Summary: Reverberation brings about speech signal time and frequency faintness, serious when speech becomes understand hard. To improve speech quality, put forward a kind of least square beam that has wave of filter of dimension accept postposition to form reverberation to restrain algorithm. This algorithm believes the speech after reverberation a cent to be nonstop part and reverberation part, get gain of filter of dimension accept postposition estimates; Be in in the light of speech signal the issue with stronger coherence of noise of low frequency part, use least square beam to form algorithm to undertake scale down is handled, seek solution finally best power worth. Emulation experiment makes clear, the algorithm that puts forward has better reverberation to restrain character.
TN912.16
A
10.16157/j.issn.0258-7998.172043
Chinese cites format: Zhang Xiu again, wang Ting, guo Yecai, etc. The LS beam that is based on wave of filter of dimension accept postposition forms reverberation to restrain algorithm [J] . Electronic technology applies, 2018, 44(3) : 131-134.
English cites format: Zhang Xiuzai, wang Ting, guo Yecai, et Al. Reverberation Suppression Algorithm For LS Beamforming Based On Wiener Post-filter[J] . Application Of Electronic Technique, 2018, 44(3) : 131-134.
0 foreword
Microphone array can accurate get spatial information, apply at telephone conference widely, avoid carry speech of speech communication, man-machine to be mixed interactively aid listen to the system such as equipment [1] . Beam is formed is technique of a kind of main microphone array processing, its are main the useful signal that the purpose is pair of specific direction forms beam, the interference signal that checks other way at the same time and noise [2] . What study as communication of people antonym sound in recent years is thorough, microphone array gets more apply widely.
In the speech communication application that has many distributed microphone, normally expectation quantifies every sensor to handle the reverberation volume that feels signal, so that the choice has the channel of top quality or the smallest reverberation. Assume irrelvant premise falls between the noise on different channel, the beam entity become a useful person that ZELINSKI R[3] offers to have wave of filter of dimension accept postposition, used spatial information to solve the estimation problem of dimension accept filter. But this kind incoherent noise field is encountered rarely actually, especially field of low frequency noise. BERKUN R[4] and MARRO C[5] put forward what array and filter of dimension accept postposition are united in wedlock to be based on Mack fall a confusion of voices and go reverberation is algorithmic. MCCOWAN I[6] is based on noise the more general expression that the filter of hypothesis knowledge development of field answer be concerned with estimates, but this algorithm asks to get function of noise be concerned with beforehand, suitable scope is restricted. The filter of accept of much passageway dimension that SIMMER K U[7] offers (MCWF) , its can be decomposed do not have for the smallest variance lack fidelity answer beam entity become a useful person and filter of only channel postposition, beg optimum solution expression, apparent to improvement of quality of the speech in reverberation. LUEBS A[8] increases a dot to disturb processing on the foundation of white noise and diffuse reflection noise, through offerring the least square solution with optimum overall situation, use the information that microphone array gathers effectively, improve speech quality.
The least square beam that the article puts forward to have wave of filter of dimension accept postposition forms reverberation to restrain algorithm. This algorithm mixes noisy speech signal to get through scale down to containing high frequency with low frequency heft, the beam that use LS forms the overlay after algorithm undertakes handling respectively, have wave of filter of dimension accept postposition, again by go against Fourier commutation to get going the speech signal after reverberation. The computer emulates result proof, the is based on wave of filter of dimension accept postposition LS beam that the article puts forward forms reverberation to restrain algorithm to have better reverberation to control capacity.
1 signal model
Beam of 1.1 least square forms algorithm
Set base blast for M yuan even wire blast, every blast yuan there is filter of FIR of a L rank after the passageway, the frequency bandwidth Nf that receives broadband signal is [Fl, fh] , the transmission way of signal and base blast the included angle of normal direction is θ , blast yuan of span D. The design method in become a useful person of entity of least square beam [2, in 9] , cost function the definition is:
The 1.2 beam entity become a useful person that have wave of filter of dimension accept postposition
ZELINSKI R[3] puts forward if the graph is shown 1 times,get used to postposition filter structure oneself. Among them, a(i) expresses I of the passageway from power spectral density, c(i) expresses each other power spectral density, y(t) is output signal. Compensation is delayed when the speech signal that this system receives to microphone array above all undertakes. I the signal Xi(t) that the passageway receives expresses to be:
The 1.3 LS that are based on wave of filter of dimension accept postposition improve beam to form
The hypothesis that puts forward according to LEBART K and BOUCHER J M [10] , be based on room pulse to answer H(k) namely is a random process, express to be:
Hypothesis Sd(k) and Sr(k) state pure speech believes date S(t) and Hd(k) and Hr(k) respectively roll indigestion, criterion the nonstop signal share that Sd(k) treats speech signal to wait for, sr(k) is the reverberation part that needs processing speech signal.
The estimation gain that by above the analysis gets improvement dimension accept filter [5] :
Because attenuation factor Δ and existence of reverberation time T60 are contacted, and different frequency paragraph reverberation time is different, reason Δ is as specific room the attenuation to different frequency is mixed reflex degree and change, differ namely the reverberation that the acoustical signal of frequency produces has certain difference, and be in actual sound field coherence of noise of low frequency part is stronger, because this introduces the idea that scale down handles, the signal after Fourier alternates cent is high frequency with low frequency heft, dot of frequency dividing line is taken for 1 KHz. Your α is coefficient of the matrix that increase advantageous position, it is normal number, tall, low frequency authority vector HL, HH parts addition, be about to the signal after scale down, after the beam that use LS forms algorithm to undertake handling respectively again sue for peace, the signal Y(ω that will get) have wave of filter of dimension accept postposition, undertake signal going against Fourier commutation gets going finally the speech signal after reverberation. The LS that is based on wave of filter of dimension accept postposition improves beam to if the graph is shown 2 times,form a structure.
By type (5) , type (12) can get the best authority vector that improves beam entity become a useful person to be:
2 experiments and evaluate
Emulation environment: Use 6 microphone sensor even wire blast, among them acoustical source position (- 5, 1, 0) , the leftest position of the first microphone is (- 5, 2, 1) , microphone blast yuan of removed 1 M. The reflection coefficient of all sides wall, floor, ceiling is respectively 0.9, 0.95, 0.9, 0.95, 0.2, 0.4. The diagonally coordinate of the room is (- 7, - 3, 0) , (1, 3, 1.24) . This experiment believes a paragraph of pure speech date (the first class welcomes new fellow student) answer function to do with room pulse roll accumulate the signal after getting reverberation, fs of frequency of sampling of pure speech signal is 8 000 Hz, continuously 2.3 S. Undertake test and verify to the algorithm that put forward from two index.
(1) words music pursues: If pursue 3 show with three-dimensional means speech spectrum is characteristic, it is a kind of dynamic spectrum. Represent the energy bulk of specific frequency band with color depth.
Graph 3(a) is the words music of pure speech signal to pursue; Graph 3(b) is the words music of reverberation speech signal to pursue, from this graph knowable formant around overlaps more serious, in cross axle (time axis) on speech signal appears the defer that makes an appointment with 0.17 S, speech quality damage is serious; Although algorithm of wave of filter of accept of dimension of passageway of the sheet in graph 3(c) has certain inhibition to reverberation, but signal is lack fidelity still serious; It is better that the algorithm that the article in graph 3(d) puts forward restrains the effect to reverberation, lack fidelity to signal improvement is clearer.
(2) uses speech quality to feel an evaluation (Perceptual Evaluation Of Speech Quality, PESQ) to going the effect of reverberation undertakes assessment, PESQ expresses to be:
In type, dind, Aind is the symmetrical interference value that average interference value and average linear combination obtain respectively. From the watch 1 be pleasant to the eye piece, the PESQ of signal of the speech after joining reverberation drops for 1.9, the quality of speech signal drops badly. The PESQ after putting forward algorithm to undertake reverberation is restrained through the article is achieved 2.3.
3 summary
The signal cent after the has wave of filter of dimension accept postposition least square beam that the article puts forward forms reverberation to restrain algorithm general reverberation is nonstop part and reverberation part, get postposition filter gain estimates improvement Wiener, form algorithm to undertake signal scale down is handled with least square beam next, seek solution finally best power worth, part through emulating an experiment the words music chart from signal and index of speech quality perception undertake assessment to algorithm. Emulation test result makes clear, relative to Yuweina reverberation calculates restrain algorithm, the article puts forward algorithmic reverberation to restrain the effect more remarkable.
Bibliographical reference
[1] LOIZOU P C.Speech Enhancement: Theory And Practice[M].CRC Press, inc.2013.
[2] DOBLINGER G.Optimization Of Wideband Fixed Beamformers With Adaptive Sensor Calibration[C].Signal Processing Conference, 2010, european.IEEE, 2010: 2062-2066.
[3] ZELINSKI R.A Microphone Array With Adaptive Post-filtering For Noise Reduction In Reverberant Rooms[C].International Conference On Acoustics, speech, and Signal Processing.IEEE, 2002, 5: 2578-2581.
[4] BERKUN R, COHEN I.Microphone Array Power Ratio For Quality Assessment Of Reverberated Speech[J].EURASIP Journal On Advances In Signal Processing, 2015(1) : 1-11.
[5] MARRO C, MAHIEUX Y, SIMMER K U.Analysis Of Noise Reduction And Dereverberation Techniques Based On Micro-phone Arrays With Postfiltering[J].IEEE Transactions On Speech&Audio Processing, 2002, 6(3) : 240-259.
[6] MCCOWAN I, BOURLARD H.Microphone Array Post-filter Based On Noise Field Coherence[J].IEEE Transactions On Speech&Audio Processing, 2003, 11(6) : 709-716.
[7] SIMMER K U, BITZER J, MARRO C.Post-filtering Techniques[M].Microphone Arrays. Springer Berlin Heidelberg, 2001: 39-60.
[8] HUANG Y A, LUEBS A, SKOGLUND J, et Al.Globally Optimized Least-squares Post-filtering For Microphone Array Speech Enhancement[C].IEEE International Conference On Acoustics, speech And Signal Processing.IEEE, 2016: 380-384.
[9] Wang Yan, wu Wenfeng, liang Guolong. Beam forms the rash club that is based on dovish least square [J] . Electronics signs up for, 2013, 41(12) : 2321-2326.
[10] LEBART K, BOUCHER J M, DENBIGH P N.A New Method Based On Spectral Subtraction For Speech Dereverberation[J].Acta Acustica United With Acustica, 2001, 87(3) : 359-366.
Author information:
Zhang Xiu again 1, 2, wang Ting 1, guo Yecai 1, 2, chen Xiaoyan 1
(1. Electron of university of Nanjing information project and information project institute, jiangsu Nanjing 210044;
Environment of air of 2. Jiangsu province and equipment technology innovate in coordination center, jiangsu Nanjing 210044)
