonalen Komponenten in Rauschen, Deutsche Jahrestagung für Akustik, DAGA (2010). VERHEY, JL. AND, STEFANOWICZ, S. Binaurale Tonhaltigkeit, Deutsche Jahrestagung für Akustik, DAGA (2011). KAMP, F. Modellierung der Wahrnehmung tonaler Geräuschkomponenten (Modeling the perception of tonal sound components), Master thesis (2012). 52 23
Pure tones (f = 1 khz) of different level in pink noise category scale 12 10 8 6 4 2 noise level = 60dB 12 10 8 6 4 2 noise level = 70dB 12 10 8 6 4 2 noise level = 80dB 0 35 40 45 55 65 75 tone level [db] 0 45 50 55 65 75 tone level [db] 0 55 60 65 75 tone level [db] The listening test has shown, that mean value tonality median value tonality mean value loudness of tonal components median value loudness of tonal components the perceived tonality and the loudness of the tonal content are strongly correlated (for high noise levels). 70 65 60 55 50 45 40 SOTTEK, Pegel des R., KAMP, Rauschsignals F. AND FIEBIG, [db] A. A new hearing model approach to tonality, Internoise 2013, Innsbruck, (2013). 53 Hearing model (Sottek, dissertation, 1993) Channel 1 Time Signal a 1 Channel i Auditory Sensation a i Channel n a n Specific Roughness Specific Loudness Specific Fluctuation Outer and Ear-related Filterbank Lowpass Nonlinearity Middle Ear Formation of Envelope Filtering 20 Bark 15 10 5 0 0 0.1 0.2 0.3 0.4 s Aurally-adequate Spectral Analysis of a Door Slam Noise SOTTEK, R. Modelle zur Signalverarbeitung im menschlichen Gehör, Dissertation, RWTH Aachen (1993). 54 24
Extended nonlinearity of the hearing model (Sottek) BIERBAUMS, T. UND SOTTEK, R. Modellierung der zeitvarianten Lautheit mit einem Gehörmodell, DAGA 12, Darmstadt (2012). EPSTEIN, M. AND FLORENTINE, M. A test of the Equal-Loudness-Ratio hypothesis using cross-modality matching functions, J. Acoust. Soc. Am., vol. 118(2), pp. 907 913 (2005). ZWICKER, E. Über psychologische und methodische Grundlagen der Lautheit, Acustica, vol. 8, pp. 237-258 (1958). 55 Double (half) loudness of a 1-kHz tone Level increment (solid line) and decrement (dashed line) DL of a 1 khz tone, required for double and half-loudness perception (Zwicker). Mean values of 12 subjects and interquartile ranges of the measurements are shown. Starting both parts of the experiment with the lowest level! 56 25
Two-dimensional pitch sensation Analysis of excitation pattern in (x, τ)-plane allows for modeling psychoacoustic phenomena (e. g. difference tones, residual components) LICKLIDER, J.C.R. A Duplex Theory of Pitch Perception (1951). 57 Hearing model (Sottek) (basis) time signal s(t) Example: door slam noise t outer and middle ear filtering auditory filter bank with i=1..n critical bands 1 i n one-way rectification attenuation autocorrelation function (ACF) more compressive non-linearity consideration of threshold in quiet 1 i n SOTTEK, R. Modelle zur Signalverarbeitung im menschlichen Gehör, Dissertation, RWTH Aachen (1993). 58 26
Autocorrelation function of periodic signals Pure tone, f = 1 khz White noise, f = 20 Hz 20 khz 5 x 10-4 4 E S E S E S E S E S 3 2 1 0-1 -2-3 -4 5 x 10-4 4 3 2 1 0-1 -2-3 -4 E S -5-2 -1.5-1 -0.5 0 0.5 1 1.5 2 t / ms -5-2 -1.5-1 -0.5 0 0.5 1 1.5 2 t / ms E s = φ E ss 0 = φ E ss n T E s = φ E ss 0 59 Delay time window autocorrelation function of a tone in pink noise 8 delay time window 4 0-4 -8 t 0 t Start t Ende 0 20 40 60 80 delay time t / ms E s = φ ss 0 E tonal = y (φ ss τ start f, τ end f E s E tonal detection of periodicities separation of periodic signals and noise 60 27
Frequency-modulated tone at 2 khz (f mod =2 Hz, mi=150, L=30 db) 61 28
Roughness calculation procedure 63 Extension to the third dimension New algorithm: (two dimensional: time-frequency) auditory filter bank is extended by modulation spectral analysis to a tensor frequency t 3 rd dimension: time structure time hearing model spectrogram frequency weighting weighting vs. t Tensor contains information about time, frequency and time structure (modulation rate). 64 29
Model calibration according to Zwicker and Fastl Roland Sottek Internoise 2016, Hamburg - A hearing model approach to roughness 66 Roughness of tones (f=1 khz, f mod =70 Hz, m=1) Roland Sottek Internoise 2016, Hamburg - A hearing model approach to roughness 67 30
Roughness of engine sounds R I Measurement Simulation Number of engine noise 68 Application of the hearing model roughness calculation Comparison of calculated roughness with experimental results roughness 1 0,8 0,6 a 0,4 0,2 0 1 2 3 4 5 6 7 8 Number of engine noise M1 M2 M3 M4 SGNB Attia & Okker (DAGA 95) Comparison of different models for the prediction of engine roughness Klemenz (Dissertation 05) Roughness compared to dissonance 69 31
Blind source separation 70 Investigations of methods for detecting acoustic patterns Spectral or temporal structures (patterns) are significant for the hearing impression. Complex sound events are constructed from several of these patterns. In listening tests it can be observed that in the evaluation of complex sound events often large variations occur, which point to an individual focusing on one of these patterns. For a better analysis and exploration of psychoacoustic phenomena, therefore, an automatic detection and separation of these patterns is sought by mathematical methods. 71 32
Blind source separation using spectrograms Spectrograms are level representations as a function of time and frequency. Better frequency resolution means worse time resolution, and vice versa, i.e., depending on the resolution temporal structures are recognizable and the representation in terms of frequencies can be smeared. The two-dimensional representation is thus often inadequate for the description and especially for the detection of patterns. The quality of the results depends greatly on the signal and the selected resolution. 72 Extension to the third dimension New algorithm: (two dimensional: time-frequency) auditory filter bank is extended by modulation spectral analysis to a tensor frequency t 3 rd dimension: time structure time hearing model spectrogram frequency weighting weighting vs. t Tensor contains information about time, frequency and time structure (modulation rate). 73 33
Separation of a mixture of modulated tones (example 1) original detected pattern 1 detected pattern 2 detected pattern 3 detected pattern 4 Good reconstruction of the modulated tones! detected pattern 5 74 Separation of a the patterns of a hard disc noise (example 2) original detected pattern 1 detected pattern 2 detected pattern 3 detected pattern 4 Different meaningful patterns can be detected. detected pattern 5 75 34
Conclusions and outlook Summary of psychoacoustic parameters Status of standardization Hearing model approach of Sottek Two-dimensional representations for the detection and separation of the components are often not sufficient. New method is based on three-dimensional modulation tensor. Additional information about temporal structure Based on a signal processing model of human perception First results indicate that the method can extract patterns in technical noise similar to the perceived patterns by humans. Validation shall be performed in listening tests. Outlook: possible application for soundscape projects 76 Thank you for your attention. Dr.-Ing. Roland Sottek Manager Research NVH Roland.Sottek@head-acoustics.de www.head-acoustics.de Copyright HEAD acoustics GmbH 77 35