Journal of ELECTRICAL ENGINEERING, VOL. 66, NO.6, 2015, 301–310 A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS Radek Martinek — Michal Kelnar — Jan Vanus ∗ Petr Bilik — Jan Zidek The authors of this article deals with the implementation of a combination of techniques of the fuzzy system and artificial intelligence in the application area of non-linear noise and interference suppression. This structure used is called an Adaptive Neuro Fuzzy Inference System (ANFIS). This system finds practical use mainly in audio telephone (mobile) communication in a noisy environment (transport, production halls, sports matches, etc). Experimental methods based on the two-input adaptive noise cancellation concept was clearly outlined. Within the experiments carried out, the authors created, based on the ANFIS structure, a comprehensive system for adaptive suppression of unwanted background interference that occurs in audio communication and degrades the audio signal. The system designed has been tested on real voice signals. This article presents the investigation and comparison amongst three distinct approaches to noise cancellation in speech; they are LMS (least mean squares) and RLS (recursive least squares) adaptive filtering and ANFIS. A careful review of literatures indicated the importance of non-linear adaptive algorithms over linear ones in noise cancellation. It was concluded that the ANFIS approach had the overall best performance as it efficiently cancelled noise even in highly noise-degraded speech. Results were drawn from the successful experimentation, subjective-based tests were used to analyse their comparative performance while objective tests were used to validate them. Implementation of algorithms was experimentally carried out in Matlab to justify the claims and determine their relative performances. K e y w o r d s: adaptive neuro fuzzy inference system, background noise, colored noise, noise cancellation, voice communication 1 INTRODUCTION Audio communication [22] through modern communication interfaces (PSTN, IP phones, mobile communications, satellite phones, radios, etc) is an integral part of human lives in today’s over-technologized world. In real conditions, audio communication is often degraded by the noise in the background [12]. Such a noise can be, for example, a noise from motor vehicle engines, a noise from machines in production halls, city sounds (traffic, etc), a noise at sports and social events, etc. The authors of the article base their work on the results of their own studies published in [1, 2]. The work in this article builds on these studies and extends them. In these studies, the authors address the issue of removing the disturbing components from useful signals (noise suppression during audio communication in the cockpit of a fighter airplane [1], noise suppression in foetal ECG [2, 32, 38]) using linear adaptive filtration [5]. Linear adaptive filters play an important role in statistical signal processing [6]. Currently, in commercial applications for suppressing noise during audio communication, the most widespread algorithm is the LMS adaptive one [6]. The LMS algorithm is mathematically simple and undemanding [29]. In real applications, however, it achieves a lower convergence rate [6] and a higher error of the filtration process [6]. Better results are achieved using the RLS family of algorithms [6, 35] (an extremely low error rate and extremely high speed of convergence). The results of the experiments carried out by the authors [1, 2] show that the basic RLS algorithm [18] is mathematically very demanding in terms of actual implementation [18]. Many studies [8, 13, 19] further indicate that we encounter the nonlinear nature of disturbance [4, 13, 21], which cannot be completely eliminated using a linear filter [1, 4]. Due to the reasons discussed above, the authors focus on the use of a neuro-fuzzy system [19, 20]. A number of studies suggest [8, 9, 17] that such a system could achieve better results than commercially used systems. 2 THE ADAPTIVE NEURO–FUZZY INFERENCE SYSTEM ANFIS is an adaptive feed forward neural network that is functionally equivalent to a fuzzy inference system of the Sugeno type (Takagi-Sugeno) [19, 28, 32, 34, 36–38]. A Two Rule Sugeno ANFIS has rules of the form [19] IF x is A1 and y is B1 , THEN f1 = p1 x + q1 y + r1 . (1) IF x is A2 and y is B2 , THEN f2 = p2 x + q2 y + r2 . (2) Where x and y are the inputs, Ai and Bi are the fuzzy sets, fi , i = 1, 2 are the output fuzzy system, and pi , qi ∗ Department of Cybernetics and Biomedical Engineering, VSB – Technical University of Ostrava, 17. listopadu 15/2172, 708 33 Ostrava, Czech Republic, {radek.martinek, michal.kelnar, jan.vanus, jan.zidek, petr.bilik}@vsb.cz c 2015 FEI STU DOI: 10.2478/jee-2015-0050, Print ISSN 1335-3632, On-line ISSN 1339-309X Unauthenticated Download Date | 1/3/16 5:52 PM The parameters in this layer (pi . 28. Layer 2 Every node in this layer is a fixed node labelled Π. 2 . qi . These sources allow working with any signals in the wav format. 2.302 R. 34]: Layer 1 The output of each node is O1. Layer 5 The single node in this layer is a fixed node labelled Σ. i = 1. the O1.i = w i fi = w i (pi x + qi y + ri ) . A is a linguistic label and µAi (x) and µBi−2 (y) can adopt any fuzzy membership function.A. 40PP CPP Free-field QA microphones as reference and primary MIC and NI DAQ 9234. Layer 4 Every node i in this layer is an adaptive node with a node function O4. As figure illustrates. bi .i = µAi (x) for i = 1. 1 [19. ANFIS architecture consists of five layers [19]. 28] in which a circle indicates a fixed node whereas a square indicates an adaptive node. 2 . ci } is the parameter set which changes the shapes of the MF degree with maximum value equal to 1 and minimum equal to 0. The experimental system includes two signal sources: the Speech Source and the Noise Source.R. (7) w1 + w2 (8) Amplifier Noise speaker Amplifier Speech speaker Ambient space of test room Reference MIC Primary MIC Fig. whose output is the product of all incoming signals O2. The ANFIS architecture to implement these two rules is shown in Fig. which computes the overall output as the summation of all incoming signals: P X wi fi . NI DAQ 9234 (6) Layer 3 The i -th node of this layer.i is essentially the membership grade for x and y . Where w i is the output of layer 3. (9) Overall output = O5. (5) Where {ai . calculates the normalized firing strength as wi O3. labelled N . 2 . 4 . (4) Where x and y are the inputs to node i . The membership functions could be anything but for illustration purposes we will use the bell shaped function given by µA1 (x) = 1 1 + |(x − ci )/ai | 2bi .i = w i = .i = w i fi = Pi i wi i 3 SUPPRESSING BACKGROUND NOISE DURING AUDIO COMMUNICATION IN A NOISY ENVIRONMENT Figure 2 shows a schematic diagram of the measurement workplace where the experiments were conducted. The functioning of the ANFIS is described as [19.S. Experimental measuring workplace Unauthenticated Download Date | 1/3/16 5:52 PM . Using G. The architecture of ANFIS and ri are the design parameters which are determined during the training process [19]. i = 1. (3) O1.i = µBi−2 (y) for i = 3. So. ri ) are to be determined and are referred to as the consequent parameters. 1. Martinek et al : A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS Layer 1 Layer 2 A1 Layer 3 w1 x Layer 4 x y w1 A1 N e Layer 5 w1 f1 A2 f ? B1 y e w2 B2 N w2 f2 A1 w2 x y Fig.i = wi = µAi (x)µBi (y) . which contains the unwanted background noise. In the experiments performed. 5. delay [7].303 Journal of ELECTRICAL ENGINEERING 66. Signal recorded by the primary microphone — signal measured — m[n] Fig. this approach will not work since the noise n1 [n] picked up by the reference microphone is not the same as the noise that contaminates the voice signal m[n] captured by the primary microphone. 2015 x1 x 0. Source signal x[n] — original speech signal m n 30000 n 150 Fig. 5. The system designed has two entry points (a primary microphone and reference one). how- Unauthenticated Download Date | 1/3/16 5:52 PM . see Figs. 4 (records only noise). However. If we then applied this filter to the signal from the reference microphone. Noise signal n1 [n] Noise 3 2 n1 2 1 20000 n2 1 0 0 -1 -1 -2 -2 -3 0 10000 20000 n 30000 0 50 100 Fig. 1 channel mono). a recorded engine noise was used. it is evident that the clarity of radio communication in such a noisy environment would be very bad. used standard microphone (omnidirectional). As shown in Fig.4 2 0. 501 [10] was used (Test signals for the use in telephonometry). see Fig. Another possible way was the use of a linear filter (frequency selective filters) [5]. 4. 6 and 7 (n1 [n] →Non-Linear Unknown System (NLUS ) → n2 [n]). The first entry is a reference microphone that registers unwanted noise and this signal is denoted n1 [n]. we could apply any adaptive algorithm [6] that will teach the FIR filter [18] to recognize the characteristics of the channel. 2. Therefore. Unfamiliar environment properties apply here (signal distortion due to the environment — interference [5]. see Fig. These source signals were reproduced by means of reference and primary loudspeakers. Many studies [8. 3. we could subtract this unwanted noise successfully. However. a test voice according to ITU-T P. 2. this signal is denoted m[n].8 4 0.0 0 -2 -0. 13]. the easiest way how to suppress the unwanted background noise would be direct subtraction of the signal from the reference microphone n1 [n] from the signal captured by the primary microphone m[n]. 6. bit rate 128 kbps.8 -4 0 10000 20000 n 30000 0 10000 Fig. The second source signal is noise n1 [n]. etc). The human voice used reaches values of 30 dB to 40 dB while the engine noise used reaches the value of about 80 dB to 100 dB. see Fig. NO.4 -0. see Fig. If we considered the unknown environment (system) as linear. In the experiments conducted. 6. 3. The second input is a primary microphone (signal measured) that picks up the useful signal (voice) plus unwanted background noise. Recording parameters: sampling frequency of 8 kHz. Noise difference ( n1 and n2 — detail) The first source signal is the human voice x[n]. 7]. The aim of the authors was to create a system that would reduce unwanted background noise n1 [n] that contaminates the useful speech signal x[n]. For the purposes of the experiments carried out. these conventional filtration techniques cannot be used because of the spectrum time variability of the interfering and useful signals [5. PCM audio format. 9. the first n = 30000 samples of audio recordings were used. This response was determined on the basis of ITU-T – P34 [11]. but it can not be obtained directly in practice. In Fig. type LMS and RLS. m[n] = x[n] + n2 [n] . · · · )k2 (14) Mathematical expectation is calculated in the equation.03 0.01 n2 [n] = f (n1 [n]. and the membership function of every variable is a Gaussian function. so a frequency filter cannot be realized. Figure 8 shows the impulse response [23] of the experimental room. x[n] and n2 [n] are not correlated [8]. the input is noise n1 [n] and n1 [n − 1]. The impulse response experimental room — RIR ever. The authors deal with the study of linear adaptive filters. 7. indicate that the real environment in a number of applications shows a nonlinear nature [4]. Note the expectation of x[n] is zero (the condition suits for many problems or simple transform). Here. the linear adaptive filtration techniques do not have satisfactory results here (in particular the LMS algorithm families [18]). (12) The output of the ANFIS can be the estimated value y[n]. When the ANFIS network approaches colored noise. · · · )k2 . n1 [n − 1].01 0 0. letting ANFIS approach the colored n2 [n] so estimated x[n] can be estimated [8]. The experiments were performed in a standard office room. its frequency range is usually overlapped with the frequency range of m[n]. n1 [n − 2]. so we can get e2 [n] = kx[n]k2 + kn2 [n] − y[n]k2 = kx[n] + n2 [n] − fˆ(n1 [n]. Then. (10) kx[n] + n2 [n] − fˆ(n1 [n]. 7 signal n2 [n] can be seen (coloured noise). Here. x[n] and Unauthenticated Download Date | 1/3/16 5:52 PM . 2] use this publication as a foundation. Signal recorded by the primary microphone — signal measured — m[n] Where m[n] is polluted by noise.10 Time (s) 0.02 0. which can approach the nonlinear function. this signal represents noise n1 [n] after running through an unknown environment (interference. whose parameters are defined in Fig. delay). In practice. (13) Where fˆ is the nonlinear approach produced by the ANFIS network. Martinek et al : A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS n [n] Noise source Reference MIC n1 [n] e [n] Adaptive filter NLUS y [n] + n2 [n] + m1 [n] x [n] Speech source Primary MIC Signal propagation model + - xe [n] m [n] Noise canceller Fig. 7 (Details of the selected environment). we use the ability of ANFIS network. [8] in the publication [1. n2 [n] is the delay and deformation of n1 [n].304 R. n1 [n2 ]. 8.00 -0. n1 [n − 1]. 0. n2 [n] is nonmeasurable. n1 [n − 1]. The above equation can be expanded. n1 [n − 1]. 4 IMPLEMENTATION OF ANFIS INTO THE SYSTEM OF ADAPTIVE SUPPRESSION OF UNWANTED INTERFERENCE (11) Where function f is unknown and nonlinear. it can be replaced with a measurable signal m[n] = x[n] + f (n1 [n]. Obviously. it is suitable that n2 [n] is estimated. as follows RIR 0. the objective of the training ANFIS is to make the following error minimum e2 [n] = km[n] − y[n]k2 = kx[n] + n2 [n] − y[n]k2 = Figure 7 shows a principal diagram of the designed adaptive system on which the experiments were conducted. Signal m[n] is then given by the relation. x[n] is estimated from the polluted signal m[n]. the output sample should be colored noise. · · · ) .20 Fig. · · · ) . n1 [n2]. 11.4 0. 6.4 0. NO. that is.8 0 10000 20000 n 30000 Fig.0 0 -1 -0.8 ANFIS model B Fig.8 ANFIS model A 0.0 -0.4 -0.8 0 10000 20000 30000 n 0 10000 Fig.4 -2 -0. and minus n2 [n] in 2 5 THE RESULTS OF THE EXPERIMENTS CONDUCTED Various ANFIS networks (structures) were investigated in the experiments performed. 2015 m x 0. Results of noise cancellation — ANFIS model C Fig.4 -0. the output signal estimated x[n] equals the useful signal x[n] because of complete cancellation of the noise. y[n] = n2 [n] so estimated x[n] = x[n].8 ANFIS model D 0 10000 20000 n 30000 Fig. ANFIS adjusts its parameter automatically and makes n1 [n] process into n2 [n].8 0. which is equivalent to E[(n2 − y) ] minimum. Results of noise cancellation — ANFIS model A 10000 20000 n 30000 Estimated x ANFIS model A 0. Noisy speech signal Estimated x 0.4 0 0 0. 9. Thus. (15) the original input signal m[n]. 14.4 0. Makes E(e2 ) minimum.0 0.8 30000 n Fig. 12. function fˆ produced by ANFIS will approach the practical noise transmission function with minimum mean square error.4 -0.4 -0. so n2 [n].305 Journal of ELECTRICAL ENGINEERING 66. In ideal conditions. make y[n] approach n2 [n] as much as possible in other words. Results of noise cancellation — ANFIS model D n1 [n] is noncorrelation.8 -0.0 -0. Unauthenticated Download Date | 1/3/16 5:52 PM . y[n] is noncorrelation. An overview of the network models used is provided in Table 1.8 2 0. 13. Results of noise cancellation — ANFIS model B Estimated x 0. 10. so E 2 [e2 ] = E(x2 ) + E[(n2 − y)2 ] .0 0.4 1 0.8 20000 10000 20000 n 30000 -0. Original speech signal Estimated x 0. 314 9. ni (n) = n(mi+n) — segments of length M selected step m. . the reverse is true. .551 16. .4705 32. Table 2 shows the ability of the analysed ANFIS models to improve the value of SNR (Signal to Noise Ratio [26]).8876 7.2] RLS [1. . .4961 –21. . . obtains the local cost matrix C ∈ RN ×M .749 Ps [dB] . . mℓ ) ∈ {1.630 0.689 7. the signal output is bigger than the noise output. (21) A path which runs through the low-cost areas of cost matrix is called warping path.. .3837 29.3832 SNRout SNR Improvement (dB) –0. . ℓ ∈ {1. . N }×{1. (19) Evaluating the local distance for each pair of elements of these time-series using Manhattan distance (ie absolute value of difference). A detailed description of work with ANFIS and EVALFIS functions in MATLAB environment can be found in [14. yM ) .306 R. L}.617 0.9591 –9.2] RLS [1. .6858 17. (20) Where elements of cost matrix are CI. M ). Unauthenticated Download Date | 1/3/16 5:52 PM .692 0. .3832 –21. the noise has a much greater output then the speech signal. Resulting values of the SSNR and DTW Building the ANFIS Model ANFIS Model A ANFIS Model B ANFIS Model C ANFIS Model D LMS [1. More information in [30– 33]. M }. Martinek et al : A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS ANFIS functions were used for building of the ANFIS model [13] as well as for training (estimating) of the EVALFIS function [13]. 2. it is clear that all the structures achieved a significant improvement in the SNR values. . The SNR ratio is defined by the following relation [26] SN R = 10 log If we look at the acquired values stated in Table 2. Resulting values of the SNR SN Ri Improvement of the SNR Value Building the ANFIS Model SNRin ANFIS Model A ANFIS Model B ANFIS Model C ANFIS Model D LMS [1. x2 .598 19. (23) satisfying the following conditions i Boundary condition: p1 (1. When SN R > 0 .j = |xi − yj | .6184 –21.754 18. (17) SSN R = 10 log PM−1 2 K i=0 n=0 ni (n) {z } | Table 2. L) . . . the SNR value of the contaminated speech signal and the signal after passing the system designed was determined. If we evaluate the results of filtration by means of the tested ANFIS structures.3427 27. Figures 9 to 14 compare the time flows of the analysed signals.891 0.8662 5. i ∈ {1. 39]: • reference sequence (original speech signal) x = (x1 . The SSNR ratio is defined by the following relation [30.2] SSNR (dB) improvement Where L is the number of segments of speech signal. .. d (DTW distance) 15. (18) • test sequence (output speech signal from ANFIS) 0. If the SN R = 0 dB. V ADi is information about speech activity (values 0 and 1). Pn N ∈ N. .725 0. . j ∈ {1. the SSNR value of the contaminated speech signal and the signal after passing the system designed was determined. y2 . . . .3832 –21. When SN R < 0 . 1) ∧ pL = (N. The difference between these values showed what improvement was achieved by each model.6974 4.4765 –21. M } .8521 10. Table 1. . further xi (n) = x(mi+n). The DTW (Dynamic Time Warping) criterion in Tab. 3 is used to compare the similarity of two sequences of the speech signal to calculate the distance d between them [30. 2. .2] (dB) –21. The difference between these values showed what improvement was achieved by each model. 2. M ∈ N. it means that both the signal and the noise have the same output.. p2 . . N } . In the experiments performed. . 33] (16) where the values Ps and Pn indicate the power of the signal (s is signal) and noise (n bis noise). Table 3 shows the ability of the analysed ANFIS models to improve the value of SSNR (segmental signal to noise ratio) [30].0191 Table 3.3641 (dB) 22. defined as sequence P = (p1 . 31. K is the number of segments in speech activity. Figures 9 to 23 show the results of experiments conducted. .3423 11. 15]. xN ) . we can see that in the signal of the mixture of the speech signal and the noise SNR. (22) pℓ = (nℓ . .797 y = (y1 . In the experiments performed. . Information about anfis structures used ANFIS info Building the ANFIS Model Number of nodes Number of linear parameters Number of nonlinear parameters Total number of parameters Number of fuzzy rules A 21 12 12 24 4 B 35 27 18 45 9 C D 53 75 48 75 24 30 72 105 16 25 ! PM−1 2 L−1 1 X n=0 xi (n) ·V ADi . L ∈ N. This is a 3D graph that has two axes with independent variables — frequency and time (order of sections spectra). at a certain time. A 2D spectrogram is used here. (01). Figures 21 to 23 show 3D spectrograms of the analysed signals. . 0). we put intensity of individual frequencies on axis z . it is a top view of the original 3D graph. 16. 18. 20. it is clear that all the structures Figures 15 to 20 show spectrograms of the analyzed signals. If. 19. Original speech signal 4000 f (Hz) ANFIS model A 4000 3000 2000 2000 1000 1000 1 2 Time (s) f (Hz) ANFIS model C 4000 3000 2000 2000 1000 1000 1 2 Time (s) 1 Time (s) 3 f (Hz) ANFIS model D 0 3 2 Fig. . ℓ ∈ {1. A 3D spectrogram is a kind of a spectrogram that is displayed in a three-dimensional space. If we evaluate the results of filtration by means of the tested ANFIS structures. Results of noise cancellation — ANFIS model C Fig. . 2. 2015 f (Hz) f (Hz) 4000 4000 3000 3000 2000 2000 1000 1000 0 1 2 3 Time (s) 0 1 Fig. Compared to the traditional spectrogram. N ×M d = min{ΣL }. Noisy speech signal 3000 0 2 1 2 Time (s) 3 Fig. . 17. L − 1} . Results of noise cancellation — ANFIS model A 4000 Time (s) Fig. NO. 1)} . The optimal path is selected from the set of all available warping paths by distance function achieved a significant improvement in the SSN R and DT W values.307 Journal of ELECTRICAL ENGINEERING 66. we make a Unauthenticated Download Date | 1/3/16 5:52 PM . 15. iii Step size condition: pℓ+1 −pℓ ∈ {(1. 6. l=1 pl . (1. Results of noise cancellation — ANFIS model B 3000 0 3 f (Hz) ANFIS model B 0 3 Fig. Results of noise cancellation — ANFIS model C ii Monotonicity condition: n1 ≤ n2 ≤ · · · ≤ nL ∧ m1 ≤ m2 ≤ · · · ≤ mL . p ∈ P (24) Where P N ×M is the set of all possible warping paths for cost matrix C . In this paper. 22. at the expense of higher computational demands for the algorithm. ANFIS Model B seems to be the most convenient. therefore. achieved. the same results. see Tab. In the future. This article has clearly investigated the use of adaptive neuro-fuzzy inference system (ANFIS) for effective noise suppression in speech. In practice. Results of noise cancellation — ANFIS model B cross section of the 3D spectrogram with a plane parallel to the frequency axis and axis z .I0(A) 308 R. D) achieve slightly better results. Other ANFIS models examined. 21. The authors also conducted listening tests of clarity. 23. 1) and results achieved (Tabs. Original speech signal 40 0 -40 -80 3000 1. 1. 2]. This type of adaptive noise cancelling technology can be used when there is a characteristic of an unknown external interference source and the background noise is similar to Unauthenticated Download Date | 1/3/16 5:52 PM . More complex ANFIS models (C. The adaptation algorithm is based on the learning ability of ANFIS. however.2 2000 f (Hz) 0. required while maintaining a high quality of unwanted noise suppression. Here. more complex and more efficient algorithms can be implemented. Hence. Noisy speech signal Modul (dB) 2000 f (Hz) 0. The experiments conducted confirmed the functionality of the technique designed.4 0 Fig. The increase of the SSNR is the most important objective criterion for comparison of the effectiveness of different algorithms.2 0. the lowest possible costs for implementation of adaptive systems are. the requirements for a low computational complexity and memory consumption of the algorithms will be decreased and. In experiments. 3. An effective noise cancellation in speech was undertaken with the use of adaptive noise cancellation techniques because fixed filters would necessitate the redesign of the filter once the variations exist in the communication channel.8 t (s) 0. SSNR and DTW) indicate that systems using ANFIS show better experimental results than conventional systems based on adaptive algorithms of the LMS and RLS families [1.4 1000 3000 0 Fig. we obtain a spectrum of signals at this time.8 t (s) 0.4 1000 1000 0 Fig. From the perspective of production costs. This model is a good compromise in terms of mathematical performance (Tab. Martinek et al : A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS Modul (dB) Modul (dB) 40 40 0 20 -40 -20 -80 -40 0 3000 1. The experiments showed that the technique designed can successfully extract the speech signal even if it is completely contaminated by the background noise. An adaptive canceller which can automatically adjust itself to the environment is therefore employed. ANFIS model A achieved the worst results (worse clarity).8 t (s) 2000 f (Hz) 1. adaptive noise cancellation using ANFIS has been implemented on audio speech signal (voice communication). from the subjective point of view of the listening tests. The noise which interferes with speech signals in transmission varies depending on the propagation path which is mostly unknown. we can expect a steady increase in performance and quality in the area of computer technology. B. for a long time to come. see Tab. 2 and 3). C and D. The results (SNR. With the advent of more efficient hardware. it is obvious that the area of adaptive filtration still is and will be. the most convenient solution will be to construct system for ANFIS Model B. 6 DISCUSSION 7 CONCLUSION Various ANFIS structures were examined during the experiments conducted. As for actual implementation in commercial equipment. we have achieved better performance.2 0. of course. a wide open area for scientific research as well as commercial applications. dio Processing 1 No. [20] STAVROULAKIS. [29] MARTINEK. : The Real Implementation of NLMS Channel Equalizer into the System of Software Defined 2010. [17] JANG. 2015 the measured noise. Berlin.html. T. Przeglad Elektrotchniczny (Electrical MIRY. J. Demonstration tronics Letters 51 No. A. : The Use of the Adaptive Noise Cancellation for http://www.aspx.. pp.—KOUDELKA. R.-KALA. Czech Republic. and the high signal-to-noise ratio signal is obtained. H. 1-st International Conference on Energy. Proc. on Artificial Intelligence (AAAI-91). References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] ITU-T Recommendation P. tive Colored Noise Cancellation Based on ANFIS. J. 2008. J. G. cessing: Principles. InSHUKLA. [14] MathWorks. 3 (May 1993).—KELNAR. [31] VANUS. [21] ZAFAR. In Conference [22] PATEL. of Artificial Intelligence. [25] MAKINO. : Advanced Digital Signal Processing and Noise Reduction. H. 2012. http://www. 4th.—ZIDEK. Speech and AuUSA. L. J. 22 (Oct 2015).—CHARBIT.[32] MARTINEK. C. J. : Use of Adaptive Filtering for Noise Reduction in Communication systems. : Application of Variations of the LMS Adaptive Filter for Voice Communications with Control Group XII (1988-1993) and was approved by the WTSC (HelSystem. International Journal on Artificial Intelligence and Machine Learning No. C. 1738–1740. 665–685. pp. : Refining the Diagnostic Quality Conference on Image Analysis and Signal Processing. Springer (24 Mar 2004).com/help/fuzzy/index. K. Matlab R2012b – Fuzzy Logic Toolbox. R. J. : EnELECOMP). 35–40. 501. http://www. J.—HERNANDEZ. M. Man. 2697–2700. ITU-T P. 2008. R : ANFIS Control for RoboProceeding: The International Conference Applied Electronics tic Manipulators: Adaptive Neuro Fuzzy Inference Systems for (AE). The McGraw-Hill Companies. 289–296. 762–767. 5 (Dec 2012). to Generate Auditory Virtual Environments. 555–580.int/net/itu-t/sigdb/genaudio. 330–336. R. ANFIS: Adaptive-Network-based Fuzzy Inference Systems : IEEE Transactions on Systems. J. New [26] SILZLE. Tehnicki Vjesnik 18 (2011). P.—GARCIA. MARTINEK. ference on Electrical Communications and Computers (CONIJ.mathworks. J. MARTINEK. Lisbon.-S. 2004. http://www. Pilsen. Voice Communication with the Control System. John Wiley & Sons Inc (E).—TIWARI. Algorithms.34 was revised by the ITU-T Study [30] VANUS. : IKA-SIM: A System Jersey. Power and Control (EPC-IQ). V.—ZIDEK. A.3 (2006). Cholula. Matlab R2012b – genfis1 and anfis – Generate Fuzzy Inference System structure from Data using Grid Partition. R : Towards Hybrid and terspeech conference. of the Ninth National Conf. P. R. Experiments have been evidently performed on these three distinct approaches and results drawn.LAP. [15] MathWorks. of the Abdominal Fetal Electrocardiogram using the Techniques IASP 2009. Acknowledgment This work is partially supported by the project SGS SP2015/181 VSB — Technical University of Ostrava. [16] KURIAN. Przeglad Elek.itu. of Adaptive Nonlinear Noise Cancellation using the Fuzzy Logic Unauthenticated Download Date | 1/3/16 5:52 PM . International [12] REYES. P. CASTILLO. This technology can cancel the external disturbance. J. R.com/products/fuzzy-logic/examples. J. 54–70. S : Autopilot System for an Unmanned Aerial Vehicle (UAV): Adaptive Neuro Fuzzy Inference Based Control System. 101–108.—ZIDEK. Adaptive Computing. P (editor) : Neuro-Fuzzy and Fuzzy-Neural Applications in Telecommunications (Signals and Communication Technology). 6. D. July 1991.—FINSTER. Intelligent Control. R : Fuzzy Modeling Using Generalized Neural Networks and Kalman Filter Algorithm. M.—ZIDEK. A concise analysis and constructive comparison was afterward documented in order to achieve the aim and objectives of the article. J : Signals and Systems: Analysis Using TransWeighted Step-Size NLMS Adaptive Filter Based on the Statisform Methods and MATLAB.309 Journal of ELECTRICAL ENGINEERING 66. it is more efficient to eliminate noise. 2005. ISTE USA..—NOVO.—KOIYUMI.—GIANCHANDANI. M : Digital Speech: Coding for Low Bit Rate trotchniczny (Electrical Review) R 88 No. 1993). One. H. A. Radio. 4 edition (15 WENNA. sinki. Y. 2008.et al : ANFIS Model for the Time Series Prediction of Interior Daylight Illuminance. : Exponentially ROBERTS.—KANEDA. N. The future work includes the optimization of algorithms for all kinds of noises and to use the optimized one in the implementation of FPGA and LabVIEW. LAP Lambert Academic Publishing (25 Nov 2011). Puebla. : The Simulation of Realistic Applications. 6. : Digital Signal Pro92663.—VANUS. pp. Newport Beach. H. CA [27] PROAKIS. ANFIS has two advantages over the LMS and RLS algorithms in low SNR environment.—KHLEAF. [19] VASEGHI.—STRAUSS. J. March 1-12. Z. S. tics of a Room Impulse Response. : Type 2 Fuzzy Logic: Theory and [24] HIRSH. 386–388. 12b (2012). DSP-Based Oversampling Adaptive Noise Canceller for Background Noise Reduction for Mobile Phones. A. S : Journal of Computer Science and Applications 8 (2011). H. International [28] MARTINEK. Wi164–173. 8-9 September 2010. Czech Republic. pp. G. Pearson. ITU-T Test Signals for Telecommunication Systems. and Applications.—ZIDEK.-S. O. M.—BILIK. 2001. 1 (1993).—JANKU. P. E. J. it offers faster convergence time.—YANG. NO. Prentice Hall. 4 edition. 155–160. http:// www.mathworks. HAYKIN.html. W.—MANOLAKIS. 2009.—MELIN.—NAZERAN. L : A Method of AdapFeb 2006). Communication Systems.html. ToolboxTM Functions ANFIS and GENFIS1. 2nd Edition (24 Sep 2004).mathworks. H. for Speech Employing Fuzzy and Neural Network..com/help/fuzzy/genfis1. P. Taizhou. pp. S. Elec[13] Matlab R2012b — Adaptive Noise Cancellation. A. 116th Convention BLANCHET. P. G. Processing using MATLABr.—ZAID. Lambert Academic Publishing (10 Mar 2011).—JAVED. K : Adaptive Noise Cancellation Review) 88 No. [18] JANG. MathWorks. V. S : Adaptive Filter Theor.int/en/Pages/default. Germany. The second. : A System for Improving the Diagnostic Quality of Fetal Electrocardiogram.itu. 1st Edition. R. IEEE Trans. ley-Blackwell. 2006. Advances in Electrical and Electronic Engineering 10 No.—PEIFENG. Portugal. 5b (May 2012). R. pp. USA.[23] KONDOZ.—STYSKALA. 2010. 215–220. 327–332. International Con. and Cybernetics 23 No. Acoustic Input Scenarios for Speech Recognition Systems. hanced Processing and Analysis of Multi-Channel Non-Invasive Abdominal Foetal ECG Signals during Labor and Delivery. M : Digital Signal and Image of the Audio Engineering Society. E. and measurement systems design. 69–84. In 38th International Conference on Telecommunications and Signal Processing (TSP). In 2014. P. R. Modern Communication Technology (SoftwareDefined Radio. J. Petr Bilik was born in 1968. power quality measurement. and Optical Wireless Communication) and Graphical Programming (Virtual Instruments). Since 2014 he has worked at the VSB-TU Ostrava (Technical University) as an assistant professor. Unauthenticated Download Date | 1/3/16 5:52 PM . His main interests are graphical programming. He received his Master’s degree in Power Engineering from VSB-TU of Ostrava in 1982. P. J. automated test. : Dynamic Time Warping. P. 157–161. pp. R. Adaptive Neuro-Fuzzy Inference Systems).—BILIK. P. [37] MARTINEK. 1-2 July 2013. J. Jan Zidek was born in 1957 in Opava. Received 1 December 2014 Radek Martinek was born on 10 August 1984 in Nove Mesto na Morave. data acquisition. 382–386. 2 (2014). M.—KELNAR.—BILIK. Los Angeles. 363–366. : Virtual Simulator for the Generation of Patho-Physiological Foetal ECGs during the Prenatal Period. His main interests are automated test and measurement systems design. P.—BILIK. From 1991 until 2003 he worked as the head of Department of Electrical Measurement. : Power Quality Improvement by Shunt Active Performance Filters Emulated by Artificial Intelligence Techniques. IETE Journal of Research 60 No.—MANAS.—BILIK. P.—BILIK. he successfully defended his dissertation thesis on ”The use of complex adaptive methods of signal processing for refining the diagnostic quality of the abdominal fetal electrocardiogram”. J. [38] MARTINEK.—ZIDEK. : Control Methods of Active Power Filters Using Soft Computing Techniques. measurement in communication systems. Wireless Channel Equalization. H. He completed studies at the Faculty of Electronics and Computer Science at the Technical University of Ostrava in the field of Measuring and Control Technologies in 1996. USA. [35] MARTINEK. P.—KOUDELKA. : The Real Implementation of ANFIS Channel Equalizer on the System of Software-Defined Radio. In 8th International Scientific Symposium on Electrical Power Engineering (Elektroenergetika).—VANUS. Currently he is vice-head of Department of Cybernetics and Biomedical Engineering. M. 344–347. [39] MULLER. Stara Lesna. 9-11 July 2015.—VANUS. He works as the vice-dean of this faculty from 2003 and is a member of the Department of Cybernetics and Biomedical Engineering. Jan Vanus was born in 1972 in Rymarov.—KOUDELKA. Stara Lesna. J. [34] MARTINEK. 16-18 September 2015. He received his Master’s degree in Information and Communications Technology from VSB-TU of Ostrava in 2009. signal processing and design of distributed measurement-control systems. pp. Information Retrieval for Music and Motion (2007).—JANKU. Since 2001 he has worked at ˇ -TU Ostrava (Technical University) as an assistant the VSB professor. R. he is working on a dissertation thesis based on progressive methods of data processing in embedded systems. he has been working at the Department of Cybernetics and Biomedical Engineering at the Technical University of Ostrava within the Postdoc Grant called an opportunity for young researchers. His main interests are text-based and also graphical programming. P. pp. Since 2013.—ZIDEK. Czech Republic. J.—KOZIOREK. In 8th International Scientific Symposium on Electrical Power Engineering (Elektroenergetika).—ZIDEK. M. graphical programming. R. pp. Czech Republic. J. he finished his PhD study in the field of Technical Cybernetics in 2004.—NAZERAN. which was established at the new Faculty of Electrical Engineering and Computer Science on VSB-TU of Ostrava. Czech Republic. CA. the author focuses on the design and implementation of operational and technical management functions in Smart Home and in Smart Home Care. Prague. [36] MARTINEK. Biological Signal Processing (Fetal ECG).—VANUS.310 R.” Since 2012. Michal Kelnar was born in 1988 in Novy Jicin.—VANUS. P. R. His current research interests include Digital Signal Processing (Adaptive Filtering. : Adaptive Noise Suppression in Voice Communication Using a Neuro-Fuzzy Inference System. 22 (Oct 2015). J. From 2000 until 2012 he worked as the head of Power Quality Measurement System department in commercial company. J. Slovakia.—VOJCINAK. Slovakia. He received his Master degree in Power Electronics from VSB Technical University in 1991. He received his master’s degree in Measurement and Control Engineering at the Technical University of Ostrava in 2013. M. 2nd International Conference on Advances in Computer Science and Engineering (CSE 2013). 183–193. 16-18 September 2015.—KELNAR. Czech Republic. At present. Martinek et al : A ROBUST APPROACH FOR ACOUSTIC NOISE SUPPRESSION IN SPEECH USING ANFIS [33] MARTINEK. R. : New Strategies for Application of Recursive Least Square Algorithm in Active Power Filters. In 2010. Electronics Letters 51 No.—KELNAR.—ZIDEK. 1744–1746. he successfully defended his dissertation thesis on “Voice communication with control system.. J. finished his PhD study in 1986.