TW200843541A - Noise reduction system and method - Google Patents

Abstract

A noise reduction system and a noise reduction method are provided. The noise reduction method estimates directions of arrival of signals by directly using a signal subspace of the signals. Noise of the signals is suppressed at directions other than the directions of arrival. In one embodiment, the signals include audio signals. The signals may be multiple wide-band signals and/or coherent signals in multipath environment with a low signal-to-noise ratio.

Description

200843541 • Nine, invention description: - [Technical field to which the invention pertains] The present invention relates generally to noise reduction techniques, and more particularly to reducing the complexity of signals detected by a line (four) detector array: law. Cloth, dead and square [Previous technology] For portable communication devices such as mobile phones, walkie-talkies, etc. *, "Near microphone array" has long been used as an audio ^^

Slgnal detector). When the user uses the portable communication = set to talk to another person, the linear microphone array detects the audio signal sent by the ^, so that the measured signal can be transmitted when the detection is sent. The audio signal is linear and cool = : The nuisance of the noise signal in the environment. Change: To achieve the quality of the audio signal at the receiving end, the noise signal appearing in the detected audio n needs to be suppressed. ° Code =, linear microphone array is usually a microphone that includes a plurality of separate settings. The microphone of the linear microphone array is the same as the :::! signal. The audio signals detected by the microphones in the -instant (1) -?) or :::PShot will be in the direction of arrival of the audio signal (D0A). Accurately estimate the profit test = such as 'Multiple Signal Classification (Music) algorithm has been (4) 砟 ° mi ^ CMultiple Emitter Location and Signal 93988 5 200843541 * Parameter Estimation", IEEE Transactions on Antennas and • Propagation, Vol. AP- 34, No 3, pp. 276-280, 1986), DOA used to estimate the narrowband signal received by the array sensor. In general, the MUSIC algorithm constructs a spectral density matrix from a snapshot vector ( Spectral density matrix ), and performing eigen-decomposition of the spectral density matrix to obtain the eigenvalues and eigenvectors of the spectral density matrix. Then, the MUSIC algorithm uses the eigenvalues and eigenvectors to calculate the DOA The spatial spectrum is used to estimate the DOA. Due to the miniaturization of modern portable communication devices, the microphones of the linear reference array are separated by only a small distance. The audio signal source and the linear microphone array are only used for very short distances. Separated. For example, microphones in modern portable communication devices may be only two centimeters apart from each other, while linear The distance between the wind array and the audio source may be less than 10 cm. Under the above miniaturization conditions, the audio signal may be reflected in the microphone I and/or between the linear microphone array and the audio signal source. The reflection of the audio signal may cause a multi-path condition, which may cause the coherent of the audio signal. However, the MUSIC algorithm often cannot accurately predict the DOA of the homophonic signal. Overcoming the multipath condition One method of limiting the MUSIC algorithm is to use the spatial smoothing method proposed by TJ Shan et al. ("Adaptive Beamforming for Coherent"

Signals and Interference" , IEEE Transactions Acn Acoustics, Speech and Signal Processing, Vol. ASSP-33, 6 93988 200843541

No. 3, pp. 527-538, 1985). However, although this spatial smoothing method can be used to estimate the DOA of a coherent signal, it requires a linear microphone array to contain a large number of microphones, resulting in a spatial optical error with a lower resolution. Furthermore, because the MUSIC algorithm uses only one snapshot vector, the MUSIC algorithm can only process narrowband signals. In order to extend the MUSIC algorithm to process broadband signals, multiple snapshot vectors are required. SUMMARY OF THE INVENTION In one illustrative embodiment, a noise reduction system is provided. The noise reduction system can include an input unit, a first converter, a signal converter, a second converter, and an output unit. The input unit can include a linear array of detectors for detecting analog signals at a plurality of time snaps, merging the analog signals constructed in the time domain. The first converter is coupled to the input unit for receiving an analog signal in the time domain and converting the analog signal in the time domain into a digit in the time domain ^I: the signal processor further includes The transform unit (tranSf〇rmati 2 is used to convert the digital signal in the time domain into a digital signal in the frequency domain (10) cy domain); the noise suppression unit is used for == weight: quantity (weightingv kiss) multiplied In the frequency domain, the noise in the digits (4) in the frequency domain is suppressed, thereby obtaining the digital signal with reduced noise in the frequency or medium; and the inverse transform ', transf〇rmatiGn unit), Xiao Yu /nverse栌 丝 丄丄 Τ Τ Τ Τ 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂 杂Signal, and will be mixed in the time domain, salty, divination. The digital signal transition of the reduced H is less. 93988 7 200843541 Analog signal with reduced noise in the time domain. The analog signal of the noise reduction in this time domain.兀•可轮出: In another exemplary embodiment, a method is provided. • The noise reduction method can reduce the two methods by the linear microphone array. Noise in the signal. The method may include a plurality of snapshot vectors of the audio signal, from the snapshots to the 准= tiger quasi-injury matrix; the feature decomposes the spectral density matrix === and the plurality of eigenvalues, and (4) obtains the signal subspace between the direct use The signal subspace (4) spatial spectrum is the "DQA of the tiger; based on the view to prepare the weight vector, the weight vector to obtain the noise reduced noise] with reduced audio signal. ... Tiger's and output the noise should It is to be understood that the foregoing general description and the following detailed description are merely illustrative and not limiting as to the present invention. The W and/or variations of the gossip are in addition to those provided herein. It can also be provided. For example, the present invention can be applied to various groups of the disclosed features.

Sub-combinations and/or combinations and a number of combinations and sub-combinations of the features disclosed in the following detailed description. S [Embodiment] Reference will now be made in detail to the preferred embodiments of the invention, in which It does not represent all possible implementations of the claimed invention. Where possible, the same component symbols will be used in the entire figure of 93988 8 200843541 for the use of the button - + irrigation; Μ曰 not the same or similar components. Decrease =:: Salt explains the noise of a wide variety of noise... The mic array is detected at 222, and the ::: Detector array may be linear, and the resulting signal may be an audio signal. While the number and linear microphone array' should be understood, it is possible to make μk#u, such as electromagnetic radiation signals, as well as other types of linear__ columns, such as linear arrays. Tea Looking at Figure 1, the linear predator array 11〇 contains a plurality of detectors that are linearly arranged and spaced apart from one another. In one implementation, the linear detection array 110 can include three detectors 112, m, and 116. It should be understood that in other embodiments, the 'Linear Detector' may include any number of detectors. In one embodiment, the detectors 112, 114, 116 may include a microphone for detecting the tone signal. For convenience of explanation, the detectors j ^ Μ, 116 疋 are organized in a two-dimensional plane characterized by a horizontal axis 120 and a vertical axis 13 垂直 perpendicular to the horizontal axis 12 〇. The horizontal axis intersects the vertical axis 130 to define the origin. As shown in FIG. 1, the detector 114 is at the origin, the detector 112 is on the horizontal axis 120 and is located on the left side of the detector 114, and the detector 116 is on the horizontal axis 12. The upper side is positioned on the right side of the detector 114. The descents 112, 114, 116 are separated from each other by a separation distance D. In one embodiment, the separation distance D can be about 2 cm. The linear detector array 110 is a group of receiving broadband analog signals. 9 93988 200843541 - Because the broadband analog signal received by the linear controller array 11G: contains the noise signal, therefore, in order to simulate the received broadband analogy, the signal source u can be used to generate the linear detector array to be generated. (10) • Alternatively, the noise source 12 can also be used to generate a signal that is not intended to be received by the linear detection "array 1U", as shown in Figure i. To be received with a ^ number together with a signal that is not intended to be received and simulated The linear detection: the broadband analog signal received by the array 110. In one embodiment, the frequency analog signal includes an audio signal. The signal source 11 can be a user's mouth (m〇uth), the mouth The portion produces an audio signal emitted by the user. In one embodiment, 11 is positionable about 6 cm away from the linear detector array m and = the first angle 01 of the positive direction of the horizontal axis 120. It should be understood that 10 200843541 contains a controller for detecting or wheeling people from p sounds to generate the signal "纟中Μ and P are positive integers. p sound generators can include: on source 11 And/or noise source 12. The tamping sound generator will generate The analog signal detected by the array 110. The first analog detector of the linear detector array 110 can form an input signal h(7), ρ 兄 (〇=Σ a (4, 〇® ~ % , slice equation 1 where α'heart) represents the ith detector (1<=i<=M) pair and has the jth angle at the jth angle A and the d0A at the instant t The impulse response between the generators; 'represents the analog signal generated by the jth sound generator at the instant ^; _ represents the noise detected by the second detector at the instant t; and 3 represents the convolution operation By instantiating all the input signals, we can construct the snapshot vector x(t), ie y(t) = A(t) ® uif) + n(t), the equation 2 where x(t) and n(t) are respectively input 仏 and the row vector of the noise signal, IL(t) is the Ρχ1 line vector of the generated analog signal, and 八(1) is the ΡχΜ matrix of the impulse response. For example, Ζ(7)七(〇,"·,/(0 = [~(0, ··.,~(0],!>〗〇),...,~(?)], and Α(ί) == ..., this (^ρ, ί)] (10)1, 〇« «称,,) Equation 3 Equation 4 Equation 5 Equation 6 93988 11 200843541 where T in the private 3 to 5 represents the transpose operation of the vector or matrix. The -person' can also be the snapshot vector of the other program 2 (t Perform a transformation to obtain the snapshot vector z(z) of the Z_transition, .(^^(zmz),^), where Equation 7 f where J(Z) —[^'z),.··.·,妳〆)], and z is a complex number represented by Z = exp(J0), where J is the number of imaginary units, defined as the square root of negative j ' and p is the azimuthal angle of the complex plane (azimuthal angle ). By using the snapshot vector χ(Ζ) of the z_transition given in Equation 7, the spectral density s(z) can be constructed, 5(Z) = E[y(Z)yT (Z^)] = A( Z) E{u(Z)uT (Z^1 )]Ar (Z'1) + E\n(Z)nT (Z^1)], Equation 8 where Ε[·] represents the expected value. The spectral density s(z) contains the signal (no noise) spectral density and the spectral density of the noise. According to z = exp(j0), Equation 8 can be expressed as S.S_hp, Equation 9 where sNF(6) is the signal (no noise) spectral density, is the spectral density of the noise, and ~ is the proportionality constant. To calculate the eigenvectors and eigenvalues of the spectral density 5(6) of the B-transition given in Equation 9, by multiplying (6) by the spectral density will be on the left side and (Σ-1/2 (state " multiply by the spectrum) The right side of the density willow) and the characteristic decomposition ζ-transition spectral density (10)), where [-1/2(0) is the reciprocal of the square root of the spectral density Σ@), and (ς-1/2_〃 is 2 : ~ 1/2w Hermitian conjugate. Therefore, the spectral decomposition of the characteristic decomposition is obtained 12 93988 200843541 degrees, that is, Σ-" matter, (paper, production + from equation q where I is the unit Matrix. Because Σ, _)(Σ-"w ranks (and k) is P, so you can get p non-disaster pull n push non-recognition values (indicated by eight ice) and (MP) zero features The eigenvectors corresponding to the p-non-disaster-like-like 4-inch convergence values constitute the signal subspace, and the eigenvectors corresponding to the (Μ_ρ) hot-drawing ^u 4-inch eigenvalues constitute the noise sub-sense. In addition, the prior density of the knives of the feature decomposition can lead to normalized eigenvectors, that is, chat,) =. So, you can get: A corpse (沴)0 0 ΕΗ{φ) Equation 9 can therefore be rewritten as 特征"1/2(狮)(Σ-1/2,=·λ_ Λ〆卢) + /?wI 0 with the eigenvalue m=

The Pyf equation 10 〇 and the eigenvector take form W, where the eigenvalue october contains the signal source 11 and the noise source 12 eigenvalues. According to Equation 1〇, Z can be obtained, and the signal first spectral density s and the signal spectrum factor of the Z_transition are (6). That is ~ (multi) = Σ 1/2 (do S'(4) eight p(8)#(6) (£ still (multi)) 丨 and, 3]1, \φ)-Συ\φ)ΕΡ(Φ)Αιί\φ)1 Equation 11 Equation 12 where chat is a feature vector containing Ρ elements corresponding to ρ non-love eigenvalues. By performing the inverse Fourier transform on the '93988 13 200843541 stomach spectral density sNF(8) and the spectral factor W(6) given by the equation 11#σ12, the signal spectral density can be obtained. The signal spectral factor is 5^2(z), which is the 'J' of J, and the equation 13 *=-〇〇上兀Ο幻=ΣΖ^士!. Equation 14 can be obtained by using the moving average model Calculating the spectral density of the signal '(z) in Equation 13 by inserting a point on the C unit circle. In one embodiment, 2n + 1 points can be used on the unit circle and can be pulled by Lagrange interpolation and uniquely determines the spectral density of the signal '(7), ie 4(7) Equation 15 where \ is the spectral density matrix (L = A), which is defined as 〇b£(z) = -^±(wyzk ^ ^ T ^ Z = W^Qxp[J2n-^-£] is an interpolation function from 2^ + 1 k=_n, and +1. The interpolation point can be placed evenly in the unit circle for estimation Signal subspace. Signal light can be obtained by decomposing the spectral density of the signal given by Equation 15: (7) The eigenvalue of the density '= (z) and the eigenvector, thus estimating the dimension of the signal subspace. The Euclidean distance J(6) between the noise subspace and the direction vector is defined as 14 93988 200843541 Equation 16 ^CV (^)=- where the Μ is the noise subspace formed by the row eigenvectors of the noise subspace: the matrix, the direction vector (discussed later), and the spectral weighting function (, > ;〇) (also discussed later) The spatial spectrum of D0A can be defined as 1 Equation 17 J ^ f〇h〇^)^f+2p(\^ie)Etcf II. For accurate estimation, it is used in many A plurality of wideband audios in the environment and a D0A with the same js number can be used to form a covariance matrix using a plurality of snapshots at each instant. In one embodiment, consider ◎ snapshots inward. Q is a positive integer. The qth snapshot vector is given as _, Equation 18, where lSqgQ. Using a plurality of snapshot vectors defined in Equation 7, a covariance matrix Rt can be constructed, which is given as is(q)

^ Q Σ灼(4)是—无),·· Σ乃(4)~(“一无) g {qk) Qk (4)+1 g=k+l Qk : , : ^ 2 Σ yu (Φι Y^yM (q) yM (q - k) 籴 +1 Equation 19 where and the subscript moxibustion is an integer from i to ", and 2 = 2" +1 (" is an arbitrary integer). Figure 2A schematically illustrates the delay along time ( Time lag) direction 24 〇 a plurality of covariance matrices R 〃. As shown, each covariance matrix & square 210 indicates that the square 21 〇 represents across the first axis 22 〇 15 93988 200843541 = second axis Spatial correlation of 23G (spatial(10)elation). Using ρ snapshot vectors to form + i covariance matrices in the case of 贝 也, using equation 19, the spectral density matrix & can be defined as Π = Σ w(k) Rit Exp[-J 2; 2n + l Equation 20 ^ where w(4) is a weighted vector. The eigenvalues and eigenvectors of n+1 spectral density matrices can be obtained by decomposing spectral density matrices & The eigenvalues and eigenvectors of the matrix A can distinguish and identify the signal: space and noise subspace. If the noise subspace matrix ～ includes the noise For the two (Μ-P) eigenvectors, the spatial spectrum '(9) can be calculated using the equation 丨7, without regard to the direct current (DC) component (ie, the term). Then the spectral weighting function can be Defined as (if Ρ unweighted) and < § & (if weighted), its center is the eigenvalue used for the signal subspace I. In addition, the direction vector 仏(9) is given as {〇) = [1 ? exp(J2^.3 exp(J2^(M -l)k)], Equation 21

^ _ DisinO where % + 1 and D is the separation distance between the two debt estimates of the linear detector array 11〇. The direction vector meaning (9) may be a complex sinusoid vector (complex sinusoid vector) used to calculate the Euclidean distance d(8) in the signal subspace and/or the noise subspace. Figure 2B schematically illustrates a plurality of spectral density matrices along the temporal frequency direction 260 as shown, 93988 16 200843541 The pupil mountain matrix & is represented by a square 250, which is 25 = = horizontal The spatial correlation between the first car by 270 and the second car by 28 。. In one embodiment, the spectral density matrix & can be constructed from the covariance matrix & In order to accurately estimate the D〇A for the homology signal and overcome the shortcomings of the spatial smoothing method, the spatial spectrum can be controlled by directly using the signal subspace (DUSS). The noise subspace of the t z_ transition can be expressed as Equation 22 (10). (8) Z - (4) = secret) exp exp exp [for, / = 1 八 旦 (10) represents the first eigenvector of the noise subspace Component: buys the ζ_ polynomial of the representative (Μ_ρ) components, and is called the incident angle parameter. The angle of incidence parameter ω (which may be defined as the human angle information containing the i-th sound generator at the center frequency /.) 3 In one embodiment, for 3 coherent signals (ie p=3) and 8 - (ie M = 8) detector with signal-to-noise ratio (SNR) $: The root of the polynomial coffee. Polynomial. The root of (7) is a complex number, which can be represented by a point on the number plane. In Fig. 3, the points representing multiple hearts are scattered evenly in the unit circle of the complex plane. The polynomial (10) = the root of the uniform spread suggests that the (four) child is estimated to be used for the DOA of the same week 5 k. Considering the spatial correlation matrix, the AO 珂 珂 于 于 苓 苓 苓 ^ ^ ^ ^ 向量 向量 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 空间 空间 空间 空间 空间 空间 空间 空间Zero and satisfy the homogeneous matrix equation 4 (h〇m〇genec) US matr 93988 17 200843541 equation), ie U铋厶=0, ,. Equation 23 π π Κ ΓΓ Correlation matrix ~ is (M_K+1)xK The matrix, and the zero-inward-疋X 仃 vector. If the inner product of the spatial correlation matrix % and the zero vector 疋 is not 疋τ △ is not the characteristic of the spatial correlation matrix 1^ to the zero vector. For the sake of simplicity, the special emblem inner (v) private (-) / 7 % 攸 inward - to know that the table is not ν (·), and the nullity matrix Uw is given as interphase v(l) v(2) υΛ v(4) Meal 1:)... v(m) V(jT)... Equation 24 ·: : '· V(W VC^ —1)...v(Af - 尤+ ι)"w - To calculate the zero vector ι, the feature decomposition can be performed on the inner product f of the spatial correlation matrix %. The inner product & is defined as, k Σ, ^(ι + Κ) ν(ι + Κ) ... ΛΓ - iT ' · Σν*(ζ + Ι)ν(/ + Χ) ... Μ-Κ ΣνΦ (/ + Γ)ν(ζ + 1) >〇Μ-Κ η Σν*(ζ + 1)ν(/ + ΐ) 7=0 Equation 25 • where ρ is the real dimension of the spatial correlation matrix (real dimension ) 'K is a parameter determined by the rule of thumb, and ·) is a conjugate complex of <·). In one embodiment, the eigenvector "Z_" of the signal subspace is calculated for three coherent signals (ie, p=3) and eight (ie, $8) detectors having a signal to noise ratio of 1 〇 dB. The root of the polynomial L(z) formed by the transformed zero vector & and represented by a point in the complex plane, as in the first graph. As shown, the points are approximately around the unit circle. Therefore, 93988 18 200843541 The spatial spectrum obtained from L-Dragon II can be better estimated for the DOA of the homology signal. By directly using the k-space (Duss), the spatial spectrum rDUS can be obtained, )==- -^^ 1 l~t/t{e)EmcEiat{0) In Equation 26 2, I represents the signal subspace matrix 'which includes a plurality of rows corresponding to the non-valued 7 sign vector (10), and (10) is the direction of Equation 21. In the latter (Κ(9)=β(θ) is the Euclidean distance between the signal subspace (the sentence and the direction vector (8). In order to suppress the array from the D0A array 1 1 〇, , the weighting of the linear detector vector in / is ..., and is given in addition to the weight. Using the minimum variance method (minimum variah 々 direction is less added to the -_ = 1 limit e method), By itiplier )) take W峨二 (Lagrangian multiplier (Lagrange " I is the source of the signal source 11 to obtain the weight vector 咐), its covariance matrix. Therefore, can be buried * Derogatory to _) in Equation 19, that is, the weighting vector W(Jt) = -^jlak(0l) - ak. Equation 27 Once calculated The weight of the square (4) 27 (4) Bay J can be calculated by adding 19 93988 200843541 weight vector ', multiplied by the frequency domain' input signal to calculate the noise reduction input signal, ie equation 28, in the frequency domain eight The input signal in Equation 1 is the input signal after Fourier transform of Equation 1. (7). As a result, the inverse discrete Fourier transform (Zhu Dewen U^T) can be performed on the input signal of the noise reduction of the frequency domain to the card by the accumulation field. The noise-reduced input signal is also transmitted, so that the noise-reduced input signal is transmitted to the receiver, because those signals entering the linear detector array 110 in the direction other than the inspection are significantly suppressed to noise. The reduced input signal is cut, the receiver can only receive the desired signal to be transmitted. Therefore, a high quality audio signal can be transmitted from the transmitting end to the slot through the communication device including the linear side 11 array 11G. The communication device may include a portable communication device, such as a mobile phone, etc. Referring to Figure 5, a noise reduction system according to an embodiment of the present invention will be described in detail. 500. As shown, the noise reduction system can include an input unit 5H), a first-converter trace, and a signal processor trace noise reduction system 5GG that can include a second converter 54() and an output unit 550. In one embodiment, the 'input unit 51' may include a linear detection benefit array having a first detector 5U, a second detector 514, and a third detector 516. The input unit it 510 is in a plurality of transient debts The analog signal is measured, thus constructing an analog signal in the time domain. In one embodiment, detectors 512, 514, and 516 can be audio or microphones, and analog J can be an audio signal. In one embodiment, the first controller, the 512, the second debt 93988 20 200843541, the *, and the second detector 516 are linearly aligned and spaced apart from each other. Although three prescalers 512, 514, and ... are shown in Fig. 5, it should be understood that the input unit 51G may include any number of detectors. It should be understood that the 'predictors 512, 514, and 516 can include electromagnetic radiation signals that can be included. The first converter 52 is connected to the input unit 唬 as shown in Fig. 5. Coupled for receiving an analog signal in the time domain and converting the analog signal in the time domain into a digital signal in the time domain. In one embodiment, X, the first converter 520 can be analog to digital conversion crying (anal °g such as digital (A / D) (3) please ne, such as four-channel A /; converter or two-channel stereo Decoder (c〇dec), and can be sampled at 16 kHz. Another hole, the gentleman signal processor 530 is connected with the first turn (four) 52 ,, for receiving the digital signal after the conversion. The processor 53G converts the digital signal in the time domain into a digital signal in the frequency domain, and suppresses the noise in the digital signal in the frequency domain by multiplying the vector by the digital signal in the frequency domain. Obtaining a digital signal with reduced noise in the frequency domain. In one embodiment, the 'signal processor 53' may include commercially available signal processing 1 (off), for example by Texas Instruments (τ^ Instnnnems Inc.) Made by Ti Dsp 6713. It should be understood that

The processor 530 can further convert the digital noise reduction in the frequency domain back to the noise reduction digital signal in the time domain. The U signal processor 530 can include a transition unit 531, a weighted unit milk, a plurality of multipliers 537, another milk, and an inverse transition 93988 21 200843541 element 535' to perform the functions described above. • For example, the signal processor may include a transition sheet & 531, which is converted to a digit signal in the frequency domain by a digit (4) in the time domain. In an embodiment, the transform unit 531 can perform a Diffracted Fourier Transform (DFT) in the time domain. The progressive message processor 530 may also include a weight vector preparation unit 5%. The weight vector preparation unit 533 receives the digit signal in the frequency domain and calculates a weight vector based on the digital signal received in the frequency domain. The weighted vector preparation unit 533 forms a plurality of snapshot vectors from the received digital signals in the time domain according to Equation 18 = two 9 constitutes a covariance matrix from the snapshot vectors. Then, the weighting is prepared early το 533 to calculate the spectral density moment according to Equation 2〇 and the feature is decomposed into the spectral density matrix to obtain the value and the eigenvector of the spectral density matrix. The spectral density matrix can be decomposed into a noise subspace using the eigenvalues of the spectral density matrix and the extra 榷 vector preparation unit 533. The signal subspace may contain the eigenvalues of the := degree matrix, and the diurnal moment may include a minus moment (four) dragon value corresponding to the == value. By directly ❹ the signal sub-* addition 榷 vector preparation unit 533 can calculate ^ 根据 according to the equation %, thus accurately estimating D0A. Furthermore, the weight vector preparation list ^ milk = thank you to prepare the weight vector. In one embodiment, the gain of the analog signal is maximized) and the weighting is less than the === analog signal (or the ❹ of the analog signal is minimized: Γ/δ 22 93988 200843541 鬌. Once the weight vector is calculated , weighting. The weighting vector is sent to the multiplication method and the digits in the frequency domain; and sum to multiply the weight vector f, k, and t. The weighting vector is in the frequency domain. In the implementation of the helmet, in the implementation - in the middle = less _ ^ ^ ^ is ready to be sent to the receiving end. (four) less digital signal can be - in one embodiment, the signal processor 53 can be included 535, for receiving a digital signal of 2 in the frequency conversion, early το frequency domain, and converting the digit n in the §5 tiger into a noise reduction in the time domain. In one embodiment, the inverse transform unit 535 reduces the amount of noise in the frequency domain by reducing the number of noises in the frequency domain by ^^^^ 〇DFT. The digital signal. The digital signal in the noise of = can be prepared (four) sent to receive Han drag: L5 picture, noise reduction The less system 500 can further include a second second converter (10) coupled to the signal processor $. The 540 receives the digital signal in the time domain and reduces the noise in the % domain: The analog digital signal is converted into an analogy of the noise in the time domain. In one embodiment, the second converter 54 can be clamped to an analog (four) device (eg, (4) o-analog (D/A). The analog signal of the noise reduction in the day T domain may be ready to be sent to the receiving end. In addition, the noise reduction system 500 may include an output unit 550, which is coupled to the second converter. 54〇coupled. Output unit receives 23 93988 200843541 I Noise reduction in the time domain • Reduced analog signal. The first and second sub-rings emit the noise in the time domain; the sounder. In an embodiment, the output unit 550 includes a plurality of "Phoenix & rs?" detailed description of the noise reduction: - Guanglu is in accordance with an embodiment of the present invention, reducing the linear microphone array _ to = using f noise The method is to suppress the noise from the number in the step (10). Γ· Signal quasi-injury It is given in a fast Zhaozhong. In a ^:, two in. The # snapshot vector is in the equation! 8 and / or in the signal (four) contains a plurality of broadband audio signal. The linear microphone; ^ column 7 = path environment Same tone signal. detected ^ ί 1 __ detected the audio message

Wt 9 ° 5 is the audio signal in the time domain. The equal tone I: at: using discrete Fourier (10) In step 620, the covariance matrix is composed of fast: density matrix is composed of the covariance matrix. The co-party is given in 19, and the spectral density matrix is in the equation ^2 spectral density, which may include a weighting vector. The weighting vector may be determined by a method that is suitable for (for example, a minimal variance method). In 630, the spectral density matrix is decomposed into features to obtain a complex eigenvector and a plurality of eigenvalues. The corresponding::vector is used to construct the signal subspace. In addition, the corresponding:: eigenvector of the defensive value Used to construct the noise subspace. In step 640, the dqa of the audio signal is estimated by the spectrum between 93988 24 200843541 derived from the direct use of the signal subspace - 'spectral system is given to equation I in equation 26 The middle space and the direction vector are based on the Euclidean distance between the signal subspace and the door. • In step 6 50, the force avoids ab〆doa to prepare. In one, Using minimal variance method

Where more is given, the weighting vector can be weighted at the DOA and less weighted. Given in a direction other than μ ^ In step 660, the noise subtraction vector is obtained. In a ^ ^ 曰 "Tiger hunting by multiplying the weighted audio signal, 1: two in the 'weight vector can be associated with the frequency domain. Then, by using the inverse two noise to reduce the coffee audio The signal is converted into noise in the time domain 2 to reduce the audio signal. The noise in the field is reduced to ΐ: Γ::, the noise signal in the time domain is reduced by the signal signal. Receiver The sound that can be received by the noise is obviously reduced. The computer that performs the above-mentioned noise reduction process is also light. Bu detects 々 Consider the orientation detector (omini-directi〇nal 22 has two _ _ _ linear line ❹ each other w The 1st phenotype has the same frequency characteristics. The π brain simulation system considers 2 日 日 门 ^ ^ In this example, the electric vector and the Hamming window / 曰 'delay (ie the heart 20) 400 I, ammmgwmdow) determined in the snapshot to calculate the spectral density matrix of the equation 20 r corp / τ ', σ 。. " 知 知 pull to consider three signal sources, each source contains 93988 25 200843541 Extra white Gaussian noise with band pass filter n noise ). In this example, the signal sources are delayed by an array spacing parameter (array spacing) that is substantially equal to five (ie, 乂-5). Therefore, the sources can be represented as follows: Source 1 : Source 2: Source 3: 1 + 0.371 Ζ^+036 Ζ-2 ? __1_ , Ϊ+0Α33Ζ"1 +0.49 Ζ^1 and 1__ ΐ+0.994 Ζ^1+0.64 Ζ'2 〇 In this example Signal source 1 inputs a signal from q = 1 〇. The first angle of incidence & signal source 2 inputs a signal from % = 〇. The second angle of incidence %; and source 3 input from θ3 = + ι第二. The signal of the second incident angle %. The amplitudes of the signal sources 1 to 3 in the frequency domain are shown in Fig. 7. Here, the signal sources 1 to 3 are generated with the center frequency bit at 〇·3 Ηζ In this computer simulation, Y"), where σ^ is the covariance of X and y. Signals of the same power. The spectra of signal source U 3 may overlap each other. In computer simulations, the signal-to-noise ratio (SNR), which is defined as the ratio of the dispersion and dispersion of the signal, is considered to be zero. The DC component of the spectral money matrix, (b 0), removes the Dc component (heart 〇) of the spectral density matrix from this calculation because it does not affect the calculation of the spatial spectrum used for the orientation detector. Consider the correlation coefficient Y〃 (which is defined as the covariance of X y, while the ~ and σ 〆 are respectively in the first case, the covariance algorithm of \=()·585 is considered to have a correlation coefficient. 'Ya and calculate the spatial light according to Equation 17 93988 26 200843541. The k 唬 subspace is four-dimensional, and the white Gaussian noise correlation matrix is given as follows: 1 0.585 0.585· 0-585 1 0.585 >585 〇 · 585 1 空间 The spatial spectrum produced in this first case is shown in Figure 8. Since the signal in this first case is weakly correlated, Equation 17 is used to calculate the spatial spectrum. The covariance algorithm is sufficient to accurately predict DOA. The computer simulation considers the correlation coefficient and calculates the spatial light according to Equation 17 and the correlation matrix of white Gaussian noise is in the second case, \= The covariance algorithm spectrum of α9. The signal subspace is given in four dimensions as follows: ^ 1 0-9 0.9' 0.9 1 〇·9 〇·9 0.9 1 j. ί In the second case, because in the second The correlation coefficient in the case is greater than in the first case Correlation coefficient, so the signal is more relevant than the signal in the first case. Therefore, the signal in the second case can be referred to as intermediately correlated. In the second case The spectral density is shown in Figure 9. As shown, the DOA of sources 1 to 3 are still clearly distinguishable in the spatial spectrum. However, the amplitude of the spatial spectrum at SDOA has been significantly reduced. In three cases, the computer simulation first considers the correlation coefficient 93988 27 200843541 and calculates the covariance algorithm spectrum of the spatial light number Y", = 1.0 according to Equation 17. The correlation matrix becomes

In the second case, there is not a multi-path ring ^ r — the brother-in-law input k唬 is the same as the 5 tiger. If the younger brother does not have a picture, the DOA of the L唬 source 1 to 3 is no longer distinguishable in the spatial aperture. However, in the case of phase + 1 + 1 under the relative conditions, the computer simulation calculates the spatial spectrum for the two cases again according to Equation 26 by directly using the signal subspace. The root market is very similar to the space light produced by Wan Hong 26 and is shown in Figure 11. If 筮&quot;曰口 τ刘昂11 is not, DOA is now clearly distinguishable in this spatial spectrum. Therefore, the computer simulation of the μ 1 p J prisoner has shown that the spatial spectrum of Equation 26 can be used to break the DOA of the coherent signal and/or signal in the eve path. Other embodiments consistent with the present invention will become apparent to those skilled in the <RTIgt; The description is intended to be illustrative, and is only illustrative, and the scope of the invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS The various embodiments and aspects of the invention are set forth in the accompanying drawings, and the description Figure 1 illustrates a linear microphone array for receiving audio signals from the source and the noise source 28 93988 200843541 #. - The 2nd and 2nd sects, the α-one-fast-vectors constructed by the ^ 方 方 方 方 方 方 方 方 方 方 方 方 方 以及 以及 以及 以及 以及 以及 以及 以及 以及 以及 。 。 。 。 。 。 。 。 。 。 。 。 Figure 3 illustrates the root of the polynomial consisting of the eigenvectors of the noise space in the complex plane (_ coffee piane). Figure 4 illustrates the root of the polynomial consisting of #梦办H w in the complex plane. The feature vector between the two is illustrated in Fig. 5 to illustrate a noise reduction system in accordance with the present invention. Figure 6 illustrates a method of noise reduction in accordance with the present invention. Figure 7 shows the amplitude of three model source sources according to computer simulations in accordance with the present invention. ' a Figure 8 illustrates the spatial spectrum of the correlation signal based on the use of the covariance algorithm. Day Riding Figure 9 illustrates the spatial spectrum of a medium-relevant signal based on a computer model using a covariance algorithm. & Figure 10 illustrates the spatial spectrum of a homology signal that is suspected by a computer using a covariance algorithm. Figure 11 illustrates the spatial spectrum of a coherent signal suspected by a computer using the signal subspace (1) um), 寅 algorithm. ~ [Main component symbol description] 11 Signal source 12 Noise source 110 Linear detector array 112, 114, 116 Detector 93988 29 200843541 120 Horizontal axis • 130 Vertical axis * 210, 250 Square 220 ^ 270 First axis 230 280 second axis 240 time delay direction 260 time frequency direction 500 noise reduction system 510 input unit 512 first detector 514 second detector 516 third detector 520 first converter 530 signal processor 531 transition Unit 533 weight vector preparation unit 535 inverse transformation unit. 537 ^ 538 &gt; 539 multiplier 540 second converter 550 output unit steps 610, 620, 630, 640, 650, 660, 670 30 93988

Claims (1)

200843541 •, the scope of application for patents: A noise reduction system, including: Instantaneous two inclusion: linear debt detector array 'for analog signals in the time domain in multiple 盥^k仏; wrong with this input; a first converter of unit switching, the receiving said to replace the m-medium 1 apostrophe in the time domain and to convert the digital signal in the time domain to a digital signal in the time domain; In addition, the signal processor is configured to receive the i:bit signal of the 5, the signal processor further comprising: 'for converting the digital signal in the time domain into the frequency domain The digital signal; and the noise suppression unit are configured to suppress the noise in the digital signal in the frequency domain by multiplying the weight vector by the digital signal in the frequency domain to obtain the noise signal. The noise signal in the frequency domain is reduced by the digital signal 0. 2. The system of claim </ RTI> wherein: - the signal processor further comprises an inverse transform unit for receiving the digital signal of the noise reduction in the frequency domain and for the frequency domain + The noise-reduced digital signal is converted into a digital signal with less noise in the time domain. ° ' 3. The system of claim 2, further comprising: a second converter coupled to the processing of the number B, the second converter receiving the digital signal of the noise reduction in the time domain And the digital signal reduced by the noise in the time domain t is converted into an analog signal of the 93988 31 200843541 noise reduction in the time domain. 4. 5. The system of claim 3, wherein: ^ The second converter comprises a digital to analog converter. For example, in the system of claim 3, the further steps include: 3 out units for taking turns out the analog signals in the time domain.讯减减# 6. The system of claim 5, wherein: the wheel unit comprises a speaker. 7. The system of claim 1 wherein: 8. The detector array comprises a plurality of detector states that are linearly arranged and equally spaced from one another. "A system for applying for a patent scope, item 7, and a medium. 9. A communication device, including: a system as claimed in claim 8 of the patent application. For example, the system of claim 7 of the patent scope, wherein: The apparatus includes a green magnetic radiation signal. The analog signals include electricity. 11. The first converter to digital converter according to the patent application. The system of the first item, wherein: includes a class having a sampling rate of about 16 KHz. 12. If the patent application model is to change the system of item 2, where: the discrete Fourier transform is performed, and the inverse transform 93988 32 200843541 unit performs the inverse discrete Fourier transform. 13. If the patent application scope is the first item The system, wherein: • the chowder suppression unit further comprises weighting the weighting vector of the car saw to stay-plus μ #里+备早儿, .w Shaofeng early 7° according to the hunting by using the analog signal space! For example, apply for special wealth ^ direction (job) "" the weight vector. The system of the 13th item of the target of the search for the target, where: the weight vector preparation unit is used by the meter, the dry quiz 1 is connected with the h number Subspace Between the #出μ inter-spectral spectrum, the letter is decomposed. The 二丁二间攸攸攸 density matrix 15. As in the patent application scope of the "item of the system, and the transmission vector preparation unit according to from the time domain Let the number two and one number of snapshot vectors form (4) the variance of the spectral density matrix.啲 啲 ^ 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16铗ia thousand is multiplied by b to obtain a digital signal that is reduced by 颂4 f. 17. A signal processor, comprising: a transform unit, the group constituting a digit received in the time domain, 乜旎, and the transition unit group is configured to convert the under-order L旒 in the time domain into the frequency domain a digital signal; and a noise suppression unit, the group is configured to receive the digital bits in the frequency domain and to suppress the noise in the digital signals in the frequency domain, and obtain the interference in the frequency domain with 93988 33 200843541 The reduced digital signal is obtained by multiplying the weight vector by the digital signal of the noise reduction in the frequency domain. 18. The signal of claim 17 The processor further includes: a reverse transfer unit, configured to convert the digital U tiger whose noise in the frequency domain is reduced into a digital signal reduced in the time domain A, and output, and the like in the time domain The noise signal of the reduced noise is further processed. 19. The signal processor of claim 17 wherein: the noise suppression unit is further configured: dan. according to the time domain digit The signal constructs a plurality of snapshots to construct a nine-density matrix according to the covariance matrix of the snapshot vectors (fourth); interpolating the pupil matrices into signal subspaces and noise subspaces = directly using the signals The spatial spectrum obtained by the space is used to estimate the direction of arrival; and 20 is calculated by applying the direction of arrival to the direction of the weighting vector. In the case of: Shen, the signal processing of the 19th patent range is in the frequency domain. The gain of the digital signal in the most is 21. As in the patent application, the signal processing of the cage target 阗罘19 is suppressed by 7 degrees by 7° in the arrival direction (DOA) by using the weight vector. , 93880 34 200843541 The noise suppression unit minimizes the gain of the bit signal by using the weighted direction of the direction other than the direction of the direction: &lt; the number of .22 communications Equipment, including · · Signal processing as claimed in item 19 of the patent scope. Prepare the complex snapshot vector of the lungs; construct a covariance matrix from these snapshot vectors, construct the difference matrix a spectral density matrix; , 'decompose the covarian feature to decompose the spectral density matrix and a plurality of Wei values, because (4) obtains == mesh ^ vector by directly using the signal sub, :, , subspace; The direction of arrival of the audio signal; the spatial spectrum of the V is based on the arrival direction of the weighting vector. ^ The weight vector obtains the noise reduction of the noise output and the noise reduction of the noise and 24 · as claimed in the 23rd The abundance includes a plurality of broadband signals, wherein the audio signals 25 are as in the method of claim 23, including homology in a multipath environment, and the audio signals 26 are as claimed. The 23rd party #&quot;, converts the audio signals into advances including: 27, as in the 23rd of the patent application scope, the audio signal in the ', domain. Going to 'where' to obtain the noise 93988 35 200843541, the reduced audio signal further includes: multiplying the weighting vector with the audio signals in the frequency domain to obtain an audio signal with reduced noise in the frequency domain . 28. The method of claim 27, further comprising: transducing the noise signal of the noise reduction in the frequency domain to obtain the noise signal of the noise reduction in the time domain. 29. If applying, the method of item 23 of the scope of interest, further includes: (The Euclidean I distance between the subspace and the direction vector of the Λ彳° number. 30. For the method of claim 29, Where: the spatial spectrum (8) is given as DUSS (sentence 1 - d2 (0), which corresponds to the direction of arrival (D0A) and _ is the angle of the signal subspace disk (0A), the distance. /, the χ 之间 之间 之间 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 The space includes corresponding to two vectors, and the noise subspace package: the feature vectors for the sign of the temple. The sub-segment should be such that the zero feature value is 33. The use of the 23rd item of the patent application is extremely small. Variance method to prepare the addition:: quantity. 93988 36