Low Delay Wind Noise Cancellation for Binaural Hearing Aids

Transcription

1 INTER-NOISE 6 Low Deay Wind Noise Canceation for Binaura Hearing Aids Nobuhio HIRUMA ; Ryousue KOUYAMA ; Hidetoshi NAKASHIMA 3 ; Yoh-ichi FUJISAKA 4, 4 Rion Co., Ltd, Japan, 3 Nationa Institute of Technoogy, Kumamoto Coege, Japan ABSTRACT Wind noise annoys hearing aid users, and it is undesirabe to attach a windscreen to a hearing aid microphone for cosmetic reasons. Accordingy, some hearing aid devices reduce the ow frequency components of input signas by using high-pass fiters to suppress wind noise. Athough wind noise can be attenuated by using this approach, the perceived binaura information of the desired signa is aso simutaneousy degraded; therefore, ocaization cues are partiay ost. We previousy proposed a FFT-based binaura wind noise canceation agorithm to preserve the cues. This agorithm required a 3 ms frame ength in order to maintain a high-frequency resoution. However, it is nown that the toerabe deay time for mid hearing oss or better shoud be beow about 5 ms in the high frequency region. In this study, we propose a ow deay binaura wind noise canceation agorithm using a frequency warping fiter. The deay time caused by this agorithm is shorter than the toerabe one. The objective evauation resuts, SNR and PESQ scores, were improved whie maintaining a ow deay time. As a resut of subjective experiments, the proposed method produces amost the same score as our conventiona FFT-based method for the directionaity of output signas. Keywords: Hearing aids, Wind noise, Binaura processing, Low deay, Frequency warping fiter, HRTF I-INCE Cassification of Subjects Number(s):.4. INTRODUCTION With an increasing number of active hearing aid users in recent years, opportunities for speech communication have increased even in outdoor situations. Therefore, the deveopment of optima hearing aid processing appropriate to various environments is desired. According to a survey by MareTra (), wind noise sti remains as a major cause of dissatisfaction for hearing aid users in istening environments. Wind noise in hearing aids is generated by the air-fow around the head, by the fow over and around the pinna, and by the fow past the tragus and heix(,3). Generay, a windproof is utiized to reduce the noise by attaching it to the microphone. However, for hearing aids, wind noise is reduced by signa processing for cosmetic reasons. Thus, severa agorithms have been proposed using digita signa processing to reduce the discomfort of wind noise (4 7). In singe channe wind noise suppression in hearing aids, high-pass fiter processing has been utiized to reduce wind noise since the main components of wind noise exist in the ow frequency region. Athough wind noise can be attenuated by using a high-pass fiter, the desired signa is aso degraded simutaneousy. Furthermore, when using hearing aids on both the eft and right ears that wor independenty, the perception of the arriva direction and distance of the desired signa wi be uncertain. Therefore, we previousy proposed a binaura wind noise canceation agorithm using the characteristics of head reated transfer functions (HRTFs) in order to construct a future binaura t ype of hearing assistant system (8,9) to avoid the above issue. In our previous study, it was shown that the spatia information (directionaity and externaity) was maintained, compared with the conventiona high-pass fiter method by performing psychoacoustic experiments. Incidentay, open-fit styes () have been appied to reduce occusion effect, so that especiay mid hearing oss often uses it. However, a processed sound of hearing aids is heard as echo if the n-hiruma@rion.co.jp ae5ouyama@g.umamoto-nct.ac.jp 3 naashi@umamoto-nct.ac.jp 4 y-fujisaa@rion.co.jp 45

2 ILD [db] IPD[rad.] INTER-NOISE 6 sound has some ong deay, since a direct sound is aso heard through a arge vent in hearing aids for open-fit styes. Our proposed method has a required frame processing of 3 ms to maintain a high frequency resoution in the ow frequency region using a Fourier transform. However, the deay time with this method is intoerabe for hearing aid users (). Therefore, the processing deay must be as short as possibe. In this paper, we propose a binaura wind noise canceation agorithm using a frequency warped fiter as a substitute for FFT. With this approach, it is possibe to eep the deay time toerabe whie maintaining a high frequency resoution in the ow-frequency region.. BINAURAL WIND NOISE CANCELLATION SYSTEM. Characteristics of HRTFs (IPDs and ILDs) The characteristics of the frequency response from a sound source to both ears can be described as HRTFs. Figure shows (a) interaura phase differences (IPDs) and (b) interaura eve differences (ILDs) cacuated by using HRTFs (), where the horizonta axis shows the direction of arriva (DOA) of a sound source. It is apparent that IPD and ILD vary depending on the DOA and frequency of the sound. These characteristics are stored to memory as a database () for wind noise detection and canceation. Thus, DOA can be estimated by comparison between the database and the characteristics that are derived from the observed binaura signas. The database of each frequency band is (, ) for IPDs and (, ) for ILDs, where φ and represent the azimuth ange of a sound source and frequency index, respectivey Hz 5Hz 75Hz 5Hz 75Hz 35Hz 375Hz azimuth [deg.] Hz 5Hz 375Hz 5Hz 65Hz 75Hz azimuth [deg.] (a) (b) Figure Database of proposed system. (a) IPDs and (b) ILDs (3).. System Boc Diagram Figure shows a boc diagram of the proposed system. In our system, the use of binaura hearing aids is assumed; therefore, binaura input signas are utiized for signa processing. Binaura hearing aids have two microphones put on each ear. The observed signas on both the eft and right sides, (n) and r(n), are defined as ( n) s( n)* h, ( n) wn ( n), () r( n) s( n)* h, ( n) wn ( n), () r where h,φ (n) and h r,φ (n) denote HRTFs from the direction φ, and wn (n) and wn r (n) denote the wind noise signas. The asteris symbo [*] represents the convoution operation. We assume that wn (n) and wn r (n) are uncorreated and that the IPD and ILD of the noise wi be random in the time domain. r 46

3 INTER-NOISE 6 Figure Boc diagram of proposed system.3 Frequency Warping Fiter In our previous wind noise canceation agorithm (8,9), a 3-ms frame ength was required for FFT anaysis with high resoution. In the case of 6-Hz samping rates, it needs 5 sampes for frequency anaysis in order to satisfy the above frequency resoution Δf of 3.5 Hz, but the processing deay time, which means the group deay of the system, is 3 ms. Stone and Moore (3) showed that the toerabe hearing aid deay time must be ess than about 5 ms. Therefore, the group deay of our system shoud aso be shorter than the toerabe processing time, defined as toerabe deay, for the anaysis and the synthesis of audio signas. In our FFT based agorithm, it is necessary that the frame ength must be shorter than the toerabe group deay. However, when the number of signa sampes is reduced in FFT anaysis, the frequency resoution is degraded. As a resut, effective wind noise suppression cannot be possibe because the anaysis frequency bandwidth is arger than the fine harmonic structure of speech. We propose a binaura wind noise canceation agorithm using a frequency warping fiter (FWF) to preserve a ow group deay and the intended frequency resoution. A boc diagram of anaysis and synthesis using FWF is shown in Figure 3, where M represents the number of a-pass fiters (APFs). The signa vector, p (n,) to p (n,m) in Figure 3, is a warped signa, of which the frequency components are warped from ow to high frequency region with an appropriate ratio decided by the coefficient γ in equation (3) (4,5). Figure 3 Boc diagram of portion of wind noise canceation system using frequency warping for eft channe 47

4 group deay time [msec] INTER-NOISE 6 The a-pass fiter is given by A( z) ( z ) /( z ), (3) where γ is the warping parameter. The frequency warped sequence of the eft channe is obtained by p ( n,) ( n), (4) p ( n, m) { p ( n, m ) p ( n, m ) p ( n, m ), where m ~ M, and that of the right channe is obtained with the same cacuation. The frequency warped output of the th a-pass stage is given by { p ( n, m), pr ( n, m)} in both the eft and right channes. Figure 4 shows an exampe of the group deay characteristics of FWF compared with the toerabe deay. The soid ine represents the toerabe group deay time, derived from reference (3), as a function of frequency. The broen ine represents the FWF group deay time in the case that γ equas.8. We can see that the FWF group deay time is shorter than the toerabe group deay time frequency [Hz] toerabe group deay time FWF group deay time Figure 4 Group deay of proposed system (γ =.8) and toerabe deay time ().4 IPD and ILD Cacuation of Input Signa The IPD of the input signas, (), is cacuated by using the cross spectrum of each input signa, C (), represented as where ( ) tan Im{ C ( )}/ Re{ C ( )}, (6) * C ( ) L( ) R( ). (7) Note that the operator [*] represents a compex conjugate. The same as IPD cacuation, the ILD of the input signas, (), is cacuated as ( ) og C ( ) / C ( ), (8) where C () denotes the power spectrum of the input signa (n)..5 Wind Noise Detection Agorithm Wind noise detection is performed by comparing databases ( ) () are the absoute difference between them, defined as and ). () ( (5) and 48

5 INTER-NOISE 6 ( ) (, ) ( ), (9) ( ) (, ) ( ). () Here, ψ denotes a speech signa direction. Normay, this direction ψ is in front of a hearing aid istener. However, isteners can seect the direction freey for speech enhancement in our system. The minimum vaue of () and () is, whereas the maximum vaue varies depending on the frequency, so normaized differences are introduced as foows. D ( ) P( ) ( ) L( ) (for ow frequency region), () (for high frequency region) where P() is the maximum difference of (, ) when varies, defined as L() is aso defined as (, ) min (, ) P( ) max. () (, ) min (, ) L( ) max. (3) We assume that wind noise exists in the frequency region if D (n)>tw, where Tw is a predetermined wind detection threshod. This comparison procedure is performed for whoe frequency bands, and c(n) is the wind counter for determining the amount of suppression regarding wind noise, given by K c( n) u ( n), (4) where u (n) represents the detection fag of wind noise reated to each frequency band. The fag is described by ( if D ( n) Tw) u (n). (5) ( otherwise).6 Binaura Wind Noise Canceation Fiter The instantaneous binaura wind noise canceation (BWNC) fiter W i (n) is defined as foows. W i ( n) exp( ( n) D ( n)), (6) where β (n) is a gain contro parameter that varies depending on the wind counter c(n) and is appropriatey determined. ( n) c( n). (7) Thus, c(n) and β (n) wi be arge under windy conditions, and W (n) approaches. On the contrary, W (n) approaches under cam conditions. To avoid sudden changes in the fiter gain, the actua BWNC fiter W (n) is defined as W ( n) W ( n ) ( ) W i ( n), (8) where λ is a forgetting factor. The impuse response form of the time domain is obtained by the inverse Fourier transform of the wind noise suppression fiter coefficient, w ( n) IFFT W ( n). (9) The processed output signas, (n) and r (n), are obtained by convouting both warped signas and the noise suppression fiter w (n). 49

6 db (Reative) INTER-NOISE 6 3. NUMERICAL SIMULATION M ' ( n) w ( n) p ( n, m), () m, M r' ( n) w ( n) p ( n, m). () m, 3. Numerica Configuration A numerica test was performed to confirm the abiity of the wind noise detection and the contro of the fiter gain w (n). In this simuation, the samping frequency was 6 Hz, and 3-point FFT was utiized to anayze warped signas. The warping fiter coefficient γ in equation (3) was set to.8 in order to anayze the wind noise components for the high resoution in ow frequency. The fiter ban derived from FWF in this simuation is shown in Figure 5 (a). The frequency characteristics of the fiter ban between to Hz are simiar to that of /3 octave fiter ban. Figure 5 (b) shows the IPD database of the warped HRTFs for each ange, which were cacuated via FWF. The frequency bands under the 75 Hz center frequency were utiized for IPD cacuation, so that wind noise estimation is mainy performed by using IPD. r frequency [Hz] (,) [rad.] 4 3-5Hz 5Hz - 45Hz -3 75Hz azimuth [deg.] (a) (b) Figure 5 Characteristics of frequency warp fiter-ban (a) and warped domain IPDs (b) Figure 6 shows an image of the reative ocations among a istener, a desired speech source, and wind noise in our numerica simuation. The wind noise fow is ocated at degrees, and the desired speech is ocated at -6 degrees. HRTFs donated by the MIT Media Lab () were utiized to ocate the speech source at -6 degrees and the distance between the sound source and dummy head. In addition, wind noise sound that was pre-recorded by using the binaura microphones of the dummy head was used in a condition where wind veocity was approximatey 4m/s. There were two sient intervas in the wind noise test sound. The intervas were set to time periods of - 3 and 5-6 sec. The SNR of the input signa was set to - db in front of the hearing aid istener. Here, the speech signa direction ψ was set to -6 degrees, assuming that the direction of the desired signa is nown. Figure 6 Reative ocation among istener, desired speech source, and wind noise. Desired speech source was ocated at -6 degrees, whie wind noise fowed from degrees. SNR was set to db. 5

7 gain wind counter ampitude ampitude ampitude ampitude INTER-NOISE 6 3. Resut of Simuation Figure 7 shows the waveforms and interna parameters obtained from the simuation in 3.. Each figure shows the (a) eft channe input signa and (b) right channe input signa. For the eft channe, the origina speech signa (represent by the bac ine) and the speech in wind signa (represented by the red ine) are shown in (a), and the processed output signa is in (b). In the case of the right channe, the above mentioned waveforms are shown in Figures 7 (c) and (d), which are simiar to the eft channe. We can see that the wind noise in the processed output signa graduay decreased as described in (b) and (d) when wind noise existed. Figure 7 (e) shows the wind counter vaue as a function of time. This vaue varies corresponding to wind noise strength. Figure 7 (f) is the variation of the fiter gain W (n) for the 3, 9, and 3th bands. It is ceary seen that the gain controer W (n) for a frequencies was cose to when wind noise disappeared, and it controed the gain to reduce the noise graduay when it appeared. These resuts revea that the gain controer wored precisey according to wind noise conditions (a) speech signa and speech in wind noise (Lch) speech in wind signa speech signa (b) output signa (Lch) output signa(lch) time[sec] (c) speech signa and speech in wind noise (Rch) speech in wind signa speech signa (e) wind count (f) fiter gain for each frequency time[sec] (d) output signa (Rch) output signa(rch) time[sec] 5Hz (=3) 6Hz (=9) 5Hz (=3) Figure 7 Simuation resuts. (a) Input speech signa and speech in wind noise at Lch, (b) noise canceed output signa at Lch, (c) input speech signa and speech in wind noise at Rch, (d) noise canceed output signa at Rch, (e) wind count c(t), and (f) fiter gain for each frequency band. SNR was set to - db. 4. EVALUATIONS 4. Objective Evauations Two inds of objective evauations were performed to verify the effectiveness of the proposed method. The first one was SNR improvement, and the second one was an evauation of the sound quaity using the PESQ measure. The sound eve of wind noise was set to SNRs of db and - db for the eft channe when the speech signa came from degrees. The wind noise eve was fixed, and the direction of the speech signa varied from -9 to +9 degrees in 3-degree steps in this simuation. To evauate the SNR improvement and PESQ score, there was a comparison between the previous FFT method with a 3-ms frame ength (FFT3 method), the FFT method with a 4-ms frame ength (FFT4 method), and the proposed method (FWF method). 5

8 PESQ PESQ SNR improvement[db] SNR improvement[db] INTER-NOISE 6 (I) SNR Improvement Figure 8 shows the resut of SNR improvement, which was obtained from the difference between the SNR of the input signa and its output signa. Figures 8 (a) and (b) represent the SNR improvement as a function of the degree for the speech arriva direction for the SNRs of - db and db. The resuts show that the SNR of the signa after processing with our previous FFT3 method was significanty improved compared with the input signa, whereas it was sighty improved with the FFT4 method. The figure demonstrates that it is aso indicated good performance in the case of using the FWF method since the curve was cose to the FFT3 resut. 3 5 input SNR: -db FWF Based Method FFT Based Method (3ms frame ength) FFT Based Method (4ms frame ength) 3 5 input SNR: db FWF Based Method FFT Based Method (3ms frame ength) FFT Based Method (4ms frame ength) azimuth azimuth (a) (b) Figure 8 Resut of SNR improvement. (a) Input SNR was - db. (b) Input SNR was db. 4.. PESQ Scores The measurement resuts for PESQ are shown in Figures 9 (a) and (b), which represent the PESQ score as a function of the degree for the speech arriva direction for the SNRs of - db and db. There was enough improvement in each direction in case of FWF compared with FFT4, whie it was represented by the thinnest margin in the ow SNR environment. 3.5 input SNR: -db FWF Based Method FFT Based Method (3ms frame ength) FFT Based Method (4ms frame ength) input signa 3.5 input SNR: db FWF Based Method FFT Based Method (3ms frame ength) FFT Based Method (4ms frame ength) input signa azimuth azimuth (a) (b) Figure 9 Resuts of PESQ score for each method. (a) Input SNR: - db. (b) Input SNR: db. 4. Subjective Evauations To evauate the preservation of binaura information, we performed two types of subjective evauation: directionaity and externaity tests. The directionaity test is to examine the infuence on the perception of the arriva direction of sound. The externaity test compares the proposed FWF method with the conventiona high-pass method in order to evauate whether spatia information is retained. These tests were carried out in a way seected on in a way seected on the MATLAB GUI to present pre-processed sound from headphones (Sennheiser HD-6) in an anechoic chamber. 5

9 perceved ange [degree] perceved ange [degree] INTER-NOISE 6 (I) Directionaity Test As stimuus, we used two types of the word. The first one is a 4-mora Japanese words to which seven inds of directiona characteristics were added by using HRTFs. The second one is the words in wind noise which were processed in the FWF proposed method. The seven directions were from -9 degrees to +9 degrees in 3-degree steps. In this experiment, three subjects from 9 - years od were recruited. The stimui were presented at random for each subject. The subjects were ased to answer the perceived direction of the words. Athough the stimui were presented in 3-degree steps, the subjects seected their perceived direction on a scae of -degree steps. Figure shows the resuts of the directionaity test for the (a) target signa and (b) FWF processed signa. It indicates that the perceived direction of unprocessed target signas was not affected. It aso indicates that the signas output by the FWF process preserved the directiona cues. 9 target signa (without wind noise) 9 9 FWF method presented ange [degree] presented ange [degree] (a) (b) Figure Resut of directionaity test, (a) target signa and (b) proposed method. Here, horizonta and vertica axes indicate presentation and perceived anges, respectivey. (II) Externaity Test The output signas obtained from FWF and conventiona high-pass fiter methods were used as stimui, when the input signas of Japanese sentences have each seven directiona characteristics with wind noise. In this experiment, five subjects from 9 - years od were recruited. The subjects were presented randomy with two types of stimui that were derived from the processed signas of these methods. The question to the subjects was, which one is farther from your position?. Then, the subjects had to seect an answer from first one, second one, or even. The subjects had to answer even if he did not fee any difference. The experiments were performed for five subjects, and four inds of sentences were used, so the number of resuts was twenty per ange. Figure shows the resut of the externaity test. The horizonta and vertica axes show the ange of the stimui and the percentage of the seected externaity sound, respectivey. The resuts of a directions impy that the proposed method was superior from an externaity standpoint. 53

10 percent of each seection (%) INTER-NOISE presented ange [degree] FWF Method Even Conventiona Method Figure Resut of externaity test. Horizonta and vertica axes show presentation ange and percentage of each seection. 5. CONCLUTIONS We proposed a ow deay wind noise canceation agorithm for binaura hearing aids. The conficting probem of deay time and frequency resoution was soved by using a frequency warping fiter. The numerica simuation resut showed that wind noise can be canceed with accuracy by using a BWNC fiter. The objective evauation resuts revea that the SNR and sound quaity were improved whie maintaining a ow group deay time. As a resut of subjective evauation, a directionaity test indicated that the BWNC fiter did not infuence the perceived direction. Furthermore, the perceived spatia impression was sighty superior to the conventiona high-pass fiter method from an externaity standpoint. The resuts of our research show that the processing deay time of the proposed BWNC system can be shorter than the toerabe group deay time using a frequency warping fiter whie maintaining the ocaization cue of the desired signa. REFERENCES. S. Kochin. MareTra VIII: Consumer satisfaction with hearing aids is sowy increasing. Hearing Journa; ;63():9 3.. H. Dion, Roe I, Katch R. The sources of wind noise in hearing aids. Lae Tahoe, USA: IHCON; August. 3. Kates, J. M. Digita hearing aids. San Diego: Pura Pubishing; 8. p G. W. Eo. Reducing noise in audio systems. US Patent 7 7 8, Kates, J. M. Hearing aid with suppression of wind noise. US Patent 8 9 3, K. S. Petersen, G. Bogason, U. Kjems, T. B. Emedyb. Device and method for detecting wind noise. US Patent , J. A. Zais, C. M. Tan. Robust wind noise detection. Forence, Itay: IEEE ICASSP; 4. p Hiruma N, Naashima H, Fujisaa Y. Binaura wind noise canceation agorithm in hearing aids using HRTF cues. Lae Tahoe, USA: IHCON; August Naashima H, Kouyama R, Hiruma N, Fujisaa Y. Binaura wind noise detection, canceation and its evauation for hearing aids. IEEE IECON; 9- November 5. Yoohama, Japan: p H. Dion. Hearing Aids. Turramurra, Austraia: Boomerang Press;. p Gardner, B., Martin, K. HRTF measurements of a dummy-head KEMAR dummy-head microphone. 8 Juy ; Avaiabe from: Naashima H, Chisai Y, Usagawa T, Ebata M. Frequency domain binaura mode based on interaura phase and eve diffences. Acoust. Sci. & Tech. 3;4(4): p

11 INTER-NOISE 6 3. M. A. Stone, B. C. J. Moore, K. Meisenbacher, R. P. Dereth. Toerabe hearing aid deays. V. estimation of imits for open cana fittings. Ear Hear. 8;9(4): A. Härmä, M. Karjaainen, L. Savioja, V. Väimäi, U. K. Laine, J. Houpaniemi. Frequency-warped signa procesing for audio appications. J Audio Eng Soc. ;48(): Kates, J. M., Arehart, K. H. Mutichanne dynamic-range compression using digita frequency warping. EURASIP J App Signa Process. 5;8: