Noise covariance beam forming algorithm based on direction gain and minimum value control

Abstract

The invention relates to the field of array signal processing, in particular to a noise covariance beam forming algorithm based on directional gain and minimum value control, which comprises the following steps of: 1, acquiring a signal; step 2, an MVDR beam forming algorithm module; step 3, a noise covariance matrix estimation module; step 4, a time-frequency mask estimation module; step 5, a time-varying smooth parameter estimation module; according to the algorithm, through a unique two-parameter noise covariance estimation method, interference noise in the non-target sound source direction can be powerfully filtered out, noise existing in the target sound source direction can be effectively eliminated, the algorithm has remarkable advantages especially in a low signal-to-noise ratio environment, and the quality of voice signals is greatly improved; structural characteristics and spectrum sparsity of voice signals are fully considered, different noise environments and voice scenes can be flexibly coped with, and the method has wide applicability; the directional gain is organically combined with the time-frequency mask and the minimum value control recursive average algorithm, and many defects of a traditional beam forming algorithm are overcome.

Description

Technical Field

[0001] The invention relates to the field of array signal processing, and in particular to a noise covariance beamforming algorithm based on directional gain and minimum value control. Background Art

[0002] With the rise of artificial intelligence, microphone arrays have also been widely used. This is because microphone arrays can obtain more effective information than single channels, which is conducive to the processing of later signals. To this end, many researchers have proposed various beamforming algorithms in the past decade. The beamforming algorithms are mainly divided into two types: fixed beamforming and adaptive beamforming. Although the fixed beamforming algorithm has low computational complexity and is easy to implement, its performance is affected by the number of microphones, and the noise reduction effect is not ideal in a reverberant environment. Based on the defects of the fixed beamforming algorithm, some researchers have proposed adaptive beamforming algorithms. This type of algorithm can achieve adaptive nulling of noise in other directions while retaining the target sound source without distortion, thereby achieving the purpose of noise reduction. However, the adaptive beamforming algorithm is affected by the accuracy of statistics, and inaccurate estimation will lead to a decrease in the noise reduction performance of the algorithm and distortion of the target signal. Since then, some researchers have proposed various improved methods of adaptive filtering. The improved methods can achieve better noise reduction effects in coherent noise fields, but the noise reduction performance is poor when interference sources and reverberation are added.

[0003] Based on the above problems, some people have proposed a beamforming method based on neural network masking. This method estimates the value of the time-frequency mask through a neural network model. Its value represents the probability of the existence of speech, thereby achieving noise reduction for noisy speech. Although this method has a good noise reduction effect, it is based on the effect when the training data and the test data are not much different. Therefore, the quality of the neural network model is affected by the training data. Researchers have further proposed a method based on spatial clustering to address the above problems. Although it has a good noise reduction effect and is not affected by generalization, the complexity of this method is high and it is difficult to apply in practice. For this reason, some researchers have proposed a time-frequency masking method based on directional gain. Although this method has better performance than the traditional time-frequency masking method and can effectively filter out interference noise, this type of method cannot filter out the noise in the direction of the target sound source.

[0004] Therefore, those skilled in the art provide a noise covariance beamforming algorithm based on directional gain and minimum value control to solve the problems raised in the above background technology. Summary of the invention

[0005] To solve the above technical problems, this paper proposes a noise covariance beamforming algorithm based on directional gain and minimum value control. This method first uses the information of the target sound source signal direction to estimate the time-frequency mask of the noise, which is used to estimate the noise in the non-target sound source direction contained in the noisy speech. Then, the time-varying smoothing parameter is estimated in combination with the minimum control recursive average to estimate the noise in the target sound source direction in the noisy speech. Finally, the two are combined to estimate the noise covariance matrix, and the MVDR beamforming algorithm is used to complete the denoising of the noisy speech signal. This method can not only filter out interfering noise from other directions, but also filter out the noise in the target sound source direction, thereby achieving better noise reduction performance and further improving the clarity and intelligibility of the speech.

[0006] The following steps are involved:

[0007] Step 1: Signal acquisition;

[0008] Step 2: MVDR beamforming algorithm module;

[0009] Step 3: The noise covariance matrix estimation module estimates the noise covariance according to the two characteristics of the speech signal. On the one hand, the dynamic smoothing parameters of the noise covariance matrix are estimated by combining the characteristics of the existence of silent segments (no speech) and non-silent segments (speech presence) in the speech signal; on the other hand, the sparsity of the speech signal frequency domain after short-time Fourier transform is utilized, and the application of the second-order statistics of the beamforming coefficient is calculated by using the time-frequency mask. The mask value of each time-frequency point represents the probability of the existence of the expected speech, so as to estimate the time-frequency mask value of the noise covariance matrix;

[0010] Step 4: Time-frequency mask estimation module. The time-frequency mask method based on directional gain uses the guidance vector information of the target sound source direction to estimate the time-frequency mask to perform noise reduction processing on the noisy speech signal. Compared with some classic noise reduction algorithms, it can more effectively filter out noise and improve the intelligibility of the signal, so it is suitable for noise reduction processing of speech signals.

[0011] Step 5: The time-varying smoothing parameter estimation module uses the minimum control recursive averaging algorithm to estimate the time-varying smoothing parameters of the noise covariance.

[0012] Preferably, in step 1, it is assumed that the signal collected by the microphone array and the noise as well as the signals between the channels are independent of each other, so the noisy speech signal collected by the array is represented by y(n):

[0013] y(n)=x(t)*q(n)+n(t); (2.1)

[0014] Where * represents convolution, y(n) represents the observed signal and is an M×1 column vector y(n)=[y1(n), y2(n), …yM(n)]T, and T represents transpose, and M represents the number of channels;

[0015] x(t) represents the target signal, and is an M×1 column vector x(t) = [x1(n), x2(n), …xM(n)]T; q(n) represents the impulse response of the target sound source to each microphone, and is an M×1 column vector, and v(n) represents the M×1 noise signal.

[0016] Preferably, formula (2.1) is subjected to short-time Fourier transform processing, and y(n), s(n), q(n), and n(n) are represented by Y(k, t), S(k, t), Q(k), and N(k, t) after Fourier transform:

[0017] Y(k,t)=STFT[y(n)]=X(k,t)Q(k)+N(k,t) (2.2)

[0018] Among them, Q(k) represents the steering vector of the target sound source, and the specific expression is as follows:

[0019]

[0020] where τ m is the delay between the other microphones and the reference microphone, and the specific formula is:

[0021]

[0022] Where R is the radius of the circle, θ is the azimuth of the target sound source, is the pitch angle of the target sound source;

[0023] The output of MVDR beamforming is:

[0024] X(k,l)=HH w H (k,t)Y(k,t); (2.5)

[0025] Where H represents the conjugate transpose, H w (k, l) represents the filter coefficient matrix; the MVDR beamforming algorithm is based on certain constraints to obtain the optimal filter coefficients. The specific constraint formula is as follows:

[0026] argmin(H w H (k,t)R n (k,t)H w (k,t)), stH w H(k,l)Q(k)=1 (2.6)

[0027] Among them, R n (k, t) is the noise covariance. The Lagrangian function is constructed by Lagrangian multiplier method for the constraint formula (2.6), and H w (k, t) derivative, the final optimal filter coefficient H w The results of (k,l) are as follows:

[0028]

[0029] in is the inverse of the noise covariance.

[0030] Preferably, in step 3, the noise covariance estimation is based on the assumption of the presence or absence of speech, using Indicates that the voice does not exist. Indicates the presence of speech, as shown in formula (10):

[0031]

[0032] in, is the conditional probability of speech not existing, is the conditional probability of speech existence;

[0033] Based on the minimum mean square error criterion, the noise covariance is estimated as follows:

[0034]

[0035] in, is the noise covariance in the absence of speech, is the noise covariance under the condition of speech; the noise covariance matrix is ​​derived according to formula (2.9), as shown in formula (2.10):

[0036]

[0037] However, formula (2.10) only estimates the noise covariance in the direction other than the target sound source, so the general form of directional gain time-frequency mask is introduced to estimate the noise covariance in the direction of non-target sound sources; the noise covariance estimation in the state of speech presence and speech absence is shown in formula (2.11):

[0038]

[0039] Where C represents the time-frequency mask of the noise, a n is a time-varying smoothing parameter based on the conditional probability of speech presence.

[0040] Preferably, in step 4, the specific formula of the time-frequency mask is as follows:

[0041]

[0042] Its G n (k, t) represents the gain value of the noise, G S (k, t) is the gain value of the target sound source, which involves two fixed beamformers, divided into the main beamformer and sub-beamformer h s (k); the specific formula is as follows:

[0043]

[0044] The specific formula of function D(·) is as follows:

[0045]

[0046] where r(·) represents the real number of the complex number; from formula (2.13), it can be seen that the directional gain It is not only related to the covariance matrix and variance of the observed signal, but also to the beamformer and h s (k) is related, so the main beamformer The specific formula is as follows:

[0047]

[0048] in, represents the diagonal loading factor, I M is an M×M unit matrix; in order to maximize the white noise gain, the sub-beamformer h s (k) A delay-sum beamformer is used; the specific formula is as follows:

[0049]

[0050] Preferably, in step 5: the specific calculation formula is as follows:

[0051]

[0052] Among them, μ represents the smoothing factor, which is usually taken as 0.85. P(k, t) is the conditional probability of speech existence, which can be obtained through two iterations of smoothing and minimum tracking. The first iteration is to roughly determine whether the speech signal exists. The specific formula is as follows:

[0053]

[0054] When I(k, t) is equal to 1, it means that the speech signal exists, and when I(k, t) is equal to 0, it means that the speech signal does not exist; γ0 and ζ0 are constant threshold parameters, whose values ​​are 4.6 and 1.67 respectively, B min is the minimum noise estimation bias compensation, whose value is 1.66, It is to track the minimum value of S(k,t) within a window of finite length b;

[0055]

[0056] Among them, S(k,t) is obtained by first-order recursive smoothing, and the specific formula is as follows:

[0057]

[0058] Among them, α s is the smoothing parameter, its value is 0.8, w(i) is the window function, W FL is the window function length; based on the first iteration, the second iteration smoothing is performed to filter out the part with larger voice signal power and improve the robustness of the minimum tracking of voice activity detection. The specific formula is as follows:

[0059]

[0060] right Perform first-order recursive smoothing, the expression is as follows:

[0061]

[0062] Similar to the first iteration, Tracking in a window of finite length b Minimum value of

[0063]

[0064] according to The value of union B min To calculate and The specific formula is as follows:

[0065]

[0066] However, the probability of speech not existing depends on and The specific formula is:

[0067]

[0068] Among them, γ3 is a threshold value of 3, and then according to the prior probability of speech absence And the Bayesian formula is used to solve P(k,L), the P(k,l) formula is as follows:

[0069]

[0070] in, The posterior signal-to-noise ratio, tr(·) is the trace of the matrix, ζ2(k,t) represents the prior signal-to-noise ratio. However, the prior signal-to-noise ratio is unknown, so it is estimated using the direct decision method and is used Symbolically, the expression is as follows:

[0071]

[0072] Among them, υ represents the weight factor, its value is 0.92, G H (k,t) represents the log spectrum amplitude gain function

[0073]

[0074] Finally, formula (2.28) is substituted into formula (2.27) to calculate the result of P(k, t), thus obtaining the dynamic smoothing parameter

[0075] Technical effects and advantages of the present invention:

[0076] 1. Excellent noise reduction performance: The algorithm of the present invention uses a unique dual-parameter (time-frequency mask and dynamic smoothing parameter) noise covariance estimation method, which can not only strongly filter out interference noise from non-target sound source directions, but also effectively remove noise from the target sound source direction. It has significant advantages, especially in low signal-to-noise ratio environments, and greatly improves the quality of speech signals.

[0077] 2. Strong adaptability: It fully considers the structural characteristics of speech signals (silent segments and non-silent segments) and spectral sparsity, can flexibly respond to different noise environments and speech scenarios, and has wide applicability.

[0078] 3. Innovative combination: The directional gain is organically combined with the time-frequency mask and the minimum control recursive averaging algorithm, which overcomes many drawbacks of the traditional beamforming algorithm and provides a new solution for the field of array speech signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a structural diagram of a noise covariance beamforming algorithm based on directional gain and minimum value control provided in an embodiment of the present application. DETAILED DESCRIPTION

[0080] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

[0081] Example 1

[0082] See also Figure 1 In this embodiment, a noise covariance beamforming algorithm based on directional gain and minimum value control is provided, including the following steps:

[0083] Step 1: Signal acquisition;

[0084] Step 2: MVDR beamforming algorithm module;

[0085] Step 3: The noise covariance matrix estimation module estimates the noise covariance according to the two characteristics of the speech signal. On the one hand, the dynamic smoothing parameters of the noise covariance matrix are estimated by combining the characteristics of the existence of silent segments (no speech) and non-silent segments (speech presence) in the speech signal; on the other hand, the sparsity of the speech signal frequency domain after short-time Fourier transform is utilized, and the application of the second-order statistics of the beamforming coefficient is calculated by using the time-frequency mask. The mask value of each time-frequency point represents the probability of the existence of the expected speech, so as to estimate the time-frequency mask value of the noise covariance matrix;

[0086] Step 4: Time-frequency mask estimation module. The time-frequency mask method based on directional gain uses the guidance vector information of the target sound source direction to estimate the time-frequency mask to perform noise reduction processing on the noisy speech signal. Compared with some classic noise reduction algorithms, it can more effectively filter out noise and improve the intelligibility of the signal, so it is suitable for noise reduction processing of speech signals.

[0087] Step 5: The time-varying smoothing parameter estimation module uses the minimum control recursive averaging algorithm to estimate the time-varying smoothing parameters of the noise covariance.

[0088] Preferably, in step 1, it is assumed that the signal collected by the microphone array and the noise as well as the signals between the channels are independent of each other, so the noisy speech signal collected by the array is represented by y(n):

[0089] y(n)=x(t)*q(n)+n(t); (2.1)

[0090] Where * represents convolution, y(n) represents the observed signal and is an M×1 column vector y(n)=[y1(n), y2(n), …yM(n)]T, and T represents transpose, and M represents the number of channels;

[0091] x(t) represents the target signal, and is an M×1 column vector x(t) = [x1(n), x2(n), …xM(n)]T; q(n) represents the impulse response of the target sound source to each microphone, and is an M×1 column vector, and v(n) represents the M×1 noise signal.

[0092] Preferably, formula (2.1) is subjected to short-time Fourier transform processing, and y(n), s(n), q(n), and n(n) are represented by Y(k, t), S(k, t), Q(k), and N(k, t) after Fourier transform:

[0093] Y(k,t)=STFT[y(n)]=X(k,t)Q(k)+N(k,t) (2.2)

[0094] Among them, Q(k) represents the steering vector of the target sound source, and the specific expression is as follows:

[0095]

[0096] where τ m is the delay between the other microphones and the reference microphone, and the specific formula is:

[0097]

[0098] Where R is the radius of the circle, θ is the azimuth of the target sound source, is the pitch angle of the target sound source;

[0099] The output of MVDR beamforming is:

[0100] X(k,l)=H w H (k,t)Y(k,t); (2.5)

[0101] Where H represents the conjugate transpose, H w (k, l) represents the filter coefficient matrix; the MVDR beamforming algorithm is based on certain constraints to obtain the optimal filter coefficients. The specific constraint formula is as follows:

[0102] argmin(H w H (k,t)R n (k,t)H w (k,t)), stH w H(k,l)Q(k)=1 (2.6)

[0103] Among them, R n (k, t) is the noise covariance. The Lagrangian function is constructed by Lagrangian multiplier method for the constraint formula (2.6), and H w (k, t) derivative, the final optimal filter coefficient H w The results of (k,l) are as follows:

[0104]

[0105] in is the inverse of the noise covariance.

[0106] Preferably, in step 3, the noise covariance estimation is based on the assumption of the presence or absence of speech, using Indicates that the voice does not exist. Indicates the presence of speech, as shown in formula (10):

[0107]

[0108] in, is the conditional probability of speech not existing, is the conditional probability of speech existence;

[0109] Based on the minimum mean square error criterion, the noise covariance is estimated as follows:

[0110]

[0111] in, is the noise covariance in the absence of speech, is the noise covariance under the condition of speech; the noise covariance matrix is ​​derived according to formula (2.9), as shown in formula (2.10):

[0112]

[0113] However, formula (2.10) only estimates the noise covariance in the direction other than the target sound source, so the general form of directional gain time-frequency mask is introduced to estimate the noise covariance in the direction of non-target sound sources; the noise covariance estimation in the state of speech presence and speech absence is shown in formula (2.11):

[0114]

[0115] Where C represents the time-frequency mask of the noise, a n is a time-varying smoothing parameter based on the conditional probability of speech presence.

[0116] Preferably, in step 4, the specific formula of the time-frequency mask is as follows:

[0117]

[0118] Its G n (k, t) represents the gain value of the noise, G S (k, t) is the gain value of the target sound source, which involves two fixed beamformers, divided into the main beamformer and sub-beamformer h s (k); the specific formula is as follows:

[0119]

[0120] The specific formula of function D(·) is as follows:

[0121]

[0122] where r(·) represents the real number of the complex number; from formula (2.13), it can be seen that the directional gain It is not only related to the covariance matrix and variance of the observed signal, but also to the beamformer and h s (k) is related, so the main beamformer The specific formula is as follows:

[0123]

[0124] in, represents the diagonal loading factor, I M is an M×M unit matrix; in order to maximize the white noise gain, the sub-beamformer h s (k) A delay-sum beamformer is used; the specific formula is as follows:

[0125]

[0126] Preferably, in step 5: the specific calculation formula is as follows:

[0127]

[0128] Among them, μ represents the smoothing factor, which is usually taken as 0.85. P(k, t) is the conditional probability of speech existence, which can be obtained through two iterations of smoothing and minimum tracking. The first iteration is to roughly determine whether the speech signal exists. The specific formula is as follows:

[0129]

[0130] When I(k, t) is equal to 1, it means that the speech signal exists, and when I(k, t) is equal to 0, it means that the speech signal does not exist; γ0 and ζ0 are constant threshold parameters, whose values ​​are 4.6 and 1.67 respectively, B min is the minimum noise estimation bias compensation, whose value is 1.66, It is to track the minimum value of S(k,t) within a window of finite length b;

[0131]

[0132] Among them, S(k,t) is obtained by first-order recursive smoothing, and the specific formula is as follows:

[0133]

[0134] Among them, α s is the smoothing parameter, its value is 0.8, w(i) is the window function, W FL is the window function length; based on the first iteration, the second iteration smoothing is performed to filter out the part with larger voice signal power and improve the robustness of the minimum tracking of voice activity detection. The specific formula is as follows:

[0135]

[0136] right Perform first-order recursive smoothing, the expression is as follows:

[0137]

[0138] Similar to the first iteration, Tracking in a window of finite length b Minimum value of

[0139]

[0140] according to The value of union B min To calculate and The specific formula is as follows:

[0141]

[0142] However, the probability of speech not existing depends on and The specific formula is:

[0143]

[0144] Among them, γ3 is a threshold value of 3, and then according to the prior probability of speech absence And the Bayesian formula is used to solve P(k,l), the P(k,l) formula is as follows:

[0145]

[0146] in, The posterior signal-to-noise ratio, tr(·) is the trace of the matrix, ζ2(k,t) represents the prior signal-to-noise ratio. However, the prior signal-to-noise ratio is unknown, so it is estimated using the direct decision method and is used Symbolically, the expression is as follows:

[0147]

[0148] Among them, υ represents the weight factor, its value is 0.92, G H (k,t) represents the log spectrum amplitude gain function

[0149]

[0150] Finally, formula (2.28) is substituted into formula (2.27) to calculate the result of P(k, t), thus obtaining the dynamic smoothing parameter

[0151] Noise reduction processing of array speech signals can improve the intelligibility of speech signals and plays an important role in the fields of video conferencing and speech recognition. According to the characteristics of speech signals, this paper proposes a beamforming algorithm that jointly estimates the noise covariance by directional gain and minimum control recursive averaging.

[0152] On the one hand, based on the structural characteristics of speech signals, the part where speech does not exist is mainly noise components, and the part where speech exists is mainly composed of two components: speech and noise. Therefore, the minimum control recursive method is adopted to estimate the conditional probability of speech existence in the noisy speech signal through two recursions and minimum tracking, so as to obtain the time-varying smoothing parameter, and use this parameter to estimate the noise covariance in the direction of the target sound source from the noisy speech covariance.

[0153] On the one hand, according to the spectral sparsity of speech signals, the directional gain is introduced to obtain the time-frequency mask of noise, and the noise covariance in the direction of non-target sound source is estimated from the noisy speech covariance using this parameter. Finally, the obtained time-frequency mask and dynamic smoothing parameters are recursively smoothed to obtain the noise covariance estimate, and the MVDR algorithm is combined to achieve denoising of noisy speech signals.

[0154] The algorithm can not only effectively filter out the interference noise from the non-target sound source direction, but also effectively filter out the noise in the target sound source direction through the dual-parameter noise covariance estimation, especially in the low signal-to-noise ratio environment.

[0155] The electrical components appearing in this article are all connected to an external main controller and 220V AC power, and the main controller can be a conventional known device that controls a computer, etc. The specific implementation method of the present disclosure omits the detailed description of known functions and known components. To ensure the compatibility of the equipment, the operating methods used are consistent with the parameters of marketed equipment.

[0156] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.

Claims

1. A noise covariance beamforming algorithm based on directional gain and minimum value control, characterized in that: The following steps are involved: Step 1: Signal acquisition; Step 2: MVDR beamforming algorithm module; Step 3: The noise covariance matrix estimation module estimates the noise covariance according to the two characteristics of the speech signal. On the one hand, the dynamic smoothing parameters of the noise covariance matrix are estimated by combining the characteristics of the existence of silent segments (no speech) and non-silent segments (speech presence) in the speech signal; on the other hand, the sparsity of the speech signal frequency domain after short-time Fourier transform is utilized, and the application of the second-order statistics of the beamforming coefficient is calculated by using the time-frequency mask. The mask value of each time-frequency point represents the probability of the existence of the expected speech, so as to estimate the time-frequency mask value of the noise covariance matrix; Step 4: Time-frequency mask estimation module. The time-frequency mask method based on directional gain uses the guidance vector information of the target sound source direction to estimate the time-frequency mask to perform noise reduction processing on the noisy speech signal. Compared with some classic noise reduction algorithms, it can more effectively filter out noise and improve the intelligibility of the signal, so it is suitable for noise reduction processing of speech signals. Step 5: The time-varying smoothing parameter estimation module uses the minimum control recursive averaging algorithm to estimate the time-varying smoothing parameters of the noise covariance.

2. The noise covariance beamforming algorithm based on directional gain and minimum value control according to claim 1, characterized in that: In step 1, it is assumed that the signal collected by the microphone array and the noise as well as the signals between the channels are independent of each other, so the noisy speech signal collected by the array is represented by y(n): y(n)=x(t)*q(n)+n(t); (2.1) Where * represents convolution, y(n) represents the observed signal and is an M×1 column vector y(n)=[y1(n), y2(n), …yM(n)]T, where T represents transpose and M represents the number of channels; x(t) represents the target signal, and is an M×1 column vector x(t) = [x1(n), x2(n), …xM(n)]T; q(n) represents the impulse response of the target sound source to each microphone, and is an M×1 column vector, and v(n) represents the M×1 noise signal.

3. The noise covariance beamforming algorithm based on directional gain and minimum value control according to claim 2, characterized in that: Formula (2.1) is processed by short-time Fourier transform. After Fourier transform, y(n), s(n), q(n), and n(n) are represented by Y(k, t), S(k, t), Q(k), and N(k, t): Y(k,t)=STFT[y(n)]=X(k,t)Q(k)+N(k,t) (2.2) Among them, Q(k) represents the steering vector of the target sound source, and the specific expression is as follows: where τ m is the delay between the other microphones and the reference microphone, and the specific formula is: Where R is the radius of the circle, θ is the azimuth of the target sound source, is the pitch angle of the target sound source; The output of MVDR beamforming is: X(k,l)=H w H (k,t)Y(k,t); (2.5) Where H represents the conjugate transpose, H w (k, l) represents the filter coefficient matrix; the MVDR beamforming algorithm is based on certain constraints to obtain the optimal filter coefficients. The specific constraint formula is as follows: argmin(H w H (k,t)R n (k,t)H w (k,t)),s.t.H w H (k,l)Q(k)=1 (2.6) Among them, R n (k, t) is the noise covariance. The Lagrangian function is constructed by Lagrangian multiplier method for the constraint formula (2.6), and H w (k, t) derivative, the final optimal filter coefficient H w The results of (k,l) are as follows: in is the inverse of the noise covariance.

4. The noise covariance beamforming algorithm based on directional gain and minimum value control according to claim 1, characterized in that: In step 3, the noise covariance estimation is based on the assumption of speech presence or absence, using Indicates that the voice does not exist. Indicates the presence of speech, as shown in formula (10): in, is the conditional probability of speech not existing, is the conditional probability of speech existence; Based on the minimum mean square error criterion, the noise covariance is estimated as follows: in, is the noise covariance in the absence of speech, is the noise covariance under the condition of speech; the noise covariance matrix is ​​derived according to formula (2.9), as shown in formula (2.10): However, formula (2.10) only estimates the noise covariance except for the direction of the target sound source, so the general form of the directional gain time-frequency mask is introduced to estimate the noise covariance in the direction of the non-target sound source; the noise covariance estimation in the state of speech presence and speech absence is shown in formula (2.11): Where C represents the time-frequency mask of the noise, a n is a time-varying smoothing parameter based on the conditional probability of speech presence.

5. The noise covariance beamforming algorithm based on directional gain and minimum value control according to claim 1, characterized in that: In step 4, the specific formula of the time-frequency mask is as follows: Its G n (k, t) represents the gain value of the noise, G S (k, t) is the gain value of the target sound source, which involves two fixed beamformers, divided into the main beamformer and sub-beamformer h s (k); the specific formula is as follows: The specific formula of function D(·) is as follows: where r(·) represents the real number of the complex number; from formula (2.13), it can be seen that the directional gain It is not only related to the covariance matrix and variance of the observed signal, but also to the beamformer and h s (k) is related, so the main beamformer The specific formula is as follows: in, represents the diagonal loading factor, I M is an M×M unit matrix; in order to maximize the white noise gain, the sub-beamformer h s (k) A delay-sum beamformer is used; the specific formula is as follows:

6. The noise covariance beamforming algorithm based on directional gain and minimum value control according to claim 1, characterized in that: In step 5: the specific calculation formula is as follows: Among them, μ represents the smoothing factor, which is usually taken as 0.

85. P(k, t) is the conditional probability of speech existence, which can be obtained through two iterations of smoothing and minimum tracking. The first iteration is to roughly determine whether the speech signal exists. The specific formula is as follows: When I(k, t) is equal to 1, it means that the speech signal exists, and when I(k, t) is equal to 0, it means that the speech signal does not exist; γ0 and ζ0 are constant threshold parameters, whose values ​​are 4.6 and 1.67 respectively, B min is the minimum noise estimation bias compensation, whose value is 1.66, It is to track the minimum value of S(k,t) within a window of finite length b; Among them, S(k,t) is obtained by first-order recursive smoothing, and the specific formula is as follows: Among them, α s is the smoothing parameter, its value is 0.8, w(i) is the window function, W FL is the window function length; based on the first iteration, the second iteration smoothing is performed to filter out the part with larger voice signal power and improve the robustness of the minimum tracking of voice activity detection. The specific formula is as follows: right Perform first-order recursive smoothing, the expression is as follows: Similar to the first iteration, Tracking in a window of finite length b Minimum value of according to The value of union B min To calculate and The specific formula is as follows: However, the probability of speech not existing depends on and The specific formula is: Among them, γ3 is a threshold value of 3, and then according to the prior probability of speech absence And the Bayesian formula is used to solve P(k,l), the P(k,l) formula is as follows: in, The posterior signal-to-noise ratio, tr(·) is the trace of the matrix, ζ2(k,t) represents the prior signal-to-noise ratio. However, the prior signal-to-noise ratio is unknown, so it is estimated using the direct decision method and is used Symbolically, the expression is as follows: Among them, υ represents the weight factor, its value is 0.92, G H (k,t) represents the log spectrum amplitude gain function Finally, formula (2.28) is substituted into formula (2.27) to calculate the result of P(k, t), thus obtaining the dynamic smoothing parameter

Images