Microphone array (MA) technology commonly implemented in almost all acoustic equipments, such as, hearing aid, surveillance devices, smart phone, cochlear implant, voice - controlled device, teleconferencing system for extracting the desired target speaker while suppressing background noise, annoying recording scenario without speech distortion. MA beamformer use the priori spatial information about the designed distribution of MA, the characteristics of environment and the ability of combination with single-channel approach to obtain a steerable beampattern in a specified direction of sound source for recovering the clean speech data from the noisy mixture with high directivity index and high satisfactory speech quality. A superdirective (SDB) beamformer is one of the most helpful beamforming methods for preserving the original speech component in the diffuse noise field. SDB beamformer applied to numerous speech applications because of its advantages and easy implementation. Under realistic recording scenario, due to the complex environment, the inaccurate estimation of preferred direction of arrival (DoA) of useful signal, the error of sampling rate, the different speech sensitivity, the overall SDB beamformer's performance often corrupted. The speech distortion, musical noise, the remained noisy component usually exists and degrades the speech quality of the final output signal. In this contribution, the author proposed method for increasing robust speech enhancement of SDB beamformer in adverse environments. The numerical simulation has confirmed the effectiveness of the author's suggested technique in improving the speech quality by removing musical noise, background noise at SDB beamformer's output signal.
Nowadays, the using of MA technology has been popular and significantly increases the overall speech enhancement of various types of speech applications. MA use the spatial information about designed geometry, the preferred steering vector, the characteristic of surrounding noise, the properties of background noise, the coherence of captured MA signals to obtain high - directional beampattern towards the sound source while suppressing interference, third - party talker, as in Figure 1. MA beamforming, which owns the advantage of noise reduction and speech enhancement simultaneously, allows achieving the original speech component without distortion. Compared with single - channel approach, MA beamforming has the flexible working with different recording scenarios, such as, coherent/incoherent, diffuse and other complex situations.
MA beamforming technique can be categorized into two groups: the fixed beamformer and
adaptive beamformer. The fixed beamformer bases on the prior information of incident angle of
helpful signal to achiveve the beamformer’s coefficient. Fixed beamformer includes Delay - and
sum (DAS) beamformer [
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
As in Figure 2, SDB beamformer uses the properties of diffuse noise field to extract the desired talker at certain location with high directivity index. However, due to the complex and annoying environment, the error of estimation the preferred steering vector, the displacement of MA distribution, the overall SDB beamformer’s performance often decreased. Consequently, speech distortion, musical noise still be challenging problem. There are numerous efforts were studied for dealing this task. The scheme of MA beamforming is given by Figure 3.
Atkins A [
Berkun R [
In [
Wang J [
Yang X [
Gong P [
In [
A data-driven approach, which controls WNG threshold for obtaining robustness speech
enhancement, was presented [
Huang G [
However, these above approaches performed in laboratory condition with proper environmental factors. Due to the complex and annoying recording situation, the displacement of MA geometry, the inaccurate estimation of preferred steering vector, the error of sampling frequency, the microphone mismatches, the different microphone sensitivities, SDB beamformer’s evaluation usually degraded. In this paper, the author introduces modifying covariance matrix of complex diffuse noise field and enhanced steering vector to obtain more robustness of SDB beamformer. The author’s method improves the real-time SDB beamformer’s performance without processing large database.
In this section, the principal working of SDB beamformer will be presented. The author uses dual microphone array (DMA2) system to describe the representation of observed MA signals in Figure 4. In the currently considered frame , the frequency
, the received array signals can be formulated in the short-time Fourier transform (STFT) as: Where is the original speech component, is the additive noise, ,
is the preferred steering vector of the interest useful signal to axis of DMA2, , is the distance between two mounted microphones, propagation in the fresh air. is sound speed (2) (3) (4) (5) denote , ,
is transpose operator, the equation (1)-(2) can be rewritten as: The essential core problem of speech enhancement is determining an optimum coefficient for obtaining approximate clean speech data. where
is conjugate operator.
And .
In the diffuse noise field, SDB beamformer’s weight can be computed as: Where .
SDB beamformer especially effective in diffuse noise field with high directional beampattern towards the sound source. Because of the complex and annoying environment, the displacement of MA distribution, the inaccurate estimation of incident angle of the impinging helpful signal, the microphone mismatches, the different microphone quality, the error of sampling frequency, the moving head of speaker during a conversation, the existence of non-directional noise or undetermined reason, SDB beamformer’s performance often degraded. Consequently, the musical noise or speech distortion occurs and significantly effects speech quality.
For overcoming this drawback, in the next section, the author proposed an effective modified SDB beamformer for decreasing the speech distortion, musical noise, noise level and increasing the perceptual metric listener and speech intelligibility.
The author’s ideal is enhancing the accurate estimation of steering vector and modifying the covariance matrix , which according to the rapidly changing environmental factors.
Due to the complex environment, the steering vector can be expressed in the way: With the improved steering vector and modified coherence matrix in complex diffuse noise field according to the rapidly changing recording scenario, the author’s proposed technique can be applied for adaptively achieving the optimum coefficient.
where
presents the distortion of clean speech data at second microphone in comparison with first microphone and
. In this paper, the author uses the information of standard deviation
of diagonal loading of noisy covariance matrix, and
[
As we know that, the coherence between two points in diffuse noise field can be formulated as . The author proposed using the speech presence probability [22] to adjust
as the following equation: where [23].
is covariance of uncorrelated noise, is spectral density of diffuse noise field
The equation (7) contains the steering vector and ensure the robustness of the coherence between two points in diffuse noise field. Therefore, the coherence matrix can be modified as the following equation: (7) (8) (9)
In this section, the author aims at demonstrating the effectiveness of the suggested technique (SDBsgt) in enhancing SDB beamformer’s robustness speech enhancement with increasing the speech quality in the term of the signal - to - noise ratio, reducing musical noise and noise level at SDB beamformer’s output signal. The experiment was conducted in living room with the existence of interference, non - directional noise, third - party talker, washing machine and other undetermined sources. A talker stands at distance
, the incident angle of helpful signal to the axis of dual - microphone system (DMA2) is
, the distance between two mounted microphones is . For capturing the clean speech data, these parameters were set: , the sampling frequency , overlap . An objective measurement [24] was implemented for calculating the speech quality in the term of signal-to-noise (SNR) ratio. The scheme of demonstrated experiment is shown in Figure 5.
For further signal processing, smoothing parameter was used for implementing SDB beamformer to extract the target speaker. The obtained results are given in Figure 8 and Figure 9.
Because of the error of sampling rate, the different microphone sensitivities, the displacement of designed microphone distribution, the inaccurate estimation of preferred steering vector, the moving head of speaker, the adverse environment, SDB beamformer’s evaluation often degraded. The existence of speech distortion, musical noise, remained noise corrupt the speech quality. Therefore, the author suggested improving the accurate calculation of steering vector and modified the noisy coherence matrix increase SDB beamformer’s performance in adverse environment.
The promising result of SDB-sgt is given in Figure 10 and Figure 11. Figure 12 describes the comparison of energy between the observed microphone array signals, the processed signals by SDB-beamformer, SDB-sgt.
The advantage of SDB-sgt has been confirmed with the decreasing the musical noise, noise reduction to 9.5 dB and Table 1 shows the increased the SNR from 10.9 to 12.7 dB.
The effectiveness of the author’s proposed method was illustrated through the numerical simulations. The approach modifies the steering vector and the formulation of coherence between two points sources, which exploit the speech presence probability and standard covariance matrix of observed microphone array signals. The modified covariance matrix and enhanced steering vector adaptively change according to the characteristics of surrounding noise, the complex recording scenario. Therefore, the updating these necessary parameters allows improving the robustness of SDB beamformer in complex diffuse noise field. Steering vector, which contains the direction and source location, plays an important role in SDB beamformer in extracting the desired target speaker. Accurate estimation of steering vector and modified covariance matrix of complex diffuse noise allow increasing SDB beamformer’s performance in realistic recording environment.
In this contribution, the author proposed an effective modified steering vector and covariance matrix of diffuse noise field in presence of annoying recording scenario. The obtained numerical results have showed the effectiveness of musical noise reduction, noise level suppression and increasing the speech quality in the term of signal-to-noise ratio from 10.9 to 12.7 dB. The author’s approach based on adaptive tracking and adjusting the steering vector and covariance matrix according to the simultaneous changing environmental factors. The suggested technique incorporates the speech presence probability to enhance the robustness of computing the necessary parameter for determining SDB beamformer’s coefficient. The above method can be integrated into multi-channel system for dealing other complicated problems, such as speech recognition, reverberation.
During the preparation of this work, the authors used Grammarly in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [22] T. Gerkmann, M. Krawczyk and R. Martin, "Speech presence probability estimation based on temporal cepstrum smoothing," 2010 IEEE International Conference on Acoustics, Speech and Signal Processing, Dallas, TX, USA, 2010, pp. 4254-4257, doi: 10.1109/ICASSP.2010.5495677. [23] J. Bitzer, K. . -D. Kammeyer and K. U. Simmer, "An alternative implementation of the superdirective beamformer," Proceedings of the 1999 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics. WASPAA'99 (Cat. No.99TH8452), New Paltz, NY, USA, 1999, pp. 7-10, doi: 10.1109/ASPAA.1999.810836. [24] SNRVAD. [Online]. Available: https://labrosa.ee.columbia.edu/projects/snreval/.