Real-time spectrum estimation–based dual-channel speech-enhancement algorithm for cochlear implant
- 3.9k Downloads
Improvement of the cochlear implant (CI) front-end signal acquisition is needed to increase speech recognition in noisy environments. To suppress the directional noise, we introduce a speech-enhancement algorithm based on microphone array beamforming and spectral estimation. The experimental results indicate that this method is robust to directional mobile noise and strongly enhances the desired speech, thereby improving the performance of CI devices in a noisy environment.
The spectrum estimation and the array beamforming methods were combined to suppress the ambient noise. The directivity coefficient was estimated in the noise-only intervals, and was updated to fit for the mobile noise.
The proposed algorithm was realized in the CI speech strategy. For actual parameters, we use Maxflat filter to obtain fractional sampling points and cepstrum method to differentiate the desired speech frame and the noise frame. The broadband adjustment coefficients were added to compensate the energy loss in the low frequency band.
The approximation of the directivity coefficient is tested and the errors are discussed. We also analyze the algorithm constraint for noise estimation and distortion in CI processing. The performance of the proposed algorithm is analyzed and further be compared with other prevalent methods.
The hardware platform was constructed for the experiments. The speech-enhancement results showed that our algorithm can suppresses the non-stationary noise with high SNR. Excellent performance of the proposed algorithm was obtained in the speech enhancement experiments and mobile testing. And signal distortion results indicate that this algorithm is robust with high SNR improvement and low speech distortion.
Signal to noise ratio
First-order differential microphone
Adaptive null-forming method
Multiple input/output inverse method
Minimum-variance distortionless-response technique
Finite impulse response
Continuous interleaved sampling strategy
Advanced combined encoder strategy.
The clinical cochlear implant (CI) has good speech recognition under quiet conditions, but noticeably poor recognition under noisy conditions . For 50% sentence understanding [2, 3], the required signal to noise ratio (SNR) is between 5 and 15 dB for CI recipients, but only −10 dB for normal listeners. The SNR in the typical daily environment is about 5–10 dB, which results in <50% sentence recognition for CI users in a normal noise environment.
Most previous studies on recognition improvement have focused on the coding strategy, design of the electrode array, and stimulation adjustment of pitch recognition, as well as on the virtual electrode technique [4, 5] and optical CIs . More recent efforts have focused on the microphone array technique [7, 8]. This array beamforming method promises to be more effective for situations in which the desired voice and ambient noise originate from different directions, the usual work environment for CI devices.
Speech-enhancement methods include single- and multichannel techniques. Spectral estimation methods are the most widely used single-channel techniques. Typical single-channel approaches, such as the spectral subtraction [9, 10], Wiener filtering , and subspace approach , are based on estimations of the power spectrum or higher- order spectrum, assume the noise to be stationary, and use the noise spectrum in the nonspeech frame to estimate the speech-frame noise spectrum. Algorithm performance sharply weakens when the noise is non-stationary, or under typical situations with music or ambient speech noise.
The microphone array technique considers the signal orientation information and focuses on directional speech enhancement. Specifically, the generalized sidelobe canceller  and delay beamforming [14, 15] use multiple microphones to record signals for spatial filtering. For CI devices, the generalized sidelobe canceller is overly complicated and requires too many microphones, conditions that exceed the capabilities of current CI devices. Delay beamforming technologies, such as the first-order differential microphone (FDM)  and adaptive null-forming method (ANF) [17, 18], are adopted in hearing aids. These methods need only 2 microphones, which is an appropriate set-up for the CI size constraint and real-time processing.
CI devices are similar with the hearing aids in size constraint and the requirement of front-end noise suppression. So, for CI speech enhancement, one simple solution for CI speech enhancement is to directly utilize the microphone-array–based noise-reduction methods from the present hearing aids, in which the sensor-array techniques have been more widely used. However, the difference between CI devices and hearing aids is prominent, and a direct application of these algorithms to CI speech processing is not appropriate. Firstly, the principle is very different. CI devices transfer the acoustic signal to electrical stimulation into the cochlea wirelessly, and then the electrical pulses are used to directly stimulate the acoustic nerve to yield the auditory perception. But the hearing aids only need to change the corresponding gains in different subbands for multi-frequency signal loss. In brief, the hearing aid is only an amplifier with adjustable gain in different frequency band. Secondly, the application of the microphone array technique is different. Many algorithms for speech application were borrowed from the narrowband methods in radar and antenna. Algorithms for front-end enhancement are indispensable to match the CI speech strategy. Thirdly, the solution for low frequency roll-off may be different. The hearing aids need to calibrate and preset the subband gain based on user’s hearing loss. Therefore, in the hearing aid, one solution is to directly preset the subband gains in the filter banks in the processor by both taking the hearing loss and signal loss in microphone array algorithm into account. However, for CI devices with the modulated electrical pulse directly stimulate the cochlear nerves, we only need to adjust the algorithm loss. Finally, the signal distortion is different. When the enhanced signal is modulated by the CI speech strategy, the signal distortion will noticeably decreased (detailed analysis was given in the result section). Therefore, an array for a cochlear implant is similar to a hearing aid in the speech-enhancement situation, but is different for the actual algorithm design, such as the tradeoff between speech distortion and noise suppression.
The Frost algorithm [19, 20], multiple input/output inverse method (MINT) , minimum-variance distortionless-response technique (MVDR) [22, 23] and the binaural frequency-domain minimum-variance algorithm  are proposed presently, with excellent performance in some specific situations. Kates  used a novel five-microphone end-fire array, with an MVDR included, to construct an adaptive frequency-domain noise-reduction algorithm with higher SNR improved. However, this algorithm is overly complicated, and the five-microphone array also exceeds the CI size constraint.
In daily environment, we previously proposed a low-complexity beamforming with optimal parameter to suppress the environmental stationary noise . But for the music and speech noise, we need a higher SNR to weaken these ambient noises, aiming to obtain more than 10 dB SNR for the CI front-end signal acquisition. This paper focuses on directional noise suppression with one directional ambient interference for CI devices. In typical situations in which CI users want to talk with a nearby person in a conference hall or a theater, the directional voice from the lecturer or film screen must be suppressed. To weaken the directional noise in such situations, a dual-channel CI speech-enhancement algorithm was introduced that combines the single-channel power spectrum estimation and the first-order differential microphone technique of the microphone array, for beamforming and noise prediction. Our algorithm uses the dual-channel power spectrums in the noise-only intervals, including the nonstationary noise, to estimate and update the noise directivity coefficient. For noise changing in normal human walking velocity, the proposed algorithm can avoid the noise leakage and is robust to mobile noise. For spectrum estimation based speech enhancement, the speech distortion is also unavoidable in our algorithm. But when the signal is modulated in the CI speech strategy, the speech distortion will sharply decrease and the speech quality noticeably improves. The experimental results indicate that the proposed algorithm successfully achieves the desired speech reconstruction and enhancement.
Figure 1 shows the flow chart of the proposed dual-channel speech-enhancement algorithm. Signals recorded by two omnidirectional microphones and two delaying signals by the delay filters are summed to yield dual-channel outputs. Firstly, the signal frequency response in each channel is extracted to obtain the power spectrum and the cepstrum distance. The cepstrum distance differentiates the desired-speech and noise-only power spectrums. Then the noise-only power spectrums are adopted to estimate the noise directivity coefficient. The narrowband signal magnitude is estimated by these power spectrums, including the desired-speech and noise-only segments, and the directivity coefficient. Furthermore, the narrowband signal magnitude, also named as single-frequency magnitude, is adjusted to yield the multifrequency magnitude for the desired speech (broadband signal with the compensation for low-frequency loss). And the phase information in channel 1 is extracted and inserted for signal reconstruction to obtain the enhanced speech signal. Our proposed algorithm is theoretically depicted as below:
where and .
Seen from Eq. (10), the power spectrum for each framed data set in channel 1 includes the power spectra of desired speech and the ambient noise. Equation (11) only contains the power spectrum of the ambient noise in channel 2. In addition, the power spectra of the noise in channels 1 and 2 are different, which are functions of the noise azimuth.
Eq. (14) indicates the algorithm for magnitude estimation must obtain the power spectrum of the framed data set of each channel, as well as the power spectrum of each channel at the noise-only frame.
For an actual CI size constraint of d≈ 0.01 m, the directivity coefficient approaches to cot 4(φ/2). This result indicates that the estimation of the directivity coefficient is robust, because it only depends on the noise direction φ. This simplified form of noise directivity coefficient also implies that, for noise direction with slowly varying, the adjusted gain for noise reduction can be accurately obtained with excellent algorithm stability.
In Figure 1, the recorded signals MIC1(t) and MIC2(t) are sampled at the 44.1 kHz sampling rate by the AD converter as MIC1(n) and MIC2(n). These digital signals are then delayed by the fractional delay filter with an algorithm offset of d/c. In our hardware platform, the system design specifies the intermicrophone distance d to be at or near 1 cm, corresponding to 1.297 sampling points.
For fs = 44.1 kHz, the digital signals MIC1(n) and MIC2(n) are delayed, with d/c offsets of MIC1(n-1.297) and MIC2(n-1.297), respectively.
The proposed fourth-order Maxflat FIR filter agrees well with the ideal filter at the signal range of 0–6000 Hz, a range that includes most of the subbands of the CI filter bank. The maximal error for the magnitude response and phase response are less than 0.3% and 0.4%, respectively. Additionally, this filter can obtain the required delaying signal easily, with low computation complexity.
This Maxflat digital filter is used to cover frequencies between 0 Hz and 6000 Hz as the CIS strategy does. The range of the CIS actually depends on the corner and center frequencies of the filters, therefore, the frequency range will change based on different channel quantities. But both the present CI filter banks (8, 16, 24 channels etc.) are primarily within these range. So, the required signal delaying is accurate.
Directivity estimation and noise-frame identification
The directivity coefficient (Figure 1 and Eq. (15)) is obtained from the power spectrums of the 2 channels at the noise-only frames, which correspond to the time-domain signals of ch1(t)|s(t)=0 and ch2(t)|s(t)=0, respectively. We used the cepstrum [31, 32, 33] method to differentiate the desired speech frame and the noise frame.
Equation (18) with a predefined threshold value is used to differentiate the noise frames for both channels, ch1(t)|s(t)=0 and ch2(t)|s(t)=0. Then, the corresponding power spectra are used to obtain the real-time noise directivity coefficient.
Broadband signal adjustment
The proposed filter for multi-frequency adjusting is highly consistent with the required coefficient function between 100 and 6000 Hz. Because λ(f) = 30λ’(f), the filtered signal needs an additional 30 times gain (or 29.54 dB) for signal energy rebalancing.
Parameters of each sub-band in CI filter bank and the corresponding adjusting coefficients
Band edge (Hz)
Center frequency (Hz)
Figure 4 describes the transmission of multi-frequency adjusting coefficients to the CI processor. This method of directly transmitting the parameters to the CI filter bank requires very little additional calculation, which is suitable to the situation of a filter bank–based strategy. For the situation in which the CI processor uses the speech-processing strategy without a filter bank, the proposed Butterworth filter (Figure 3) should be adopted for the coefficient adjusting.
In our CI speech-enhancement platform, the sampling rate is 44.1 kHz and the Hamming window is used for framing, with a window length of 1024 sampling points. Each frame is about 23 ms in duration, with 50% overlap.
The speech magnitude is estimated by Eq. (14). The signal phase of the original data (channel 1) is used directly for signal construction, because the human cochlea is relatively insensitive to phase information. The gain of the single-frequency signal is adjusted by the proposed Butterworth filter or by directly transmitting the adjusting coefficients in the filter-bank based CI processor. This broadband signal, expressed in the form of frequency response, is processed by the subsequent processing of the inverse Fourier transform and deoverlapping to reconstruct the enhanced speech signal.
The experiments were carried out in a chamber, or to be extract, an actual office measured by 8 m× 8 m× 4 m with the room reverberation time T60 = 450 ms. Two microphone modules, placed at the center (O), recorded the signal. P1-P12, which represent 12 testing points at 15° intervals, were marked 1.5 m from the microphones arranged in a semicircle. P1 indicates the forward direction for playing the desired speech by Speaker 1; the other locations (P2-P12) indicate the directions for playing the ambient noise by Speaker 2. The recorded signals by this hardware system, after amplification, filtering, and analogue to digital conversion, were transmitted to the computer via a USB interface for further analysis.
Figure 6 (a) is the original waveform of the desired speech, played by Speaker 1 in P1. Speaker 2 located at P7 plays the noise. Figure 6 (b-1) and (b-2) are the plots recorded by the hardware platform (located at O) corresponding to the situations of music noise and speech noise, respectively.
where the output SNR uses the estimation of ŝ(n) and to obtain.
For the music noise, use of the single-channel method (panel c-1) based on spectral subtraction weakened most of the music noise, but much of the transient impulse at the nonstationary part of the music noise remained. Panels d-1 and e-1 plot the signal outputs in channel 1 and channel 2 in our dual-channel system. Comparison of (d-1) and (e-1), the magnitude attenuation or enhancement were different for the desired speech and music noise. And this characteristic can remarkably be seen in (d-2) and (e-2) for the ambient speech noise. The waveforms of the two channels were similar in time domain, but the gains in channel 2 were discrepant, with respectively about 0.3 and 2.8 gains for the desired speech and ambient speech compared with channel 1. These gains changed when the noise moved. The previous directivity coefficient in our algorithm was used to estimate the noise gain between channels 1 and 2 in the noise-only intervals. For accurate noise gain estimation, two channels’ noise power can be adjusted to nearly the same but noticeably discrepant for the power of the desired speech. Then the desired speech can be extracted from the subtraction of the adjusted signals in these two channels. Our proposed method suppressed the overall music noise, including the instantaneous noise (panel f-1). This method also suppressed the ambient speech noise, with nearly 20 dB SNR improvement (panel f-2). The single-channel method did not adequately suppress the ambient speech noise (panel c-2). The comparison indicates that the proposed dual-channel speech-enhancement algorithm successfully suppresses the nonstationary noise, which adds to its practical value.
For this test, the original signal, recorded by the platform, contained desired speech and ambient music noise. The signal duration was 7.5 sec and the desired speech was located in the time axis approximately between 3 sec and 5 sec. Figure 7 (a) describes the time-frequency energy distribution of original signal, which was dispersed in the frequency range between 0 and 6000 Hz.
Figure 7(b) describes the energy distribution of the signal after the modulation of the CIS strategy. This speech strategy divided the original signal by the filter bank and then modulated the subband signal, by using the center frequency of each band to characterize its corresponding information for further speech synthesis. The plot indicates that the CIS modulation only changed the frequency-domain energy distribution, but maintained the time-domain energy distribution. As a result, the signal energy concentrated in the corresponding center frequency of each subband, and the ambient noise in the time domain was not suppressed.
Figure 7(c) describes the energy distribution of the signal after enhancement by the proposed dual-channel algorithm and modulation of the CIS strategy. The energy distribution changed in the time domain, primarily between 3 sec and 5 sec, and the ambient noise was sharply weakened.
Together, the plots in Figure 7 indicate that the speech enhancement can achieve the following 2 purposes. First, the desired speech remained while the ambient noise was sharply suppressed, which improved the CI speech recognition. Second, the global signal energy was lowered, and the CI battery life was prolonged, because information from the signal and the energy both are transmitted wirelessly to the inner part of the CI device.
Algorithm robustness and signal distortion analysis
For the tests in these 2 situations (moving noises from 90° and 135°), the original signals recorded by the omnidirectional microphone were plotted in (a) and (b), respectively. The original signal consists of the desired speech and the ambient speech noise. The noise suppression results are plotted in (c) and (d), respectively. A comparison of these plots reveals that the proposed algorithm also effectively weakens moving noise, with an SNR improvement of about 15 dB. The conventional noise-reduction methods need to reconvergent in algorithm for coefficients updating, and will always result in noise leakage and noticeable SNR decrease. The mentioned MVDR method, one of the most widely used adaptive beamformer, can choose and adjust the filter coefficients to minimize the output power with the constraint that the desired signal is not to be distorted. For moving noise, the MVDR method will also partly result in noise leakage, which attenuates the algorithm performance. The proposed algorithm calculates the noise directivity coefficient for moving noise, and also remains excellent performance with a few attenuation of SNR. As shown in result, the proposed algorithm is advantageous to avoid the noise leakage and is robust to mobile noise.
Figure 9 describes the SNR improvement (in dB) for all situations in which the noise comes from 30° to 180° and the desired speech is played at the azimuth of 0°, 5°, 10°, 15°, or 20°. In the office environment with T60 = 450 ms, for a fixed speech direction, the improved SNR (Figure 9 (a)) was higher when the noise azimuth approached 180° (backward), and was lower when the noise approached the desired speech (forward). For the situation of speech deviation to the 0° azimuth, greater offset resulted in less SNR improvement. The plot also indicates that, for a speech deviation range of 0° to 20° and noise range of 60° to 180°, the SNR improvement was >10 dB. For a noise azimuth of 180° to 300°, the expected analogous and mirror-reversed result was obtained. For comparison, experiments were also carried out in an anechoic chamber (T60 = 100 ms), and the SNRs improvement for head deviation are plotted (Figure 9 (b)). A set of similar SNR results are obtained, with only 1 to 3 dB globally SNR increased in situation of anechoic environment. The room reverberation actually influences the algorithm performance, but in an acceptable constraint.
where λ i is the element of diagonal matrix given in . Actually, the speech distortion index presents the attenuation between speech power and the original clean speech.
Figure 12 describes a graph of the speech distortion index for the enhancement of the desired speech after the CIS processing. The speech distortion indexes are sharply small, mainly between −18 and 21 dB. A smaller value of the speech distortion index means the desired signal is less distorted. Compared with figure 10, it prominently depicts a set of low speech distortion for the application of CI devices, which is in an acceptable range of signal distortion. For CI front-end signal acquisition, our algorithm is advantageous for a large amount of SNR improvement, but disadvantageous in a little bit of greater speech distortion when comparing with other low-distortion algorithms. However, for signal modulation and transmission by the CI CIS strategy, the speech distortion is sharply decreased.
These results, including the test of moving noise, head deviation (Figure 9) and algorithm evaluation of speech distortion (Figures 10, 11 and 12), indicate that the proposed algorithm is robust and flexible for CI speech-enhancement, with great SNR improvement and low speech distortion.
Approximation of directivity coefficient
The approximation of the directivity coefficient in Eq. (15), corresponding to the noise azimuth φ, is important to algorithm performance. The hardware platform was also constructed for the experiments to evaluate the approximation. Firstly, the test was carried out in an anechoic chamber (T60 = 100 ms). The loudspeaker played music as the ambient noise at 90°, 180° and 270° orientations respectively. The target speech located at 0° orientation, with equal-power signal play from the loudspeaker (SNR = 0). The calculated orientations for the noise azimuth φ in Eq. (15) are 81°, 192° and 280°. The orientation error is about 10° for this situation. And the corresponding results are 77°, 171° and 285° respectively for the test in office environment (T60 = 450 ms), with about 15° errors. These errors is acceptable for the CI application, therefore, the directivity coefficient can be applied in the estimation of desired signal power spectrum in Eq. (14).
Algorithm constraint for noise estimation
There are noise-only segments before and after the desired speech segment. In our algorithm, the anterior noise segment (before the desired speech) is used to estimate the directivity coefficient, and the length of which will influence the algorithm performance. For situations of the ambient noise with length no shorter than 450 ms (panels a-1 and a-2), the enhanced signals (panels b-1 and b-2) still remain great SNR improvement. For shorter length of ambient noise (panels a-3 and a-4), the SNR decreases noticeably (panels b-3 and b-4), as well as more speech distortion. To remain algorithm performance and low speech distortion in CIS modulation for CI devices, the minimal length of the noise-only period before the speech segment is about 0.5 second. This delaying time is acceptable for CI users in daily conversation. Specially, if we do not need a great SNR improvement, the length of the noise-only segment for pre-estimation can decrease to be about or less than 200 ms.
Distortion in CI processing
The previous results indicate that the proposed algorithm introduced a bit larger distortion to the desired speech but with noticeable low distortion after the CI processing. This phenomenon may result from the speech processing strategy of CI devices. The CIS strategy is a speech processing strategy to extract the signal information from the time-domain envelope and then be transferred to the electrode array. This CIS speech strategy primarily includes the process of window-adding, frame-dividing, pre-emphasis, sub-band dividing, envelop extraction and signal compression. As the CIs only have several channels, correspond to a few specific stimulating rates, the extracted envelope may lose lots of signal information. In addition, in each band of the CI channel, only one frequency (correspond to the center frequency of the CI filter bank) is applied to modulate the desired signal. That is, a set of sinusoidal signals (only 16 or 24 for CI devices) are modulated by the corresponding envelope of the band-pass signals in the CI filter bank. Therefore, the single-frequency modulation processing seems to be a smoothness process to reduce the distortion. For example, taking the distortion in the band of [1653, 1960] (channel 9) into account. If the 1700 Hz signal is strengthened but the 1900 Hz signal is weakened, the distortion will be introduced in this band. But after the CI modulation, both frequencies in this band, the center frequency 1807 Hz sinusoidal signal is applied to modulate the envelope based on the whole band energy. Therefore, the difference between the same bands will be smoothed, and the speech distortion after CI processing may come from the difference between different bands. Consequently, the CI speech strategy can reduce speech distortion and more aggressive algorithm can be applied in the CI application.
The proposed algorithm is an aggressive noise-suppression method with high SNR improved but a bit large distortion. But the distortion can be reduced in the CI processing and we can obtain excellent performance. The prevalent Frost algorithm based methods, such as linearly constrained minimum-variance and MINT algorithms, can suppress the noise with less signal distortion. These methods use iterative-adaptive technique to update the filter coefficients by gradient estimation and are advantageous in minimizing the ambient noise with low or no distortion for desired signal. Whereas, when the moving noise (or in the situation that the noise changes its azimuth) is present, these methods will weaken their noise-reduction performance. To obtain the optimal filter coefficients, if the algorithm does not reconvergent at the beginning or too slow to update the new coefficients, the noise will leakage and the desired performance will be attenuated. Other methods, MVDR and the binaural frequency-domain minimum-variance algorithm etc., present effective ways for noise suppression. Though these algorithms can converge more quickly, they will also cause noise leakage and the algorithm performance will be weakened. The approximation of directivity coefficient in our algorithm is tested, and about 10° error can be found. The directivity coefficient estimation is accurate enough to separate the desired speech and the ambient noise. For the trade-off between SNR improvement and speech distortion, the prevalent optimal-filter methods minimize the speech distortion while guarantying a certainly level of SNR improvement or maximize the SNR improvement while guarantying a certainly level of speech distortion. So it is hard for these algorithms to obtain both high SNR and low signal distortion. However, in the application of cochlear implant, the CI processing helps to obtain excellent speech enhancement while guarantying low speech distortion in the proposed algorithm.
The proposed speech-enhancement algorithm based on a dual-channel microphone array and spectral estimation technique aims to suppress the directional noise and improve the speech recognition of CI devices. A hardware platform was constructed and the experiments were carried out in an office to evaluate the algorithm performance in a real working environment for CI users. The experimental results indicated the excellent algorithm performance for speech enhancement. For stationary and moving noises, in orientations from lateral to rear, the improvements in SNR were 20 and 15 dB, respectively. For the situation of ± 20° speech deviation and a broad range of noise azimuths from 60° to 300°, SNR improvement of >10 dB was maintained. Also, the speech distortion was very low when evaluating the modulated signal in CIS processing. The proposed algorithm is robust to mobile noise and signal orientation deviation and is applicable to the improvement of the front-end signal acquisition and speech recognition for CI devices.
The authors are grateful to the support of National Natural Science Foundation of China under the grant No. 61271133, Beijing Natural Science Foundation under the grant No. 3082012, Basic Development Research Key Program of Shenzhen under the grant No. JC201105180808A.
- 10.Kamath S, Loizou P: A multi-band spectral subtraction method for enhancing speech corrupted by colored noise. Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing, Orlando, Florida 2002, 675–678.Google Scholar
- 11.Scalart P, Filho JV: Speech enhancement based on a priori signal to noise estimation. Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 2, Atlanta, GA 1996, 629–632.Google Scholar
- 13.Bitzer J, Simmer KU, Kammeyer KD: Theoretical noise reductoin limits of the generalized sidelobe canceller (GSC) for speech enhancement. Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing, Phoenix, AZ 1999, 2965–2968.Google Scholar
- 16.Elko GW, Pong ATN: A simple adaptive first-order differential microphone. Proceedings of IEEE workshop on Applications of Signal Processing to Audio and Acoustics, New Paltz, NY, USA 1995, 169–172.Google Scholar
- 20.Van Veen BD, Buckley KM: Beamforming: a versatile approach to spatial filtering. IEEE Trans Acoust Speech Signal Process 1988, 5: 4–24.Google Scholar
- 28.Cooklev T, Nishihara A: Maximally flat FIR filters. Proc. IEEE Int. Symp. Circuits Syst. Chicago, USA 1993, 96–99.Google Scholar
- 29.Valimaki V, Laakso TI: Principles of fractional delay filters. IEEE ICASSP’00 2000, 3870–3873.Google Scholar
- 30.Martin E, Arild L: Maximally flat FIR and IIR fractional delay filters with expanded bandwidth. European Signal Processing Conference. Poznan, Poland 2007, 1038–1042.Google Scholar
- 31.Oppenheim AV, Schfer R: Dig ital Signal Processing. Prentice-Hall, Englewood Cliffs, NJ; 1975.Google Scholar
- 33.Viikki O, Bye D, Laurila K: A recursive feature vector normalization approach for robust speech recognition in noise. In Proceedings of ICASSP' 98. IEEE Acoustics, Speech and Signal Processing Society, Seattle, WA, USA; 1998:733–736.Google Scholar
- 34.Yousheng C, Qin G: A normalized beamforming algorithm for broadband speech using a continuous interleaved sampling strategy. IEEE Trans Audio Speech Language Process 2011, 20(3):868–874.Google Scholar
- 35.Wilson BS, Lawson DT, Zerbi M, et al.: Design and evaluation of a Continuous Interleaved Sampling(CIS) processing strategy for multichannel cochlear implants. J Rehabil Res Dev 1993, 30(1):110.Google Scholar
- 37.Greenwood DD: A cochlear frequency-position function for several species-29 years later. J Acoust Soc Am 1990, 87(6):2593–2605.Google Scholar
- 39.Acoustical Society of American: American national standard mehthods for calculation of the speech inteligibility index. 1997. ANSI S3.5Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.