Different approaches for the equalization of automotive sound systems

Transcription

1 Auio Engineering Society Convention Paper Presente at the 112th Convention 2002 May Munich, Germany This convention paper has been reprouce from the author's avance manuscript, without eiting, corrections, or consieration by the Review oar. The AES takes no responsibility for the contents. Aitional papers may be obtaine by sening request an remittance to Auio Engineering Society, 60 East 42 n Street, New York, New York , USA; also see All rights reserve. Reprouction of this paper, or any portion thereof, is not permitte without irect permission from the Journal of the Auio Engineering Society. Different approaches for the equalization of automotive soun systems Angelo Farina 1, Alberto ellini 2, an Marco Romagnoli 3 1 Inustrial Eng. Dept., Università i Parma, ia elle Scienze Parma Italy [email protected] 2 Information Eng. Dept., Università i Parma, ia elle Scienze Parma Italy [email protected] 3 ASK Automotive Inustries, via Fratelli Cervi n. 79, Reggio Emilia Italy [email protected] ASTRACT The paper iscusses the implementation of various possible DSP algorithms on low-cost platforms suitable for mass series prouction of automotive soun systems. The analysis takes into account traitional IIR an FIR filtering schemes, ual-rate an hybri approaches, an new algorithms such as Warpe FIR (WFIR) or frequency-omain partitione convolution (rutefir). The comparison is critically mae checking the implementation, cost an performance of ifferent processing schemes, but also aressing the hot problem of computing the optimal filtering coefficients for each of these schemes starting from measurements taken insie the car cockpit, making use of examples taken from the real life. The results show that both IIR an FIR structures are capable of goo results on low-cost DSP systems; the more avance algorithms will probably become competitive as soon as a new generation of floating-point DSP processors will also be available for low-cost applications. INTRODUCTION One of the major application of Digital Signal Processors (DSPs) in auio systems is the acoustic equalization. This is not limite to the flattening of the overall frequency response, but it is a more extensive concept. In certain case the goal of the equalization is to provie a prescribe frequency response curve (which in general can be very ifferent from the flat one). In other cases, also the time-omain response is important, an the goal of the equalization can be escribe as a re-alignment in time omain (or the linearization of the phase response, which are equivalent options). Special issues are to be aresse when the equalization is to be applie to a multi-channel soun system, an when there are many listeners in ifferent positions. All these issues are relate to car soun systems also, where further specific hot topics nee to be aresse: The optimal frequency response requires a substantial boost of the low frequencies, in orer to ensure a reasonable S/N ratio in presence of lou backgroun noise, which has approximately a brown spectrum. The frequency response shoul be ajuste accoring to the vehicle spee an the engine RPM, since the backgroun noise is strongly variable an epens on these variables. The listeners are locate off-axis with respect to the louspeakers. Moreover the louspeakers are often place in strange positions. This results in a blurre stereo image an may cause perceive coloration because of the (Hea Relate Transfer Function) HRTF filtering of wavefronts coming from not-symmetrical irections. The available electrical power is limite. Therefore a too heavy equalization causes an unacceptable rop in signal amplitue, making the reprouce soun too weak, an causing istortion if

2 the user is force to raise too much the gain control of the amplifier, in orer to compensate the loss of ynamic range. Dealing with an entry-level soun system, the car manufacturer fixes strong constraints on the total cost of the soun system, an also restrains the size, weight an power consumption. In summary it is impossible to achieve at the same time all these requirements: some trae-offs are neee. This paper iscusses in etail the esign of DSP-base equalizers to be inclue in automotive soun systems, base on practical experience one employing both fixe-point an floating-point processors, an making use of ifferent algorithms for performing the equalization. DSP ALGORITHMS FOR AUTOMOTIE SOUND SYSTEM EQUALIZATION In this paper, the following DSP algorithms are taken into account an compare: IIR structures (typically in the form of bi-qua 2 n orer filters). These recursive filters are functionally equivalent to their analog counterparts, an they allow the implementation of general frequency response by cascaing ifferent sections, each being configure as a typical filter type (LP, HP, P, R), an specifying for each section the cut-off frequency f c, the sharpness factor Q, an the gain G (the latter being relevant only for fixepoint implementation, because with floating-point processing a single overall gain can be applie at once, either before or after the cascae of IIR sections). FIR structures. These structures are also efine convolution filters, an are usually implemente as repetitive sum an multiplication, by means of a tightly optimize MAC (Multiply- Accumulate Cycle), which is usually taking just one DSP clock cycle for each coefficient on moern architectures. The length of the impulse response (usually calle the number of taps) is in the orer of some hunreths. Dual-rate FIR structures. These filters are obtaine employing a short FIR for equalizing the higher frequencies, while simultaneously the signal is own-sample an processe with another FIR filter for the low frequencies. Finally the two processe signals are merge back together. This topology aresses specifically the fact that a short FIR filter is not capable of an effective equalization at low frequencies, while the rawback is that the overall structure is more complex an that some errors appear in the cross-over region [1]. Warpe FIR. These structures closely reprouce a stanar FIR filter, except that the elay unit is replace by an all-pass section. y proper choice of the warping coefficient λ incorporate in the all-pass sections, the frequency axis is istorte, giving more resolution to the low frequency range, an compressing the high frequency range in a little number of spectral lines. This allows an accurate low-frequency equalization, at the expense of a coarser equalization at high frequencies. Moreover the computational cost of each cell of the computing network increases [2]. Frequency-omain convolution (Select-Save). This approach is base on the use of large FFT blocks, which allows the implementation of long filters, thus overcoming the constraints impose by processor size. This is usually one with the Select- An-Save algorithm, which requires the use of FFT blocks being long at least twice the length of the impulse response (typically several thousans of samples). This results also in a large processing latency, so that the filtere signals is substantially elaye with respect to the input signal. Another rawback of this algorithm is that it requires very large memory blocks, for storing the input an output buffers, the intermeiate ata an the coefficients [3]. Partitione frequency-omain convolution (rutefir). This algorithm is quite ol, since it was first applie in 1966, when the limite memory available was preventing from using the theoretically more efficient Select-Save algorithm. It was reimplemente only very recently, because it was iscovere that, on moern processors, this turns out to be more efficient than the Select-Save, giving the aitional avantages of require much less memory, an reuce input-output latency [4]. In summary four of the above six structures are conceptually similar, being always ifferent implementation of FIR convolution. The others instea (IIR an WFIR) are recursive structures, which require more care for esigning the filtering coefficients so that the require transfer function is create, without causing instability. In the following chapters an applicative example of the performances of the first two structures is given. All the examples are referre to the equalization of the soun system of the same car, an consequently the results can be compare irectly. DSP PLATFORMS ASK Inustries evelope two inepenent multichannel power amplifiers for automotive applications, incorporating a DSPcontrolle filtering section. The first one is a 4-channels amplifier equippe with a Texas Instruments TMS320C54 processor (16 bits, fixe point, 100 MIPS). This unit is shown in the following figure 1. Figure 1 ASK 4-channels integrate DSP amplifier (DIGIcar). The secon one is a 6-channels amplifier, with igital input an 6- channels output, equippe with an Analog Devices SHARC 21065L processor (32 bits, floating point, 60 M, 180 MFLOPS). This secon system was not employe in this work, but in the next future it will be evaluate in strict comparison with the first one. SOFTWARE TOOLS An automatic proceure for the synthesis of filter coefficients was evelope. It relies on state-of-the art algorithms of signal processing an is base on a frienly user interface, which allows the user to check the effectiveness of its equalization before the actual measurements. This tool, name DIGItools is a stan-alone Win32 application, which allows the generation of 4-channels 7-ban IIR filters, an of 4-channels FIR filters. As far as IIR filters are concerne, for each of the 7 bans the tool allow to specify the center frequency, the filter Q, an the filter gain. Each of the above mentione parameters can be moifie by simple sliers (fig. 2). AES 112 TH CONENTION, MUNICH, GERMANY, 2002 MAY

3 Figure 2 DIGItools IIR esign screenshot. As an example in the following some screenshots from DIGItools are presente. They escribe step by step the efinition of a parametric equalizer, starting from car acoustics measurements. In the screenshots the measurements of the car SPL (Soun Pressure Level) is reporte together with the frequency response of the parametric equalizer an the estimation of the car frequency response with the current equalization. Figures 3 report the above-escribe screenshots for FL, FR, RL, RR channels of a FIAT Stilo, use as a test car. Figure 5 DIGItools IIR equalizer for FIAT Stilo RL channel. Figure 6 DIGItools IIR equalizer for FIAT Stilo RR channel. Figure 7 shows the FIR equalizer tool. It allows to loa the car measure SPL an to choose among a few inversion algorithms, like Kirkeby [5], Neely-Allen [6], Mourjoupolos [7]. Figure 3 DIGItools IIR equalizer for FIAT Stilo FL channel. Figure 6 DIGItools FIR equalizer for FIAT Stilo FL channel. EXPERIMENTAL RESULTS Figure 4 DIGItools IIR equalizer for FIAT Stilo FR channel. Experiments were performe on a test car to valiate ifferent equalization methos with the 1 st eicate harware esigne by ASK (DIGIcar, shown in fig. 1). The test car was a FIAT Stilo, equippe with a stanar auio system. The DIGIcar DSP amplifier AES 112 TH CONENTION, MUNICH, GERMANY, 2002 MAY

4 was inserte in the system riving irectly the louspeakers an each of the four channels was operate separately. Stanar acoustics measurements were performe with the ai of the Aurora suite of plugins for CoolEit [8], an they were processe by DIGItools to obtain suitable inverse filters, an operate to realize a both parametric IIR equalization an Neely & Allen FIR equalization. The SPL measurements (before equalization) obtaine are shown in the following figures. Fig. 8 shows the impulse response of the FL channel. Figs show the SPLs of the four channels in the frequency omain (with a resolution of 1/3 of octave). Figure 11 Frequency Response of FIAT Stilo RL channel Figure 8 Impulse Response of FIAT Stilo FL channel. Figure 12 Frequency Response of FIAT Stilo RR channel Figure 9 Frequency Response of FIAT Stilo FL channel The inverse filters compute by means of the DIGItools software, starting from the measure car SPLs, were implemente on the DIGIcar platform by means of an assembly coe, highly optimize program. Each of the 4 FIR filters was 387 taps long, an each of the 4 IIR filter was mae of 5 2 n orer sections. Figs show the electrical measurements of the DIGIcar unit programme to realize the four inverse FIR filters. An Auio Precision System 2022 was use for measuring these 4 electrical transfer functions Figure Frequency Response of FIAT Stilo FR channel Figure 13 Electric measurements of DIGIcar FL channel programme to synthesize the FIAT Stilo inverse FIR. AES 112 TH CONENTION, MUNICH, GERMANY, 2002 MAY

5 Figure 14 Electric measurements of DIGIcar FR channel programme to synthesize the FIAT Stilo inverse FIR. Figure 17 Frequency Response of FIAT Stilo FL channel with (thick line) an without DIGIcar FIR equalization Figure 15 Electric measurements of DIGIcar RL channel programme to synthesize the FIAT Stilo inverse FIR. Figure 18 Frequency Response of FIAT Stilo FR channel with (thin line) an without DIGIcar equalization Figure 16 Electric measurements of DIGIcar RR channel programme to synthesize the FIAT Stilo inverse FIR. Figs. 170 show the results of the SPL measurements mae insie the FIAT Stilo while the DIGIcar equalizer was operating. The target curve was set as the flat one, an a noticeable flattening of the spectrum can be seen, especially at meium an high frequency. Each figure reports single channel measurements with an without the igital equalization. Figure 19 Frequency Response of FIAT Stilo RL channel with (thick line) an without DIGIcar equalization. AES 112 TH CONENTION, MUNICH, GERMANY, 2002 MAY

6 Figure 22 IPA Subjective results for the IIR equalization. CONCLUSIONS Figure 20 Frequency Response of FIAT Stilo RR channel with (thick line) an without DIGIcar equalization. As far as the IIR equalization is concerne a set of parametric equalizers was esigne with DIGItools. It resulte in 5 biqua IIR filters for each channel, which were implemente on the DIGIcar platform. Figure 21 show the electric measurements of the four channel programme to realize the parametric equalizers Figure 21 Electric measurements of DIGIcar channels programme to synthesize the FIAT Stilo IIR parametric equalizer. Subjective results were performe also. Listening tests were performe within the car (Fiat Stilo) equippe with DIGIcar equalization. The IPA proceure was use [9]. Five questions are mae to the listeners, an for each of them the answer range from a minimum to a maximum: 1) Spatial Soun, Flat Soun; 2) Insufficient Level, Sensible Level; 3) Clear Treble, Unclear Treble; 4) Weak bass, Clean bass; 5) Pleasant Soun System, Unpleasant Soun System. 13 listeners were use for the IPA experiments. The final results were IPA = 7.2 with FIR equalization an IPA = 6.9 with IIR equalization (fig. 22). The same car will also be available uring the 112 th AES Conference in Munich for performing aitional listening tests, which will make it possible to enlarge the number of subjects. An automatic proceure for the computation of the coefficients to be employe in both IIR an FIR igital equalization for car soun systems is presente. It allows to compare the two ifferent structures from both implementation an algorithmic point of views. Experiments were performe on a commercial car to show how it is possible to obtain very quickly a nice automatic equalization an to test it with both subjective an objective tests. The example presente here refers to an equalization aime to flattening the frequency response, but the evelope system allows also to equalize for a ifferent target curve, as it is common to o regaring the optimal equalization of a car soun system. The research will therefore prosecute, with the goal to assess the performance improvement obtainable by the use of a DSP-controlle power amplifier uner normal usage conitions (i.e., with the backgroun noise of the car running on the roa), an to select the equalization algorithm which will give the better results at the listening tests. References [1] G. Cibelli, A. ellini, E. Ugolotti, A. Farina, C. Morani, Experimental valiation of louspeaker equalization insie car cockpits, 106th AES Convention, Munich, 8-11 may [2] A. ellini, G. Cibelli, E. Armelloni, E. Ugolotti, A. Farina, Car cockpit equalization by warping filters, IEEE Transactions on Consumer Electronics, vol. 47, no. 1, pp , February 2001 [3] Oppenheim, A.. an R. W. Schafer. Discrete-Time Signal Processing. Englewoo Cliffs, NJ: Prentice Hall, 1989 [4] A. Torger, rute-fir an open-source general-purpose auio convolver, [5] O.Kirkeby, P.A. Nelson, P. Rubak, A. Farina, Design of Crosstalk Cancellation Networks by using Fast Deconvolution, preprint of 106th AES Convention, Munich, 8-11 may 1999 [6] S.T. Neely, J.. Allen, Invertibility of a room impulse response, J.A.S.A., vol. 66, pp (1979). [7] J.N. Mourjopolos, Digital Equalization of Room Acoustics, Journal of AES, vol. 42, n. 11, November 1994, pp [8] A. Farina, F. Righini, Software implementation of a MLS analyzer, with tools for convolution, auralization an inverse filtering, 103 r AES Convention, New York, September [9] E. Ugolotti, G. Gobbi, A. Farina, IPA - A Subjective Assessment Metho of Soun Quality of Car Soun Systems, Proceeings of the 110th AES Convention, 2001 May Amsteram, The Netherlans. AES 112 TH CONENTION, MUNICH, GERMANY, 2002 MAY