Broadband beamforming compensation algorithm in CI front-end acquisition

Transcription

1 Chen an Gong BioMeical Engineering OnLine 23, 2:8 RESEARCH Open Access Broaban beamforming compensation algorithm in CI front-en acquisition Yousheng Chen an Qin Gong * * Corresponence: gongqin@mail. tsinghua.eu.cn Department of Biomeical Engineering, Tsinghua University, Beijing 84, P.R. China Abstract Backgroun: To increase the signal to noise ratio (SNR) an to suppress irectional noise in front-en signal acquisition, microphone array technologies are being applie in the cochlear implant (CI). Due to size constraints, the ual microphonebase system is most suitable for actual application. However, irect application of the array technology will result in the low frequency roll-off problem, which can noticeably istort the esire signal. Methos: In this paper, we theoretically analyze the roll-off characteristic on the basis of CI parameters an present a new low-complexity compensation algorithm. We obtain the linearize frequency response of the two-microphone array from moeling an analysis for further algorithm realization. Realization an results: Linear metho was use to approximate the theoretical response with ajustable elay an weight parameters. A CI ual-channel harware platform is constructe for experimental research. Experimental results show that our algorithm performs well in compensation an realization. Discussions: We iscuss the effect from environment noise. Actual aily noise with more low-frequency energy will weaken the algorithm performance. A balance between low-frequency istortion an corresponing low-frequency noise nee to be consiere. Conclusions: Our novel compensation algorithm uses linear function to obtain the esire system response, which is a low computational-complexity metho for CI real-time processing. Algorithm performance is teste in CI CIS moulation an the influence of experimental istance an environmental noise were further analyze to evaluate algorithm constraint. Backgroun Over 2, recipients with severe hearing loss currently use a cochlear implant (CI) to improve their soun sensation []. Cochlear implant recipients generally require higher signal to noise ratios (SNRs) than normal listeners to obtain similar speech recognition [2]. In noisy environments, such as the cocktail party situation, the practical performance of the CI evice is ramatically weakene [3,4]. Microphone array technologies, which use multiple sensors to obtain aitional spatial information for noise suppression, have been applie in hearing ais (HAs) [5-8]. Given the size limits in CI proucts, the ual-channel array is most suitable for front-en signal acquisition. Cochlear implants are similar to HAs in their front- 23 Chen an Gong; licensee BioMe Central Lt. This is an Open Access article istribute uner the terms of the Creative Commons Attribution License ( which permits unrestricte use, istribution, an reprouction in any meium, provie the original work is properly cite.

2 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 2 of 2 noise-suppression methos from HA applications. Many theoretical an experimental stuies of array speech enhancement have been performe to improve CI speech recognition [9-]. A ual-microphone CI was recently esigne by Cochlear Lt. to improve the front-en SNR. However, the clinical CI evice is still equippe with a single omniirectional microphone, an no microphone array-base metho is applie in aily usage for the CI evice. Microphone array methos [2,3] a spatial information to the algorithms an are especially effective in suppressing irectional noise. These array-beamforming methos inclue fixe [4,5] an aaptive beamformers [6-2]. Present complicate algorithms combine beamforming technology with single-channel signal-filtering an noise-suppression methos [2-23]. The elay an sum beamforming metho, especially with two microphones, is more suitable for application in the CI evice, given its appropriate size constraints an low complexity in real-time processing [24]. Microphone array technologies are very effective to narrowban signals, whereas speech is a broaban signal. The irect application of narrowban technologies to acoustic an phonic fiels may result in low frequency roll-off. Stuies have reporte the practical implications of low frequency roll-off for HA evices [25], comparing sensatory performance results between situations with an without compensation for low frequency roll-off. Theoretically, slopes for the first- an secon-orer low frequency roll-off are 6 an 2 B/octave, respectively []. A solution for the low frequency roll-off is gain ajustment in each sub-ban. Corresponing gains nee to compensate for the energy loss, an the sub-ban signal is reajuste to be the same with the omniirectional microphone [26]. A first-orer ifferential low-pass filter is conventionally applie [7,9] for compensation of the low frequency roll-off. However, these methos cannot accurately compensate the esire gains for the low frequency roll-off. Increase focus has been given to broaban beamforming [27,28]. Post-filter technology uses maximum likehoo filter for spectral shaping an noise suppression [29]. However, these methos are very complicate, require large computational complexity, an excee the capabilities of current CI speech processors for portable application. We use the array metho for CI speech enhancement an further nee to compensate for the signal istortion introuce from the low frequency roll-off. For a simple ual-microphone evice with a elay parameter, we previously propose the normalize beamforming algorithm [3] base on the Taylor approximation. However, this metho is not applicable for a general array with elay an weight parameters. In this paper, we theoretically analyze the low frequency roll-off feature uner ifferent parameters. A ual-microphone harware platform, base on actual CI parameters, was constructe to obtain practical experimental results. From theoretical analysis an experimental observation of the roll-off, we propose a novel compensation algorithm for the low frequency roll-off with ajustable elay an weight parameters. The propose algorithm can accurately compensate the signal istortion an requires very few calculations. Further testing of this ajustment metho in the CI speech strategy reveale the efficiency an applicability of the propose compensation algorithm.

3 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 3 of 2 Methos Dual-microphone system in CI front-en Given the size constraint of CI evices an the limite calculation capability in the speech processor, the ual-microphone array is most suitable for use in front-en signal acquisition. Figure presents the system sketch for a ual-microphone array with elay an weight parameters base on the elay-an-subtract metho. Accoring to practical requirements, theinter-microphone istanceisaboutcm. Dual-channel signals recore by these 2 microphones are elaye an subtracte to yiel the esire irectional output signal. The elay-an-subtract metho is one of the most effective technologies in sizeconstraint evice, hearing ai an cochlear implant etc. In Figure, the two microphones each recor the signal, is the inter-microphone an c is the spee of soun in air. If we assume that the recore signal in Microphone is x(t), then the recore signal in Microphone 2 is asynchronous ue to the spatial ifference, with a aitional time elay c cosθ. This signal is elaye an weighte with the corresponing parameters τ an β to yiel a new signal, which is combinate with the signal in Microphone to obtain the irectional output y(t). The system magnitue response is given by Eq. (). H e jω ¼ βe 2jπf ð c cosθþτ Þ ðþ where the parameters f correspons to the narrowban frequency. For a ifferent orientation θ, the system response is also ifferent. The algorithm elay τ correspons to the beam pattern (ipolar pattern: τ = ; supercario pattern: τ =.342/c; cariois pattern: τ = /c) [3]. A fixe value of the elay τ can yiel a set of similar beam patterns. The magnitue response in Eq. () is base on a specifie narrowban frequency to obtain the irectional beams. Therefore, the system response is a function of the signal frequency f, an the use of ifferent frequencies results in ifferent magnitue responses. Low frequency roll-off characteristic We use the actual parameters of a CI evice, =. m an c = 34 m/s, in the theoretical analysis. We initially analyze a simplifie situation, in which the weight parameter β is x( t cos θ ) c xt () θ +τ { Delay + β Weight + yt () Figure Sketch of elay-an-subtract metho in the ual-microphone system.

4 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 4 of 2 chosen to be, an the 2 microphones are equally weighte. Figure 2 shows the ipolar, supercario, an carioi beam patterns for ifferent frequencies in this situation. Figure 2 escribes changes in the ual-microphone beams for ifferent frequencies (correspon the center frequencies of 8-channel CI filter bank, shown in Table ). The system magnitue response in Eq. () is a function of signal frequency. In each beam pattern (ipolar, supercario, an carioi), the system response increases as the corresponing frequency increases. Because speech is a broaban signal incluing various frequencies, the corresponing beams are not consistent with each other. In aition, the system response for low-frequency beams is smaller than that for highfrequency ones. Observe from these panels, the main lobe an sie lobe amplitue both increase at higher frequency. Therefore, we observe low frequency roll-off in the broaban application, which will introuce istortion to the esire speech. We can theoretically analyze the reason of low frequency roll-off. To simply the analysis, let β =, an the correspon response simplifie in Eq. (2). H e jω ¼ 2 sin πf c cosθ þ τ 2πf c cosθ þ τ ðif f Þ ð2þ When signal frequency is low, the system response is approximately proportional to f. Therefore, low frequency correspons to small response. As an example, this roll-off feature can easily be observe when we use white noise as testing signal, shown in Figure (a) Channel 8 j H (e ω ) Channel 6276 Hz 3869 Hz 2376 Hz 462 Hz j H (e ω ) (b) Channel 8 5 Hz 76 Hz 57.5 Hz 274 Hz Channel j H (e ω ) (c) Orientation Channel 8 Channel Figure 2 System responses base on ifferent frequencies for (a) ipolar, (b) supercarioi, an (c) carioi beam patterns.

5 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 5 of 2 Table Frequency parameters in 8-channel CI filter bank Ban ege (Hz) Center frequency (Hz) Channel [56, 396] 274 Channel 2 [396, 639] 57.5 Channel 3 [639, 883] 76 Channel 4 [883, 27] 5 Channel 5 [27, 797] 462 Channel 6 [797, 2955] 2376 Channel 7 [2955, 4783] 3869 Channel 8 [4783, 7769] 6276 The spectrogram of white noise contains with information from all frequencies. Corresponingly, the first-orer output will noticeably weaken the low-frequency signal in (b-2). Base on Eq. (), we can euce the following inequality. βe 2jπf c cosθ þ τ ¼ β cos 2πf c cosθ þ τ jβ sin 2πf c cosθ þ τ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ þ β 2 2β cos 2πf c cosθ þ τ þ β 2 2β ¼ j βj ð3þ Amplitue Amplitue Magnitue Magnitue Figure 3 Waveform of the original white noise signal (panel a-) an the first-orer ifferential output signal (panel a-2); an the corresponing spectrum are presente in panels of (b-) an (b-2) respectively.

6 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 6 of 2 It inicates that, for the situation of β =, the system response will approach zero, which correspons to frequency f =. Therefore, the amplitue response of the lowfrequency signal is smaller an will approximate to zero. Parameter analysis The most common situation uses a weight parameter of β =. Use of ifferent weights in the algorithm will change the corresponing beam patterns. We analyze beam features with elay times of τ =,.342/c, an /c (Figure 4). Figure 4 escribes the influence of ifferent weights on the system response (Channel No., 2, 3, 4, 5, 6, 7 an 8 correspon to 274, 57.5, 76, 5, 462, 2376, 3869 an 6276 Hz respectively). For β =, the aforementione analysis inicates that the lowfrequency response will approach zero. For β, the system response in the lowfrequency ban oes not approach zero, but approximates to a specific amplitue. For β =.2, the low-frequency ban approaches approximately.8. For β =.5, 2, an 4, the corresponing low-frequency responses are approximately.5,, an 3, respectively, consistent with the theoretical analysis of H(e jω ) min = β in Eq. (3). In the previous analysis, the low-frequency response oes not approach zero, but low frequency roll-off an response istortion are observe. In each panel, the system response increases with increasing signal frequency. In Figure 4, a-, a-2, a-3, an a-4 refer to erivative patterns base on the stanar ipolar beam (τ = ). In the range of j H (e ω ) H (e j ω ) H (e j ω ) H (e j ω ) Figure 4 Effect of the weight parameters on the system response for (a) ipolar, (b) supercarioi, an (c) carioi beam patterns, with corresponing weight parameters. Panels a-, b-, c-: β =.2; a-2, b-2, c-2: β =.5; a-3, b-3, c-3: β =2;a-4, b-4, c-4: β =4.

7 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 7 of 2 ± 8, we observe the same amplitue-changing feature, such that the high-frequency ban increases. The ipolar pattern in Figure 2(a) shows zero-response values (Nulls) at the ±9 orientations. In contrast, the erive ipolar beam (β ) in the 4 panels of Figure 4(a) shows Minimal values (Mins) rather than Nulls at ±9. For τ =.342/c an /c, the corresponing Mins are obtaine at ± an ±8, respectively. In Figure 4, the system response is locate at ifferent amplitue ranges for ifferent weights. Because the overall amplitue levels are ifferent, low frequency roll-off cannot be quantitatively observe irectly from the figure. To compare each roll-off characteristic on a unifie stanar, we normalize system responses at Hz to be the same. For a fixe weight, the amplitue at each orientation iffers from the others. Therefore, we nee to normalize the average amplitue from -8 to 8 to be. The Hz frequency is an infinitesimal case of low frequency, as the corresponing frequency response approximates to β (shown in Eq. 3). Figure 5 shows the normalize system responses. Figure 5 escribes the beam patterns on the basis of the response normalization at Hz. In each panel, the system response increases with increasing frequency, but the magnitue of the increase iffers for ifferent parameters. After normalization, for β =.5 (a-2, b-2, an c-2) an β = 2 (a-3, b-3, an c-3), the system responses can be more easily istinguishe an the high-frequency response increases more. In contrast, for β =.2 (a-, b-, an c-) an β = 4 (a-4, b-4, an c-4), the system response changes more smoothly. The situation with β > correspons to a larger gain ae to Carioi Supercarioi Dipolar Figure 5 Normalize beam patterns for (a) ipolar, (b) supercarioi, an (c) cariois, with corresponing weight parameters. Panels a-, b-, c-: β =.2; a-2, b-2, c-2: β =.5; a-3, b-3, c-3: β =2; a-4, b-4, c-4: β =4.

8 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 8 of 2 Microphone 2, an β < correspons to a larger gain ae to Microphone. Similar to the previous comparison, when the weight approaches, the high-frequency beams grow more rapily, which results in more low frequency roll-off. When the weight parameter is far from ( or ), the roll-off an corresponing istortion are less than the previous ones. The system response for β < can be given with the following equation transform. β e 2jπf c cosθ þ τ ¼ β e ¼ β e β 2jπf c cosθ þ τ 2jπf e c cosθ þ τ c cosθ þ τ ¼ β 2jπf e c cosθ þ τ β 2jπf ð4þ The equivalent equation in Eq. (4) inicates that the system response for β < correspons to the case of β 2 ¼ β : >, with only an aitional gain paramete, β. Therefore, the normalize system response presents ientical roll-off features for both situations. Equation (4) inicate that the weight value has reciprocal an symmetrical features; thus, we only nee to consier the case of β <. Realization an results Low frequency roll-off in ual-channel CI platform In this paper, we evelop a ual-channel harware platform, base on actual CI size constraints an esign requirements, for signal acquisition an algorithm analysis. The harware system inclues microphone moules, a signal acquisition circuit, signal transmitting evice, computer, an accessory evices (holer frame, etc.), as shown in Figure 6. The inter-microphone istance is ajuste to be cm, an the louspeaker is.5 m apart from the harware. The experiment is conucte in an orinary room ( m 8 m 3.5 m) with a reverberation time of about 4 ms. As the experiment begins, the louspeaker plays a voice recoring of a speaker reaing a paragraph of English material (Native American English). Two microphones recor this signal from the louspeaker, an the collecte signal is amplifie an filtere by the harware circuit. The USB soun car transfers the analog signal to igital signal, which is transmitte to the computer. The ata are store on a PC har isk for further analysis. To simplify the analysis of low frequency roll-off with a unifie stanar, an aitional gain is ae to the output signal to maintain a consistent inputoutput energy. On the basis of the previous theoretical analysis, the weight only nees to be. We choose the weight parameter to be.2,.4,.6,.8, or an the elay parameter τ to be,.342/c, or /c, to analysis the practical spectrums of the recore signal by the harware (Figure 7). Figure 7 shows signal spectra for ifferent elay times an weights. To simplify the analysis, the corresponing spectrum at each panel is normalize to maintain the same energy with the original signal. The spectrum of the original signal is shown for a etaile comparison of the low frequency roll-off. For a fixe elay time, as the weight

9 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 9 of 2 Figure 6 Dual-channel CI front-en platform. increases, the high-frequency response strengthens an the low-frequency response weakens. For a weight parameter of, the low frequency roll-off is most noticeable. For this speech signal, after energy normalization, signals in the ban between an 8 Hz are relatively weak, whereas the corresponing signals are relatively strong in the high-frequency ban (>8 Hz). The elay parameter also influences the spectrum istribution, although this effect is not as obvious as the influence of the weight parameter. Compensation algorithm We next attempt to compensate for the signal istortion base on the low frequency roll-off feature. For ifferent weight an frequency parameters, the system response is a function of β an f, H (e jω )=H (β, f). Using a weight range of β (similar to the previous section), the system responses for a signal between an 6 Hz are shown in Figure 8 (ipolar-base parameter of τ = ). Figure 8 presents the system responses for ifferent weights an frequencies, in which the weight is increasing at an interval of.2. The response curves are equally space, with goo linearity (especially in the high-frequency ban). We use (ƒ, H(β, ƒ)) to escribe points in the 2-imensional plane for the f-h function. For points on the left vertical axis (corresponing to f = ), the system responses are equally space istribute. For a fixe weight β, the response function H (β, f) f = = β, such that the coorinate of the left enpoint in the f-h function is (, β). For f = 6 Hz, the right enpoints in the response curves are also equally space, but with a turning point of H 2 corresponing to β =.45. From Figure 8, we calculate three response values: H 3 = H (β, f) β =,f = 6 =, H 2 = H (β, f) β =.45,f = 6, an H 3 = H (β, f) β =,f = 6. For the case of β..45, the system response at f = 6 Hz ecreases in an equally space manner from H to H 2. For the case of.45 < β., the system response increases in an equally space manner

10 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page of 2 Original spectrum τ = τ =.342 / c / c β =.2 β =.4 Magnitue β =.6 β =.8 β = Figure 7 Signal spectrums for ifferent elay times an weights. from H 2 to H 3. Consequently, we obtain the following system response at the frequency f = 6 Hz. 8 >< β ð Hðβ; f Þ f ¼6 ¼ :45 H 2Þ; β :45 ð5þ >: β :45 H 2 þ ðh 3 H 2 Þ; :45 < β :55 The previous analysis inicates the approximate linearity of the system response. To reuce the computational complexity in CI evices, we linearize the f-h function. We use the curve enpoints (, β) an (6, H (β, f) f = 6 ) to yiel the approximate system response H eva (β, ƒ), as shown in Eq. (6). H eva ðβ; f Þ ¼ 8 >< >: β ð :45 H 2Þ β þ 6 β :45 ðh 3 H 2 :55 H 2 þ β þ 6 ð β Þ f ; β :45 Þð βþ f ; :45 < β ð6þ

11 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page of 2.2 β = β = = H H f β 3 (, ) f H = β =, = 6 = H H f β 2 (, ) f β = =.45, 6 H( β, f ) We use the expression βe 2jπf ð c cosθþτ Þ, the ipolar-base parameter τ =, an the actual CI parameters =. m an c = 34 m/s to calculate the values of H 2 an H 3 an to obtain the linear expression of the f-h function, given by Eq. (7). H eva ðβ; f Þ ¼ frequency (Hz) Figure 8 System-response function for ifferent weights an frequencies. β þ ð:279 4 βþf ; β :45 β þ 2:45 4 β 3:896 5 f ; :45 < β ð7þ For a fixe beam pattern, we use Eq. (7) to obtain the linear system response. The corresponing response curves are shown in Figure 9 (a). Panels (a) an (b) present the linear f-h function an the relative errors calculate by Eq. (8), respectively. E ¼ H evaðβ; f ÞHðβ; f Þ % Hðβ; f Þ ð8þ As seen in Figure 9(b), the relative error ranges from -% to 4%. A comparison of Figures 8 an 9(a) inicates that the theoretical f-h function (ieal system response) is primarily a set of concave curves, such that most of the response amplitues in the linearization curve are greater than the ieal ones. Therefore, the concave feature of the linearization curve inicates that most of the relative errors are positive, but negative relative errors are lacking. Therefore, the whole response amplitues are larger than the ieal ones. This fining implies that the total energy of the output signal is strengthene..2 β = β = 4 3 H( β, f) (a) frequency (Hz) Enor ( %) (b) frequency (Hz) Figure 9 Linearization curves of the f-h response function (a) an corresponing errors (b).

12 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 2 of 2 We only nee to analyze the relative amplitue ifference for ifferent signal bans because the overall an coorinative enhancement for the compensation will not introuce istortion to the esire signal. To evaluate the compensation error in a etaile an accurate manner, we normalize the relative error to balance the input an output signals. The normalize error is efine by Eq. (9). E nor ¼ H evaðβ; f ÞGðβ; f Hðβ; f Þ Þ Hðβ; f Þ % ð9þ where the normalize coefficient G (β, ƒ) is given in Eq. () to maintain the energy equilibrium for a signal between an 6 Hz. When N is large, the frequency ban is fractionize enough, an the calculate normalize coefficient can approximate the ieal value. vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X N i 2 u H β; 6 Gðβ; f Þ ¼ lim u t i¼ N X N i¼ N i H β; 6 eva N The corresponing normalize error is presente in Figure. Accoring to Figure, the normalize error ranges between -% an 4%. However, most of these errors are concentrate in the ±% range (corresponing to the error istribution in the high-ensity ark part of the figure), with only a few large errors. Therefore, most of the errors are very small. Base on a fixe weight parameter, figure also inicates that the error will change for ifferent frequencies. To evaluate the total error further, we use equation () to calculate the average error. 2 ðþ E ave ¼ lim N N X N i¼ E nor j fi j ðþ For ifferent weight parameters, the corresponing average error (in re) is shown in Figure (N is 5 in this paper). 4 3 Enor ( %) frequency (Hz) Figure Normalize errors of the f-h linearization curves.

13 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 3 of 2 τ = τ =.342 c τ = c Figure Average errors base on ifferent weights an elay times. β For ifferent weight values, the average error (in re) ranges between % an 7%, corresponing to the ipolar-base parameter τ =. For other elay values, supercarioi- (blue), an carioi-base (green) parameters, analogous results can be obtaine, with average errors ranging from % to 7%. These errors are acceptable an sufficiently small for the CI application. For a set of fixe parameters, the system beam pattern is fixe. The propose linear algorithm for the compensation of low frequency roll-off can approximate the ieal system response with few errors. Aitionally, this algorithm is low-calculation for the real-time processing in the CI speech processor. The broaban signal is ivie into several sub-bans uring signal processing of the CI speech strategy. Each sub-ban inclues the signal in its corresponing frequency ban, in which the signal envelope information is extracte an transferre to the next process. Because the CI filter bank ivies the signal into many bans, the ivie signal in each ban is approximately narrowban. The CI speech strategy can moulate the information in each ban an sen it to the electroe array that, with a specific stimulating rate, stimulates the auitory nerve to generate acoustical sensation. The stimulating rates in the electroe array correspon to the center frequencies of the filter bank. The propose algorithm uses the center frequency f cen-i in each sub-ban for the compensation. One of the center frequencies is applie as the reference frequency, f cen-ref. Because most speech signals are concentrate in the low-frequency ban peaking aroun Hz, the reference frequency is chosen in the ban near Hz. For compensation of low frequency roll-off in the i-channel, the corresponing center frequency f cen-i an reference frequency f cen-ref are use in the gain ajustment, as shown in Eq. (2). G channel i ¼ H eva β; f cenref H eva ðβ; f ceni Þ ð2þ The reference an i-channel center frequencies in this gain compensation equation are base on the linearization response function; therefore, the gain ajustment nees only a few aitional calculations.

14 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 4 of 2 Compensation results in the ual-channel CI platform The low frequency roll-off experiments are conucte on our ual-channel harware platform. We a compensation gains to the filter bank of the CI speech strategy to ajust the amplitue for each channel. Recore signals are moulate by the Continuous Interleave Sampling (CIS) strategy [32]. We use sinusoial moulation to actualize this speech strategy. A set of sinusoial signals are moulate by the corresponing envelope of the ban-pass signals after frame-ivision in the filter bank. Then, the original continuous spectrum becomes several iscrete frequency components (line spectrum), each of which correspons to the sub-filter an CI electroe. The electroe array, base on the corresponing stimulating rate, stimulates the nerve to yiel auitory perception. In this paper, the frequency components after CIS moulation match the center frequency of the CI filter bank. The Welch metho [33] is use to calculate the Power Spectral Density (PSD), to compare results with an without algorithm compensation, as shown in Figure 2. The experiment is conucte in quiet environment (The SNR is about 5 B an the performance of the compensate beamformer is evaluate for ifferent noise levels in the iscussion section). For ifferent weight parameters, Figure 2 escribes the signal PSD (8 line spectrums, corresponing to the 8-channel filter bank) after CIS moulation of the CI evice. The top panel presents the original CIS spectrum obtaine by an omniirectional microphone without signal istortion. Figure 2 shows the corresponing CIS spectrums with (b) an without (a) algorithm compensation when the signal is recore by the ual-channel array. Both CIS spectrums are normalize to balance the overall energy with the original signal. Without compensation of the low frequency roll-off, the signal PSD istribution changes noticeably. The low-frequency β =.2 β =.4 β =.6 β =.8 β = Magnitue Magnitue Magnitue Magnitue Figure 2 PSD for the CIS signal (a) without ajustment or with ajustment by (b) the propose algorithm or (c) ieal coefficient for low frequency roll-off.

15 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 5 of 2 amplitue is weakene, whereas the high-frequency amplitue is relatively strengthene for both weight parameters. A greater weight results in more obvious low frequency roll-off. Panel (b) shows that, after algorithm compensation, amplitues in these 8 channels match the original CIS spectrum well. Therefore, the propose ajustment algorithm can accurately compensate the signal istortion in array beamforming. For further comparison, Figure 2(c) shows the compensation results base on the ieal compensation coefficients from the theoretical system response. The ieal compensation uses the i-channel center frequency f cen-i an reference frequency f cen-ref to obtain the gain ajustment, given in Eq. (3). H e jω β; fcenref G Iealchannel i ¼ Hðe jω ð3þ j Þj β; fceni Equation (2) use the accurate response amplitue, base on a set of fixe frequency an weight parameters, to calculate the ieal ajustment coefficients for the compensation of low frequency roll-off. A comparison of panels (b) an (c) shows that the compensation results by our algorithm are very consistent with the ieal response-base gain ajustment. For etaile comparison, we calculate the corresponing 8-channel spectrum amplitues for situations (a) without compensation, (b) with compensation by the propose algorithm, an (c) with ieal compensation. The spectrum amplitues are compare to the original signal spectrum amplitue (in B) to obtain the relative enhancement or attenuation results (Figure 3). Figure 3 shows that low frequency roll-off is obvious when the array output signal is not ajuste (a). The low-frequency signal is weakene, ranging from to -2 B, an the high-frequency signal is relatively enhance, ranging from to +2 B. For the whole frequency ban, the signal overall istortion is extremely large, between -2 Error (B) (b) 2 3 β =.2 β =.4 β =.6 β =.8 β = Error (B) Channel i ata ata2 ata3 ata4 6 (a) 7 8 (c) Channel i Figure 3 PSD errors for the (a) non-ajuste, (b) algorithm-ajuste, an (c) ieal-ajuste CIS signal, compare to the original CIS signal (in B)

16 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 6 of 2 an +2 B. Panel (b) presents the compensation results obtaine with our propose algorithm. The signal istortion is very small, with only a few amplitue ifferences between -2 an 3 B. When the signal is compensate by the ieal coefficients (c), the error ranges from - to 2.5 B. Therefore, the compensation accuracy by our algorithm can approach the ieal response-base ajustment. The ajuste signal matches the esire CIS spectrum well, with little istortion. Discussions In the previous harware experiments, the louspeaker is.5 m apart from the harware. Actually, a common rule of for the range at which the transition from spherical waves to planes waves occurs for a monopole source is at least two times the wavelength, so the corresponing minimal istances for 8 channel of CI filter bank is given in Table 2. Seen from Table 2, we fin that signal acquisition will be influence in Channel as the minimal istance Distance between louspeaker an microphone is 2.48 m. however, the practical usage an aily face-to-face communication for CI users is.5 m, so we finally ajuste the louspeaker.5 m apart from the harware platform. An.5 m correspon to 453 Hz, therefore, signal with frequency higher than 453 Hz will not be influence. The environment noise is ifferent from the white noise, with more noise in the low frequency ban. In aily usage, the car & fan noise (low-frequency signal) were recore in microphones. We use the harware platform to recor the environment noise, shown in Figure 4. Observe from panel (b-), most environmental noises contain most of the energy concentrate at low frequencies. Panel (b-2) presents the signal spectrum after a firstorer ifferential processing in the ual-microphone array. The corresponing low frequency energy was sharply weakene; however, the amplitue was still very large. Therefore, the low frequency roll-off was not always a ba thing an boosting up low frequency contents might increase internal microphone noise together. A balance between low-frequency istortion an corresponing low-frequency noise nee to be consiere. This compromise nees an actual test of a set of suitable attenuation coefficient base on omniirectional microphone, an the coefficients were ae in the usage of array beamforming, to obtain low istortion an less low frequency noise. Table 2 Minimal istance for 8-channel CI filter bank Center frequency (Hz) Distance between louspeaker an microphone (m) Channel Channel Channel Channel Channel Channel Channel Channel

17 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 7 of (a-) waveform (b-) spectrogram (a-2) t (s) (b-2) Normalize frequency Figure 4 Waveform of (a-) original backgroun noise, (a-2) first-orer ifferential output signal, an the corresponing spectrum (b-) an (b-2) respectively. For ifferent noise levels (, 5 an B respectively), the performance of the compensate beamformer is given in Figure 5. Figure 5 escribes the signal PSD for ifferent noise levels. When noise increase, with smaller SNR, more low-frequency energy was ae in. Particularly, the case of SNR = B presents noticeable results that the signals in low-frequency are enhance from channel to channel 3. Seen from these panels, lower SNR will introuce more istortion in CI evices. Conclusion The microphone array noise-suppression metho can separate the esire speech an ambient noise on the basis of their spatial ifferences. Use of a ual-channel array with appropriate size constraints is more suitable for CI evices. However, irect application of the narrow-ban metho in broaban speech will yiel low frequency roll-off an noticeable signal istortion. Low-frequency loss from the speech signal can be observe in first- an secon-orer ifferential systems [34]. To compensate for low frequency roll-off, conventional methos use only a simple low-pass filter to enhance the lowfrequency signal an weaken the high-frequency signal. These methos are not sufficiently accurate to match the original signal. The broaban beamformer was recently introuce as a metho to obtain precise compensation. However, these algorithms require extensive calculation, preventing their actual application in CI evices. In our previous work, we construct a microphone array base platform for signal acquisition. To suppress the environmental noise, we use elay-an-subtract metho an propose the optimal parameter section methos of elay an beamforming for CI speech enhancement [24]. In our later work, we aim to compensate the low frequency roll-off in speech application an propose the normalize beamforming algorithm using a continuous interleave sampling strategy [3]. However, this work only contain

18 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 8 of 2 Magnitue Original PSD Magnitue Algorithm-ajuste PSD (a) SNR= B (b) SNR=5 B (c) SNR= B frequency (khz) frequency (khz) frequency (khz) 7 β =.2 β =.4 β =.6 β =.8 β = 8 Figure 5 PSD of the CIS signal with the compensate beamformer for ifferent noise levels (a) SNR = B, (b) SNR = 5 B an (c) SNR = B. elay parameter, but without weight parameter. In this paper, we propose a novel CI filter bank-base algorithm for the compensation of low frequency roll-off. This metho, with ajustable elay an weight parameters, uses a linear function to approximate the esire system response, with very low computational complexity. Theoretical an experimental results inicate that our algorithm can accurately compensate the signal istortion an is easy to embe in the CI speech strategy, supporting its practical application in the CI evice. Abbreviations CI: Cochlear implant; SNR: Signal to noise ratio; CIS: Continuous interleave sampling strategy. Competing interests The authors eclare that they have no competing interests.

19 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 9 of 2 Authors contributions YC initiate an conceive the algorithm, esigne experiments an analyze the ata. QG is the corresponing author. This stuy was solve uner QG s irection. QG rafte this paper s manuscript, incluing content an overall arrangement. QG was also responsible for revising this manuscript. All authors rea an approve the final manuscript. Acknowlegements The authors are grateful for support from the National Natural Science Founation of China (grant no ). Receive: 26 October 22 Accepte: 2 February 23 Publishe: 27 February 23 References. Chung K, Zeng FG: Using hearing ai aaptive irectional microphones to enhance cochlear implant performance. Hearing RES 29, 25: Nelson PB, Jin SB, Carney AE: Unerstaning speech in moulate interference: cochlear implant users an normal-hearing listeners. J Acoust Soc Am 23, 3(2): Ginger Stickney S, Zeng FG: Cochlear implant speech recognition with speech maskers. J Acoust Soc Am 24, 6(2): Hu Y, Loizou PC: A new soun coing strategy for suppressing noise in cochlear implant. J Acoust Soc Am 28, 24(): Leeuw AR, Dreschler WA: Avantages of irectional hearing ai microphones relate to room acoustics. Auiology 99, 3(6): Walen BE, Surr RK, Cor MT, Ewars B, Clson L: Comparison of benefits provie by ifferent hearing ai technologies. J Am Aca Au 2, : Surr RK, Walen BE, Cor MT, Olsen L: Influence of environmental factors on hearing ai microphone preference. J Am Aca Auiol 22, 3(6): Xun H, Chi X, Long Bai: Is the cochlea coile to provie soun localization. Europhys Lett 22, 98(5):582(p p5). 9. Wouters J, Vanen BJ: Speech recognition in noise for cochlear implantees with a two-microphone monaural aaptive noise reuction system. Ear Hear 2, 22: Chung K, Zeng FG, Acker KN: Effects of irectional microphone an aaptive multichannel noise reuction algorithm on cochlear implant performance. J Acoust Soc Am 26, 2(4): Spiet A, Van Deun L, Eftaxiais K, Laneau J, Moonen M, van Dijk B, van Wieringen A, Wouters J: Speech unerstaning in backgroun noise with the two-microphone aaptive beamformer BETM in the nucleus freeom cochlear implant system. Ear Hear 27, 28(): Carter G: Variance bouns for passively locating an acoustic source with a symmetric line array. J Acoust Soc Am 977, 62(4): Huang YT, Benesty J, Elko GW: Passive acoustic source localization for vieo camera steering. Proc IEEE International Conf Acoust, Speech Signal Process, Istanbul, Turkey 2, 2: Flanagan JL: Computer-steere microphone arrays for soun transuction in large rooms. J Acoust Soc Am 985, 78(5): Mahieux Y, Le Tourneur G, Saliou A: A microphone array for multimeia workstations. J Auio Eng Soc 996, 44(5): Griffiths LJ, Jim CW: An alternative approach to linearly constraine aaptive beamforming. IEEE Trans on Antennas Propagation 982, 3: Elko GW, Pong ATN: A simple aaptive first-orer ifferential microphone. New Paltz, NY, USA: IEEE ASSP workshop on Applications of Signal Processing to Auio an Acoustics; 995: Elko GW: Microphone array systems for hans-free telecommunication. Speech Communication 996, 2: Luo FL, Yang J, Pavlovic C, et al: Aaptive null-forming scheme in igital hearing ais. IEEE Trans Acoust, Speech Signal Process 22, 5(7): Matsumoto M, Hashimoto S: A miniaturize aaptive microphone array uner irectional constraint utilizing aggregate microphones. J Acoust Soc Am 26, 9(): Zelinski R: A microphone array with aaptive post-filtering for noise reuction in reverberant rooms, proceeings of IEEE international conference on acoustics, speech, an signal processing. New York: NY, USA; 988: Chou T: Frequency-inepenent beamformer with low response erro. 5th eition. Detroit, MI, USA: Proceeings of IEEE International Conference on Acoustics, Speech, an Signal Processing; 995: Aarabi P, Shi G: Phase-base ual-microphone robust speech enhancement. IEEE T SYST MAN CY 24, 34(4): Qin G, Yousheng C: Parameter section methos of elay an beamforming for cochlear implant speech enhancement. Acoust Phys 2, 57(4): Cor MT, Surr RK, Walen BE, Dyrlun O: Relationship between laboratory measures of irectional avantage an everyay success with irectional microphone hearing ais. J Am Aca Auiol 24, 5: Kuk F, Keenan D, Lau C-C, Luvigsen C: Performance of a fully aaptive irectional microphone to signals presente from various azimuths. J Am Aca Auiol 25, 6: Goara LC, Jahromi MRS: Limitations an capabilities of frequency omain broaban constraine beamforming scheme. IEEE Signal Processing Lett 27, 4(5): Yan SF, Hou CH, Ma XC, et al: Convex optimization base time-omain broaban beamforming with sielobe control. J Acoust Soc Am 27, 2():46 49.

20 Chen an Gong BioMeical Engineering OnLine 23, 2:8 Page 2 of Marro C, Mahieux Y, Simmer KU: Analysis of noise reuction an ereveraberation technologies base on microphone arrays with postfiltering. IEEE Trans Speech Auio Process 998, 6: Yousheng C, Qin G: A normalize beamforming algorithm for broaban speech using a continuous interleave sampling strategy. IEEE Trans Auio, Speech, Lang Process 2, 2(3): Chung K: Challenges an recent evelopments in hearing ais part I. Speech unerstaning in noise, microphone technologies an noise reuction algorithms. Trens Amplif 24, 8(3): Wilson BS, Lawson DT, Zerbi M, et al: Design an evaluation of a continuous interleave sampling(cis) processing strategy for multichannel cochlear implants. J Rehabil Res Dev 993, 3():. 33. Welch PD: The use of fast Fourier transform for the estimation of power spectra: a metho base on time averaging over short moifie perioograms. IEEE Trans Auio Electroacoust 967, 5: Thompson SC: Tutorial on microphone technologies for irectional hearing ais. Hear J 23, 4(3):2 35. oi:.86/ x-2-8 Cite this article as: Chen an Gong: Broaban beamforming compensation algorithm in CI front-en acquisition. BioMeical Engineering OnLine 23 2:8. Submit your next manuscript to BioMe Central an take full avantage of: Convenient online submission Thorough peer review No space constraints or color figure charges Immeiate publication on acceptance Inclusion in PubMe, CAS, Scopus an Google Scholar Research which is freely available for reistribution Submit your manuscript at