Home

loudness compensation of music in a car audio system

1. Chop suey 2 00 2 30 Hard Rock 7 Beethoven 5 symphony 0 00 0 30 Classical 8 Trentem ller Snowflake 2 41 3 12 Electronic TABLE 5 3 CHOSEN SOUND SOURCES FOR PROGRAM MATERIAL The first period is pink noise which is intended for level adjustments It s allows us to reproduce the levels in different measurements using a SPL meter The silence is necessary for noise floor recording 9 1 3 Noise measurements in car The other sound sources are different kind of music and speech The Pavarotti and Beethoven sounds sources are highly dynamic compared to the Coldplay and System of a down sounds sources which does almost have no dynamic And Trentemgller is a sound source with huge information in the lowest frequencies Each part of the program material has a length of about 30 seconds and will have a fade in and fade out of 1 second They are individually normalized using DVD Codes Matlab codes Loudness normalizing for wave files Main m based on recommendation UIT R BS 1770 2 This recommendation is based on LKFS Loudness K weighted relative to nominal full scale The program material is normalized to 24dB LKFS which gives us headroom and possibility to gain frequencies if needed in e g the loudness compensation system The sound sources were put together with the software Adobe Audition CS5 one after each other and exported to one mono 16bit file DVD Program material Car project mixdown MONO wav This allows us to play
2. Frame no 1 fps FIGURE 6 13 NOISE ESTIMATION FOR EACH SLICE IN OCTAVE BANDS SLICE SIZE 1S SIMULATION IS GAINED BY 1 5dB Although the differences become smaller some band estimations are too far away from the noise floor average By looking at individual slice bar graph e g Figure 6 11 we saw that when the noise estimation is bigger than the simulation of the material the estimation is very close to the noise floor We concluded that the estimation for an octave band cannot be trusted when it is smaller than the simulated playback signal SPL value for the same octave band the estimation is bigger than the real value following the playback signal content and went on analyzing how well the estimation performs in the presence of a more powerful masker The following analysis will only take into account noise estimations higher than the simulation for each individual octave band The analysis was done with 1 second slice in the front position of the microphone and with the transfer function gain of 15 35 dB without the noise floor subtracted when no engine was running 6 2 4 1 O0 KM H RECORDING ENGINE RUNNING Implementation Noise estimation um o ol N o 31 0Hz g 63 0 Hz Frame no 1 fps 125 0 Hz 250 0Hz 1 1 1 at 1 1 I ms WON PWON 100 Frame no 1 fps T 1 1 FIGURE 6 14 ESTIMATION OF THE NOISE 6 2 F
3. Extract Noise Estimate Tevels dB SPL Transfer function from rec position raw signal to recording Octave Band Raw signal PA Position Pa Filters levels dB SPL time convolution rec position FIGURE 6 4 NOISE ESTIMATION BLOCK DIAGRAM 6 2 1 TESTING FOR RELIABILITY To trust the method described above we needed to test that the simulation of the program material was close enough to a recording under the same conditions Thus we compared the simulation of the program material in the front position of the microphone with the recording of the same playback signal played and recorded inside the car in the same position with no engine running The comparison was done each second the slice length was 44100 samples long and the signals were adjusted as mentioned in 5 5 Car transfer functions The signals were firstly analyzed without the subtraction of the noise floor RMS value in the transfer function gain a gain of 18 75 dB was computed The error graph was plotted for each second for each band calculated as abs Level_recording Level_simulation Seconds 1 500 0 Hz NN uu onono Hm un dB SPL dif for each band ouno Seconds 1 FIGURE 6 5 ERROR BETWEEN SIMULATION AND RECORDING FOR EACH SLICE SLICE SIZE 1S Implementation On the graph the dotted black vertical lines represent separation of periods Figure 6 6 depicts the a
4. o 30L i i i 250 0 Hz 2 20 i 1 1 i sor i 1 1 1 1 10 1 1 1 1 1 L Ll L L 1 Ll 1 L L LL 06 50 100 150 200 250 300 90 T TT r n Frame ng l fps_ ee rr 80 i i 1 i J 5 70l 1 1 1 1 1 1 1 I 500 0 Hz 5 I 1 1 1 K RTI AN e yf YU E 90 pem w x puc hara d with iN NT r Hr aai se pe 1 0 kHz ao kw ti x dec we Hes adi tl 2 0 kHz i i 1 i E I I I i F I 1 10 1 1 1 1 1 1 I f 1 La 06 50 100 150 200 250 300 90 i Frame no 1 fps i 80 i I 1 J 5 70L 1 4 0 kHz z 1 i 1 1 1 i g 60 1 1 1 1 1 1 1 1 8 0 kHz 5 50f 1 he 1 I I 1 16 0 kHz be Vr etit hte i 2 2 f I f x a f I I I F i 1 1 i s 10r 1 1 1 06 50 100 150 200 250 300 Frame no 1 fps FIGURE 6 16 ESTIMATION OF THE NOISE 6 2 FOR EACH SLICE ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S The average values for each band Pink Pop Hard Band Hz Noise Speech Opera Rock Rock Classical Electronic 31 81 19 82 33 81 44 82 68 83 75 85 52 84 66 63 73 66 76 8 74 25 75 47 75 27 78 44 77 3 125 67 29 69 42 69 79 6937 68 47 69 56 69 63 250 66 42 67 4 65 9 66 53 65 62 67 15 65 84 500 57 5 58 92 58 34 60 18 59 32 61 27 60 1 1000 53 99 53 75 52 6 53 44 54 02 54 88 53 98 2000 0 44 79 51 55 46 86 45 63 47 97 43 89 4000 0 35 86 35 04 35 48 37 35 36 16 35 45 8000 0 38
5. 6 3 2 8 INPUT SIGNAL FREQUENCY OCTAVE BANDS In order to be able to compare the playback material with the measured one there is a point in the chain where a bank of filters applied to the input signal in time playback signal in order to get a signal representation in octave bands At the same time an octave band combined transfer function is needed In ANSI S3 4 2005 standard values in third octave bands are present In order to get the correct values for octave bands two possibilities have been studied e Values of center frequencies of octave bands in ANSI S3 4 2005 standard These are the values corresponding to the center frequencies of each octave band that have been computed according to IEC 61260 standard taking from the values given in the ANSI S3 4 2005 e Average of frequencies values contained in an octave band These are the averages in SPL of the values in third octaves bands given in the ANSI S3 4 2005 standard In Figure 6 37 differences between these two options can be seen Implementation 10 Center of bands value Average of bands value SPL difference L o T 15F 20 25l L L 1 L 1 29 31 0Hz 63 0 Hz 125 0Hz 250 0Hz 500 0 Hz 1 0 kHz 2 0 kHz 4 0 kHz 8 0 kHz 16 0 kHz Frequency log Hz FIGURE 6 37 CENTER FFREOUENCY VALUES AGAINST AVARRAGE FREQUENCY VALUES OF THE COMBINED TRANSFER FUNCTION As it can be seen the differences between them are ve
6. COMPARISION BETWEEN SIMULATION AND RECORDING FRONT MIC POSITION LEVELS SIGNAL NOISE RECORDING LEVELS LEVEL SIGNAL RAW SIMULATED LEVELS SECOND 11 Interesting to mention in this comparison study is that the 31 Hz band is almost always higher in the recording than the simulation This is because of the shape of this filter see 6 3 8 Octave band filter and equalizer which could not be fitted well inside the IEC 61260 1995 specifications without a down sampling it picks not even playback signal content from other bands but also noise floor content from other bands All the bar graph analysis of the 1 second slices in this comparison were put together with the program material LeftChannel the recording RightChannel simulation uncompressed sound both converted from Pa to some DU in the same manner both gained by 12 dB in a movie which can be found on the DVD Video Recording vs Simulation Front 12dB 1 0S slice wmv The process was repeated without the delay adjustment mentioned in 5 5 Car transfer functions and the average errors values did not change expected for such a small delay given the slice length about 200 samples compared to 44100 samples It should also be mentioned that the recorded signal should always be higher or equal to the simulation because of the noise floor In the presented graphs the analysis was done without the subtraction of the noise floor that is why the blue bars are usually higher Implementati
7. position which is the more realistic and any possible position that could be implemented in a real system 5 4 2 VELOCITIES Several different car velocities for each position have been tested Changes in the behavior of the noise due to the car s velocities are studied The chosen velocities are 50Km h 80Km h 110Km h These represent the most common used velocities while driving inside cities roads outside cities and highways 1 Ref to 20uPa and applies to all following SPLs Analysis 5 4 3 OCTAVE BAND ANALYSIS The noise analysis is done in octave bands to best fit other parts of the project Some parts in the loudness and masking analysis later are analyzed using octave bands and parts of the implementation will be done in octave bands It is there reasonable to study the noise behavior with the same frequency representation technique Besides Python module DVD Code Python _ codes AnalysisWoise Analysis py module DVD Code Python codes BandAnalysis Band Analysis py has been used in this analysis 6 3 8 Octave band filter and equalizer An octave band bank of filters is applied to the measured signals Afterwards the signal is converted from digital units to Pascals 6 3 8 2 Converting from DU to Pa For each filtered signal an RMS value is computed and converted to dB re 20uPa 5 4 4 RESULTS AND ANALYZING The results will be shown depending on the parameters velocity and position First we compare the noise at different ve
8. 079 1Pa h n Pa Some tests were done to check if the new transfer function was reliable The recording in one position front position without engine of the entire measurement signal was converted to Pascals 6 3 8 2 Converting from DU to Pa and was compared to the transfer function h n p convolved with the measurement DVD Measurements Car measurements front mic no motor wav max recording Pa 0 82Pa RMS recording Pa 0 11Pa min recording Pa 0 86Pa max raw Wave h Pa 0 10Pa RMS raw Wave h Pa 0 01Pa min raw Wave h Pa 0 11Pa Also the RMS value in Pa of the noise floor which can be measured in the 30 seconds of silence in the program material no engine running was calculated RMS noise floor recording Pa 0 03Pa Analysis We know that the recording of the transfer function was done with the amplifier set on O dB and the recording of playback material was done with the amplifier set on 20 dB Calculating the dB difference between the RMS values RMS ecoraing Ji RMSnoise RMS awwT F 5 4 0 11 0 035 20108 E 0 012 1535 dB 20l08 0 and taking into account that the two recordings transfer function measurement and program material measurements were done in different days and both the software used and soundcard gains were changed the values seem reasonable However gain compensation will need to be done for noise extraction to compensate for the mentioned gai
9. 2 Up 29cm 74cm both left and right 29cm TABLE 9 5 MICROPHONE POSITIONS 9 1 3 4 EQUIPMENT SETTINGS Power amplifier 208GB gain on amplifier using modified input with static gain Soundcard 7596 output gain 7596 input gain Laptop and Fl studio recording software Asio4all drivers with 512 samples latency setting for soundcard Fl studio 10 producer edition used with the project file DVD Measurements Setup for music playing and recording in car FL studio Setup with chosen listening level flp 9 1 3 5 PROCEDURE 1 Setthe listening level The level should be the preferred level for the listener which is normally close to the Appendices level of the original speech or music Use the playback material Table 9 6 and take a test run in the car to be sure that the level is ok The playback signal should not be too loud or too low which will cause that noise and playback signal will mask each other 2 Measure the level of the pink noise period using the SPL meter and note the result A weighted and linear 3 Record while the calibrator excites the microphone with the 94dB 1KHz signal 4 Record the noise while playing program material at OKm h Velocity 1 Table 9 7 5 Repeat step 4 for all velocities Number Music sound source Genre type 1 Music for archimedes track 3 0 00 0 30 Pink noise 2 Silence Silence 3 Music for archimedes track 4 and 5 0 00 0 15 Speech
10. 6 33 52 4 230 5 we H s where X gt O control the behavior of the system w is called the natural frequency is called the damping ratio and controls the overshot defined as the maximum output value of the system stationary value of the output of the system for a step input Figure 6 45 depicts the step response amplitude of 1 from second 1 of a second order system for different values of damping ratios 0 16 0 33 1 00 I Step 3525549 Transfer Fcn gt j a 5252599 Scope Transfer Fcn1 9 s2 85 9 Transfer Fon2 FIGURE 6 45 OVERSHOT AND DAMPING OF SECOND ORDER SYSTEM It is reasonable to choose a damped or over damped system for our purpose and not include additional oscillations Thus the damping ratio was set to 1 Another important quantity we are interested in is called the settling time which is defined as the time when the step response of the system stabilizes within a band around the step value of the input For a second order system the time for the output to settle within a band of 2 0 02 of the input is 6 34 i N gt 4 Therefore for a desired settling time Ts expressed in seconds a given natural frequency is calculated a Ue Se Since the system which is actually a low pass filter will be applied in the discrete world the system can be expressed as a discrete function using the bilinear transfo
11. 68 36 5 36 83 36 65 37 12 36 88 36 45 TABLE 6 4 AVARAGE OF THE ESTIMATION OF THE NOISE 6 2 FOR EACH PERIOD ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S THE VALUES ARE SPL dB Implementation 6 2 4 5 RESULTS We define the velocity noise floor as the SPL values for each octave band calculated during the silence period of the playback material at constant car velocity The results show that the estimation is consistent with the velocity noise floor in the second period and the deviations from the velocity noise floor were calculated for each band Yperiodestimation velocity noise floor 6 4 7 AVY_EYTOT pana For all velocities Figure 6 14 to Figure 6 17 and Table 6 1 to Table 6 4 engine running the averages were found for 1s slice length See Table 6 5 Band Hz V Velocity OKm h 50 Km h 80 Km h 110 Km h 31 2 53 3 3 4 24 1 68 63 2 27 3 27 3 1 2 84 125 3 22 3 59 3 37 3 98 250 3 82 2 9 4 44 3 57 500 2 2 2 7 3 81 3 63 1000 4 9 68 2 95 2 93 2000 13 47 10 63 5 6 3 12 4000 2 6 13 97 2 55 32 8000 1 84 12 76 4 04 7 47 16000 4 4 91 4 68 2 62 TABLE 6 5 AVARAGE ERROR OF THE NOISE ESTIMATION SLICE LENGTH 1S THE VLAUES ARE IN SPL The same comparison was done by lowering the slice size to 0 1 seconds An estimation graph is presented for 50 Km h recording 100 a a 80 hh ARN un a j fr POM LP ily MI M PM o
12. ERROR 0 15S SLICE 110 KM H Implementation 6 2 5 3 FIRST COMPARISON The back of listener head has the biggest deviation from the velocity noise floor The lowest error in low frequencies is achieved by the ear level listening position except for the velocity of 50 Km h The front position shows relatively small deviations from the velocity noise floor and outperforms other microphone positions at the velocity of 50 Km h 6 2 5 4 SECOND COMPARISON Second comparison was done to see how close the estimation for each band is to the velocity noise floor as calculated in 5 4 Noise in the car where for each band the RMS value of the entire silence period was computed Again additional averaging needs to be done in order for the comparison to be done First of all an average across periods needed to be done for each band the silence period was ignored for a better comparison at constant velocity 1 6 8 Noise Average Estimation velocity 7 M NOiS p and period j j71 34 8 From the graphs in 5 4 Noise in the car we can see that the noise floor in the back position always exceeds the noise floor at ear level sometimes as high as 10 dBs more Also taking into account the deviation in the first comparison for this microphone positioning the back position recordings were discarded for this comparison Secondly the velocity noise floor was calculated for four different positions three because we leave out the back microphone
13. IMPROVEMENTS Although not changing the ideas and behavior of the system additional system improvements can be investigated and implemented 8 4 5 Improve transfer functions by approximating the time domain impulse response which would result into improved computational efficiency Shorten slice size if the above improvement is put into practice the slice size can be shorten even further and the performance of the noise estimation block is expected to improve Improve noise estimation block the case when the noise is poorly estimated can be addressed and a solution provided For instance the noise estimations will maintain its values when the estimation cannot be trusted Improve the gaining system the gains calculated for each band could be based on additional information like spectrum of the playback signal With further analysis the gain values can be better controlled against elevating playback signal noise or the most important bands to be unmasked depending on playback material genre Improvement in computational complexity currently the system is quite pretentious when it comes to processing power and this can optimized further on since such a system should not require a very powerful CPU IMPROVED NOISE ESTIMATION SYSTEM For the current noise estimation system the recording system could be improved by using additional microphones placed inside the car cabin and the recordings set as inputs to the noise estimation block Alt
14. OCTAVE BANDS NOISE LEVELS AT 50 KM H Analysis 5 4 4 8 VELOCITY 80 KM H In this case we can see a higher difference between the back position and the rest with SPL differences around 6 8 dBs in low frequencies Noise at 80 km h Different Positions 100 0 MEE um 60 0 3 in cr Noise Car_Back_80 5 40 0 SS f Noise Car_Front_mic_80 30 0 Noise Chest level 80 20 0 gt Noise Ear 80 10 0 0 0 T T T T T T T T T 1 31 63 125 250 500 1000 2000 4000 8000 16000 Freguency Hz FIGURE 5 11 OCTAVE BANDS NOISE LEVELS AT 80 KM H 5 4 4 9 VELOCITY 110 KM H As can be seen from Figure 5 12 the behavior of the noise at 80 Km h and 110 Km h is very similar The only difference is a little increasing in the SPL values in all freguency range Noise at 110 km h Different Positions 100 0 90 0 80 0 70 0 60 0 50 0 40 0 30 0 20 0 10 0 0 0 T T T T T T T T T 1 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz Noise Car Back 110 SPL dB Noise Car Front mic 110 Noise Chest level 110 lt Noise Ear 110 FIGURE 5 12 OCTAVE BANDS NOISE LEVELS AT 110 KM H Analysis 5 4 5 CONCLUSIONS From the previous section some conclusions about the behavior of the noise can be extracted In general it can be seen that the low frequencies 31 250 Hz are much higher than middle and high frequencies for all different velocities and different positions Regarding velo
15. SLICE 0 1S VELOCITY 80 KM H Implementation All data AVG so AV Gi so AVG 110 was then collected for three microphone positions Front Ear Level Back For all values collected also for 1 second slice see the project s DVD Extra Docs Noise comparison xlsx For the three microphone positions the average errors for each band and velocity AVG velocity are plotted Average Noise Estimation Error 0 1s Slice 50 Km h 14 00 12 00 4 10 00 8 00 6 00 Back Position 4 00 X E Ear level Position co o pe o ps o y 1 m x E lt 2 00 A Front Position 0 00 31 63 12525005 1 2 Hz Hz Hz Hz kHz kHz kHz kHz kHz kHz Frequency FIGURE 6 19 AVERAGE NOISE ESTIMATION ERROR 0 15 SLICE 50 KM H Average Noise Estimation Error 0 1s Slice 80 Km h 8 00 7 00 6 00 5 00 4 00 3 00 2 00 1 00 Front Position Back Position E Ear level Position Average error dB 0 00 31 63 125 250 05 1 2 4 8 Hz Hz Hz Hz kHz kHz kHz kHz kHz kHz Frequency FIGURE 6 20 AVERAGE NOISE ESTIMATION ERROR 0 1S SLICE 80 KM H Average Noise Estimation Error 0 1s Slice 110 E Back Position gt Ear level Position Average error dB Front Position 31Hz63Hz 125 250 0 5 1kHz2kHz4kHz8kHz 16 Hz Hz kHz kHz Frequency FIGURE 6 21 AVERAGE NOISE ESTIMATION
16. aai 28 6 2 2 Ana VS S O da d eee eo qe n Mock etude lace eee es CR Mus o GW are Dee a HER ae d e ORTA 29 6 253 Decreasing slice sizZe eee RR WYF E Er teu ES ee PR T OR ARA 31 6 2 4 OS AA T NN 32 6 2 5 Microphone position for noise estimation o ooccncconononooocnnononononnnnnncnnnnnononnnnnnnnnnnnnnnnnnnnnnnnnnnonnnnnnncaninnnnns 39 Preface 6 3 Loudness and masking compensation cccesesecccecessesessececececseseeeeeeeececeeeaaeseeeesceeseaeeseeeesceeseaasaeseeseeeseaeaeees 44 6 3 1 Signal to diffuse field transfer function ccconcococcnncnnnonononnnnncnnnononnnnnnncnnonnnnonnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnccnnnnnnns 45 6 3 2 Diffuse field to cochlea transfer function ooonnoccconccnnocaconnnononcnonanononccnnnnonnncconn nono cnn nn entere ennt nennen enne 45 6 3 3 Noise threshold levels NN 55 6 3 4 Signal threshold level i ter reete et de e e o eee do de E Od y FFR 57 6 35 iHearing threshold s t et e e bobcat te cedes WG TOR ee eo 58 6 3 6 nGallicalculationsc sid detti eet ehe yd YO edt e ehe oae Dd Ne 59 6 3 7 Gain smoothing s ecd rrt ERREUR ETUR TERR ERE RR USTED Fe ERR YF Y WYE E 61 6 3 8 Octave band filter and equalizer o ooooccccconocooncnnonnconenonnnnncnnnnnnnnnnnnnnnnnnnononnnnnnnnonnnnnonnnnnnnnnnnnennnnnancnnnnnnnns 65 6 4 Total implementation of the loudness compensation system in Python LY LL iie 69 6 41 Insidethe main applicatiOn eet eerta nee eth
17. best possible The speakers were placed symmetrically to the listener in the cabin and the speakers pointing at the listener ears The microphone does not have a certain position This is because different positions will be analyzed later in the report Temperatures and humidity is not taken into account Car driver person Laptop driver person Microphone Speaker L Speaker R Listener person Equipment FIGURE 5 3 EQUIPMENT POSITIONS IN THE CAR Analysis 5 4 NOISE IN THE CAR This section is about the study of the behavior of the noise in different scenarios It is clear that noise from different sources wind engine traffic etc is present during driving activity We want to know how the noise is distributed and the SPL to know which frequencies of the playback signal we can expect to be masked or have decreased loudness while driving In the loudness compensation system we want to develop and described in chapter 6 Implementation one recording position should be chosen Since the recorded signal will be used solely for noise estimation the positioning of the microphone should best estimate the noise in the car as close as possible to noise at the listener position and should be robust enough to playback signal and car velocity The noise is recorded at 4 positions 9 1 3 Noise measurements in car The noise is extracted from the silence period of the program material 5 7 Chosen program materia
18. grows more rapidly than the loudness function and if the noise is wider than an octave band the loudness of the signal will grow more slowly Analysis 5 7 CHOSEN PROGRAM MATERIAL To analyze the behavior of loudness in a car we need some playback signals which are normally played in a car audio system These playback signals will be used during measurements implementation of the loudness compensation system and finally used for evaluation of the system In order to choose some useful playback signals we have followed recommendations given in the technical report IEC 60268 13 part 13 listening tests on loudspeakers for program material e The chosen sounds should present differences between them allowing the study of different important sound perception aspects dynamic range frequency content etc e At least six different sections should be included in the program material covering from human speech to modern music e High sound quality of the program material is needed Based on these recommendations we have chosen the following materials See Table 5 3 Album titles in 9 4 Appendix D References Number Music sound source Genre type 1 Music for archimedes track 3 0 00 0 30 Pink noise 2 Silence Silence 3 Music for archimedes track 4 and 5 0 00 0 15 Speech 4 Pavarotti O sole mio 2 50 3 20 Opera 5 Coldplay Clocks 0 10 0 40 Pop rock 6 System of a down
19. noise coming from environment is picked up Noise threshold level The level where a signal will be just masked by the noise OBF One Octave Band Filters One Octave Band Filter Bank Period A bigger slice of approximate length of 30 s in the playback material containing the same type of material eg pink noise speech electronic etc Pink noise Noise signal with decreasing PSD 10 dB decade with equal amount of power in each octave Playback signal the signal to be played in the car while the system is working SPL Sound pressure level re 20uPa PSD Power Spectral Density PTC Psychophysical tuning curves Re sampling changing the sampling frequency for a discrete signal Settling time Time in seconds for the smoothing filter to go from 0 1 to 0 9 time the amplitude of a step input Signal threshold level The level where a signal masks itself Some octave bands may mask others Slice A subpart of a signal in the time domain Slice size the size of a slice measured either in seconds or in samples Appendices Smoothing ratio Number of subslices in a slice Subslice A subpart of a slice in the time domain used in gain smoothing where each subslice will be amplified by one gain only Threshold shift The level between a masker SPL and the level where a signal is just masked Velocity Noise floor SPL level inside the car when no sound is played through loudspeakers This type of noise floor is dependent on car
20. playback signal and the noise And a noise block which will estimate the noise in the car Idea 1 Figure 6 1 using a noise model controlled by some input parameters to calculated the noise in the car The input parameters could e g be velocity engine rpm accelerometers etc However there are a lot of hard measurable parameters which also influences the noise in the car and they are therefore not easy to take into account in a model These parameters could e g be road type tire type car type car condition weather conditions traffic conditions open closed windows open closed sunroof etc Music from Loudness Loudness a preamp compensation compensated music Car noise model Measured speed engine rpm etc FIGURE 6 1 IDEA 1 Idea 2 Figure 6 2 is using a microphone to measure the noise in the car cabin This ensures that all noise will be registered All the mentioned parameters from idea 1 are actually measured using 1 sensor the microphone However there is one problem The microphone will also measure the played and loudness compensated playback signal and registers this as noise It is therefore necessary that the noise block somehow subtract the loudness compensated playback signal from the microphone measurements Implementation Music from Loudness Loudness X preamp compensation compensated music Microphone Noise extraction noise and music FIGURE 6 2 IDEA 2
21. section is based in the standard ANSI S3 4 2005 The transfer function to take in account from the eardrum to the cochlea is defined in the standard ANSI S3 4 2005 as the SPL in the cochlea in relation with the SPL in the eardrum This is H2 SPL Cochlea SPL Eardrum SPL 6 13 The transfer function is given in SPL in frequency The frequency range corresponds to the audible frequency range this is 20Hz 20KHz expressed in third octaves According to ANSI S3 4 2005 an interpolation of this curve based on a second order polynomial function which fits to sets of three adjacent data points in a linear frequency scale is defined 0 Third octaves Values ANSI S3 4 2005 Interpolation Values SPL dB 60 10 10 10 10 10 10 Frequency FIGURE 6 27 EARDRUM TO COCHLEA TRANSFER FUNCTION SETS OF THREE ADJACENT VALUES IS SHOWED IN THE RED BOXES Implementation It is worth to mention that the interpolation is made in a linear frequency scale but the Figure 6 27 is presented in a logarithmic scale for viewing purposes The function used for the interpolation in python has been scipy interpolation interp1d which allows to make an approximation of a function in 1 dimension in the form yzf x The kind of interpolation used has been cubic which is actually a third order approximation of the curve This kind of interpolation has been decided for its better result than a linear interpola
22. to ISO 389 7 SPL dB 10 1 10 10 10 10 10 Frequency Hz FIGURE 6 41 TRESHOLD HEARING SPL dB THIRD OCTAVE VALUES This hearing threshold parameter is used in 6 3 4 Signal threshold level This study is done in octave bands and in order to get this octave band values the same procedure as in section 6 3 2 4 Implementation of the Combined Transfer Function to the input signal is done In order to increase the resolution of the curve for computing the average values for each center frequency an interpolation of the curve is made This interpolation is made using the python function scipy interpolate interp1d in linear scale Once the interpolation is made and average of the frequency values which are inside of the frequency range determined by the octave band filters is made The average values used in the present project given in SPL Frequency Hz 31 5 63 125 250 500 SPL dB Average 59 77 37 67 22 17 11 4 3 9 Frequency Hz 1000 2000 4000 8000 16000 SPL dB Aveage 1 23 0 78 3 4 6 73 27 1 TABLE 6 9 SPL dB AVERAGE VALUES FOR HEARING TRESHOLD Implementation 6 3 6 GAIN CALCULATIONS The gain calculations are based on the loudness model by Lochner amp Burger 1961 which are described in 5 6 2 Partial masking of loudness This model calculates the perceived loudness in sones for a signal in noise and because the experiments by Lochner and Burger for this loudness m
23. yw Yb WA o O N A Pu hd Ez pt MW EAS v 4 AMA e Wie c THY WN s Y mM ND 5 col SLON Mee MER MAN AHA yw deri Mini NY Slc UON AMAA MOR IA AN AM Ww ida CW L I I L I 40 63 0 Hz 1 1 1 1 1 1 3 125 0 Hz v 9 20 250 0 Hz 1 1 I 1 I I I I L I I LI 2 0 0 500 1000 1500 2000 2500 a 100 g Frame no 10 fps u I I I LI I a I j LI 1 a I L I L I c o U a E 3 v 3 0 500 1000 1500 2000 2500 a u 3 S 8 0kHz w 16 0 kHz A ELMO n WWF s v 2 o z 1 1 1 0 500 1000 1500 2000 2500 Frame no 10 fps FIGURE 6 18 ESTIMATION OF THE NOISE 6 2 FOR EACH SLICE ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 0 1S Implementation In this recording the highest peaks in the low frequencies estimation 31 amp 63 Hz represent movement of the microphone which will add to the average error Additionally because the recording was done twice the program material recorded at 50Km h was put together manually from two individual recordings and because of this there is a delay between the recording and the simulation can be clearly heard since from Coldplay Clocks period An animation was built with the simulated program material recorded program material and estimated velocity noise floor for front microphone position 50 Km h but no gain added to the playback see DVD Video Noise Floor Estimation 50 Km OdB O 1S slice wmv This delay will account for some high freque
24. 17 The new values of H are computed by addition of H1 and H2 values given in SPL The new values are computed for the frequency range given in the ANSI S3 4 2005 These new values in SPL are shown in Figure 6 29 Implementation Combined Transfer Function Third octaves Values Combined Transfer Function Interpolation Values Cubic Interpolation Values linear 60 i 10 Frequency Hz FIGURE 6 29 COMBINED TRANSFER FUNCTION SPL Then the values in SPL of the combined transfer function are converted to gain using 6 12 in order to be able to apply it to the input signal See Figure 6 30 25 Gain Combined Transfer Function 0 0 MM tiit 10 10 10 Freguency Hz FIGURE 6 30 DIFFUSE FIELD TO COCHLEA TRANSFER FUNCTION IN GAINS 6 3 2 4 IMPLEMENTATION OF THE COMBINED TRANSFER FUNCTION TO THE INPUT SIGNAL As it is mentioned in the section before the combined transfer function is converted into gain for its application to the input signal In the project this implementation has been applied into two different signal input formats Two different functions have been made in python for this purpose Both can be found module called DVD Codes Python codes Head and Torso Transfer FunctionMHTF Project py HTTF is a function to compute the filter regarding to the combined transfer function in time domain DVD Codes Python_codes Head_ and Torso Transfer FunctionVHHT
25. 20 Class 2 Min Attenuation 1 30 Butterworth pass band of order 3 S 40 50r 60 10 10 10 10 10 f Hz FIGURE 6 49 PLOT OF THE THREE FILTER CLASSES WITH THE DESIGNED FILTER FOR 1KHZ BAND A zoom around O dB on attenuation axis confirms that such a filter is within the imposed limits Class 0 Min Attenuation Class 0 Min Attenuation Butterworth pass band of order 3 Class O db 10 f Hz FIGURE 6 50 ZOOM AROUND 0 dB ON ATTENUATION AXIS FOR CLASS 0 FOR THE 1KHZ BAND The last filter for octave band analysis centered at 16 kHz was constructed as a high pass filter since a digital Butterworth band pass could not be constructed with sampling frequency of 44100 Hz To fit the attenuation curves in IEC 61260 a 5 order high pass filter was constructed 20 Frequency Hz M I an o 50 80 10 Implementation 166i 10 10 10 10 10 Class 0 attenuation db FIGURE 6 51 PLOT OF THE 16KHZ BAND A second order digital Butterworth filter was fitted for the filter with the smallest center frequency in the octave band filter bank fO 31 25 because a third or higher order could not be fitted due to errors in the calculation of polynomial coefficients as the frequencies get further and further away from the Nyquist frequency the errors become larger due to the bilinear transformation The filter s response among with the IEC c
26. 4 Pavarotti O sole mio 2 50 3 20 Opera 5 Coldplay Clocks 0 10 0 40 Pop rock 6 System of a down Chop suey 2 00 2 30 Hard Rock 7 Beethoven 5 symphony 0 00 0 30 Classical 8 Trentem ller Snowflake 2 41 3 12 Electronic TABLE 9 6 PROGRAM MATERIAL 30SECEND OF EACH ARE MIXED IN ONE FILE AND NORMALIZED TO HAVE THE EQUAL LOUDNESS PERCEPTION 24 dB LFSK LOUDNESS K Velocity setting Velocity Additional notes 1 0 Km h Engine off 2 0 Km h Engine on 3 50 Km h 4 80 Km h 5 110 Km h TABLE 9 7 VELOCITIES 9 1 3 6 RESULTS The recordings are available in DVD Measurements Car Measurements Listening level pink noise 70 8dB A weighted and 77 9dB linear Appendices 9 1 4 MEASURING FINAL RESULT 9 1 4 1 PURPOSE The purpose is to record the loudness compensation system in action for later analysis and judgment 9 1 4 2 USED EQUIPMENT Description Manufacture and type AAU serial number Power amplifier Pioneer A 616 08249 00 Soundcard Edirol UA 25 64681 00 DC AC converter 12V to 230V EA TWI 220 12 2155 00 Battery 12V Biltema 80 416 12V 35Ah N A Microphone B amp K 4134 61447 00 Preamp B amp K 2639 8639 00 Phantom power supply B amp K 2804 6998 00 Speakers 2 pcs B amp W DM601 S2 2144 02 and 2144 03 SPL meter B amp K 2238 33948 00 Dummy head Valdemar Sejr 20010516 02150 00 Re
27. 5 Yes Up hands 88cm 88cm 50cm 29cm 6 Yes Up 72 5cm 72 5cm 12cm 15cm 7 No Up 72 5cm 72 5cm 12cm 15cm 8 No Up 72cm 72cm 18 5cm 60cm 9 No Up 58 5cm 87cm 18 5cm 60cm 10 No Up 86 5cm 64cm 18 5cm 60cm 11 No Up 72cm 72cm 17cm 67cm 12 No Up 72cm 72cm 18 5cm 53cm 13 Yes Up 72cm 72cm 11 5cm 69cm 14 Yes Up 72cm 72cm 6cm 92cm 15 No Up 74cm 74cm 29cm 29cm TABLE 9 3 MEASUREMENT POSITIONS IN THE CAR IN POSITION 5 THE LISTENER IS HOLDING THE SPEAKERS WITH HIS HAND POSITION 8 12 IS MICROPHONE PLACEMENT IN THE LISTENER POSITION Appendices BE FIGURE 9 4 SPEAKER POSITIONS FROM LEFT UP SIDE AND ANGLED 9 1 2 4 EQUIPMENT SETTINGS Amplifier O dB gain using modified input on amp Soundcard Max output gain 75 input gain Laptop and Holmimpulse Asio4all drivers with 512 samples latency setting for soundcard Logarithmic sine sweep with 20Hz start frequency Signal length M equal to 16 44 1Khz sampling frequency 9 1 2 5 PROCEDURE 1 UY ps qe Measure the microphone sensitivity using the calibrator and RMS meter Read the level of the RMS meter while the calibrator excites the microphone with the 1KHz 94dB calibration tone Do also read the level in Holmimpulse using the recording meter Note down the result for later use Use measurement position 1 Use the laptop with the software Holmimpulse to measure the car transfer function Save the result and repeat for all positions 9 1 2 6 RESU
28. 6 0 4 03 02 Amplitude 01 0 1 0 1000 2000 3000 4000 5000 6000 Samples FFT Convolution 20 Desired Gain Amplitude 10 10 10 10 10 10 Frequency Hz FIGURE 6 35 CONVOLUTION OF A FILTER OF 3000 SAMPLES INCLUIDING A DELAY OF 250 SAMPLES WITH A DIRAC DELTA TIME AND FREQUENCY DOMAIN 6 3 2 7 HIGH FREQUENCY RESOLUTION OF IMPULSE RESPONSE In order to study the behavior of the impulse response with high frequency resolution a vector of zeros has been appended to the output of the HTTF function The amount of zeros is nine times the length of the impulse response computed Then a FFT with NFFT length Impulse Response zeros vector 6 20 is computed The result is compare with the theoretical desired impulse response behavior Gain A plot of the result is shown in Figure 6 36 The length of the computed impulse response output of HTTF function has been 3000 samples including 250 samples of delay Implementation 55 FFT of the Combined transfer function impulse response 30000 NFFT FFT of Combined transfer function impulse response NFFT 30000 Desired Gain 2 0 Gain 1 0 0 5 0 0 10 10 10 Frequency Hz FIGURE 6 36 FFT OF THE COMPUTED COMBINED TRANSFER FUNCTION IMPULSE RESPONSE WITH NFFT 30000 As can be seen from Figure 6 36 the behavior of the impulse response is acceptable for the purpose of the present project
29. 73 34 62 34 65 34 59 34 34 17 16000 41 63 36 98 37 06 36 7 37 95 36 78 36 58 TABLE 6 3 AVARAGE OF THE ESTIMATION OF THE NOISE 6 2 FOR EACH PERIOD ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S THE VALUES ARE SPL dB Implementation 6 2 4 4 110 KM H RECORDING EN 100 T T m 80 he c 9 e0r E E 40 p 125 0 Hz o 20r 250 0 Hz o z 06 amp 100 7 80 3 60 Poot td E dE x 500 0 Hz 40 3 1 0 kHz g 20 2 0 kHz o z 06 gi 100 r r a rT un 1 I 1 I 1 1 a 80 f I f 4 0 kHz 1 i 1 1 1 S sol i i i i 8 0 kHz e 1 1 I 1 1 I E AO nnmero mat site I e S 5 1 1 1 h eds TES a 20r 1 1 1 1 I 1 1 1 E i i 1 1 1 i i 2 o fi 1 1 1 1 1 f 1 LL 1 0 50 100 150 200 250 Frame no 1 fps FIGURE 6 17 ESTIMATION OF THE NOISE 6 2 FOR EACH SLICE ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S The average values for each band Pink Pop Hard Band Hz Noise Speech Opera Rock Rock Classical Electronic 31 85 92 86 43 86 14 85 57 84 44 85 21 85 96 63 79 02 79 66 79 45 77 07 77 15 79 15 77 29 125 74 44 74 73 75 3 72 09 71 95 74 88 71 73 250 70 09 69 28 70 37 68 29 68 23 70 35 68 69 500 63 89 63 47 63 72 62 37 62 85 64 23 63 23 1000 59 72 61 08 59 7 59 1 59 71 60 14 59 69 2000 49 83 51 27 53 03 50 64 51 64 51 6 51 28 4000 0 42 54 43 22 44 37 42 7 44 15 43 47 8000 0 35 58 34 76 36 58 43 93 35 12 36 44 16000 39
30. ALYZED BY SKOVENBORG 2004 CLASS 1 ESTIMATES BEST THE LOUDNESS OF SPEECH AND MUSIC Analysis All these models are able to estimate some better than other the perceived loudness However there is one problem with the models for this project point of view They don t take noise into account which for sure affects how loud a signal will be perceived We want to know how loud the signal alone is perceived Not the total loudness of signal and noise the partial masking of loudness 5 6 2 PARTIAL MASKING OF LOUDNESS Investigations and experiments for loudness of a signal in noisy environments are performed by Lochner amp Burger 1961 and their results is used to create a function which describe the perceived loudness depends on noise and signal intensity 5 7 They played a 1KHz pure tone in the presence of an octave band 700 1400 Hz of random noise for different test subjects The pure tone noise and the pure tone alone was played alternately through earphones for periods of 1 3 sec and the test subject then had to adjust the level of the pure tone to match the level of pure tone presence in noise The results from these experiments were used to create and validate the function and later experiments by other authors confirm their results Florentine Popper amp Fay 2011 The function is based on Stevens power law The loudness in sones for a signal in noise is y kd Ij 5 7 Where I is the signal intensity and ly is the t
31. Common for both ideas is that SPL or intensity levels for the playback signal and noise shall be known at the listener position to correctly calculate the perceived loudness and then compensate if needed This means that gains transfer function etc for the used equipment including the car is needed We want to know what the playback signal in e g 16bit values correspond to in intensity level at the listener when played through the audio system in the car Likewise for the microphone levels in idea 2 Both ideas allow different volume and user sound settings if they are applied in the preamp before the loudness compensation Change in volume or sound after loudness compensation will give rise to wrong compensation of the playback signal if no corrections for these changes are added in the loudness compensation The loudness compensation shall be connected directly to the power amp for correct behavior See Figure 6 3 for intended implementation of the loudness compensation system in a car audio system Loudness Sound gt Pre amp compensation Power amp Speaker source system FIGURE 6 3 IMPLEMENTATION OF THE LOUDNESS COMPENSATION SYSTEM IN A CAR AUDIO SYSTEM The chosen solution is idea 2 because we believe we can create better noise estimations using this solution Idea 1 needs a lot of parameters to perfectly estimate the noise and even though we maybe not are able to implement idea2 perfectly we believ
32. Due to a limited time period man hours and for simplifying easier analyzing of results project limits are introduced The following points will be covered not covered by the project and project report e The loudness compensation system will only be optimized for one certain listening position even though there is room for more than one person listener in the car The listener position is necessarily not the driver position and will be chosen from a practical point of view e Only2 speakers will be used even though new cars typical have 4 or more The 2 speakers are not necessarily the speakers build into the car They can also be speakers from the laboratory Which speakers we are using depends on the audio system in the rented car and practical issues e Noise cancelation of any kind will not be included in the loudness compensation system and not discussed in the report e Equalization to flatten the response from the speakers and car cabin will not be implemented in the loudness compensation system and not discussed in the report e Cabin changes to improve the cabin acoustic or noise isolation will not be carried out analysis MEWN 5 ANALYSIS 5 1 INTRODUCTION Before development of the loudness compensation system different investigations and analysis are needed This part will therefore cover investigations and analysis of theory and practical issues to support the development of a loudness compensation system described later
33. EXA Levels signal noise SPL db 031 0 H63 0 H125 0 250 0 FEDO 0 H2 0 kHz2 0 kHzt 0 kH28 0 kH26 0 kHz 937 0 Hz3 0 H225 0 250 0 Hi00 0 H2 0 kHz2 0 kHzl O kHz8 0 KH26 0 kHz Frequency log Hz Frequency log Hz Speech period second 60 Classical music second 200 FIGURE 6 7 COMPARISION BETWEEN SIMULATION AND RECORDING FRONT MIC POSITION LEVELS SIGNAL NOISE RECORDING LEVELS LEVEL SIGNAL RAW SIMULATED LEVELS Implementation However where the period contained more low frequencies the estimation is close to the recorded playback material S 128 S 224 FE Levels signal raw FU Levels signal raw M5 Levels signal noise EXA Levels signal noise So H263 0 H225 0 150 0 H 00 0 Hz 0 kHz2 0 KHz4 0 kH28 0 kHz6 0 kHz Frequency log Hz EH H253 0 H225 0 1250 0 H 00 0 Hz 0 KH2 0 kH2t 0 KHz8 0 kHz6 0 kHz Frequency log Hz Coldplay Clocks second 128 Trendermgller snowflake second 224 FIGURE 6 8 COMPARISION BETWEEN SIMULATION AND RECORDING FRONT MIC POSITION LEVELS SIGNAL NOISE RECORDING LEVELS LEVEL SIGNAL RAW SIMULATED LEVELS By looking at the above graphs we can conclude that the simulation is close enough to the recording a conclusion enforced by the small error in the pink noise S 011 70 Gl Levels signal raw ES Levels signal noise 9o Hz3 0 H225 0 H250 0 Hi00 0 H2 0 kH22 0 kH24 0 kH28 0 KH26 0 kHz Frequency log Hz FIGURE 6 9
34. F Project py HTTF Octave bands py is the function made for computing the transfer function in frequency domain octave bands It should be noted that all fourier transforms applied to different signals for the combined transfer function analysis have been computed with the same amount of points NFFT as the length of the signals for which it has been applied Implementation 6 3 2 5 INPUT SIGNAL TIME DOMAIN For the implementation of the combined transfer function a filter in time domain is built from the information in the frequency domain Some parameters of the filter should be defined These parameters were chosen after some tests which will be explained following The parameters are e Frequency Response It is clear that the frequency response of the filter should be the most close to the theoretical worked out Gain e Delay It s defined as the number of samples before the main peak of the filter e Phase Response No information about the phase in the standard ANSI S3 4 2005 is mentioned No influence of phase for loudness compensation is taken into account therefore it has not been taken in account for the filter design e Duration of the filter It s basically the length of the signal in time The number of samples is a trade off between desired freguency response and the time needed for convolution computation 6 3 2 5 1 FREOUENCY RESPONSE GAIN The freguency response should be the gain computed to be applied to the input si
35. KHZ P i Pic io e wy WN wl Aw Ni o 40 ML ata ar w 1 1 A wai wh A uw E I 1 1 2 20 1 i I I 1 14 i Frame no 1 fps 4 0 kHz 60 8 0 kHz 40 Noise estimation N o o o u o 100 150 200 250 Frame no 1 fps FIGURE 6 15 ESTIMATION OF THE NOISE 6 2 FOR EACH SLICE ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S The average values for each band Pink Pop Hard Band Hz Noise Speech Opera Rock Rock Classical Electronic 31 84 53 82 12 83 81 85 91 84 53 85 61 84 76 63 75 62 76 37 78 46 76 89 73 92 75 72 74 12 125 66 09 65 23 69 54 70 69 66 96 67 01 66 09 250 63 46 62 34 64 17 65 76 64 99 65 55 61 78 500 57 47 56 98 57 56 61 84 57 29 58 33 57 18 1000 54 32 46 82 48 68 58 58 51 05 50 46 49 65 2000 49 12 37 5 49 76 46 83 43 08 44 59 41 26 4000 0 29 42 31 65 46 74 37 14 35 07 37 85 8000 46 87 38 48 32 93 36 74 39 4 33 39 36 01 16000 41 43 36 11 36 21 36 51 36 88 36 36 36 TABLE 6 2 AVARAGE OF THE ESTIMATION OF THE NOISE 6 2 FOR EACH PERIOD ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S THE VALUES ARE SPL dB Implementation 6 2 4 3 80 KM H RECORDING 90 T Wi T a T al TT LIS T E 80 c de iac i A J d a 270 Re je wyo T Jp Mea amant 310 Hz 5 Po Loo ps Ni Pa NW E 60 1 1 1 1 1 1 1 1 63 0 HZ 5 50F 1 I 1 1 1 I 1 1 7 B aol f f 1 I 1 125 0Hz
36. LTS The measurements are available in DVD Measurements IR Measurements Microphone sensitivity 9 8mV Pa 94dB corresponds to 22 05dB and 0 079pcm in Holmimpulse Appendices 9 1 3 NOISE MEASUREMENTS IN CAR 9 1 3 1 PURPOSE The purpose is to record program material played inside the car noise at different velocities 9 1 3 2 USED EQUIPMENT Description Manufacture and type AAU serial number Power amplifier Pioneer A 616 08249 00 Soundcard Edirol UA 25 64681 00 DC AC converter 12V to 230V EA TWI 220 12 2155 00 Battery 12V Biltema 80 416 12V 35Ah N A Microphone B amp K 4134 61447 00 Microphone Calibrator B amp K 4231 78301 00 Preamp B amp K 2639 8639 00 Phantom power supply B amp K 2804 6998 00 Speakers 2 pcs B amp W DM601 S2 2144 02 and 2144 03 SPL meter B amp K 2238 33948 00 Laptop with FL studio 10 N A N A Various cables and stands N A N A Car Chrysler grand voyager LE UB 46 018 License plate TABLE 9 4 USED EQUIPMENT 9 1 3 3 MEASUREMENT SETUP The main part in the setup is the laptop which is connected to the soundcard using an USB connection The Microphone is connected to the soundcard input 1 through the phantom power supply and the power amplifier is connected to the outputs of the soundcard The power amplifier uses battery supply together with a DC AC converter To avoid hard start of the amplifier which maybe damage to the DC AC converter
37. Motor 20 0 Noise Car Front Mic No Motor 10 0 0 0 r r r r 1 N o o S o o o o 9 m JS V dv Q S SS S S S Frequency Hz FIGURE 5 8 OCTAVE BANDS NOISE LEVELS AT 0 KM H NO ENGINE Analysis 5 4 4 6 VELOCITY O KM H ENGINE ON As can be seen from Figure 5 9 the SPL values are very similar in both positions when the engine is turned on It is worth to mention the predominance of the low frequencies in the noise level as was expected This measurement was done just in two different positions Noise at 0 km h DifferentPositions Noise Car Back 0 E Noise Car Front mic 0 0 0 r r r 1 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 9 OCTAVE BANDS NOISE LEVELS AT 0 KM H ENGINE ON 5 4 4 7 VELOCITY 50 KM H From Figure 5 10 very similar values for all positions at this position is observed It can be seen how the noisiest position is the back position Also we can observe how chest and ear level positions have slightly smaller values probably due to the absorption of the listener Noise at 50 km h Different Positions 100 0 90 0 80 0 70 0 60 0 Noise Car_Back_50 50 0 amp f Noise Car Front mic 50 40 0 Noise Chest level 50 gt Noise Ear 50 30 0 200 4 10 0 0 0 r r r r r r r r 7 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 10
38. ON OF THE ELECTRICAL PART OF THE SETUP We begin the analysis based on the units presented by Holmimpulse software transfer functions from DU Digital Units to DU including the software processing normalization output type float etc Then we will move on to the desired transfer functions which are those that transform the playback DU and corresponding type to Pascals 5 5 1 TRANSFER FUNCTION PROCESSING Several decisions needed to be taken about the measured transfer function 5 5 1 1 WINDOWING Since the measurements exported from Holmimpulse software did not include the delay information in the sample number sample O was set to the highest peak of the impulse response not to the time 0 and delay uncertainties reside in different software while playback the delay will be approximated and evaluated separately and the windowing of the impulse response will start just as the exported impulse response raises above a certain threshold from noise floor before highest peak Algorithmically this was done by searching for a number of consecutive samples used 10 consecutive samples to be below a certain threshold used 0 005 highest peak of IR Impulse Response from the highest peak backwards By visual inspection of the IR in the time domain we chose a fixed length for the impulse response set to 3000 samples the impulse drops enough from maximum peak value in all measurements up to that point This is depicted in Figure 5 14 Anal
39. OR EACH SLICE ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S The average values for each band calculated as avg_estimationperioa SPL Frame no 1 fps Y estimaion s estimastionspand where only the estimations estimaion s gt music simulation s were taken into account 1s slice Band Hz 31 63 125 250 500 1000 2000 4000 8000 16000 Pink Pop Hard Noise Speech Opera Rock Rock Classical Electronic 64 22 65 74 66 17 65 62 65 94 66 57 68 62 0 57 86 57 01 57 31 56 42 57 19 58 73 0 O 52 92 53 6 55 06 54 66 55 78 0 O 4625 48 17 49 24 46 97 51 1 0 0 O 40 09 41 63 0 44 12 0 0 O 26 83 30 66 0 32 14 0 47 9 5398 42 78 29 19 39 74 30 33 0 32 36 29 08 31 522 31 5 29 06 29 75 0 36 21 32 62 34 02 34 46 32 87 33 36 40 92 37 23 36 63 36 7 37 49 36 75 36 52 6 3 TABLE 6 1 AVARAGE OF THE ESTIMATION OF THE NOISE 6 2 FOR EACH PERIOD ONLY WHEN IT IS HIGHER THAN THE SIMULATION SLICE SIZE 1S THE VALUES ARE SPL dB Implementation 6 2 4 2 50 KM H RECORDING 1 Fm c t Mm z cn ch LI NN pred tate care a cman vs o dra EA v E ety deeft 4 Jupe 5 I I 1 Y 63 0 Hz i i g 125 0 Hz E i E 2 250 0 Hz I N Lal 1 T 1 LJ 150 200 250 100 i Frame no 1 fps a y 1 1 S 80 1 1 1 I 1 1 500 0 Hz i D 1 1 mo E 60 o ek n vm tutu lb Dem ad did l 2 10kHz E WIN ria yw 4 tithe A 2 0
40. PL at the recording position in the car the total SPL Several steps are needed First we calculate the ERBn equivalent rectangular bandwidth for normal hearing ERBy 24 673 0 004368f 1 6 21 Where f is the center frequency for the chosen band Next step is to calculate P51 P51at1KHz which are values used in later calculations Af 6 22 P51 ERBy BER iio 4000 6 23 AS BE Then the value Piower is calculated P P51 0 ges PME x TotalSPL 51 Pam lower ED Oe Sake FED Where TotalSPL is the noise and playback signal SPL at recording position Finally the shape W finput can be calculated Y OC 1 E p inoue 6 25 f W finput P Where P is equal Piower if finput is less than f Otherwise P is equal P51 Last step is to convert the auditory filter to dB scale and mirror it upside down which will give us the auditory filter shape with the center frequency f AF finput 10 log10 W finput 6 26 For every octave band with center frequency f the noise threshold level given by these on the chosen band centered on frequency f is then calculated with the computed auditory filter AF having a center frequency f NoiseThresholdLevel f NoiseSPL f AF f ThresholdShift 6 27 It should be noted that f can be equal to f which calculates the threshold level in the chosen band the threshold level for a playback signal to be just masked by the noise in the chosen band Th
41. RAE PER TERRAS ER 8 5 2 Music iricreference coridition fii iecore oe erre WY WWF verenda eoa va ae than oc Y WA WY FEW a y n 8 5 3 Meas trement Set p i cocer rior eo ate eue cos eo etes en GOF ddod a Y dedu met deseen dede spaced eet ees dw FYD 9 5 4 Noise in the Causa A Y dd dnd Gh dd Fd y FD YF Cen GFTN CF 11 BALL Recording POSITIONS E E A WN TYF Re DAD TR fee be DY FR GA 11 5452 a elu FERN FFR Ee EHE ER EUREN He RECAP NN NN 11 S43 OctaVe band analysis ironia RH HER RERO GR FEBRE E GETRETEN Ge Fe EE ERE Re GEL 12 5 44 Results and analyzing iii pter Ere a P EE Pee ba E conn ddyw tak P Pe eee pet bee ea evo o meh aede Pena daed 12 IS CONCIUSIONS REDE 17 5 5 Cat transfer TUnctlons oie in redi DN A Shara ee oe E bend yn ust d 18 5 5 1 Transfer function processing ereenn LL A LLAI ILL nennen LEL nennen enti n i rasa ss ases sitara daas FL anet 18 5 6 fe 1i a rc TEE 23 5 0 1 o A A eee ertet t y d rie oh Pe nA GG NG A EE RR ea Leere Eo RES es ko dus 23 5 6 2 Partial masking of loudness neinei E E TEER ETE ETER EAE ETRE E Ear 24 5 7 Chosen program imatetiadl tee meto a e E o re er te o eee eer ud 25 6 Implementation eoe eo om nies cie aida Eee atn tede rede a e dd E ERE e ad oet 26 6 1 Introduction zi Fe GY i OA epa Foe De RENE EMEN Y abies 26 61 TIheid as c i eep aen ae HE EU RPRPE weds DUREE Re EUER PEU PERI AN ANN 26 6 2 Noise extractione inrter O RNA 28 6 2 1 Testing for reliability i i HERE DOE e Vee FR HF YI a da ii EE
42. URE 6 40 ILLUSTRATION OF THE SIGNAL THRESHOLD LEVEL CALCULATION FOR THE 1KHZ OCTAVE BAND USING AUDITORY FILTER THE BARS ARE SIGNAL SPLS IN OCTAVE BANDS AND THE 1KHZ BAND MARKED WITH GRAY IS NOT TAKEN INTO ACCOUNT FOR THE CALCULATION OF THE 1KHZ SIGNAL THRESHOLD LEVEL IN THIS CASE THE SIGNAL THRESHOLD LEVEL IS THEREFORE ONLY AFFECTED BY THE SIGNAL IN THE 31HZ OCTAVE BAND AND NOT THE 1KHZ BAND ALL OTHER SIGNAL BANDS ARE BELOW THE AUDITORY FILTER WHEN THRESHOLD SHIFT IS TAKEN INTO ACCOUNT AND DO THEREFORE NOT AFFECTS THE SIGNAL THRESHOLD LEVEL Implementation 6 3 5 HEARING THRESHOLD According to ISO 389 7 the hearing threshold is the level of a sound at which a person gives the 5096 of correct detection responses on repeated trials One of the inputs for the signal threshold level block see Figure 6 25 is the threshold hearing of the human being at each frequency band Concretely the audible frequency range is studied in octave band For this purpose the European standard ISO 389 7 has been used as a reference for the threshold hearing levels In this document two different hearing thresholds are available depending on the sound field where it is applied We assume that the sound field in a car audio cabin is a diffuse field that means that we expect that the reflections of the sound will arrived from all directions The levels in one third octave band in SPL are shown in Figure 6 41 si Treshold hearing in a Diffuse field according
43. ack recording was very sensitive to changes additional uncertainties were induced which had to be treated separately but could not eliminated completely Due to limited time and resources car bookings etc we accepted this as a known limit to our project and did not address it by redoing all the measurements transfer function recording while driving on line test of system The playback signal had to be analyzed into smaller slices and a reasonable slice size was found so as to best fit different sub blocks in the system e Gain estimation a smaller slice helps the gain estimation block dip listening on playback material e Real time convolutions the slice could not be infinitely small since it will be smaller than the transfer functions and an on line system becomes unpractical e Computational performance although dependent on the program material the slice size affects the computational speed and can stall the playback Conclusion Additionally since this slicing could not be infinitely smooth and a feedback loop was present two new issues arose distortions and signal oscillations appeared The problem was addressed by setting up a smoothing algorithm practically implemented by a second order system which was controlled by two parameters settling time and smoothing ratio These parameters were tuned to better fit the system while it was tested outside the car and can be further tuned to eliminate the described playbac
44. acoustics com cn index php page 32 Ccrma Matlab transfer function measurements toolbox May 2012 https ccrma stanford edu realsimple imp meas Marui Matlab loudness matching toolbox May 2012 http www geidai ac jp marui matlab node40 html Wikipedia 1 Settling time May 2012 http en wikipedia org wiki Settling time Wikipedia 2 Dirac delta function May 2012 http en wikipedia org wiki Dirac delta functionstGeneralizations Parkers Information about the rented car May 2012 http www parkers co uk cars reviews facts and figures chrysler voyager estate 1997 13047 Internetautoguide Information about the rented car May 2012 http www internetautoguide com 13 13 2000 17 716 331 2000 chrysler grand voyager base minivan specifications html Car_Picture Overview picture of a car May 2012 http carblueprints info eng view ford ford pinto 1973 Appendices HETH Pink Noise Both Channels Uncorrelated Music for Archimedes CD track 03 Female amp Male Speech Music for Archimedes CD track 04 05 Pavarotti O Sole Mio Pavarotti Favorite Neapolitan songs track 01 Coldplay Clocks Coldplay A rush of blood to the head track 05 System of A Down Chop Suey System of a down Toxicity track 06 Beethoven Symphony No 5 Carlos Kleiber Wiener Philharmoniker Beethoven Symphony 5 amp 7 track 01 Trentemgller Snowflake Trentemgller The last resort tra
45. and repeat the sound sources without adding unwanted changes The data was ripped and cut lossless Implementation 6 IMPLEMENTATION 6 1 INTRODUCTION The implementation and solution part covers how the loudness compensation system is developed from scratch to solution The part will include different ideas thoughts and how the solution is developed to have the desired functionality Investigation and analysis from chapter 5 Analysis is taken into account in this part and is used to form and support the chosen solution The solution is divided into smaller parts which are developed and tested individually This ensures better controlled over the loudness compensation system and makes it easier to maintain and debug It also gives the possibility to parallel development Finally all parts are put together 6 1 1 THE IDEAS Before development different ideas were discussed and analyzed Based on a brainstorm we ended up with 2 different ideas where the main difference is how to detect the noise in the car The idea is from an early stage of the project where we have a lack of knowledge to loudness masking and loudness models Due to that different ideas for the loudness compensation were therefore not possible They are formed later in the project Figure 6 1 and Figure 6 2 Illustrate the ideas for the loudness compensation system and include both two blocks A loudness compensation block which will adjust the playback signal depending on
46. antages is that we don t know the system before we have the car It could be too bad for this project point of view The amplifier maybe introduces phase and frequency changes and the speakers maybe have a bad frequency response or it will be difficult impossible to interface a computer with the car s audio system Due to that we decided to add our own system FIGURE 5 1 THE CAR AND ITS AUDIO SYSTEM IT SEEMS THAT THE DECISION TO ADD OUR OWN AMPLIFIER AND SPEAKERS WAS A GOOD IDEA THE EXISTING CAR AUDIO SYSTEM ONLY HAS A FM RECEIVER AND A TAPE PLAYER NO AUX INPUTS When adding our own system we are able to control everything and validate that our system behaves as expected but we are limited to a 12V power supply and we are not able to position the speakers where speakers are normally positioned in a car The power supply problem is solved using as much battery powered equipment as possible We are using a laptop with an USB powered soundcard and the phantom power for the measurement microphone is also battery powered Only the amplifier needs 230V but this is easily solved using a DC AC converter 12V to 230V We could have bought a new 12V car amplifier but the used amplifier and DC AC converter was available in the laboratory DC AC y converter r Left speaker Laptop m Soundcard Power amplifier Right speaker Phantom power supply Microphon
47. apped into a GUI thread the main loop was spawned as a separate thread communicating with the GUI threads through global variables the main loop can be controlled through the application interface and it displays its progress inside the GUI Test and results 7 TEST AND RESULTS 7 1 ONLINE TEST OF LOUDNESS COMPENSATION SYSTEM IN LABORATORY Before testing the system online in the car a pre test was done in the laboratory to check if the system works A picture of the set up is presented in Figure 7 1 7 mm Yn FIGURE 7 1 SYSTEM PRE TEST The setup is equal to the setup 9 1 4 Measuring final result but with some differences The playback was done in the laboratory not in the car the transfer functions from car were kept Only one loudspeaker was used for playback right one in the photo from the right laptop in the figure Figure 7 1 One loudspeaker was used to play a pure tone left one in the photo from the left laptop using RoomEqWizzard v5 software to simulate some kind of noise The microphone was placed close to the noise source can be seen in the extreme left of the photo While playing the same file that was used for testing and will be used for the on line test in the car through the implemented application a tone of a particular frequency and gain was generated o check the loudness compensation Depending on the frequency we expect the biggest gain in the octave band where tone resides dependi
48. ation system is inactive Number Music sound source Genre type 1 Music for archimedes track 3 0 00 0 30 Pink noise 2 Silence Silence 3 Music for archimedes track 4 and 5 0 00 0 15 Speech 4 Pavarotti O sole mio 2 50 3 20 Opera 5 Coldplay Clocks 0 10 0 40 Pop rock 6 System of a down Chop suey 2 00 2 30 Hard Rock 7 Beethoven 5 symphony 0 00 0 30 Classical 8 Trentem ller Snowflake 2 41 3 12 Electronic TABLE 9 9 PROGRAM MATERIAL 30SECEND OF EACH ARE MIXED IN ONE FILE AND NORMALIZED TO HAVE THE EQUAL LOUDNESS PERCEPTION 24 dB LFSK LOUDNESS K Velocity setting Velocity Additional notes 1 0 Km h Engine off 2 0 Km h Engine on 3 50 Km h 4 80 Km h 5 110 Km h 9 1 4 6 RESULTS The recordings are available on the DVD Measurements Final Dummy Measurements TABLE 9 10 VELOCITIES 9 2 APPENDIX B DVD CONTENTS Audio o Codes o o Extra O O O Recording and Simulation Back no Engine Front no Engine Matlab Codes Loudness normalizing for wave files LoudnessToolbox 1 2 e WAW Python Codes _Analysis e Example FilesN AuditoryFilters BandAnalysis Control Delaying Head_and_Torso_Transfer_Function e Loudspeakers to head TFsV Loudness_compensation Masking Noise_detection Plotting Project_Main e Log Slicing Transfer_Functions e Front e Tran
49. ayback slice and the recording slice not to overlay completely From the noise estimation tests even a hearable delay between played signal and recorded signal would not affect the estimation and the system performs well additionally the gain smoothing will minimize the problem when the delay issue will propagate further on to the gain calculation block The system was tested on line on the same machine with the onboard soundcard to test how it performs and mostly due to time convolutions of FIR transfer functions of the chosen length 6 3 2 Diffuse field to cochlea transfer function the computations needed for each slice sometimes took longer than the slice length causing interruptions in the playback signal A performance test was done on a laptop with an Intel Core I5 CPU with 8GB of RAM inside with a slice length of 1 second and various computations blocks were timed 25 slicing equalising transfer function convolution 2 0 band levels loudness compensation smooth gaining Time s Frame number 1fps FIGURE 6 59 PERFORMANCE TEST OF VARIOUS SYSTEM BLOCKS SEQUENTIAL PROCESSING Implementation The test is just a rough estimation since the machine was running under normal conditions with multiple applications running together with corresponding interrupts like network interrupts etc As can be seen the processing is dependent on periods and the most computationally expensive operations fo
50. band greater than the simulation in recording position were taken into account because of poor estimation when masking effect due to noise is not estimated see 5 4 Noise in the car The chosen time slice length for the analysis was 0 1 seconds because the estimation improves with a smaller time slice and the errors in estimation when playback signal levels are higher than the noise estimation is lower see 5 4 Noise in the car Implementation 6 2 5 1 COMPARISON OF NOISE ESTIMATIONS The comparison will be done for different positions and different velocities of the car Data extraction values will be given in SPL For a given recording at a specific velocity the noise estimated for each octave band in each slice is averaged for each period by averaging across all slices 6 5 Noise_estimationpana iperiod j gt Period Noise_estimationstice rages ted 0 155 velocity Where K represents the slices contained in a period and Period is one Part of the test signal generally one song silence or pink noise expressed in seconds This way the average in the second period should be close enough the velocity noise floor found in 5 4 Noise in the car which is computed taking in account the RMS value of an entire period and here the noise floor is done by averaging RMS values of smaller slices concretely 0 1 seconds as it is mentioned before Noise floor values found for one octave bands and different velocities can be found on DVD Ext
51. city it can be seen in general and for all positions that as expected the noise levels increase with velocity The amount of energy in low frequencies is much more important than in middle and high frequencies rising up to 90 dB SPL in some cases Also it is important to mention that values in SPL in the range 31 5 125 Hz don t change too much for velocities 50 80 and 110 Km h generally the difference between 50 and 110 Km h is about 2 5 dB Due to this an important masking of the playback signal by the noise is expected to happen in all different velocities thus in all driving activity Regarding the position parameter it can be seen how there is a big difference in the noise at 0 Km h when the engine is turned off in all positions As it is mentioned in 5 4 4 5 Velocity O Km h No engine the noise levels due to the engine is higher than the noise coming from environment conditions therefore such a big difference is not expected while engine is on Also it can be seen that back position is the noisiest position that has been tested The SPL differences between back position and the other positions is around 6 8 dBs in the frequency range 31 1000 Hz reaching in some cases 10 dBs of difference Also it can be seen how SPL in front ear level and chest position are very close for a fixed velocity Analysis 5 5 CAR TRANSFER FUNCTIONS For estimation of the noise implementation of the loudness compensation system and possible simulati
52. ck 09
53. corder Zoom H4 64694 00 Laptop with software NA NA Various cables and stands NA NA Car Chrysler grand voyager LE UB 46 018 License plate TABLE 9 8 USED EQUIPMENT 9 1 4 3 MEASUREMENT SETUP The main part in the setup is the laptop which is connected to the soundcard using an USB connection The Microphone is connected to the soundcard input 1 through the phantom power supply and the power amplifier is connected to the outputs of the soundcard The power amplifier uses battery supply together with a DC AC converter To avoid hard start of the amplifier which maybe damage to the DC AC converter let the soft start circuit inside the DC AC converter power up the amplifier In practice this task is performed when powering the power amplifier before the DC AC converter The amplifier will hum due to the non sinusoidal AC output from the DC AC converter Be aware of the DC AC converter cabinet temperature The laptop contains the developed loudness compensation system For later analysis and judgment the playback from the system is recorded using a dummy head connected to a handheld recorder Bu A ene d m gt Left speaker Laptop D Soundcard Power amplifier Dummy head Recorder Right speaker Phantom Microphone power supply FIGURE 9 7 CONNECTIONS Appendices IEEE All eq
54. d sent to the soundcard Then the entire process is repeated for the next input slices A block diagram for each time slice is depicted in Figure 6 58 Implementation al Smoothed gains Recording signal gt DUtoPa gt OBF a Extract Diffuse field to Noise threshold Y Estimate gt cochlea gt xx T o Noise Level conversion eve Signal to Diffuse field r Gn eang transfer function l gt OBF q recording position Noise levels Branch 2 cochlea Signal to Diffuse field to Signal System raw signal Diffuse field cochlea Signal levels Gain Gain Gain gt anser transfer OBF gt Ben cochlea calculations smoothing function function Hearing threshold Branch 3 levels cochlea E 7 Gained playback slice Source slice HEN sde H zi SoundCard Branch 1 FIGURE 6 58 BLOCK DIAGRAM OF THE MAIN PROGRAM The designed application has a feedback on the gains and octave band equalizer therefore the playback will be delayed 1 slice length from the microphone recording From previous analysis the slice length will be chosen small enough so this delay will not play an important role Additionally there should be a delay in the playback and recording streams intermediary buffers fetching times etc which will add to the previously mentioned feedback delay making the pl
55. dices 9 1 2 CAR TRANSFER FUNCTION MEASUREMENTS 9 1 2 1 PURPOSE The purpose is to measure transfer functions for the car plus speakers and investigate the changes due to microphone speakers and person movements 9 1 2 2 USED EQUIPMENT Description Manufacture and type AAU serial number Power amplifier Pioneer A 616 08249 00 Soundcard Edirol UA 25 64681 00 DC AC converter 12V to 230V EA TWI 220 12 2155 00 Battery 12V Biltema 80 416 12V 35Ah N A Microphone B amp K 4134 61447 00 Preamp B amp K 2639 8639 00 Microphone Calibrator B amp K 4231 78301 00 RMS meter B amp K 2417 6680 00 Phantom power supply B amp K 2804 6998 00 Speakers 2 pcs B amp W DM601 S2 2144 02 and 2144 03 Laptop with Holmimpulse N A N A Various cables and stands N A N A Car Chrysler grand voyager LE UB 46 018 License plate TABLE 9 2 USED EQUIPMENT 9 1 2 3 MEASUREMENT SETUP The main part in the setup is the laptop which is connected to the soundcard using an USB connection The Microphone is connected to the soundcard input 1 and the RMS meter through the phantom power supply and the power amplifier is connected to the outputs of the soundcard The power amplifier uses battery supply together with a DC AC converter To avoid hard start of the amplifier which maybe damage to the DC AC converter let the soft start circuit inside the DC AC converter power up the amplifier In practice this tas
56. e FIGURE 5 2 THE BASIC SETUP IN THE CAR Analysis The chosen speakers were B amp W DM601 S2 They are chosen on compromise between size and ability to produce low frequencies They fit into the car and a 6dB cutoff frequency at 50Hz is acceptable for a speaker of this size The chosen microphone is B amp K 4134 which is a pressure field microphone and chosen because the car cabin is assumed to be a diffuse field and because of its frequency range It is able to measure frequencies between 4Hz and 20KHz which covers the frequencies we are focused on 20Hz to 20Khz Frequencies we are able to hear All used equipment including serial numbers are listed in 9 1 Appendix A Measurement journals To validate the electrical part of our setup we have measured the impulse response when the amplifier output is looped to the microphone input in 9 1 1 Verification of measurement setup We expect the phase and frequency response to be flat and the impulse response to be close to a dirac delta This holds true for this setup All equipment except speakers microphone and laptop are placed in the trunk of the car The speakers are placed on the backseats and the listener in between The used car has actually 3 rows of seat where cars normally only have 2 To handle this difference the second row of seats in the used car was not used The speaker and listener position was chosen to satisfy the speaker and listener position in reference condition
57. e car we have investigated and analyzed different theories and practical issues A car is a harsh environment for the purpose of music listening and compared to the standard listening room the car is far from ideal The car cabin will influence the playback signal especially the noise floor will affect how the loudness of the playback signal will be perceived at listener position In the standard listening room the noise floor is defined to be maximum 65dB ref to 20uPa at 31 5Hz and decreasing in the following octave bands These maximum levels cannot be met in the car because of the noise engine wind tires etc To know the noise distribution in the car we measured the noise while driving and afterwards analyzed the measurements in one octave bands From this analysis we conclude that the low frequency noise 31 5Hz 63Hz and 125Hz bands is the most dominant and actually does not change much with velocity Changing the velocity from 50km h to 110km h gives an increment of 2 5dB in these bands The noise in mid and high frequencies 250Hz 500Hz 1KHz 2KHz and 4KHz bands are more dependent on the velocity and changing the velocity from 50Km h to 110Km h gives an increment of 10 15dB in these bands The noise SPL in the 31 5Hz band reaching 90 dB in some cases and compared to the maximum noise in reference condition the noise in the car is 30 50dB louder This will for sure affect the perceived loudness of the playback signal and maybe mask some fre
58. e idea2 still estimates better than idea1 Especially when parameters like road type changes idea1 will have troubles We have not investigated how much the noise is actually changing due to change of the hard measureable parameters It s only based on our own experience Implementation 6 2 NOISE EXTRACTION The idea behind the noise separation algorithm is simple compare the recording in one position with the estimated sound of the playback signal which would be the program material convolved with the transfer function in that position The difference of levels between the recording and the estimation should be given by the presence of noise in the recording position Therefore the levels in the recorded position should always be higher than the estimated levels Of course this difference can have other sources like measurement noise both on line measurement and transfer function measurement or floating point operations error but we expect these not to dramatically affect a dB of an RMS value Thus the comparison will be done in each of the 1 octave band by comparing SPL value thus both the recording and the estimation of the playback signal should be transformed to Pascals Naturally the comparison will be done by slicing the signal into smaller intervals A block diagram of the noise extraction is depicted in Figure 6 4 Recorded signal Pa gt Octave Band Filters cune rec position
59. e maximum value of these is the noise threshold level for the chosen band which is the same as selecting the maximum masked threshold for the chosen band from the PTC of all the other bands Implementation Threshold shift is the level between noise SPL and the level where a signal is just masked by the noise within the same band We have chosen a fixed value of 18 5dB by studying the figures in Moore 2012 The described calculations are for one single band and they are therefore repeated 10 times in the software Figure 6 38 and Figure 6 39 illustrates different noise spectrums and the calculations for the noise threshold levels for the 1KHz band with different noise spectrums 90 Auditory filter 80 70 60 u o SPL db E o Noise SPL 30 Threshold shift Noise threshold level 20 10 31 0 Hz 63 0 Hz 125 0 Hz 250 0Hz 500 0Hz 1 0 kHz 2 0 kHz 4 0 kHz 8 0kHz 16 0 kHz Frequency log Hz FIGURE 6 38 ILLUSTRATION OF THE NOISE THRESHOLD LEVEL CALCULATION FOR THE 1KHZ OCTAVE BAND USING AUDITORY FILTER THE BARS ARE NOISE IN OCTAVE BANDS IN THIS CASE THE NOISE THRESHOLD LEVEL IS AFFECTED BY THE NOISE IN THE 1KHZ OCTAVE BAND ALL OTHER NOISE BANDS ARE BELOW THE AUDITORY FILTER AND DO THEREFORE NOT AFFECTS THE NOISE THRESHOLD LEVEL Noise SPL Auditory filter b Threshold shift 80 70 60 Noise threshold level 10 31 0 Hz 63 0 Hz 125 0 Hz 250 0Hz 500 0Hz 1 0 kHz 2 0 kHz 4 0 kH
60. e recording was delayed the corresponding number of samples and the transfer function modified by 15 35 dB to match the RMS value without the noise floor The results after are plotted in Figure 5 18 Recording Pa 0 24 Simulation Pa i sal Mi lul VG IT ELM Pull M EC MW 10 62 10 64 10 66 10 68 10 70 10 72 10 74 10 76 time FIGURE 5 18 COMPARISON BETWEEN RECORDING AND SIMULATION AROUND SECOND 10 WITH GAIN AND DELAY COMPENSATION FRONT MICROPHONE POSITION ENGINE OFF The above operations are done inside DVD Codes Python codes Transfer_Functions Compute_transfer_function py function readAndCompute_average_time_IR_FixedWindow Tests were done inside module DVD Codes Python codes Delaying test_delay py Analysis 5 6 LOUDNESS An important part in this project is the understanding of loudness and masking and how it influences our hearing Due to our hearing organ we do not perceive loudness of a signal equal to its intensity The perceived loudness depends on frequency content and SPL of the signal background noise masking phenomena and maybe even more The mechanisms underlying the perception of loudness are not fully understood Moore 2012 All these known parameters which affect the loudness perception are combined in several different loudness models which can be used to estimate the perceived loudness of a signal 5 6 1 LOUDNESS MODELS Different loudness models are during the years develo
61. e to variance from listener movement e The noise floor in ear level position is very close to the velocity noise floor in the front position e The averaging done in equation 6 9 is not very different to the value found at ear level see section 5 4 4 Results and analyzing Implementation 6 3 LOUDNESS AND MASKING COMPENSATION From section 5 6 Loudness none of the analized loudness models and calculations seems to fit perfectly to our problem The loudness models by Skovenborg 2004 estimates the loudness of music and speech well but don t take into account noise And the loudness calulation from Lochner amp Burger 1961 which calculates the loudness of a signal present in noise have some disadvantages The calculation is only confirmed valid in the bandwidth 200 8000Hz and with pure tones Our playback signal has a wider bandwidth and contains complex tones Temporal forward and backward masking are also not taken into account in this calculation However the loudness function by Lochner amp Burger 1961 is the only approach we have found for calculation of loudness of a signal present in noise Our loudness compensation system Figure 6 25 is therefore formed around this function and used in the gain calculations block Diffuse field to Noise cochlea Noise threshold transfer level function nani Signal to Diffuse field to Loudness ignal i g Octave band Diffuse field cochlea Signal threshold Gain Gain Octave band co
62. ed to minimize the noise floor An octave bands diffuse field to cochlea transfer function has to be applied A test in order to know which method average frequency values center frequency band values fitted better in our system when trying to apply the diffuse field to cochlea transfer function in octave bands was done Best results were obtained with average values of the contained frequencies in each band To compensate for such noise it needed to be measured or determined One of the biggest challenges of the loudness compensation system was to determine the noise at listener s position since a direct measurement is not possible within a playback signal Although the noise inside the car is pretty consistent from a spatial point of view as we have seen from different microphone positions retrieving the actual value of this noise in octave bands was not flawless and proved to be quite a challenging task Because the estimation of such noise was higher when the noise was not predominant a higher gain for playback was expected under these circumstances While the system was tested inside the car a higher than necessary gain was applied by the developed application and the system gain had to be tuned down to balance for this An important setback for the current project was the different gains applied while measuring soundcard input gain soundcard output gain amplifier gain and software gains Because the gains were not the same and the playb
63. encies but also the higher freguencies This means that parts of the high dynamic periods which was masked before is now hearable Also the Trentemgller period is hearable In total it sounds like the loudness compensation system does what it should but there are some problems Because of the applied gain smoothing the gain adjustment is slow and the loudness compensation system is therefore hearable When the program material is low in level the loudness compensation system applies a high gain in the masked bands Then when the program material then changes to a high level faster than the smoothing time the gain is too big because of the low level part before and will take some time to be adjusted to the correct level In this case clipping can occur Also when changing from a period to another it is easy to hear that the loudness compensation system need some time to adjust the gains In our mind a good loudness compensation system is systems which adjust a playback signal to the correct loudness without the listener to notice This is not the case for our loudness compensation system However it applies some improvements to the experience of the program material 8 4 FURTHER DEVELOPMENT During the various stages of the project different ideas were considered but not investigated nor implemented Since the developed loudness compensation system evolved into a rather complex and detailed piece of software there is plenty of room for improvements a
64. ernatively the noise estimation block could be changed if better and easily fitted inside the application even a non acoustical block based for instance on speed and or outside conditions could be implemented Appendices METNIMME 9 APPENDICES 9 1 APPENDIX A MEASUREMENT JOURNALS This appendix includes all the measurements journals Because each measurement journals is created as individual documents some repetitions will be present 9 1 1 VERIFICATION OF MEASUREMENT SETUP 9 1 1 1 PURPOSE The purpose is to verify the electrical part of the setup used for all measurements The transfer function and impulse response are measured and verified to ensure correct functionality 9 1 1 2 USED EQUIPMENT Description Manufacture and type AAU serial number Power amplifier Pioneer A 616 08249 00 Soundcard Edirol UA 25 64681 00 DC AC converter 12V to 230V EA TWI 220 12 2155 00 Battery 12V Biltema 80 416 12V 35Ah N A Speaker B amp W DM601 S2 2144 02 Laptop with Holmimpulse NA NA Various cables and stands NA NA TABLE 9 1 USED EQUIPMENT 9 1 1 3 MEASUREMENT SETUP The main part in the setup is the laptop which is connected to the soundcard using an USB connection The power amplifier is connected to the outputs of the soundcard and the speaker output from the amplifier is connected to the speaker and input 1 on the soundcard The power amplifier uses battery supply together with a DC AC conve
65. f AALBORG UNIVERSITET LOUDNESS COMPENSATION OF MUSIC IN A CAR AUDIO SYSTEM Master Acoustics gr Semester Spring 2012 12gr860 A Pablo Cervantes Sebastian Prepelita Regnar Bonde Supervisor Pablo Faundez Hoffmann Department of Electronics Systems Preface 1 PREFACE This report is written by group 860 at 2 semester on the Acoustics Master program at the department of electronics systems Aalborg University spring 2012 Group members Pablo Cervantes Sebastian Prepelita Regnar Oxholm Bonde Supervisor Pablo Faundez Hoffmann 1 1 ACKNOWLEDGES Before developing the project explanation we d like to thank Pablo Faundez Hoffmann our supervisor who provided us help about the project conduction Moreover he took care of renting the university car used for measurements and he was our driver when needed Thanks to Peter Dissing and Claus Vestergaard Skipper for help regarding equipment which could be used in the car 12V supply Thanks to the IT staff members for assisting us on problems regarding to the group folder for storage and for the SVN Finally thanks to Aalborg University for giving us the opportunity to discover a school system and a relevant experience for international experience Preface 1 2 READING GUIDE The project documentation is divided into the following three parts Report is the main documentation for the project and is chronologically com
66. f the smoothing algorithm is depicted in Figure 6 46 Settling Smoothing time ratio Gain Gain Gain SliceGain FIGURE 6 46 OVERVIEW OF THE SMOOTHING ALGORITHM The parameters controlling the smoothing are e Settling time seconds e Smotthing ratio Implementation An example of gain smoothing for one band with 31 Hz as center frequency with a settling time of 3 seconds and a smoothing ratio of 3 done with 1 second slices 50 Km h recording mic in front position m a m Lj Roe o N Gain 31 Hz Band Gain 6 4 J 2 0 50 100 150 200 250 Frame slice number 1 fps 14 Smoothed gain 12 Gain held constant Gain 31 Hz Band B o o N 0 100 200 800 400 Frame slice number 1 fps FIGURE 6 47 GAIN SMOOTHING FOR 31 HZ BAND 1 FPS Figure 6 48 depicts the same smoothing same parameters and recording for a slice of 0 1 seconds 4 2 Mi ol ili TA 1000 1500 2000 2500 Frame slice number 10 fps 14H Gain held constant Y Smoothed gain 2 eo PR A H n o N o o Gain 31 Hz Band p S 2 o Gain 31 Hz Band 4000 8000 Frame slice number 10 fps FIGURE 6 48 GAIN SMOOTHING FOR 31 HZ BAND 10 FPS Of course a filter can be constructed for each band each with its own settling time and smoothing ratio to better control
67. frequency range from 20Hz to 20kHz 8 1 THELOUDNESS COMPENSATION SYSTEM AND IT S BEHAVIOR To analyze and evaluate the chosen loudness model we implemented it in a sub block of the loudness compensation system In this system we calculate the loudness of the playback signal in reference condition which we chose to be the perceived sound at the listener s position in the car without noise and with the help from these loudness values we calculated gains in octave bands When the gains are applied to the playback signal before playback in the car the loudness should be equal to the loudness in reference conditions for each octave band Band wise the original apparent loudness of playback signal is restored In order to calculate the correct perceived loudness we need the correct SPL at listener position Since it s not practical to mount a measuring microphone close to the listener s cochlea we calculated the SPL at listener position using transfer functions for speakers car cabin head torso and ear pinna and middle ear The transfer functions for speakers and car cabin are measured for the used car and speakers The head torso and ears transfer function are from ANSI S3 4 2005 Most of the analysis and calculations are based on one octave band analysis The design of such a system was addressed according to IEC 61260 1995 and needed some compromise filters for the lower bands became erroneous without a down sampling Although desirable for s
68. g of the program material Beethoven s 5 Symphony file DVD Audio Bethoven back 0 wav recorded while engine was not running Test can be reproduced in test Song of DVD Codes Python codes BandAnalysis test_DU2Pa py SPL found for entire wave file was 79 78 dB SPL and the 1 octave filter output is depicted in figure Figure 6 55 31 0Hz 63 0Hz 125 0Hz 250 0Hz 500 0 Hz 1 0 kHz 2 0 kHz 4 0 kHz 8 0 kHz 16 0 kHz Frequency log Hz FIGURE 6 55 OUTPUT FROM THE 1 OCTAVE BAND OF FILTERS FOR BETHOVEN Implementation EN 6 4 TOTAL IMPLEMENTATION OF THE LOUDNESS COMPENSATION SYSTEM IN PYTHON This section presents how all the subparts were fitted together inside the main application program DVD Codes Python Codes Project_main application_main py As described in previous sections the main inputs and outputs of the loudness compensation are Recording sound from the microphone INPUT for noise estimation Program material INPUT for playback and noise estimation Gained program material OUTPUT for soundcard speakers The preamplifier depicted in Figure 6 3 was included inside the main application therefore an additional INPUT is needed and will be controlled by the user through the application s interface system gain Of course the inputs and the outputs of the loudness compensation system are digital signal slices of equal length the following description will refer to slices of signal not the entire signal An o
69. ga asses enti an 79 AppendiC6S uo ctr T a o re ehe n DR eden LP it vest T UR RE Cn 80 9 1 Appendix A Measurement journals ccccccccccccecsssesesecececsesesaseecececeesesaesesecsceesesesaeseescecsesesaeseeececeeeaaeaeeeeeees 80 9 1 1 Verification of measurement setup esses LLE LLALL LLE LL ALIS LEL EL sien entire ansias s entera dana sanas enean 80 9 1 2 Car transfer function measurements sese enne nennen rennen enr en nennen nen nnne nnns en nene enne 82 9 1 3 NOISE measurements iniCar ior iret re PRre du gd nun FYN ER oso EE eae Da se NAF e dee Yo SEEE AE E p Paso EEEO 85 OA Measuringinalires lE 5 a5 eene a Ne RENTE EPI RE 88 9 2 Appendix B DWVD coritents eorr EH ROTE RH Y FT AERE UR EC A REESE ERES ERN ORE TER ERO 91 9 3 Appendix C DICtIOnIaly ce ee erre ER HARE CYFF E AREE YD FFF YA 93 9 4 Appendix D References in ccc ierit eoe aa en entis Poo eite epa cesser bia eee tae e Poe E aree euet a d esee tac Pee Ead de hv 95 Introduction 3 INTRODUCTION In today s fast moving world the car is becoming little by little the main place people listen to music to audiobooks or good old radio Despite many advantages that a car can offer compared to a standard listening room while stuck in a traffic jam things get a little complicated when it comes to listening to various playback materials while average driving velocities becomes contemporary relevant As the vehicle s velocity increases var
70. gnal for this section The resolution of this gain depends on the length of the impulse response desired samples so an interpolation as it is explained in section 6 3 2 1 Diffuse field to eardrum transfer function will be computed The amount of points in the function gain interpolated will be the same as the desired samples Once the freguency response is obtained an entire model spectrum is built This spectrum will include the freguency range 0 22050 Hz and the negative part of this spectrum is obtained computing the conjugate of the positive part Therefore the number of points obtained for the entire spectrum is the double of the number of samples introduced in the function HTTF inside DVD Codes Python codes Head and Torso_Transfer_Function HHTF_Project py minus one sample due to the repetition of the sample corresponding to O Hz As no information about the response of this frequency range is available it is decided to have the same response that 20 Hz It is worth to mention that this approximation has no influence in the filtered signal since the audible range is above 20 Hz Once the desired freguency response is built an inverse fourier transform is made numpy fft ifft function is used for that purpose The number of samples after the computation will be the same as the NFFT included in the desired spectrum so this length in samples will be controlled by the amount of samples in the interpolation of the desired spectrum as it is me
71. hose to convert the input signal usually a digital converted signal thus represented in Digital Units abbreviated by DUJ to Pa and then calculate the output of each Butterworth filter relative to Py 20 uPa Thus the N samples Fy time mean square level output for each filter in a given time T s will be calculated using the following S formula 1 a 6 39 N Zin 0 Pout n Po RMS P 20 log P 0 Lout 20 logro dB re20 uPa Where P is reference pressure of 20 uPa F is sampling frequency and P n is pressure converted from DU according to 6 40 at sample n 6 3 8 2 CONVERTING FROM DU TO PA To convert from measured DU recorded wave files to measured Pascals the calibrator recording done in the day with the measurements DVD Measurements Calibrator SECOND DAY wav was used We know the calibrator produces 94 db re20 Pa which represents 1 Pa Pays and by calculating the RMS value of the recording thus RMS of DU we could convert the digital units to Pa Sample DU 6 40 DU Calibrator measurement RMS Pa Sample Pa 6 3 8 3 TESTING FOR LEVEL INDICATORS A small number of tests were done to check if the transformations are correct and do make sense Since the calculation of the input level in SPL could not be evaluated with the designed set up and since the conversions and the indicators will be done in Pa not in Volts we compared outputs with other known measured outputs u
72. hreshold intensity for the noise l is the threshold of the signal in the presence of any noise intensity of the signal at which it will just be masked by the noise n is approximate 0 27 according to Lochner amp Burger 1961 and k is a constant depending on the used units In our case k is calculated to fit the formula when the intensity levels are converted to SPL The loudness in sones is then 5 8 y 10 19 0 27 _ 10 0 10 0 27 11 0266 Where L is the signal SPL and Lo is the noise threshold level in dB Figure 5 20 shows the function with different noise threshold levels Loudness perception of a signal in noise 20dB 30dB 40dB odg ra dam sd n J 10 2 30 40 50 60 70 80 Signal SPL dB FIGURE 5 20 PLOT OF PERCEIVED LOUDNESS OF A SIGNAL IN NOISE BASED ON THE MODEL BY LOCHNER amp BURGER 1961 5 8 THE LOUDNESS IS PLOTTED FOR NOISE THRESHOLD LEVELS AT O 20 30 AND 40dB FOR A NOISE THRESHOLD LEVEL AT OdB SIGNAL LEVELS AT 40dB CORRESPOND TO 1 SONE Since the loudness model is based on 200 8000Hz pure tones as the signal the function is not totally reliable for this project We want to predict the loudness for a complex signal music and this will maybe change the perceived loudness depends on frequency contest in the signal The width of noise does also affect how the loudness is perceived Florentine Popper amp Fay 2011 If the noise has a width of a critical band the loudness of the signal
73. ime zero of the window Once we have the impulse response in this position 0 delay samples from the left part of the impulse response or from the right can be taken in account In other words we can choose the delay just taking a number of samples before the main peak of the impulse response and to maintain the desired number of samples we can discard samples from the end of the impulse response With this method we can create a new impulse response with a desired delay Since information about frequency response of the impulse response is before the main peak a test was made in order to know the amount of samples before the main peak of the impulse response delay has to be taken in account to have an acceptable frequency response Impulse Response 2500 samples no delay 0 6 o gt Amplitude o N 0 500 1000 1500 2000 Samples Hg FFT of the impulse Response Desired Gain MEG pn hc 10 10 10 Frequency Hz Amplitude o ia e mo S 2 FIGURE 6 33 IMPULSE RESPONSE 2500 SAMPLES WITH NO DELAY Implementation As can be seen when no delay is included in the impulse response too much information about the behavior in frequency is lost and the frequency response is not acceptable for our purposes Impulse Response 2500 samples 250 samples delay Amplitude o 0 0 F eo 500 1000 1500 2000 2500 Samples FFT of the impulse Response Des
74. in this report Analysis of loudness and loudness models which can be used to predict the perceived loudness and therefore be used to restore the original apparent loudness of music presence in noise will also be analyzed 5 2 MUSIC IN REFERENCE CONDITION Various playback signals like cd radio etc are intended to be played at reference conditions or close e g in a living room The playback signal is often mixed in a studio with reference conditions and to have the same experience and sound it is recommended to play it in the same conditions or close to From IEC 60268 13 a reference conditions can be obtained using following steps e To ensure uniform distribution of low frequency eigen tones the room dimension ratios should be W H lt L H lt 4 5 W H 4 where L is length H is height and W is width The preferred size is 25m to 40m e The reverberation time should be between 0 3 s and 0 6 s for 200 4000Hz sounds The ceiling should be mostly reflective the floor mostly absorbent and additional absorption material should be uniform distributed e The background noise should in no circumstances exceed the levels in Table 5 1 Frequency Hz 31 5 63 125 250 500 1000 2000 4000 8000 Max SPL dB ref to 20pPa 65 47 35 26 20 15 12 9 7 TABLE 5 1 MAXIMUM BACKGROUND NOISE SPL FOR REFERENCE CONDITION e _ The distance between the speakers should be between 2m and 3 5m pointing
75. ious indispensable noise sources increase in loudness making the playback material from partly unhearable to indistinguishable Sound generated by the car s engine and tires by wind friction with the car body by road bumps or simply road type increase so much with velocity that from one point on the material played through the car s sound system turns out to be quite different than what was initially intended With some bad weather added to this the listener has to take action like turning up the volume which will become a strong impediment to many normal car activities chatting speaking on the phone etc The noise generated while travelling will have most of its energy concentrated at low frequencies making the middle and higher frequencies not audible Although one would expect only some frequencies to disappear the psychoacoustical effect of masking makes the masked frequency band even larger As the velocity increases the energy starts moving up in frequency and with enough care by the user car and environment the sound inside the car will become pure noise usually unpleasant to listeners This can transform travelling by car into an unpleasant stressful and unhealthy environment An expensive solution would be a better isolation of the car Another approach would be to compensate for such adverse sound companions by adjustments in the playback material in such a way that it will not be masked by the described noise and it will not affec
76. ired Gain Frequency Hz FIGURE 6 34 IMPULSE RESPONSE 2500 SAMPLES WITH 250 SAMPLES OF DELAY In Figure 6 34 an impulse response of 2500 samples and 250 samples of delay is shown As we can see the frequency response is acceptable in all the frequency range that we are interested in In this report just two examples of the different lengths and delays are shown although more tests were done A script for testing different lengths is included in DVD Codes Python codes _Analysis Test_HTTF py 6 3 2 5 3 PHASE RESPONSE Regarding phase response there is no information about it in ANSI S3 4 2005 standard After studying the different blocks which the present project consists of no phase information is taken in account for any of the mentioned blocks therefore the phase response has not been taken in account in the construction of the filter 6 3 2 5 4 DURATION OF THE FILTER The duration of the filter has been chosen taking in account different aspects The frequency response should be acceptable and convolution computation time should be fast enough for an on line loudness compensation system Considering all the aspects mentioned before a 2500 samples with 250 samples of delay filter has been chosen for use in the loudness compensation system 6 3 2 6 TEST OF THE COMBINED TRANSFER FUNCTION Once the main characteristics of the filter have been decided and the filter is built we can test the filter and check its behavior in freque
77. k is performed when powering the power amplifier before the DC AC converter The amplifier will hum due to the non sinusoidal AC output from the DC AC converter Be aware of the DC AC converter cabinet temperature DC AC Banen converter m gt Left speaker Laptop Soundcard Power amplifier Right speaker Phantom RMS meter power supply Microphone FIGURE 9 2 CONNECTIONS Appendices All equipment except speakers microphone and laptop are placed in the trunk of the car Figure 9 3 The speakers are placed on the backseats and the listener in between The used car has actually 3 rows of seat where cars normally only have 2 To handle this difference the second row of seats in the used car was not used Different microphone speaker and person placement are used and described in Table 9 3 Remember to close the doors during measurements and to avoid ear damage use earplugs Car driver person Laptop driver person Microphone Speaker L Speaker R Listener person m Equipment FIGURE 9 3 SETUP IN CAR Position Listener Speaker pos Microphone positian Left window Right window Roof Seat 1 Yes Up 77 5cm 99cm 49cm 27cm 2 Yes Up 88cm 88cm 50cm 29cm 3 Yes Side 88cm 88cm 50cm 29cm 4 Yes Angled 88cm 88cm 50cm 29cm
78. k problems if needed offering a high degree of flexibility for the described issues The equalization of the signal was done using a bank of Butterworth filters the same as the ones used for octave band separation of time signals The equalization was done into small steps because of the gain smoothing and the filters presented a known phase response the only point where the phase could play an important role inside the system Care was taken not to gain bands that could not be heard under reference conditions in each band levels below the hearing threshold or levels masked within the playback signal itself Although the low frequency roll off of the chosen loudspeakers could not cover the entire 31 5 Hz octave band the loudspeakers were kept since a typical car audio system does not have a subwoofer or speakers able to play such low frequencies and the used loudspeakers were already a bit fancy for the average car audio systems Despite this drawback we can compensate the other masked bands As an overview of the developed system the chosen noise estimation method and the loudness calculation method raised the complexity of the system and introduced additional details that needed to be addressed A trade off between simplicity flexibility and a reliable system had to be found which after extensive analysis seems as a restless endeavor 8 2 OBJECTIVE JUDGMENT OF THE LOUDNESS COMPENSATION SYSTEM All the sub blocks that are implemented in the l
79. l noise noted NM and the estimated sound in SPL for each octave band playback material noted M the noise noise noted N in each octave band can be estimated P P 6 1 NMgsp 20log 5 s ref P Msp 2010819 2 Pref Thus an estimation of the noise in each band RMS value for a certain slice 6 2 NMspL MspL Nsp 20log g 10 20 10 20 gt The value Nop was set to O if NMsp lt Msp in case of estimation errors The estimation was first tested on the program material for the front microphone position when engine was not running The estimation should approach the noise floor in all periods Figure 6 12 depicts the noise estimation for each slice 1 slice 1 second the second period represents the noise floor and the noise estimation should approach the values within that period Frame no 1 fps FIGURE 6 12 NOISE ESTIMATION FOR EACH SLICE IN OCTAVE BANDS SLICE SIZE 1S As can be seen in Figure 6 12 except maybe the 31 Hz band the estimation does not approach the noise floor in many bands the difference being as big as 40dB Some uncertainties may reside in the transfer function gain in 5 5 Car transfer functions and this could be a cause for this differences By increasing the gain of the transfer function by 1 5 dB we had the result in Figure 6 13 Implementation
80. l recorded in the measurement session and analyzed with python scripts These scripts are included in the Python module DVD Code Python codes Analysis Noise_Analysis py The analysis of the measured noise will be based on the different recoding positions and car velocity 5 4 1 RECORDING POSITIONS Following recording position was chosen e Back This position is located behind the listener s head The aim of choosing this position is to study if we have a good signal to noise ratio considering the signal desired signal as the noise and the playback signal would be considered as noise signal This will maybe improve the noise extraction e Front This position is located in the middle and top of the car A preferred position from a practical point of view if a loudness compensation system should be permanent implemented in a car e Chest Level This position has been chosen mainly for transfer functions purposes knowledge about how transfer functions changes depending on the position of the microphone Noise has been studied in this position as well in order to have a better knowledge of this scenario e Ear Level This position probably is the closest one to the reality in terms of perception but in the other hand it is also the less practical Because a microphone for recording has to be located in the car this position is not possible in a real system The purpose of this position is to study the variability of the noise between this
81. let the soft start circuit inside the DC AC converter power up the amplifier In practice this task is performed when powering the power amplifier before the DC AC converter The amplifier will hum due to the non sinusoidal AC output from the DC AC converter Be aware of the DC AC converter cabinet temperature DC AC Battery converter r Left speaker Laptop Soundcard Power amplifier Right speaker ddeuda m Microphone power supply p FIGURE 9 5 CONNECTIONS Appendices HETH All equipment except speakers microphone and laptop are placed in the trunk of the car Figure 9 6 The speakers are placed on the backseats and the listener in between The used car has actually 3 rows of seat where cars normally only have 2 To handle this difference the second row of seats in the used car was not used The microphone is placed at the position referred to in Table 9 5 Car driver person Laptop driver person Microphone Speaker L Speaker R Listener person Eguipment FIGURE 9 6 SETUP IN CAR THE RIGHT PICTURE IS THE MICROPHONE IN POSITION 3 Position Microphone Direction Microphone capsule distance from Roof Windows Seat row2 1 Back Up 6cm 72cm both left and right 92cm 2 Front Down 11 5cm 72cm both left and right 69cm other side of seat 3 Mid 1 Up 50cm 88cm both left and right 29cm 4 Mid
82. level position It can be seen how for low frequencies 31 63Hz the SPL are almost the same At this position a bigger dependence from velocity can be seen in a wider spectrum range Values in high frequencies 8000 16000Hz present small changes with different velocities Noise at differentvelocities Mic position Ear Level 90 0 80 0 70 0 60 0 50 0 Noise Ear 50 40 0 SPL dB Noise Ear 80 30 0 Noise Ear 110 20 0 10 0 0 0 E F T T T T T 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 7 OCTAVE BANDS NOISE LEVELS MEASURED IN EAR LEVEL POSITION 50 80 AND 110 REFER TO CAR VELOCITIES KM H 5 4 4 5 VELOCITY O KM H NO ENGINE This measurement was done just in two different positions From Figure 5 8 we can see how the noise levels with no movement of the car and no engine are slightly higher in low frequencies A big difference can be observed in the frequency range of 250 2000 Hz and very similar values in high frequencies 4000 16000Hz It should be mentioned that the difference between the two measurements is expected because of the variability of the environmental conditions While the engine is running we don t expect such variability due to the constant noise coming from it 0 Km h No Motor 60 0 50 0 40 0 8 z 30 0 m Noise Car Back No
83. locities for a certain microphone position and next we compare the noise at different microphone positions at a certain velocity 5 4 4 1 POSITION BACK As it can be seen from Figure 5 4 the noise levels increases in all frequency range as the velocity does If we consider the noise floor as the noise measured when the engine was turned off we can see how the engine has a big influence in the noise at low frequencies especially in the range 125 500 Hz We can see how this range is increasing proportionally with the velocity and how frequency range 1000 4000Hz start to be an important influence when the car start to move Very high frequency range 8000 16000Hz doesn t suffer a big change with velocity changes Noise at differentvelocities Mic position Back 100 0 90 0 80 0 70 0 a Noise Car Back No Motor a 60 0 50 0 E Noise Car_Back_0 6 40 0 A 7 te Noise Car Back 50 30 0 20 0 gt Noise Car Back 80 10 0 Ji Noise Car Back 110 0 0 T T T T T T T T T 1 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 4 OCTAVE BANDS NOISE LEVELS MEASURED IN BACK POSITION 0 50 80 AND 110 REFER TO CAR VELOCITIES KM H Analysis 5 4 4 2 POSITION FRONT As it can be seen from Figure 5 5 a very similar interpretation to position back scenario could be done Differences in SPL are much higher in low frequencies around 35 40 dB from noise floor to 110 km h when parameter ve
84. locity is varied We can see an important boost in the frequency range of 125 500Hz when the engine is turned on and not very important changes in level of SPL are seen in high frequencies Noise at different velocities Mic position Front 100 0 90 0 80 0 70 0 60 0 Noise Car Front Mic No Motor 2 50 0 f Noise Car_Front_mic_0 40 0 Noise Car Front mic 50 30 0 Noise Car Front mic 80 20 0 gt Noise Car Front mic 110 10 0 0 0 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 5 OCTAVE BANDS NOISE LEVELS MEASURED IN FRONT POSITION 0 50 80 AND 110 REFER TO CAR VELOCITIES KM H 5 4 4 3 POSITION CHEST LEVEL No measurements at 0 km h were done for this position As can be seen from Figure 5 6 the levels in the lower part of the frequency range studied present similar levels a fact which can be understood as a certain independence from velocity Also it can be observed an important change in SPL at middle frequencies 1000 4000 Hz Noise at differentvelocities Mic position Chest 9 Noise Chest level 50 H Noise Chest level 80 Noise_Chest_level_110 31 63 125 250 500 1000 2000 4000 8000 16000 Frequency Hz FIGURE 5 6 OCTAVE BANDS NOISE LEVELS MEASURED IN CHEST LEVEL POSITION 50 80 AND 110 REFER TO CAR VELOCITIES KM H Analysis 5 4 4 4 POSITION EAR LEVEL The behavior of the noise at this position is very similar to the chest
85. ly the input intensities are converted to SPL using 6 32 I 100719 6 32 Where I is the intensity level and L is the SPL 6 31 with 6 32 applied to all intensity levels is plotted in Figure 6 42 Implementation EW Gain when signal and reference threshold are fixed Signal 70dB Ref threshold 0dB Signal 90dB Ref threshold 0dB Signal 70dB Ref threshold 50dB Signal 90dB Ref threshold 50dB Gain 0 10 20 30 40 50 60 70 80 90 Noise threshold level dB FIGURE 6 42 GAIN WHEN NOISE THRESHOLD LEVELS ARE VARIED THE SIGNAL AND REFERENCE THRESHOLD HAVE FIXED VALUES To avoid clipping or variable overflow gain limits is applied in the developed software The limits are especially necessary at low signal levels or high noise levels In these cases the gain calculation will calculate a large gain and maybe introduce clipping or overflow if the gain is not limited The minimum gain is limited to 1 because we don t want to damp the playback signal Implementation 6 3 7 GAIN SMOOTHING After a simulation of the system the program material and the recording at 50 Km h front microphone position as inputs in slices of 1 second with the noise detection and gain blocks put together the gains for each band were analyzed Frame no 1 fps FIGURE 6 43 GAINS FOR EACH BAND FOR SLICE OF 1 SECOND 50 KM H FRONT POSITION Important to mention in this simulation is that the s
86. mpensated signal filter transfer transfer level calculations smoothing equalizer function function FIGURE 6 25 BLOCK DIAGRAM OF THE LOUDNESS COMPENSATION SIGNAL IS PLAYBACK SIGNAL The main idea in the loudness compensation system is to compare the perceived loudness of the playback signal in a reference condition signal threshold block with the perceived loudness of the playback signal in the noisy conditions noise threshold block With help from this comparison we want to calculate a gain gain calculation block which can be applied to the playback signal in noise condition octave band equalizer thus the loudness in the reference condition is equal to loudness in the noisy condition It is somehow a signal to masker ratio comparison To avoid too rapidly changing in the gain a gain smoothing block is applied The input signal is a slice with a chosen length and for every slice all calculations are repeated This gives an iteration time and averaging of the input signals depends on the slice length The transfer function blocks are applied because the calculations in the threshold level blocks are based on signal levels at cochlea at listener position in the car From 5 4 Noise in the car we know that the noise is louder in the lower frequencies than the higher frequencies The lower frequencies in the signal will therefore more often be masked than the higher frequencies Due to this we have decided to divide the loudness calcula
87. n changes 5 5 1 4 COMPENSATE FOR MEASUREMENT DIFFERENCES GAIN AND DELAY Front position the recording in front position without engine converted to Pa was compared with program material convoluted with the transfer function in the same position h n p with calculated delay A zoom around 10 seconds shows that the simulation is delayed compared to the recording Could have been caused by delay from loudspeakers to recording position differences in software when recording IR or playback material software processing of data output vs input delay of sound chain etc and that the simulation has a higher amplitude as expected from the RMS values above T T T T T T T WA ih NN MM MU Mt y TN i L Le a i 10 55 10 60 10 65 10 70 10 75 10 80 10 85 time 0 2r 0 H FIGURE 5 17 COMPARISON BETWEEN RECORDING AND SIMULATION AROUND SECOND 10 FRONT MICROPHONE POSITION ENGINE OFF To calculate the delay we took the car dimensions Parkers Internetautoguide and calculated the distance from the loudspeakers to each microphone position dim and then computed the time based on the speed of sound in air at 20 degrees Celsius c d 5 5 delay im Analysis Then the corresponding number of zeroes was added in the beginning of the transfer function based on the sampling frequency when calculating the transfer function 44100 samples s zeroes delay fs 5 6 Therefore th
88. ncies differences in this particular velocity For graphs for other velocities see DVD Extra Docs Noise comparison xlsx The average error was computed for 0 1 time slice Band Hz Velocity O0 Km h 50Km h 80 Km h 110 Km h 31 2 61 2 18 3 68 2 38 63 2 33 2 96 2 77 3 23 125 5 24 3 88 3 08 3 64 250 3 72 2 28 3 97 3 09 500 3 61 3 32 3 47 3 08 1000 21 32 7 48 1 79 2 91 2000 23 91 11 6 4 89 2 86 4000 388 19 89 5 41 2 39 8000 2 27 12 22 6 55 5 92 16000 4 4 89 4 66 2 58 TABLE 6 6 AVARAGE ERROR OF THE NOISE ESTIMATION SLICE LENGTH 0 1S THE VLAUES ARE IN SPL As can be seen the average error drops for smaller time slicing expected behavior since the algorithm is dip listening the noise within small pauses in the program material Taking into account the noise distribution see 5 4 Noise in the car which is concentrated mostly in the lower frequencies and also the error of the estimation the estimation seems reasonable enough 6 2 5 MICROPHONE POSITION FOR NOISE ESTIMATION One recording position should be chosen Since the on line recording will be used solely for noise estimation the positioning of the microphone should best estimate the noise in the car as close as possible to velocity noise floor around the listener s head and should be robust enough to playback material and car velocity It should be noted that only the noise estimations given in SPL for each
89. ncy domain Two different tests have been done First a convolution of the impulse response computed with the function HTTF and dirac delta is studied Secondly a high frequency resolution study of the impulse response behavior is made The tests are made in a python script DVD Codes Python codes Analysis Test_HTTF py Implementation A dirac delta is built for test the filter As it is known a Dirac delta is defined in time as _ i x 0 6 18 6 o x gt 0 Dirac delta has a flat response for the entire freguency domain and it is the identity element for convolution An impulse response filter is generated with a function called HTTF created in python in which can be found in the script DVD Codes Python codes Head_and_Torso_Transfer_Function HHTF_Project py The length of the filter is 3000 samples including a delay of 250 samples The filter is convoluted with a Dirac delta mentioned before A FFT is applied to the result of the convolution in order to know the freguency response As we know H x 8 x H x 6 19 Therefore the expected frequency response of the convolution should be the gain computed in section 6 3 2 3 Diffuse field to cochlea transfer function The result of the convolution and its frequency response is shown in Figure 6 35 As can be seen the behavior of the filter in frequency fits with the desired gain 07 Convolution with 3000 samples with 250 samples delay impulse response Time and Frequency domain 0
90. nd tweaks The main reason why these directions were not investigated is the lack of time or the possibility to move the project away from its scope 8 4 1 INCREASING THE AMOUNT OF OCTAVE BANDS One of the most natural improvements is to analyze the signals into more bands one third one sixth etc octave band This would make the equalization smoother and would address better the masked frequencies Also this would determine a better analysis of the noise and could be used to improve the noise estimation by off frequency listening of the estimation block However this increase in octave bands cannot be done without a cost the design of the octave filters will raise additional problems and down sampling will be mandatory raising the computational complexity of the system 8 4 2 INVESTIGATE OTHER LOUDNESS MODELS If existent other loudness models should be investigated and plugged inside the loudness compensation system The tested model was tested only for pure tones and not complex tones which is the usual playback signal inside a car The modularity of the developed system allows us to easily plug in such models for gain calculation unless additional inputs are required 8 4 3 APPLY TEMPORAL FORWARD AND BACKWARD MASKING In the project only simultaneous masking was taken into account If possible this can be extended to forward and backward masking which will be taken into account when gains are calculated 8 4 4 Conclusion SYSTEM
91. ng on the loudness of the played tone we expect higher gains in the adjacent bands but lower than the one in the main gain For a tone of 250 Hz we observed the following results the application was playing the Coldplay period Test and results E Loudness compensation la al X FIGURE 7 2 NOISE TONE OF 250 HZ COLDPLAY AS PLAYBACK The gains are behaving as expected We also checked for all octave center frequencies and the gains looked similar to the ones in Figure 7 2 The results for this raw test seem reasonable and we proceeded to the on line car test 7 2 ONLINE TEST OF LOUDNESS COMPENSATION SYSTEM IN CAR In order to test and later validate the behavior of the developed loudness compensation system measurements are performed in the car and results are recorded using a dummy head 9 1 4 Measuring final result The program material is played while driving at different velocities and the dummy head is recording the performance of the implemented loudness compensation system To be able to judge the loudness compensation system recordings are also performed with the loudness compensation system off We are therefore able to compare the recorded program material when loudness compensated on and off The recordings are on the DVD Measurements Final Dummy Measurements Conclusion 8 CONCLUSION In order to restore the original apparent loudness of music material when listening in the presence of background noise in th
92. ns and refractions are present Usually used for rooms with low normal absorption enclosures etc A diffuse field seems to be the most suitable scenario in a car cabin The American standard ANSI S3 4 2005 describes this influence as the difference of the sound pressure level in the eardrum and the sound pressure level measured in the diffuse field in the absence of a listener A transfer function H1 in third octave bands in the audible frequency range 20 Hz 20000 Hz is given where the values correspond to the difference mentioned before Therefore if a compensation due to this factor is wanted a sum of this values in SPL should be applied to the spectrum in SPL as well Eardrum SPL Diffuse Field SPL H1 SPL 6 10 H1 SPL Eardrum SPL Dif fuse Field SPL 6 11 In order to have a better resolution an interpolation is done between the data points given in this curve Since in this project this transfer function will be computed in time a signal in time domain with the characteristic frequency spectrum of the compensation curve should be computed in order to convolute this signal with the input that needs to be compensated The signals which we are going to apply the compensation are expressed in pressure Pa therefore a gain function corresponding to the compensation curve given in dB in the ANSI S3 4 2005 is computed in pressure Pa The result in frequency domain will be a multiplication of the characteristic spectr
93. ntioned before Once the IFFT is computed the output is circularly shifted rolled by half of its length in order to get an impulse response which contains the desired frequency response information and then the amount of undesired samples are removed from the extremes of the symmetric impulse response A FFT of the rolled output is computed and compared with the desired frequency response is shown in Figure 6 31 and Figure 6 32 Amplitude Pa 0 500 1000 1500 2000 2500 3000 FIGURE 6 31 IMPULSE RESPONSE 3000 SAMPLES Implementation 25 FFT of the impulse Response symmetric FFT of the impulse Response Desired Gain 2 0 Amplitude 0 5 0 0 10 10 10 10 10 Frequency Hz FIGURE 6 32 FFT OF IMPULSE RESPONSE 3000 SAMPLES As can be seen in the figure Figure 6 32 the response in frequency of the impulse response is very close to the theoretical gain It is worth to clarify that Figure 6 32 shows the positive frequency range of the FFT due to the logarithmic scale of the plot 6 3 2 5 2 DELAY Once the impulse response is built it is decided to include as less amount of delay as possible Delay is considered the samples before the main peak on the impulse response which is in the center of the impulse response respect to time The procedure to do so is First a window of time samples has to be fixed then the maximum peak of the impulse response is moved to t
94. odel use octave band noise the model fits well to our solution implementation We are also using octave bands and a gain factor is calculated for each of them In total 10 gain factors are calculated To calculate the gain we compare the perceived loudness for the playback signal in reference conditions signal threshold level with the perceived loudness for the signal in noise conditions noise threshold level The loudness in reference conditions is Pres kU Log 6 28 Where is the playback signal intensity and l e is the noise intensity threshold in reference condition k and n are constants described in 5 6 2 Partial masking of loudness The loudness in noise conditions is Wnoise k I c oigo 6 29 Where is the playback signal intensity and I ise is the noise intensity threshold in noise condition the estimated noise level in each band We want the loudness for reference and noise condition to be equal and it is therefore necessary to multiply a gain factor with the signal intensity in 6 29 The gain factor is multiplied with the playback signal intensity because we are able to adjust the signal intensity in practice 6 28 and 6 29 with the gain factor can then be combined kd x ref rZ k a 5 Gain 7 Lose 6 30 Isolating the gain 027 Ie es Lo or repo gx a es Be M 6 31 I The gain can now be calculated using signal intensity reference threshold intensity and noise threshold intensity Final
95. ome of the developed system sub blocks Conclusion e Loudness calculation a model of loudness taking into account noise in octave bands exists e Octave band equalization the equalization is done in octave bands with a gain for each e Feedback loop the time domain computation of such octave bands help the system when different parts of it are not exactly synchronized in time This is an important asset since the phase information of the filters in the loudness compensation system became less important and could be ignored without serious concerns Such an analysis raised additional issues for other sub blocks e Transfer function sub blocks for a given transfer function an exact method to apply them for an octave band input could not be done e Noise estimation the values used for the noise estimation represents a quantitative description of the noise within a certain time frame and could be only used for a rough estimate Several transfer functions were measured for some recording positions and for listener s position where the listener s head would be located The method used to measure was by sweeps which was more suitable to our needs than an MLS method However an important asset of such a measurement method the signal to noise ratio for sweep measurement was not fully taken advantage of and a more powerful sweep signal could have been used Still averaging between multiple impulse response measurements was us
96. on 6 2 3 DECREASING SLICE SIZE The slice size has been decreased to see how the simulation is working for smaller slices different results are expected due to dynamic differences in periods which would pick up the noise floor in between the playback signal content the numbers will be seen in error graphs Different slice size error graphs will be presented 31 0 Hz 63 0 Hz 125 0 Hz 250 0 Hz Al 3 3 2 2 1 1 omomdoundouo v 5 a lt U S 9 ns e v gt a 7 a D Frame no 2 fps FIGURE 6 10 ERROR BETWEEN SIMULATION AND RECORDING FOR EACH SLICE SLICE SIZE 0 5S FRAME PER SECOND FPS 2 By analyzing individual bar frames we can see that the small slices captures more music s dynamics and thus the simulation goes below the noise floor at each peak in the graph graph depicting a 0 1 second slice of speech S 708 EE Levels signal raw 80 E Levels signal noise Vio H263 0 H225 0 50 0 HEDO O H2 0 kHz2 0 kHat 0 kHzB 0 kH26 0 kHz Frequency log Hz FIGURE 6 11 BAR FRAME OF LEVELS OF SIMULATION AND RECORDING FOR SECOND 708 OF PROGRAM MATERIAL LEVELS SIGNAL NOISE RECORDING LEVELS LEVEL SIGNAL RAW SIMULATED LEVELS A video was made with all the bar analysis for slice size 0 1 s see DVD Video Recording vs Simulation Front 12dB 0 15 slice wmv Implementation 6 2 4 NOISE EXTRACTION Based on the measurements SPL value playback signa
97. on of the system a couple of transfer functions need to be measured The transfer functions were done using the software Holmimpulse The measurement for additional information see 9 1 2 Car transfer function measurements was done using a logarithmic sine sweep of 27 samples We chose sweeps over MLS for several reasons M ller amp Massarini 2001 e Sweeps perform better when it comes to distortion MLS signal has a square wave shape which cannot be tracked exactly by the loudspeaker and time variance e Sweep measurement has a better signal to noise ratio than MLS an important asset in our case because we want to measure outside in a quite noisy environment The chosen length of the excitation signal was gn samples for all transfer function measurement which for a 44100 Hz sampling frequency represents 1486 ms enough to capture the low frequency reverberations in a cabinet like a car cabin The recording was set to record an extra time of 1500 ms again more than enough for the high frequencies reverberation time A block diagram of the measurement is depicted in Figure 5 13 Out and In are processed and presented by the Holmimpulse software The test loop was done to check the system For more and additional details about the setup see 9 1 1 Verification of measurement setup uw Om te p re cae ME Microphone n PY AD y FIGURE 5 13 TRANSFER FUNCTION MEASUREMENT OVERVIEW THE TEST LOOP IS FOR VERIFICATI
98. oudness compensation system were tested individually for their correct behavior Test and analysis of the implemented system s gains showed that they behaved as expected For an objective evaluation of the system we need a quantitative value to compare the loudness of a playback signal in a reference condition and the loudness of the same signal when the loudness compensation system is running inside the car However the only loudness model found able to compute such a value is the model used inside the loudness compensation system Obviously we would not gain much from such a test since the system was designed to work according to this model and such a test was done individually on the gain sub block Other loudness model could have been used to test the system but it would have meant the comparison between the implemented model and the new one An objective way to evaluate our system would be the analysis of the gains for each octave band However this will not be an evaluation of the perceived loudness but an evaluation of the system s correct behavior 8 3 SUBJECTIVE JUDGMENT OF THE LOUDNESS COMPENSATION SYSTEM When comparing the binaural recordings of program material played in the car with and without the loudness compensation system it is easy to hear that the loudness compensation system increase the levels in some octave bands when noise is present Let us first describe how we perceive the loudness of the program material when the loudne
99. ped for use in practical situations A basic structure for loudness models Figure 5 19 proposed by Moore Moore 2012 contains 4 blocks to calculate estimate the perceived loudness First step is to filter the stimulus according to the outer and middle ear transfer functions and then transform this to excitation pattern The excitation pattern can be transformed to specific loudness and then the perceived loudness can be calculated This structure is used in the ANSI S3 4 2005 for calculation of loudness of stationary sounds Calculate area Stimulus Fixed filter for Transform Transform under specific mS transferofouter spectrumto excitationto pe EG loudness middle ear excitation pattern specific loudness pattern FIGURE 5 19 BASIS STRUCTURE FOR LOUDNESS MODELS MOORE 2012 Because the mechanisms underlying the perception of loudness are not fully understood and the variation of ears and hearing across different people none of the models are able to calculate the true perceived loudness for one specific person Outer ears have different shapes and sizes as well as the middle and inner ears and due this and for sure other factors the perception of loudness will vary across different persons The loudness models are therefore estimations of perceived loudness for the average person Some better than others depending on input stimulus and purpose Some models are developed to estimate the loudness of statione
100. posed To understand the project it is recommended to read this part The report is divided into several smaller parts A problem formulation part where the problem is described and requirements and limits for the project are chosen An analysis part where theory and practical issues are discussed and analyzed An implementation part where the development and implementation of the problem are described And finally a conclusion If a fast overview is needed read the introduction problem formulation and the conclusion Appendices include further and deeper information about the project However the appendices are not mandatory for the project understanding Measurement journals references etc are placed in the appendices DVD includes Python and Matlab codes recordings equipment datasheets etc Documents which have low importance for the project or data which are not printable The DVD does also contain the report and the appendices as PDF References for used material are written in squared brackets with author surname and year of publication The same is applicable for webpages but only the page name is in the brackets A total list of references is available in 9 4 Appendix D References References to codes and other files on the DVD are written in italic 1 3 PROGRAMMING LANGUAGE MATLAB VS PYTHON The main programming language used in this project is Python Python is a high level programming language with a lot of possibilities and some
101. position The graphs in 5 4 Noise in the car show that the noise floor in the remaining three positions are close enough but an average of these positions was done to get an overall velocity noise floor in the car 3 6 9 1 No iseFloorpana ivelocity NoiseFloorpana velocity y mic position k 1 The values for the averages are found on the project s DVD Extra Docs Noise comparison xlsx amp Noise Floor deviation xlsx The results are depicted in Figure 6 22 Figure 6 23 and Figure 6 24 Velocity Noise Floor Ear Level Position fe Front Position Frequency Hz FIGURE 6 22 ESTIMATION VS AVG VELOCITY NOISE FLOOR IN THE CAR 50 KM H 0 1 S SLICE Implementation Velocity Noise Floor Ear Level Position f Front Position Frequency Hz FIGURE 6 23 ESTIMATION VS AVG VELOCITY NOISE FLOOR IN THE CAR 80 KM H 0 1 S SLICE 110 Km h Velocity Noise Floor Ear Level Position fe Front Position a m un o mo o lun aN Frequency Hz FIGURE 6 24 ESTIMATION VS AVG VELOCITY NOISE FLOOR IN THE CAR 110 KM H 0 1 S SLICE 6 2 5 5 CONCLUSIONS We can conclude that the best microphone position for noise extraction is the front position Regarding the front position and based on the previous section comparison conclusion we can say that e The estimation can also be used for other listeners e Thetransfer function measured at this position is less susceptibl
102. quencies depend on the playback level and frequency content of the playback signal An important task in this project is to predict the loudness of the playback signal in reference condition and compare it with the predicted loudness of the playback signal played in the car When we know the differences we can restore the original apparent loudness To predict the loudness the loudness models are used We investigated different loudness models which more or less are able to predict the loudness of a signal but they all have one common problem for this project point of view They don t take into account noise and we are therefore not able to predict the loudness of the playback signal played in the car with these models To solve the problem we used a loudness function developed by Lochner amp Burger 1961 and then adapted the model to our own needs The model calculates the perceived loudness in octave bands based on playback signal SPL and the noise SPL in octave bands by taking into account simultaneous masking Temporal forward and backward masking are not taken into account in this model We are now able to predict the loudness in reference condition and in the car but the function is not perfect The function by Lochner amp Burger 1961 is based on tests using pure tones as signal and the playback signal we are using is complex tones The function is also only confirmed valid in the frequency range from 200Hz to 8kHz and we want to use it in the
103. queo aeree EE aeo eee lech ves ne ee E REA dc 70 Test ANG resul oer eu piter pp en ep RR MA poda CYNT FU gere eerie IM GAFN Nep pede GAFOD CU Ey era to ee Uo qun Reve can DO zado 73 7 1 Online test of loudness compensation system in laboratory uuu ii LLALL LLALL A ALI I nnne nni 73 7 2 Online test of loudness compensation system in Car ccconococconcnncnonononanononcnnnanenonnnnncnnonnonnnnnnnnnnnnnnnonnnnnnnncnnnnn nos 74 eon il 75 8 1 The loudness compensation system and it s behavior cccoccoooooncnncnonononnnnnonccnnnnonnnnnoncnncnnononnnnncnncnnennnnnnncnnnns 75 8 2 Objective judgment of the loudness compensation system LLALL ALL AIII LL enne nnn nnne nnne 77 8 3 Subjective judgment of the loudness compensation system esses AIALL RL LA ener nnns 77 8 4 Further developmient s n ec tr ER EE A TE YF AS 78 8 4 1 Increasing the amount of octave bands ccoconcococcnncnonononnnnnnncnnnnnonannnnncnnnnnnnonnnnnnnnnnnnnnonnnnnnnannnnonnnnnnncanananons 78 8 4 2 Investigate other loudness models cccesesssssscecececeesesnscecececeeseeeeesececesseeeeeseeeceeseseeaeseescessesesaeseesenseeseea 78 8 4 3 Apply temporal forward and backward masking nennen ener 78 8 42 System rmproVemerits uo 3 v te cons cs ines aae eee gn eaa dd A dace GW cous A ines ave du eee aes deese GG 79 8 4 5 Improved noise estimation system eeu ui i YY LALLL LLI EL LAI enne nennen nannten tires assa sense te sa
104. r each slice are the FIR convolutions On the graph whenever the sum of all the graphs are around the slice length 1 s around because only the most important blocks were timed the playback stalls and waits for processing to finish Several solutions exist shorten the length of the impulse responses code optimization with C code writing inside the main loop collapse two transfer functions into one by convolution and truncation or move to multithreading multitasking The chosen solution was to use multitasking and a pool of 4 workers was spawned on the mentioned machine The main loop was parallelized as much as possible with as long lines branches as possible to minimize inter process communication overhead The chosen three branches entailed in parallelization are graphically depicted on Figure 6 58 blue background boxes The test was repeated on the same machine under similar conditions and there were no playback stops were heard for 1 s or 0 1 s slice length The parallel threads were joined and additional sequential processing remaining blocks from Figure 6 58 was done for each slice as depicted in Figure 6 60 Worker Pool ze Parallel Playback Branch 1 Sequential I processing Seguential Parallel processing Branch 2 Parallel Branch8 GUI FIGURE 6 60 THREADING OVERVIEW Since the main looped had to wr
105. ra Docs Noise comparison xlsx This velocity noise floor will be noted as NF velocity where i represents the octave band and the velocity is the car velocity in Km h 6 2 5 2 NF veocry DEVIATON We now want to study the deviation from NF velocity for each recording position For this the difference between Noise estimation iperioa j and NFivelocity Was computed for a fixed velocity ETTOTpana Lperiod j abs Noise_estimationpang iperiod j NF velocity 6 6 In order to work with less amount of data an average across periods of course except the NF velocity was computed for each band i at a fixed velocity 6 7 7 1 AVG_Deviation velocity 72 Erroryana period j j 1 An example for such a computation 0 1 s slice velocity 80 Km h recording position at ear level front of listener is shown in Table 6 7 and Table 6 8 Period Band Pink NF g0 Speech Opera Pop Hard Classical Electronic Noise Rock Rock 31 83 14 83 09 83 13 83 28 83 53 82 69 82 98 81 18 16000 48 86 37 26 38 52 36 47 38 06 40 95 36 64 36 9 TABLE 6 7 SPL FOR EACH PERIOD TIME SLICE 0 1S VELOCITY 80 KM H The Deviation is then Band i Error Error Error Error Error g Error Errorig AVGigo 31 0 05 0 04 0 19 0 44 0 4 0 11 1 91 0 45 16000 11 6 1 26 0 79 0 8 3 69 0 62 0 36 2 73 TABLE 6 8 ERROR AND AVERAGE DEVIATION SPL FOR EACH PERIOD TIME
106. raged in time for the same position M ller amp Massarini 2001 tae h n h3 n 5 1 Such an averaging is depicted in Figure 5 16 with a zoom around 1kHz transfer function from DU to DU Analysis 10 log10 abs FFT 38 txt 45F 39txt i i i i i i i i i i 140 250 400 600 800 1100 1600 2400 3200 3900 Frequency Hz Phase response degrees o 150 Average N 200 i i i i i i i i i i 140 250 400 600 800 1100 1600 2400 3200 3900 Frequency Hz FIGURE 5 16 TRANSFER FUNCTION AVERAGING AMPLITUDE AND PHASE RESPONSE FOR MEASUREMENT 37 38 AND 39 5 5 1 3 CONVERTING IR H PU to H 25 DU Pa The desired transfer function is from output DU to Pascals For this requirement the recorded DU will be referred to as v measured in 9 1 Appendix A Measurement journals for microphone placed inside the calibrator will be used to convert any DU to its corresponding Pa value Because this value was normalized to 1 0 079 DU corresponds to 22 05 dB care must be taken in the digital signal s representation the conversion will be done dependent on maximum value of the signal with which the impulse response will be convoluted y n signal h n 5 2 Since the signals will be loaded from wave files in 16 bit signed integer format with a maximum of 32767 the new transfer function will be calculated as h n py h n py 5 3 type signal maxValue v pu 32767 0
107. re the original apparent of loudness A loudness model alone or other theory will be hard to judge and analyze if they are just formulas or a piece of code 4 2 SPECIFICATIONS Even though the main objective for this project is to investigate how to restore the original apparent loudness of music we have chosen to have big focus on development of a loudness compensation system to be able to better understand and evaluate the mentioned objective Before defining the specification we define some general terms that will be referred to throughout the report e Playback signal represents the signal played through the tested car audio system e Program material represents wave file containing various chosen playback signals mixed together used for testing the loudness compensation system to be developed e Period represents a part of approximate length of 30 s in the program material containing the same type of material e g pink noise speech electronic music etc Specifications for the loudness compensation system have been decided These are e The loudness of playback signal shall sound equal no matter the noise e The system shall allow user settings e g volume and equalizer settings If the user likes loud bass levels the loudness compensation should not overrule this user behavior e The developed system shall be able to perform online loudness compensation in a car Not only simulations Problem formulation 4 3 LIMITATIONS
108. red to as raw signal gained by applying the smoothed gains The raw signal gained and the recording signal are then fed to the noise estimation block e The raw signal gained is modified as it should sound in the recording position e One Octave Band Filters OBF is applied to the modified slice resulting in 10 levels e The recording signal is converted to Pa e OBFisapplied to the recorded signal in Pa resulting in 10 levels e Noise levels are extracted for each band e Noise levels are converted to noise levels at cochlea level e These levels are fed to the Noise threshold level block that calculates masking threshold from the noise levels result referred as noise levels cochlea The raw signal is modified as it should sound at listener s cochlea level OBF is then applied to this modified raw signal resulting in 10 levels which are processed through the signal threshold Level block that calculates masking thresholds within the signal itself not lower than the hearing threshold at cochlea level result referred as signal levels cochlea Both the results noise levels cochlea and signal levels cochlea are fed to the loudness compensation block resulting in the gains for each band of the OBF referred to as gains These 10 gains are fed into the gain smoothing block resulting into 10 smoothing ratio gains referred as smoothed gains for the equalization of the next slice The current slice is gain equalized with the smoothed gains an
109. rmation _ 22 1 _ z 1 6 35 al Bg Where fsg is the sampling frequency of the gain which in our case is the number of slices per second which will be referred to as the number of frames per second 1 _ fs 6 36 slice lengthi i Slice SiZejsamptes fps Implementation The discrete system will be z w 2zo o 6 37 H z n e n n VO CN SP nd EU S F 5 6 38 z 2z 1 4 Se 2 GE 205 eos ie ams GE 209 Because the equalization needs to be done in real time either the equalization will be done each k number of slices a minimum of 3 for the filter to be applied or the playback material slice will be subsliced into k number of subslices again a minimum of 3 subslices must be applied The second option was chosen and then a new parameter was defined for the smoothing algorithm smoothing ratio which represents the number of subslices in each slice It represents the granularity of the smoothing Due to this fact the sampling frequency of the gain changed by fps lt fps smoothing ratio Also the sampling frequency of the input had to be adapted to this by holding the input to the system the gain from the loudness compensation constant over smoothing ratio samples For instance if the gain for one band over 3 slices is 1 3 2 the input to the smoothing system will be modified as 1 1 1 3 3 3 2 2 2 length of 3 smoothing ratio A block diagram overview o
110. rter To avoid hard start of the amplifier which maybe damage to the DC AC converter let the soft start circuit inside the DC AC converter power up the amplifier In practice this task is performed when powering the power amplifier before the DC AC converter The amplifier will hum due to the non sinusoidal AC output from the DC AC converter Be aware of the DC AC converter cabinet temperature DC AC Battery converter Laptop Soundcard Power amplifier Speaker FIGURE 9 1 CONNECTIONS 9 1 1 4 EQUIPMENT SETTINGS Amplifier OdBgain using modified input on amp Soundcard Max output gain Minimum monitor gain 2596 input gain Laptop and Holmimpulse Asio4all drivers with 512 samples latency setting for soundcard Logarithmic sine sweep with 20Hz start frequency A Signal length M equal to 16 44A Khz sampling frequency 9 1 1 5 PROCEDURE 1 Usethe laptop with the software Holmimpulse to measure the car transfer function 2 Savethe results Appendices 3 Verify that the results are all right The frequency and phase response should be flat and the impulse response should be close to a dirac delta 9 1 1 6 RESULTS The measurements system performs as expected and wanted The frequency and phase response is flat between 20Hz and 20Khz and the impulse response is close to a dirac delta Appen
111. ry small so the decision is not expected to have a big influence in the behavior of the loudness compensation system As the octave band filters represents the rms value of the signal in an entire octave band frequency range it seems reasonable that the average influence of the frequency response within an octave band should be taken into account Thus we decided to apply an average of the combined transfer function frequency response for each band Implementation 6 3 3 NOISE THRESHOLD LEVELS The noise threshold level is the threshold level where a playback signal is just masked by the noise E g If the noise threshold level is calculated to 70dB a playback signal present in this noise has to be at least 70dB to be heard At 70dB the playback signal is just masked Due to the use of octave bands the noise threshold level is calculated for each of them This gives in total 10 noise threshold levels Each of the noise threshold levels are calculated using an auditory filter with a center frequency for the chosen octave band This means that noise in all bands can affect the noise threshold level for the chosen band The auditory filter is calculated using ANSI S3 4 2005 and for masking point of view the auditory filter can be mirrored upside down to present PTC Psychophysical tuning curves Moore 2005 First step is to calculate the filter shape which depends on the SPL In our case this is both the playback signal SPL and the noise S
112. ry sounds and pure tones and if these are used to estimate impulsive sounds with complex tones they fail The loudness models can be divided in two different groups Skovenborg 2004 A single band group which estimate the loudness in one band and a multiband group which estimate the loudness in several bands A single band loudness model could e g be Leq A B C D M RLB where A B C D M and RLB refers to different filter weightings and LARM by TC electronics A multiband loudness model could e g be the model by Zwicker ISO532B Moore ANSI S3 4 2005 and HEIMDAL by TC electronics The multiband loudness models are more complex than the single band because they divide the stimulus into several bands applying more filters and some of them even take into account masking Hence the multiband loudness models need more computation than the single band loudness models The question is now Which model is the best to estimate the loudness of music and speech the signals which are typical played through a car audio system Skovenborg 2004 have analyzed how good different loudness models estimate the loudness of music and speech These models are then divided into 4 groups where group 1 is the best Table 5 2 Class Models best in class listed first 1 TC HEIMDAL TC LARM 2 Leq RLB Leq C Leq Lin 3 Leq B PPM 50 Zwicker ISO Zwicker amp Fastl 9596 4 Leq D Leq A Leq M TABLE 5 2 LOUDNESS MODELS AN
113. sed as references e First the level of the calibrator recording was calculated according to 6 39 with the wave samples converted to Pa according to 6 40 and the result for the entire file was 93 98 dB SPL close enough to 94 dB SPL Test can be reproduced in test CalibratorSound of DVD Codes Python codes BandAnalysis test_DU2Pa py 100 SPL db 20 31 0 Hz 63 0 Hz125 0 H250 0 Hz500 0 Hz1 0 kHz 2 0 kHz 4 0 kHz 8 0 kHz16 0 kHz Frequency log Hz FIGURE 6 53 OUTPUT FROM THE 1 OCTAVE BAND OF FILTERS FOR THE CALIBRATOR RECORDING THE LOW FREQUENCY DEVIATION MAINLY DUE TO THE DESIGN OF THE LOWEST FREQUENCIES BUTTERWORTH FILTERS AND NOISE FLOOR Implementation 68 The level of the pink noise recording with the microphone in Position_back DVDMAudioVPink noise Back 0 wav was done and compared with the sound level meter linear measurement The SPL found was 79 24 dB SPL close to the measured 77 9 dB lin SPL with BK 2238 sound level meter Test can be reproduced in test pinkNoise of DVD Codes Python codes BandAnalysis test_DU2Pa py The output from the 1 octave filter bank can be seen in Figure 6 54 and resembles quite well a pink noise spectrum 8 31 0Hz 63 0 Hz 125 0Hz 250 0Hz 500 0HZ 1 0 kHz 2 0 kHz 4 0 kHz 8 0kHz 16 0 kHz Frequency log Hz FIGURE 6 54 OUTPUT FROM THE 1 OCTAVE BAND OF FILTERS FOR THE PINK NOISE Another test was done to analyze a small piece of recordin
114. sfer Functions Treshold hearingN Utils Car dimension Datasheets and Manual Noise estimation e Noise Detect Analysis o 0 1 sslice o 1sslice o Back listener position o Back Seat Nose height o EstimationN e Performance Test Appendices e MeasurementsV Oo o o o o Calibrator Recordings Car Measurements Engine noise measured at different mic positions Final Dummy measurements IR Measurements Setup for music playing and recording in car FL Studio e Program Material e Reportl e Videol Appendices Appendices 9 3 APPENDIX C DICTIONARY Down sampling re sampling to a lower fs DU Digital Unit s floating point values corresponding to samples of a digital signal with or without physical correspondence ERBn Equivalent rectangular bandwidth for normal hearing Fps Frames per second number of signal slices per second processed by the system eg a time slice of 100 ms is equivalent to 10 fps Fs Samplinf frequency of a continuous signal IIR Infinite Impulse Response Leq loudness equivalent Loudness compensation system The developed system for loudness compensation Nyquist frequency Fs 2 Noise Unwanted sound which will affect the loudness perception of the signal and maybe mask the signal The noise is e g wind and engine noise in the car Noise floor SPL level inside the car when the engine turned off and no sound is played through loudspeakers Only
115. similarities to Matlab The main reason why we chose Python as the main programming language over Matlab was due to the possibility to choose a part of this project to be a mini project in the course Scientific computing and sensor modeling The programming language for the mini project in this course had to be Python and to avoid a mix of Python and Matlab code which would make an on line system difficult to implement in this project we therefore decided to use Python Also Matlab is not intended for high performance computing making Python a better choice for multithreading and multiprocessing that can gain even more from GPU computing a field where Matlab still has some compatibility issues Preface 2 CONTENT 17 PACA A oY NAF FU WY ls dashed aster Does AAA FA 1 1 1 Acknowledg es on ld ERE ERN 1 1 2 Reading guiides tede tm e Yd Nd te FY eee ie edite eed ee eee etas 2 1 3 Programming language Matlab vs Python ccscccccsssscecsssceceesececsessececsssseceesseeeceessececsssseceesaeeecsesaececsseseeeees 2 3 le Lm 5 4 Problem formulation pere hd redet eth eoe beh anus be De ren a Fo neon Posh dades O Fr age 6 4 1 TR m eS OR eeu Ree eden des Ee Vol I ud b he A Rd ne RAN salen WR NINE 6 4 2 Specification Saudi GR eee eo e Hees etude Y o iot OA E IER Ru TYN 6 4 3 LimmitatiOn Ss eH UTE 7 DIMES Ta E 11 E PP HE E 8 5 1 Introd ctiOR iii reti rt Ud E RERO EGRE ER TER ERES ERE GERI FRAN XAR ER ERRARE ERSTE HER ER RE RE AER RR NR LX
116. ss compensation system is off In the case that the velocity is 50Km h only low frequency noise is present from the car which fits well to our noise analysis The program material at this velocity is clearly hearable but the low frequencies sounds weak and is in some periods masked Especially in the Pavarotti and Beethoven period which are highly dynamic periods the low frequencies are masked when the level is low When the velocity is increased to 80Km h the noise from the car becomes wider and introduces more masking We are still able to hear the program material but lot of information is lost due to masking Again it is the high dynamic periods which have the biggest Conclusion masking and are in some cases close to be totally masked by the noise Also the Trentem ller period is hard to hear because this period contains mostly low freguencies Increasing the velocity to 110Km h increases the noise levels and even more information in the program material are masked When the loudness compensation system is on it starts to gain the needed frequencies The low frequencies we needed when the loudness compensation system was off and car velocity at 50Km h is now hearable They are not gained much but improve the experience of the program material Larger differences are found between the loudness compensation system on and off at 80Km h and 110Km h When the loudness compensation system is on at these velocities it does not only gain the low frequ
117. t the expected normal activities in what has become today an indispensable comfort of our society Problem formulation EE 4 PROBLEM FORMULATION 4 1 OBJECTIVE The objective of this project is to investigate how to restore the original apparent loudness of music material when listening in the presence of background noise in the car The original apparent loudness is the same quantity an attribute of the auditory sensation to rate sounds from quiet to loud on a certain scale as in a chosen reference condition Moore 2012 In order to do this different signal processing techniques human sound perception and loudness models will be studied and finally we develop a system able to compensate for loudness of music played in a car and evaluate the performance of such system e g loudness compensation system To be able to listen to the performance of this system a recording of the loudness compensation system in action in a car will be performed with an artificial head This gives the possibility to subjectively judge and analyze the system behavior including the applied loudness model by only using headphones and the binaural recording through binaural reproduction Because we want to develop a loudness compensation system analysis and investigations is done from implementation point of view and implementation is therefore also a part in this report We need the loudness compensation system for best possible analyzing and judgment of how to resto
118. the behavior of the system but for the sake of simplicity all bands will receive equal treatment and a single smoothing filter will be used for all Implementation 6 3 8 OCTAVE BAND FILTER AND EQUALIZER According to IEC 61260 1995 the shape of each filter has to be designed within certain attenuation limits A total of two attenuation curves are given a minimum attenuation curve and a maximum attenuation curve for three types of filters class O class 1 and class 2 Figure 6 49 depicts these curves where the x scale is logarithmic An IIR digital Butterworth filter was chosen as the type of filter due to its spread and the computational speed involved in applying an IIR filter in polynomial coefficients b a form The lowest order of such a filter was found to be 3 and easily fitted inside a class O filter The comparison was done for only one filter the one with middle frequency at 1000 Hz The following frequency response Figure 6 49 depicts this where the x axis is logarithmic Ee Class 0 Min Attenuation 3 30 Class 0 Min Attenuation e Sao a Butterworth pass band of order 3 S 50 o 60 70 A 3 10 104 105 o zr Class 1 Min Attenuation 3 20 Class 1 Min Attenuation ber Butterworth pass band of order 3 50 oO F J 60 70 10 10 10 10 10 o Nd Class 2 Min Attenuation 3
119. tion As the transfer function to the eardrum this transfer function is going to be applied to the input signal in time domain so a filter should be derived from the characteristic spectrum The input signal is expressed in Pascals therefore the spectrum shown in Figure 6 27 should be transformed into gain for being applied to the input The gain is worked out as in 6 12 6 3 2 3 DIFFUSE FIELD TO COCHLEA TRANSFER FUNCTION Since two different transfer functions have to be applied to the input signal first the transfer function to take in account from diffuse field to the eardrum and then the middle ear transfer function with no intermediate computations between them a combined transfer function has been decided to be applied for computational efficiency Therefore just one convolution operation will be computed instead of two As the transfer functions shown in the sections 6 3 2 1 Diffuse field to eardrum transfer function and 6 3 2 2 Eardrum to cochlea transfer function are expressed with the same concept of SPL difference between two points the combined transfer function can be expressed with the same concept where the values of the transfer function is the addition of both H1 Diffuse Field Eardrum H2 Eardrum Cochlea FIGURE 6 28 COMBINED TRANSFER FUNCTION Where H1 Eardrumsp Dif fuse Fieldsp 6 14 H2 Cochleasp Eardrumgp 6 15 H H1 H2 Cochleasp Dif fuse Fieldsp 6 16 Y Xx H 6
120. tion and compensation into octave bands This gives us the possibility to only change the gain in the needed bands The function by Lochner amp Burger 1961 is also based on octave band noise and fits therefore well our decision The total loudness compensation can then be described as a multiband loudness model based on Lochner amp Burger 1961 which output a loudness compensated signal in slices The model only takes simultaneously masking into account and averaging the signal due to slicing When phenomena s like forward or backward masking is present due to e g passing car or highly dynamic signal the system is not expected to compensate correctly Compression effects or pumping are expected when a signal is played at low levels but depends on the length of the signal slicing iteration time gain smoothing and the dynamics in the signal Implementation 6 3 1 SIGNAL TO DIFFUSE FIELD TRANSFER FUNCTION The signal to diffuse field transfer function is the measured car transfer function for microphone position front 9 1 2 Car transfer function measurements 6 3 2 DIFFUSE FIELD TO COCHLEA TRANSFER FUNCTION 6 3 2 1 DIFFUSE FIELD TO EARDRUM TRANSFER FUNCTION This chapter is based on the American Standard ANSI S3 4 2005 In this paper we can find two kinds of transfer functions depending on the characteristics of the sound field e Free field No reflections with a frontal incidence of the sound source e Diffuse field Reflectio
121. towards the listening position with treble units at ear level The listener should be positioned symmetric in the room and 2 5m to 3 5m away from the line connecting the speakers No listener should be placed closer than 1m to a wall and 2m to a speaker In the case the playback signal is played in a car the reference conditions are almost not existent and impossible to fulfill The following will give rise to problems in the car e Reverberation time and reflections due to non uniform distribution of absorption material Soft seats and panels and hard windows e Comb filtering and strange frequency response due to the small cabin e Speaker listener position None of the distances can be obtained e Noise floor The first 3 points are due to the cabin size and cabin arrangement and because of the limits chosen for this project these will not be discussed further We will instead focus on point 4 which is due to noise from engine wind etc Analysis ENIM 5 3 MEASUREMENT SETUP In order to do measurements in the car we need a setup which consists of mainly an amplifier speakers microphone a sound source and various equipment needed for specific measurements For the speakers and amplifier we could use the car audio system which is already installed in the car or we could add our own setup The advantages of using the existing car audio system is that everything what we need is installed in the car and ready for use but the disadv
122. uipment except speakers microphone and laptop are placed in the trunk of the car Figure 9 8 The speakers are placed on the backseats pointing up and the listener in between The used car has actually 3 rows of seat where cars normally only have 2 To handle this difference the second row of seats in the used car was not used The microphone is placed 11 5cm from roof pointing down 72cm from windows and 69cm from seat Car driver person Laptop driver person Microphone Speaker L Speaker R Listener Dummy head Equipment FIGURE 9 8 EQUIPMENT AND MICROPHONE POSITION IN THE CAR 9 1 4 4 EQUIPMENT SETTINGS Power amplifier 20dB gain on amplifier using modified input with static gain Soundcard 7596 output gain 7596 input gain 9 1 4 5 PROCEDURE 1 Adjustthe output level in the loudness compensation system software while playing pink noise to 70 8dB Appendices METRI A weighted or 77 9dB linear using the SPL meter The loudness compensation system shall be inactive 2 Calibrate the input level using the calibrator and DVD Codes Python Codes Project_main calibrate_recording py The level in the software shall correspond to the level from the calibrator 1Pa or 94dB ref to 20uPa 3 Enable the loudness compensation system and record using the recorder while driving Repeat the program material and recording for each velocity Table 9 9 and Table 9 10 4 Redo point 3 while the loudness compens
123. um of the gain curve and the spectrum of the signal in which the compensation want to be applied Thereby a SPL value of 0 dB corresponds to gain of 1 The formula to work out the gain is shown in 6 12 Gain 10GPL 20 6 12 Where SPL is the amount of dB that we want to increase The resolution in the interpolated curve depends directly on the desired number of samples of the impulse signal returned by the function A plot of the transfer function from diffuse field to Eardrum given in SPL in the ANSI S3 4 2005 with cubic interpolation is shown in Figure 6 26 Implementation Transfer Function to the Eardrum Third octaves Values ANSI S3 4 2005 Interpolation Values 14 SPL dB 10 10 10 Frequency Hz FIGURE 6 26 DIFUSE FIELD TO EARDRUM TRANSFER FUNCTION SETS OF THREE ADJACENT VALUES IS SHOWED IN THE RED BOXES The curve given in the ANSI S3 4 2005 standard has been expanded from 0 to 22050 Hz As no information in 0 20Hz and 20000 22050Hz is known the gain for this frequencies has been fixed to 1 this means no change in the output for this frequencies In addition once the convolution is done a spectrum analysis from 20 20000 Hz will be the most suitable frequency range for study 6 3 2 2 EARDRUM TO COCHLEA TRANSFER FUNCTION The aim of this section is to get the filter to apply to the input signal to simulate the behavior of the middle ear concretely from eardrum to the cochlea This
124. urves are plotted in Figure 6 52 Frequency response of filters db 60 iV j WY y Ny V Designed filters EC min attenuation IEC max attenuation Y N Frequency Hz log scale FIGURE 6 52 FILTER RESPONSE AMONG WITH THE IEC CURVES THE FIGURE CAN BE OBTAINED USING PLOTFILTERBANK FUNCTION INSIDE DVD CODES PYTHON CODES BANDANALYSIS BAND_ANALYSIS PY A test was performed to see if the filter bank is working 1 2 3 4 The filter bank parameters was generated A 30 seconds wave file was input to each filter and convoluted The outputs were summed together Another wave file was created with a Left channel was the original 30 seconds song b Right channel was the summed outputs at step 3 No major differences were heard except the phase shift if both channels were played induced by the Butterworth filters The file is found on the DVD Audio Coldplay_Left_bef_Rigth_After wav and the code for generating it is found inside test sumOfFilters function of DVD Codes Python codes BandAnalysis test_octave_filters py Implementation 6 3 8 1 CALCULATING THE OUTPUT OF EACH FILTER According to IEC 61260 the output of each filter should be calculated in dB relative to an appropriate reference quantity Since the input to the noise threshold level block generated according to ANSI S3 4 2005 should be a value in SPL or Pa we c
125. velocity and microphone position therefore there will be a Velocity Noise floor for each car velocity and each microphone position Appendices 9 4 APPENDIX D REFERENCES ANSI S3 4 2005 Procedure for the computation of loudness of steady sounds American national procedure Florentine Popper amp Fay 2011 Mary Florentine Arthur N Popper and Richard R Fay Loudness 1st Edition Springer 2011 IEC 60268 13 Listening tests on loudspeakers International standard IEC 61260 Octave band and fractional octave band filters International standard ISO 389 7 Reference zero for the calibration of audiometric equipment European standard Lochner amp Burger 1961 J P A Lochner and J F Burger Form of the Loudness Function in the Presence of Masking Noise 1961 Moore 2012 Brian C J Moore An introduction to the psychology of hearing 6 edition Emerald group 2012 M ller amp Massarini 2001 Sven M ller and Paulo Massarini Transfer Function measurement with sweeps Director s cut including previously unreleased material and some corrections Skovenborg 2004 Esben Skovenborg and Sgren H Nielsen Evaluation of Different Loudness Models with Music and Speech Material Audio Engineering Society Convention Paper 2004 Holmimpulse Software for impulse response measurements May 2012 http www holmacoustics com holmimpulse php Genesis Matlab loudness toolbox May 2012 http www genesis
126. verage error for each piece in the program material Avg err 31 0 Hz 8 os Avg err 63 0 Hz 5 20 Avg err 125 0 Hz ISF Avg err 250 0 Hz Ly alr te i ed eer 100 150 200 Seconds 1 I o un je PRPNNWWSE OoOnononono ou reas roti isa in dB SPL dif for each band o u o m o o G j SH N o o N I o Seconds 1 1 1 NNWWS onono Re ouon dB SPL dif for each band ESSE St SSE 200 o u o m o o m a o Seconds 1 FIGURE 6 6 AVERAGE ERROR PER PERIOD The exact values are depicted in DVD Extra Docs Noise comparison xlsx 6 2 2 ANALYSIS OF DATA In the silence period the measurement picked only the noise floor while the simulation was constructed by convolution with zeros This accounts for the high average level in this period Figure 6 6 and for the maximums in Figure 6 5 when the periods change fade out fade in This explains the high levels of error for the 31 Hz band in many periods with little low frequency content like Speech Opera or Classical periods and it also accounts for some 63 Hz error This is also depicted in Figure 6 7 Figure 6 8 and Figure 6 9 where the red levels represent the recording s levels and the blue one the simulated ones analysis of seconds 60 and 200 corresponding to speech period and classical period respectively S 060 S 200 FCU Levels signal raw EXA Levels signal noise EX Levels signal raw
127. verview of the application depicted in the figure as Loudness compensation system is presented in Figure 6 56 with the soundcard omitted where the D A and A D converters are found System gain pre amp gain Microphone Recording slice Louanges Gained playback slice p compensation gt Power amp gt Speaker i system Sound source wav file Playback slice y FIGURE 6 56 OVERVIEW OF LOUDNESS COMPENSATION APPLICATION The Python module implements a graphical interface GUI that wraps and controls the logic behind the loudness compensation system A print screen of the GUI is presented in Figure 6 57 for user manual see DVD Extra Datasheets and Manual How to use the application pdf FIGURE 6 57 GUI OF THE MAIN APPLICATION Implementation 6 4 1 INSIDE THE MAIN APPLICATION Once the loudness compensation system is started the source file is loaded into memory and the loudness compensation system is initialized Transfer functions threshold levels internal variables output playback stream input recording stream etc Afterwards the slicing of the source file starts After each slice is loaded a slice of same length is taken from the recording chain from the soundcard buffers will be referred to as recording signal Then the system gain is read and applied to the slice taken from the program material will be referred to as raw signal The raw signal is equalized refer
128. ysis 0016 40 41 42 0 008 0 006 0 004 0 002 0 000 0 002 0 004 0 006 0 008 Amplitude 40 txt file 0 2000 4000 6000 8000 10000 Holmelmpulse sample number 0 008 0 006 0 004 0 002 0 000 0 002 0 004 0 006 Amplitude 41 txt file o 2000 4000 6000 8000 10000 o 0 008 Holmelmpulse sample number _ 0 006 5 0 004 ci 0 002 2 0 000 0 002 0 004 lt 0 006 1 o 2000 4000 6000 8000 10000 Holmelmpulse sample number FIGURE 5 14 EXAMPLE OF A TRANSFER FUNCTION CUT MEASUREMENTS 40 41 42 FROM 9 1 2 CAR TRANSFER FUNCTION MEASUREMENTS Y AXIS IS THE DIGITAL UNITS DU TO DU AS MEASURED BY THE HOLMIMPULSE SOFTWARE THE BLUE LINE REPRESENTS THE MEASURED TIME IR AND THE RED DOTS REPRESENT THE CUT SAMPLES FROM THE MEASURED IR A typical logarithmic time response of the IR would look like Figure 5 15 and it can be seen that the chosen right side cut marked with a red star falls inside the noise floor Impulse Response CUT 37 10 loglO abs h Dek Right cut after 3000 samples 10 log10 abs Normalized impulse response o 2000 4000 6000 8000 10000 samples FIGURE 5 15 EXAMPLE OF A LOGARITHM OF A TRANSFER FUNCTION H IS THE IR IN TIME 5 5 1 2 AVERAGING In order to reduce the effect of the noise floor more measurements three were done for the same position of the microphone and the windowed impulse responses were ave
129. ystem did not have any feedback playback signal was not changed from slice to slice only the gains calculated for each slice thus the gains should reflect a combination of the velocity noise floor and periods dynamics If this were done a new recording would have been needed with the recorded change into the playback signal Also the gain was capped to 15 as a maximum and to 0 as a minimum value Same analysis was done with the same parameters except the slice size The gains for 10 slices per second Frame no 10 fps 4 4 1 Frame no 10 fps i i 4 1 AR ay ALL i iet 7 7 a po r8 E A n gt n TSELE 0 500 1000 1500 2000 2500 Frame no 10 fps FIGURE 6 44 GAINS FOR EACH BAND FOR SLICE OF 0 1 SECOND 50 KM H FRONT POSITION It can be seen that because of the playback signal and the noise estimation the gains oscillate a lot and the compensation would be unpractical distortions would appear and the playback would be at least unpleasant Also a simulation with the gained program material was done for 1s slice and O 1s slice which can be found on DVDYAudio gained music 50km wav and gained music 50km O0 1s Slice wav Implementation Therefore a smoothing method was considered which should be done on line For this purpose a second order system was applied to each gain corresponding to each band For a second order continuous system defined as o2
130. z 8 0kHz 16 0 kHz Frequency log Hz FIGURE 6 39 ILLUSTRATION OF THE NOISE THRESHOLD LEVEL CALCULATION FOR THE 1KHZ OCTAVE BAND USING AUDITORY FILTER THE BARS ARE NOISE IN OCTAVE BANDS IN THIS CASE THE NOISE THRESHOLD LEVEL IS AFFECTED BY THE NOISE IN THE 31HZ OCTAVE BAND ALL OTHER NOISE BANDS ARE BELOW THE AUDITORY FILTER WHEN THRESHOLD SHIFT IS TAKEN INTO ACCOUNT AND DO THEREFORE NOT AFFECTS THE NOISE THRESHOLD LEVEL Implementation 6 3 4 SIGNAL THRESHOLD LEVEL The signal threshold level is the level where the playback signal masks itself or is not hearable due to the hearing threshold therefore the maximum between the hearing threshold and the masking threshold level within the signal itself is taken The calculations for signal hearing threshold level are almost equal to the noise threshold level calculations The difference is that the signal hearing threshold doesn t take into account the signal SPL in the octave band where the signal threshold level is calculated We are doing this to avoid that the chosen octave band masks itself and only the other bands affect the signal threshold level Figure 6 40 illustrates this The signal threshold level cannot be lower than the hearing threshold for the chosen band Signal SPL 90 Auditory filter Threshold shift 80 70 60 Signal threshold level 20 10 31 0 Hz 63 0Hz 125 0Hz 250 0Hz 500 0Hz 1 0kHz 2 0 kHz 4 0 kHz 8 0kHz 16 0 kHz Frequency log Hz FIG

loudness compensation of music in a car audio system

Contents

Download Pdf Manuals

Related Search

Related Contents