Evaluating New What are the parameters that impact the doing? - PDF document

11/1/2013 Levels of Evidence* APFs (Catherine Palmer, 2009) ◦ What does the algorithm do? Evaluating New ◦ What are the parameters that impact the doing? Efficacy of the design Technologies ◦ In a well-controlled (contrived?) environment, do we get an effect? ◦ Or, what is the effect of the feature in the lab? Effectiveness of the design RUTH BENTLER ◦ In the real-world use of this design, do we get an effect? ◦ Or, what is the effect of the feature in the real world? UNIVERSITY OF IOWA *Ala Bentler 1 http://www.uiowa.edu/~neuroerg/siren.html 1

11/1/2013 Levels of Evidence* APFs (Catherine Palmer, 2009) ◦ What does the algorithm do? ◦ What are the parameters that impact the doing? Efficacy of the design Directional Microphones ◦ In a well-controlled (contrived?) environment, do we get an effect? ◦ Or, what is the effect of the feature in the lab? Effectiveness of the design ◦ In the real-world use of this design, do we get an effect? Or, what is the effect of the feature in the real world? Efficiency (not studied in my lab) *Ala Bentler 8 APFs THE FIRST STEP IS TO UNDERSTANDING THE BLACK BOX…. 9 10 11 12 2

11/1/2013 13 14 15 16 17 18 3

11/1/2013 19 20 21 22 23 24 4

11/1/2013 25 26 27 28 29 30 5

11/1/2013 31 32 33 34 35 36 6

11/1/2013 37 38 39 40 41 42 7

11/1/2013 43 44 FF (BTE) KEMAR (BTE) Theoretical Cardioid 4.8 4.0 2.3 Hypercardioid 6.0 5.1 3.0 Supercardioid 5.7 5.0 3.3 45 0 340 350 20 10 20 330 30 15 320 40 FF (ITE) KEMAR (ITE) Theoretical 10 310 50 5 300 60 0 Cardioid 4.8 4.1 2.7 290 70 -5 280 80 500 -10 1000 270 -15 90 2000 Hypercardioid 6.0 5.6 3.3 4000 260 100 250 110 240 120 Supercardioid 5.7 5.4 3.5 230 130 220 140 210 150 200 160 190 170 180 48 8

11/1/2013 Polargram And so… o We are able to measure the acoustic and physical facts (APFs) for all possible scenarios of test; o Such APF testing is necessary to develop our hypotheses; o Newer technique for quantifying polar response patterns and directivity indices (DI) helps us understand static function in a dynamic world of noise Wu & Bentler, 2009, 2010, 2012)! 50 Data? Test Booth Field Ratings Test Booth Field Ratings 100 10 Very Good 10 Very Good 100 OMNI OMNI Plenty of efficacy data for all designs depending upon DIR DIR 8 80 80 8 ◦ Baseline used Percent Correct Percent Correct p < .0001 p < .0001 p < .0001 p < .0001 6 ◦ Speaker arrangement 60 60 6 ◦ Noise type 40 4 4 40 ◦ Etc 20 20 2 2 Effectiveness data a bit harder to come by… 0 Very Poor 0 Very Poor 0 0 60 / 0 60 / 0 75 / +2 75 / +2 Speech Speech Understanding in Understanding in CST Test Condition CST Test Condition Noise Noise Walden, Surr, & Cord, 2003 9

11/1/2013 Research Question of Study #1 Laboratory • How do visual cues affect DIR benefit? Real world Speech recognition test Speech Recognition Performance 100 Speech Recognition (%) 80 DIR Auditory-Only 60 40 OMNI OMNI-AO 20 OMNI-AV DIR-AO 0 DIR-AV -10 -6 -2 2 6 10 SNR (dB) Wu & Bentler, 2010, Ear Hear 10

11/1/2013 Speech Recognition Performance Summary of Study 1 • The advantage (benefit) of visual cues can overshadow 100 the measured benefit of directional mic schemes in real Audiovisual Speech Recognition (%) world environments. 80 Auditory-Only 60 40 OMNI-AO 20 OMNI-AV DIR-AO 0 DIR-AV -10 -6 -2 2 6 10 SNR (dB) Wu & Bentler, 2010, Ear Hear Research Question of Study #2 25 F(1, 21) = 1.21 p = 0.29 20 • How does age impact DIR benefit? DIR Benefit (%) • Laboratory 15 Laboratory • Real world 10 5 0 -5 30 40 50 60 70 80 90 Age Wu, 2010, JAAA Summary of Study #2 100 F(1, 21) = 11.78 p = 0.003 80 • Listeners of different ages obtain comparable benefits DIR Preference (%) from DIR in the laboratory. 60 Real World • Older users tend to perceive less DIR benefit than do younger users in the real world. 40 • Due to lifestyle differences, primarily • The focus of future efforts in the lab 20 0 30 40 50 60 70 80 90 Age Wu, 2010, JAAA 11

11/1/2013 Example of unexpected function… Front Forward DIR Forward DIR Backward DIR Backward DIR Back 7 Front 6 Directional Benefit (dB) p < 0.05 5 4 3 Backward DIR Backward DIR p = 0.17 2 Back 1 0 Big dogs can be Our Data Manufacturer’s Conversation Listening dangerous. Data Condition Wu, Stangl & Bentler, 2013 http://www.despicableme.com/ Big dogs can be dangerous. Front Front Forward DIR Forward DIR Forward DIR Forward DIR Backward DIR Backward DIR Back Back The boy fell from the window. 12

11/1/2013 7 Briefly, for DIR 6 Directional Benefit (dB) ◦ APFs are clear as to expected impact p < 0.05 5 ◦ Efficacy has been demonstrated 4 repeatedly; newer algorithms take special consideration 3 ◦ Effectiveness depends on many p = 0.17 2 factors 1 ◦ Environment, age, etc 0 ◦ …crud Our Data Manufacturer’s Conversation Listening Data Condition Wu, Stangl & Bentler, 2013 Analog NR (1980-90s) Early spectral approaches ◦ Switch ◦ ASP ( means low frequency compression) Digital Noise Reduction ◦ Adaptive filtering ◦ Frequency dependant input compression ◦ Adaptive compression TM ◦ Zeta Noise Blocker TM 75 Today’s versions o Most are modulation-based with some algorithm for where and how much gain reduction should occur; o At least one other (Oticon) first introduced a strategy called “synchronous morphology” treating harminic inputs like speech; APFs o Many are now implementing Wiener filters as well; o Many are now implementing impulse noise reduction; THE FIRST STEP IS TO UNDERSTANDING o Many also use some mic noise reduction, expansion, THE wind noise reduction, and even directional mics as BLACK BOX…. part of the strategy they promote. 78 13

11/1/2013 Siemens (TRIANO 3) GN ReSound (CANTA 770-D) 2 5 0 Difference (dB, 1/3 Octave) 0 Difference (dB,1/3octave) -2 -5 -4 -10 -6 -15 -8 SIREN -10 -20 TRAFFIC a DINING -12 -25 250 500 1000 2000 4000 8000 125 250 500 1000 2000 4000 8000 Frequency (Hz) Frequency (Hz) 70dB Starkey (AXENT II AV MM) 5 2 0 0 Difference (dB,1/3octave) -2 -5 ON versus OFF (output change) -4 -6 -10 -8 -15 -10 SNR00 -12 -20 SNR05 SNR10 b -14 SNR15 -25 125 250 500 1000 2000 4000 8000 -16 125 250 500 1000 2000 4000 8000 Frequency (Hz) Frequency (Hz) ICRA Speech Random Noise Babble Starkey J13 Axent AV 75 dB --SPEECH,RANDOM, MUSIC-- 85dB 5 2 0 DIFFERENCE (dB,1/3octave) 0 -2 -5 ON versus OFF (output change) -4 -10 -6 -15 -8 -10 -20 SNR00 -12 SNR05 SNR10 -25 -14 SNR15 125 250 500 1000 2000 4000 8000 -16 Frequency(Hz) 125 250 500 1000 2000 4000 8000 Guitar Frequency (Hz) Piano Saxophone with background music Random Noise Plain Speech 14

11/1/2013 What happens in the time domain? Siemens (Triano) APFs…10 years later Starkey (Axent) 88 Any reason to expect SNR-50 would change? Output SNR (re: Linear) Miller et al. 2012 15

11/1/2013 Data? Briefly, for DNR • Still, plenty of efficacy and effectiveness data for all ◦ APFs are clear as to expected impact designs if you are asking the right question: ◦ Efficacy and Effectiveness have been • Walden et al (2000) • Boymans and Dreschler (2000) demonstrated…if you are asking the right • Alcantara et al (2003) question • Ricketts & Hornsby (2005) • Marcoux et al (2006) • Mueller et al (2008) • Bentler et al (2009) • Sarampalis et al (2009) • Bentler et al (2010) • Stelmachowicz et al (2010) • Pittman et al (2011): • And those are good outcomes Not really a new concept Four (sort of) choices on the market: ◦ Frequency compression ◦ Frequency transposition Frequency Lowering ◦ Frequency “cueing” ◦ Combination of above Concept makes sense ◦ Providing the widest input bandwidth possible ◦ Data suggest this may be most important for children re: speech and language development 93 What is happening here? Frequency compression hearing aid Default settings Steeply sloping loss Freq compression: OFF APFs Assessed on 11/23/09 SN:0906H109W THE FIRST STEP IS TO UNDERSTANDING Input: 1s pure tones 100 Hz spaced with 500ms THE intervals (~75dB SPL) BLACK BOX…. Upper graph: output of Hearing aid 95 96 16

11/1/2013 1 st peak: 3661 Hz, 2 nd peak: 4306 Hz, 3 rd peak: 4927 Hz 1 st peak: 3468 Hz, 2 nd peak: 4091 Hz, 3 rd peak: 4700 Hz Input: 4306 Hz Input: 4091 Hz 97 98 1 st peak: 3765 Hz, 2 nd peak: 4392 Hz 1 st peak: 4070 Hz, 2 nd peak: 4694 Hz, 3 rd peak: 5336 Hz Input 4392 Hz Input 4694 Hz 99 100 1 st peak: 5490 Hz 1 st peak: 5598 Hz Input 5490 Hz Input 5598 Hz 101 102 17

11/1/2013 1 st peak: 5457 Hz, 2 nd peak: 6093 Hz 1 st peak: 5553 Hz, 2 nd peak: 6201 Hz Input: 6093 Hz Input: 6201 Hz 103 104 1 st peak: 5665 Hz, 2 nd peak: 5603 Hz 1 st peak: 1937 Hz, 2 nd peak: 2562 Hz Input 6309 Hz Input 6395 Hz 105 106 1 st peak: 1071 Hz, 2 nd peak: 1701 Hz, 3 rd peak: 2346 Hz Input: 4091 Hz What is happening here? Frequency compression hearing aid Default settings Steeply sloping loss Freq compression: ON Assessed on 11/23/09 SN:0906H109W Input: 1s pure tones 100 Hz spaced with 500ms intervals (~75dB SPL) Upper graph: output of Hearing aid 107 108 18