Implementation of cross-talk canceling filters with warped structures - Subjective evaluation of the loudspeaker reproduction of stereo recordings

Transcription

1 Implementation of cross-talk canceling filters with warped structures - Subjective evaluation of the loudspeaker reproduction of stereo recordings Angelo Farina, Alberto Bellini, Enrico Armelloni farina@unipr.it Dipartimento di Ingegneria Industriale University of Parma ITALY

2 Outline Sound reproduction quality; Analysis and compensation of distortion and reduction of cross-talk paths; Software and hardware implementation; Experimental results; Demonstration; 2

3 Audio processor design Audio source Digital Audio processor Amplifier Speakers + room Design a filter so that Measured SPL = Target SPL 3

4 Equalization architecture design Standard acoustic measurements DSP code (common) Room measurements Equalization filter synthesis Filter Taps reduction Filter implementation Software tools (parametric) Room measurements 4

5 Tools Hardware: Analog Devices SHARC 21061/21065 boards; Software: AURORA: measurements and acoustic characterization MATLAB for filters design; 5

6 Aurora Plug-ins of Syntrillium CoolEdit, AURORA : TIM filter Convolve with Clipboard Generate MLS Signal Generate IRS Signal Deconvolve MLS Signal Deconvolve IRS Signal Acoustical Paramenters Inverse Filter Flatten Spectrum Subtract Convolved; 6

7 Aurora GUI 7

8 Design-flow Standard acoustic measurements Software tools (parametric) Filter Taps synthesis & reduction Filter code synthesis DSP code (common) Filter implementation Acoustic validation (common) Car cockpit measurements 8

9 Warped FIR Warped FIR algorithm. You can get more info on this specific algorithm looking at:

10 Warping: Frequency mapping Applying the following bilinear transformation to the z-planez z ζ + λ = A λ ( ζ) = 1+ ζ λ Sampling-rate is not constant Consistent with psychoacoustics representations 10

11 Warping FIR Same structure as FIR, the delay unit is replaced by 1 D 1 The FIR features: Poor resolution at low frequencies Properties well defined on a linear frequency scale Linear phase z λ (z) = 1 λ z 1 Short execution time ( 0.5 clock cycles / tap ) No added quantization noise 11

12 Warping FIR LCNTR=Wfilter_taps-1, DO wmac_rr UNTIL LCE; F12=F2*F4, F9=dm(I5,M7), F4=pm(I9,M8); F10=F2*F5, F8=F8+F12, F9=dm(I5,M6); F1=F9-F10, F9=dm(I5,0); F10=F1*F7, dm(i5,m7)=f2; wmac_rr: F2=F9+F10; /* convolution */ LCNTR = Ff_taps-1, DO lpf_filter_rr UNTIL LCE; lpf_filter_rr: F8=F2*F4, F12=F8+F12, F2=dm(I5,M5), F4=pm(I9,M9); 12

13 Warping Frequency mapping λ = 0.8 λ = 0.75 λ = 0.7 λ = 0.5 λ > 0 ξ λ < 0 z 13

14 Warping Frequency mapping Frequency response of a car cockpit FIR FRF samples Amplitude (db) Normalized frequency 14

15 Warping Frequency mapping Frequency response of a car cockpit Frequency re- mapping by Warping FRF samples Amplitude (db) Normalized frequency 15

16 Warping Frequency mapping Frequency response of a car cockpit WFIR FRF samples Amplitude (db) Normalized frequency 16

17 FIR synthesis vs WFIR synthesis 30 taps WFIR 30 taps FIR Amplitude (db) Amplitude (db) Frequency (Hz) x 10 4 Frequency (Hz) x

18 Sound harmonization Equalization is not sufficient to achieve a global increase of sound comfort; Harmonization of sound image achieves a subjective improvement of binaural sound quality; Stereo-dipole systems 18

19 Stereo dipole system L C LL Cross-talk paths H C LR C RL ~ ~~ ~ S L R Digital audio processor C RR S R Ideal condition: H Design so that: S L = L C LL S R = R C RR 19

20 Stereo dipole system Cross-talk paths Audio source H LL + C LL L R H LR H RL C LR C RL ~ ~~ ~ S L S R H RR + C RR 20

21 S S Stereo dipole structure L R Target: = = C( ω) = Hence: L C R C FT(h ll LL RR ) FT(h rr In the frequency domain: InvDen ( ω) = Conj ) FT(h lr ( hrr ) ( hlr ) ( hrl) ( h ) fll = InvDen flr = InvDen frl = InvDen f = rr ll InvDen InvDen = InvFilter ) FT(h rl Conj[ C( ω) ] [ C( ω) ] C( ω) + ε( ω) ) ( h h h h ) ll rr lr rl ε(ω) is a function of frequency 21

22 ASK listening room 22

23 Digital implementation DSP SHARC EZ-LITE: 40 Mips, with which 880 Taps can be computed for each Hz DSP SHARC 21065L AD 44100/48000; 23

24 Experimental results (Measured binaural response of the room) 24

25 Experimental results (2048 FIR cross-talk cancelling filters) 25

26 Experimental results (Measured response with 2048 FIR) 26

27 Experimental results (220 FIR cross-talk cancelling filters) 27

28 Experimental results (Measured response with 220 FIR) 28

29 Experimental results (42 WFIR cross-talk cancelling filters) 29

30 Experimental results (Measured response with 42 WFIR) 30

31 Subjective tests Subjective tests have been performed by trained people, no time-limit, several choices of music available; Blind Evaluation for each of the four systems (the listener can switch between them at any time): WFIR FIR 7 question with a 0-5 score. ANOVA statistical post-processing analysis 31

32 Subjective tests Question Avg. FIR Avg.WFIR Anova's F factor Prob. Overall appreciation % Image localization % Stage amplitude % Naturality % Low frequency resp % Mid frequency resp % Hi frequency resp % 32

33 Subjective tests Averages, standard deviations and ANOVA probability results Score 6 5,5 5 4,5 4 3,5 3 2, % 4.63% 0.01% 0.28% 0.22% 21.71% 33.10% 1-Overall appreciation 2-Image localization 3-Stage amplitude 4-Naturality Question 5-Low frequency resp. 6-Mid frequency resp. WFIR FIR 7-High frequency resp. 33

34 Conclusions Multi-channel Warped filter equalization and harmonization; Automatic design of audio processors with standard acoustic measurements (AURORA); Implementation on DSP systems; Experimental results and listening tests; 34