CN118116402A - Bilinear filtering-based multichannel voice noise reduction method - Google Patents

Abstract

The present invention discloses a multi-channel speech denoising method based on bilinear filtering, including: collecting time-domain noisy speech signals, preprocessing the time-domain noisy speech signals; estimating the statistical characteristics of the time-domain noisy speech signals and additive noise signals; estimating the bilinear Wiener denoising filter based on the statistical characteristics; filtering and denoising the noisy speech signals based on the bilinear Wiener denoising filter, and obtaining an estimated value of the clean speech signal. The present invention decomposes the filter coefficients that play a denoising role in the time domain dimension and the spatial domain dimension, and converts the estimation problem of a long filter into the estimation problem of two shorter sub-filters. The shorter filter means that fewer parameters need to be estimated. Compared with the traditional method, the algorithm complexity in the present invention is significantly reduced, and only fewer observation samples are needed to estimate the filter coefficients, which improves the algorithm's tracking ability for non-stationary noise; compared with the frequency domain speech denoising method currently used in actual systems, there is no music noise.

Description

A multi-channel speech denoising method based on bilinear filtering

Technical Field

The invention belongs to the field of speech noise reduction, and in particular relates to a multi-channel speech noise reduction method based on bilinear filtering.

Background technique

Noise is everywhere in daily life. Noise can reduce the quality and intelligibility of speech signals and cause listening fatigue. Speech noise reduction technology is committed to suppressing the impact of noise and extracting clean speech signals from noise, thereby improving the quality and intelligibility of speech, and plays an important role in voice communication.

According to the different execution domains of speech noise reduction algorithms, noise reduction algorithms can be divided into time domain algorithms and transform domain algorithms (such as frequency domain, wavelet domain, etc.). At present, the most widely used speech noise reduction algorithm is the frequency domain noise reduction method. Compared with the time domain noise reduction method, the frequency domain noise reduction method requires low complexity and can be integrated in embedded systems to complete real-time noise reduction. However, the disadvantage of the frequency domain noise reduction method is that it is easy to generate musical noise. According to research, people's tolerance for musical noise is lower than their tolerance for noise. Therefore, how to reduce the musical noise generated by the frequency domain noise reduction algorithm has always been a hot topic of research. The advantage of the time domain noise reduction method is that it will not generate musical noise. However, in the time domain speech noise reduction algorithm, its filter is usually long, resulting in high complexity, which is the biggest bottleneck limiting its actual deployment. Especially for multi-channel speech noise reduction algorithms, its complexity will increase rapidly with the increase in the number of channels, making it difficult to deploy the time domain speech noise reduction algorithm in the actual system to perform real-time noise reduction processing on noisy speech signals.

In the present invention, in order to reduce the complexity of the time-domain speech denoising algorithm, a bilinear denoising scheme is proposed by updating the organization form of multi-channel signal vectors.

Summary of the invention

The purpose of the present invention is to provide a multi-channel speech noise reduction method based on bilinear filtering to solve the problems existing in the above-mentioned prior art.

To achieve the above object, the present invention provides a multi-channel speech noise reduction method based on bilinear filtering, comprising:

Collecting a time-domain noisy speech signal and preprocessing the time-domain noisy speech signal;

estimating statistical characteristics of the time-domain noisy speech signal and the additive noise signal;

estimating a bilinear Wiener denoising filter based on the statistical characteristics;

The time-domain noisy speech signal is filtered and denoised based on the bilinear Wiener denoising filter to obtain an estimated value of a clean speech signal.

Optionally, the process of collecting a time-domain noisy speech signal and preprocessing the time-domain noisy speech signal includes:

The time domain signal model is:

y _m (t) = x _m (t) + v _m (t) (1)

Wherein, t represents a discrete time point, the subscript (·) _m represents the signal received by the m-th microphone, and the microphone array has M microphones in total. x _m (t) and v _m (t) represent the clean speech signal and additive noise signal received by the m-th microphone, respectively, and y _m (t) represents the noisy speech signal received by the m-th microphone. x _m (t) and v _m (t) are independent of each other. The first microphone in the microphone array is selected as the reference microphone, that is, x ₁ (t) is taken as the expected signal.

The signal received by the mth microphone is written as a vector of length L by grouping together L consecutive sample points:

The definitions of x _m (t) and v _m (t) are similar to those of y _m (t), namely:

_{xm (} t)＝[ _xm (t)xm(t-1)… _xm (t-L+1)] ^T

v _m (t) = [v _m (t) v _m (t-1)…v _m (t-L+1)] ^T

x _m (t) and v _m (t) represent the expected signal vector and the noise signal vector of the m th channel respectively, y _m (t) represents the noisy signal vector of the m th channel, and the superscript (·) ^T represents transposition.

By concatenating M noisy signal vectors y _m (t) (m = 1, 2, ..., M) of length L, we can write:

The definitions of x (t) and v (t) are similar to those of y (t), namely:

y (t), x (t) and v (t) represent the overall noisy signal vector, the overall clean speech signal vector and the overall noise signal vector respectively.

Optionally, the process of estimating the statistical characteristics of the time-domain noisy speech signal and the additive noise signal includes:

The correlation matrix R _v (t) of the overall noise signal vector v (t) is estimated by an existing noise estimation algorithm, and the correlation matrix R _y (t) of the overall noisy signal vector y (t) is estimated by a recursive algorithm: R _y (t) = αR _y (t-1) + (1-α) y (t) y ^T (t), where α is a forgetting factor (0 < α <1); the correlation matrix R _x (t) of the overall clean speech signal vector x (t) is estimated by R _x (t) = R _y (t) - R _v (t), and the vector ρ (t) is determined based on the speech signal correlation matrix R _x (t) to obtain statistical characteristics.

Optionally, the process of determining the vector ρ (t) based on the speech signal correlation matrix R _x (t) includes:

The elements of the first row and first column and the elements of the first column of the speech signal correlation matrix R _x (t) are extracted, and the elements of the first column are divided by the elements of the first row and first column to obtain a vector ρ (t).

Optionally, the process of estimating a bilinear Wiener denoising filter based on the statistical characteristics includes:

The whole noisy signal vector is reorganized, and the filter in the traditional denoising scheme is split according to the noisy signal matrix Y(t) obtained by the reorganization to obtain a bilinear denoising scheme containing two sub-filters. The bilinear denoising scheme is equivalently transformed based on the vectorization operation of the matrix and the Kronecker product to obtain a bilinear denoising scheme based on two sub-filters;

Obtaining a definition of a mean square error of an expected signal estimate based on the statistical characteristics of the noisy speech signal, the Kronecker product of the two sub-filters h ₁ (t) and h ₂ (t), and the noisy signal matrix Y(t), rewriting the definition according to the relationship between the two sub-filters to obtain a final definition expression of the mean square error of an expected signal estimate;

The bilinear Wiener denoising filter is estimated based on the definition expression of the mean square error of the final expected signal estimate.

Optionally, the process of obtaining a bilinear denoising scheme based on two sub-filters includes:

The overall noisy signal vector is reorganized as follows:

Wherein, y (t)=vec[Y(t)], clean speech signal matrix X(t)=[ _x1 (t) _x2 (t)… _xM (t)]( x (t)=vec[X(t)]), noise signal matrix V(t)=[ _v1 (t) _v2 (t)… _vM (t)]( v (t)=vec[V(t)]), symbol vec[·] indicates matrix vectorization operation, and dimensions of matrices Y(t), X(t) and V(t) are all L×M;

Based on formula (5), the traditional noise reduction scheme is modified to obtain the following bilinear noise reduction scheme:

Among them, the two sub-filters h ₁ (t) and h ₂ (t) perform noise reduction in the time domain and spatial domain respectively. The length of h ₁ (t) is L, the length of h ₂ (t) is M, and z(t) is the estimated value of the expected signal x ₁ (t). is the filtered speech signal, /> represents the residual noise after filtering;

Transform equation (6) to obtain the following formula:

in, The symbol vec[·] represents the vectorization operation of the matrix, and the symbol /> represents the Kronecker product, and the symbol tr[·] represents the trace of the matrix;

Using formula (7), formula (6) can be written as

Where x (t)=vec[X(t)], v (t)=vec[V(t)].

Optionally, the process of obtaining a definition expression of a final expected signal estimate mean square error includes:

The two sub-filters h ₁ (t) and h ₂ (t) have the following relationship:

Among them, _IL and _IM are identity matrices with dimensions of L×L and M×M respectively;

The definition of the mean square error of the expected signal estimate obtained based on the statistical characteristics of the noisy speech signal and the noise signal, the Kronecker product of the two sub-filters h ₁ (t) and h ₂ (t), and the noisy signal matrix Y(t) is as follows:

in, is the element located in the first row and first column of the matrix R _x (t).

According to formula (10), formula (12) can be rewritten as:

in,

Optionally, the process of estimating the bilinear Wiener denoising filter based on the definition expression of the final expected signal estimate mean square error is as follows:

Step 1: According to the formula Initialize sub-filter h ₁ (t), where /> is the autocorrelation matrix of the noisy speech signal vector of the first channel, is the first L rows and first L columns of the matrix R _y (t), vector/> is the vector consisting of the first L elements of the vector ρ (t), and the superscript (·) ⁽ⁿ⁾ indicates the result of the nth iteration;

Step 2: Substitute into the formula In, get and/> The superscript (·) ⁽ⁿ⁾ indicates the result of the nth iteration;

Step 3: and/> Substitute into the formula/> In, get

Step 4: Substitute into the formula and/> In, get/> and

Step 5: and/> Substitute into the formula/> In, get

Repeat steps 2 to 5 N times to obtain and/>

According to the formula Obtain the bilinear Wiener denoising filter h _bW (t).

The technical effects of the present invention are:

The present invention decomposes the filter coefficients that play a role in noise reduction in the time domain dimension and the spatial domain dimension, thereby converting the estimation problem of a long filter into the estimation problem of two shorter sub-filters, a multi-channel bilinear speech noise reduction filter with low computational complexity and high non-stationary noise processing capability; it is applicable to the time domain, and the core idea of the present invention can be very intuitively extended to the frequency domain noise reduction framework; it can be used for intelligent voice, human-computer interaction and other systems, as well as audio and video conferencing, vehicle-mounted, immersive communication and other systems; it can be used alone or in conjunction with modules such as echo cancellation, sound source localization, dereverberation, and speech separation. The method in the present invention has the following advantages: 1) the algorithm complexity is significantly reduced; 2) fewer observation samples are required to estimate the filter coefficients, thereby improving the algorithm's tracking ability for non-stationary noise. In addition, the method proposed in the present invention is a low-complexity time domain speech noise reduction method. Compared with the frequency domain speech noise reduction method currently used in actual systems, another advantage of the present invention is that there is no music noise.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings constituting a part of this application are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

FIG1 is a flow chart of a method in an embodiment of the present invention;

FIG. 2 is a system structure diagram of an embodiment of the present invention.

Detailed ways

It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

Embodiment 1

As shown in FIG1-2, this embodiment provides a multi-channel speech noise reduction method based on bilinear filtering, including:

Step 1: Collect noisy speech signals;

Step 2, estimating the statistical characteristics of the noisy speech signal and the noise signal;

Step 3, estimating the bilinear Wiener denoising filter;

Step 4: Filter and reduce noise on the noisy speech signal to obtain an estimated value of the clean speech signal.

In speech noise reduction, the time domain signal model is:

y _m (t) = x _m (t) + v _m (t) (1)

Here, t represents a discrete time point, the subscript (·) _m represents the signal received by the m-th microphone (in the present invention, the microphone array is assumed to have a total of M microphones), x _m (t) and v _m (t) represent the clean speech signal and the additive noise signal received by the m-th microphone, respectively, and y _m (t) represents the noisy speech signal received by the m-th microphone. In the present invention, all signals are assumed to be zero-mean, broadband real signals. In the present invention, the first microphone in the microphone array is selected as the reference microphone, that is, x ₁ (t) is selected as the desired signal (the signal to be restored). However, in theory, any microphone can be used as a reference microphone.

By grouping together L consecutive sample points, the signal received by the mth microphone can be written as a vector of length L:

The definitions of x _m (t) and v _m (t) are similar to those of y _m (t), namely:

_{xm (} t)＝[ _xm (t)xm(t-1) … _xm (t-L+1)] ^T

v _m (t) = [v _m (t) v _m (t-1) … v _m (t-L+1)] ^T

x _m (t) and v _m (t) represent the expected signal vector and the noise signal vector of the m th channel respectively, y _m (t) represents the noisy signal vector of the m th channel, and the superscript (·) ^T represents transposition.

In traditional time-domain multi-channel speech enhancement, M noisy signal vectors y _m (t) (m = 1, 2, ..., M) of length L are usually concatenated together and written as:

The definitions of x (t) and v (t) are similar to those of y (t), namely:

y (t), x (t) and v (t) represent the overall noisy signal vector, the overall clean speech signal vector and the overall noise signal vector respectively.

The process of estimating the statistical characteristics of the noisy speech signal and the noise signal in step 2 is: estimating the correlation matrix R _v (t) of the noise signal vector v (t) by an existing noise estimation algorithm; estimating the correlation matrix R _y (t) of the noisy speech signal vector y (t) by a recursive method: R _y (t) = αR _y (t-1) + (1-α) y (t) y ^T (t), where α is a forgetting factor (0 < α <1); estimating the correlation matrix R _x (t) of the clean speech signal vector x (t) by R _x (t) = R _y (t) - R _v (t); based on the speech signal correlation matrix R _x (t), the vector ρ (t) (the element located in the first row and first column of the matrix R _x (t)) can be determined Divide the first column of the matrix R _x (t) by /> That is the vector ρ (t).

The specific method of step 3 is:

In traditional methods, the noisy signal vector y (t) of length ML is usually passed through a linear filter h(t), that is,

Where z(t) is the estimated value of the desired signal x ₁ (t). Therefore, in the traditional method, a filter h (t) with a length of ML needs to be estimated.

In the present invention, in order to construct a bilinear noise reduction scheme, the noisy signal vector is reorganized in the following form:

Among them, the definitions of matrices X(t) and V(t) are similar to those of Y(t), matrix X(t)＝[ _x1 (t) _x2 (t)… _xM (t)]( x (t)＝vec[X(t)]), V(t)＝[ _v1 (t) _v2 (t)… _vM (t)]( v (t)＝vec[V(t)]), the dimensions of matrices Y(t), X(t) and V(t) are all L×M, Y(t) is the noisy signal matrix, X(t) is the clean speech signal matrix, V(t) is the noise signal matrix, and the symbol vec[·] represents the vectorization operation of the matrix.

Based on formula (5), the traditional noise reduction scheme can be modified into the following bilinear noise reduction scheme:

Among them, the two shorter sub-filters h ₁ (t) (length is L) and h ₂ (t) (length is M) play a role in noise reduction in the time domain and spatial domain respectively. z(t) is the estimated value of the desired signal x ₁ (t). is the filtered speech signal, /> represents the residual noise after filtering.

Formula (6) can be transformed as follows:

in, The symbol vec[·] represents a vectorized operation of a matrix, for example, vec[Y(t)] = y (t), the symbol/> It should be noted that the filter h _b (t) here is different from the filter h (t) derived directly by traditional methods.

Using formula (7), formula (6) can be written as

Where x (t)=vec[X(t)], v (t)=vec[V(t)].

Since the speech signal and the noise signal are uncorrelated, the two parts of z(t) are uncorrelated, and the variance of z(t) can be written as

in,

R _y ( t ) = E [ y ( t ) y ^T ( t ) ] , R _x ( t ) = E [ x ( t ) x ^T ( t ) ] , R _v ( t ) = E [ v ( t ) v ^T ( t ) ] ( R _v ( t ) is a full rank matrix).

In the bilinear denoising scheme, the two sub-filters h ₁ (t) and h ₂ (t) have the following relationship:

Among them, _IL and _IM are unit matrices with dimensions of L×L and M×M respectively.

In order to derive the bilinear iterative denoising filter, it is necessary to derive the mean square error of the expected signal estimate z(t). The error of z(t) is defined as

ε(t)＝z(t)-x ₁ (t) (11)

Based on equation (11), the mean square error of z(t) can be defined as

in,

In order to fully utilize the advantages of the bilinear noise reduction scheme, equation (10) can be applied to rewrite equation (12) as

in,

From the above formulas, it can be seen that in the bilinear denoising scheme, the dimensions of the required matrices R _{y ,1} (t) (dimension is M×M) and R _{y ,2} (t) (dimension is L×L) are much smaller than the dimension of the required matrix R _y (t) (dimension is ML×ML) in the traditional denoising scheme. Therefore, fewer observation samples are usually required to estimate the correlation matrices R _{y ,1} (t) and R _{y ,2} (t). Therefore, the bilinear denoising filter can better track the changes in the statistical characteristics of the signal and is more suitable for processing non-stationary noise. In addition, in the process of solving the sub-filters h ₁ (t) and h ₂ (t), the correlation matrices R _{y ,1} (t) and R _{y ,2} (t) need to be inverted. When solving the traditional denoising filter h (t), the correlation matrix R _y (t) needs to be inverted. Since the dimensions of the matrices R _{y ,1} ( t ) and R _{y ,2} ( t ) are much smaller than the dimension of the matrix R _y ( t ), the complexity of solving the sub-filters h ₁ ( t ) and h ₂ ( t ) is much smaller than the complexity of the traditional denoising filter h ( t ).

To derive the bilinear iterative Wiener filter, h ₂ (t) and h ₁ (t) are fixed respectively and equation (13) is written as

Set the initial value of h ₁ (t) to

is the Wiener filter of the first channel, with a length of L. Among them,/> Will/> Substituting into equations (14) and (15), we can obtain

Then, substituting equations (20) and (21) into equation (19), we can obtain

The superscript (·) ⁽ⁿ⁾ represents the result of the nth iteration. Taking the derivative and setting the result to zero, we get

Substituting equation (23) into equations (16) and (17), we can obtain

use and/> Formula (18) can be written as

Reverse equation (26) Taking the derivative and setting the result to zero, we get

According to the above process, after iterating n times, we can get

in,

Based on equations (28) and (29), the bilinear iterative Wiener filter after the Nth iteration can be obtained:

The noisy signal is passed through the designed bilinear Wiener denoising filter h _bW (t) to obtain the denoised speech signal Alternatively, the formula can be used The denoised speech signal z(t) is obtained. The two methods are equivalent.

The process of deriving the bilinear iterative Wiener filter can be briefly described as:

Step 1: According to the formula Initialize sub-filter h ₁ (t), where /> is the autocorrelation matrix of the noisy speech signal vector of the first channel, is the first L rows and first L columns of the matrix R _y (t), vector/> is the vector consisting of the first L elements of the vector ρ (t), and the superscript (·) ⁽ⁿ⁾ indicates the result of the nth iteration;

Step 2: Substitute into the formula and/> In, get and/> The superscript (·) ⁽ⁿ⁾ indicates the result of the nth iteration;

Step 3: and/> Substitute into the formula/> In, get

Step 4: Substitute into the formula and/> In, get/> and

Step 5: and/> Substitute into the formula/> In, get

Repeat steps 2 to 5 N times to obtain and/>

According to the formula Obtain the bilinear Wiener denoising filter h _bW (t).

The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims (8)

1. A multi-channel voice noise reduction method based on bilinear filtering is characterized by comprising the following steps:

Collecting a time domain voice signal with noise, and preprocessing the time domain voice signal with noise;

Estimating the statistical characteristics of the time domain noisy speech signal and the additive noise signal;

Estimating a bilinear dimension accept the enemy's surrender noise filter based on the statistical characteristics;

and filtering and denoising the time domain noisy speech signal based on the bilinear dimension accept the enemy's surrender noise filter to obtain an estimated value of the clean speech signal.

2. The method of claim 1, wherein the method is characterized in that,

The process for collecting the time domain noisy speech signal and preprocessing the time domain noisy speech signal comprises the following steps:

The time domain signal model is:

y_m(t)＝x_m(t)+v_m(t) (1)

Wherein t represents a discrete time point, a subscript (·) _m represents a signal received by an mth microphone, the microphone array is provided with M microphones in total, x _m (t) and v _m (t) respectively represent a clean voice signal and an additive noise signal received by the mth microphone, y _m (t) represents a noisy voice signal received by the mth microphone, and x _m (t) and v _m (t) are mutually uncorrelated; selecting the 1 st microphone in the microphone array as a reference microphone, namely x ₁ (t) as a desired signal;

By combining L consecutive sample points together, the signal received by the mth microphone is written as a vector of length L:

y_m(t)＝[y_m(t) y_m(t-1) … y_m(t-L+1)]^T

＝x_m(t)+v_m(t),m＝1,2,...,M (2)

Wherein x _m (t) and v _m (t) are defined as y _m (t), i.e.:

x_m(t)＝[x_m(t) x_m(t-1) … x_m(t-L+1)]^T

v_m(t)＝[v_m(t) v_m(t-1) … v_m(t-L+1)]^T

x _m (t) and v _m (t) represent the desired signal vector of the mth channel and the noise signal vector of the mth channel, respectively, y _m (t) represents the noisy signal vector of the mth channel, and the superscript (. Cndot.) ^T represents the transpose;

m noisy signal vectors y _m (t) (m=1, 2,) of length L are spliced together, M can be written as:

Wherein x (t) and v (t) are defined similarly to y (t), i.e.:

y (t), x (t), and v (t) represent the overall noisy signal vector, the overall clean speech signal vector, and the overall noise signal vector, respectively.

3. The bilinear filter-based multi-channel speech noise reduction method of claim 2, wherein,

The process of estimating statistical properties of the time domain noisy speech signal and the additive noise signal comprises:

Estimating a correlation matrix R _v (t) of the overall noise signal vector v (t) through an existing noise estimation algorithm, and estimating a correlation matrix R _y(t)：R_y(t)＝αR_y(t-1)+(1-α)y(t)y^T (t) of the overall noise signal vector y (t) through a recursion algorithm, wherein alpha is a forgetting factor (0 < alpha < 1); the correlation matrix R _x (t) of the overall clean speech signal vector x (t) is estimated through R _x(t)＝R_y(t)-R_v (t), the vector ρ (t) is determined based on the speech signal correlation matrix R _x (t), and statistical characteristics are obtained.

4. The bilinear filter-based multi-channel speech noise reduction method of claim 2, wherein,

The process of determining the vector ρ (t) based on the speech signal correlation matrix R _x (t) comprises:

and extracting the elements of the first row and the first column of the voice signal correlation matrix R _x (t) and the elements of the first column, and dividing the elements of the first column by the elements of the first row and the first column to obtain a vector rho (t).

5. The bilinear filter-based multi-channel speech noise reduction method of claim 2, wherein,

The process of estimating the bilinear dimension accept the enemy's surrender noise filter based on the statistical characteristics includes:

Recombining the whole noisy signal vector, splitting the filter in the traditional noise reduction scheme according to the recombined noisy signal matrix Y (t) to obtain a bilinear noise reduction scheme comprising two sub-filters, and carrying out equivalent deformation on the bilinear noise reduction scheme based on vectorization operation of the matrix and Croneck product to obtain the bilinear noise reduction scheme based on the two sub-filters;

Obtaining the definition of the mean square error of the estimated value of the expected signal based on the statistical characteristics of the voice signal with noise, the Kronecker product of the two sub-filters h ₁ (t) and h ₂ (t) and the matrix Y (t) of the signal with noise, and rewriting the definition according to the relation of the two sub-filters to obtain the definition expression of the mean square error of the estimated value of the final expected signal;

a bilinear dimension accept the enemy's surrender noise filter is estimated based on a defined expression of the mean square error of the final desired signal estimate.

6. The method of bilinear filter-based multi-channel speech noise reduction according to claim 5,

The process of obtaining a bilinear noise reduction scheme based on two sub-filters includes:

the overall noisy signal vector is recombined as follows:

Wherein Y (t) =vec [ Y (t) ], clean speech signal matrix X (t) = [ X ₁(t) x₂(t) … x_M (t) ] (X (t) =vec [ X (t) ]), noise signal matrix V (t) = [ V ₁(t) v₂(t) … v_M (t) ] (V (t) =vec [ V (t) ]), symbol vec [ · ] represents the vectorization operation of the matrix, and the dimensions of the matrices Y (t), X (t) and V (t) are all lxm;

based on equation (5), the conventional noise reduction scheme is modified to obtain the following bilinear noise reduction scheme:

Wherein the two sub-filters h ₁ (t) and h ₂ (t) respectively perform noise reduction in the time domain dimension and the space domain dimension, the length of h ₁ (t) is L, the length of h ₂ (t) is M, and z (t) is an estimated value of the desired signal x ₁ (t), In order to filter the speech signal after it has been processed,Representing the filtered residual noise;

deforming the formula (6) to obtain the following formula:

wherein, Symbol vec [. Cndot. ] represents the vectorization operation of the matrix, symbol/>Representing the kronecker product, the symbol tr [ · ] represents the trace of the matrix;

Using formula (7), formula (6) may be written as

Where X (t) =vec [ X (t) ], V (t) =vec [ V (t) ].

7. The method of claim 6, wherein the method is characterized in that,

The process of obtaining a defined expression of the mean square error of the final expected signal estimate value comprises:

the two sub-filters h ₁ (t) and h ₂ (t) have the following relationship:

wherein, I _L and I _M are unit matrixes with dimensions of L×L and M×M respectively;

The mean square error for obtaining the desired signal estimate based on the statistical properties of the noisy speech signal and the statistical properties of the noise signal, the kronecker product of the two sub-filters h ₁ (t) and h ₂ (t), the noisy signal matrix Y (t) is defined as follows:

wherein, Is the element located in the first row and first column of the first row of matrix R _x (t);

Formula (12) is rewritten as follows according to formula (10):

wherein,

8. The method of bilinear filter-based multi-channel speech noise reduction according to claim 7,

The process of estimating the bilinear dimension accept the enemy's surrender noise filter based on the final desired signal estimate mean square error defined expression is as follows:

step one: according to Initializing a sub-filter h ₁ (t), wherein/>The autocorrelation matrix of the first channel noisy speech signal vector is the first L rows and the first L columns of matrix R _y (t), vector/>The superscript (·) ⁽ⁿ⁾ denotes the result of the nth iteration, which is a vector made up of the first L elements of vector ρ (t);

Step two: will be Carry to formulaAnd/>In (1) to obtainAnd/>The superscript (·) ⁽ⁿ⁾ denotes the result of the nth iteration;

Step three: will be And/>Carry to formula/>In, get/>

Step four: will beCarry to formula/>And/>In, get/>And/>

Step five: will beAnd/>Carry to formula/>In, get/>Repeating the steps from the second step to the fifth step for N times to obtain/>And/>

According to the formulaA bilinear dimension accept the enemy's surrender noise filter h _bW (t) is obtained.