Weighted parameterized distributed noise elimination method and system

Abstract

The invention discloses a weighted parameterized distributed noise elimination method and system, and the method comprises the following steps: carrying out the framing and windowing of E-dimensional signals received by J nodes in a wireless microphone network, and carrying out the discrete Fourier transform of each frame of signal, thereby obtaining a discrete spectrum signal; performing data compression processing on the discrete spectrum signal; each node performs weighted summation on the compressed data from all neighbor nodes to obtain a summation signal; enumerating a summation signal to the lower part of the discrete spectrum signal to construct local signal data; performing voice activity detection on the discrete spectrum of each microphone signal based on voice signal characteristics in combination with a logarithmic spectrum distance; updating the constructed correlation matrix of the local signal based on the voice activity detection result; and inverse discrete Fourier transform is carried out on the enhanced voice signal to obtain a time domain output voice signal of the current frame, and overlapping and adding are carried out on the time domain output voice signal of each frame to obtain a final output signal.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed noise elimination, and in particular to a weighted parameterized distributed noise elimination method and system. Background Art

[0002] In general speech enhancement tasks, a single microphone or a microphone array is usually used to complete noise elimination. A single microphone can use time information for speech enhancement, and a microphone array can use not only time information but also spatial information when implementing speech enhancement. However, after the microphone array is manufactured, its topology is not easy to change, which limits the application scenarios of the array. With the development of modern communications, multimedia technology, and artificial intelligence systems, the sound pickup in larger spaces and the extremely high requirements for sound quality in smaller spaces have led to an increasing demand for wireless microphone networks. A wireless microphone network consists of multiple nodes and wireless links between nodes, each of which can contain a single microphone or a microphone array, and has communication capabilities and certain computing capabilities. The network topology composed of wireless links can generally be divided into fully connected, tree-connected, and arbitrary connected topologies.

[0003] With the development of distributed signal processing technology, the wireless microphone network has evolved from the initial one with a fusion center to a wireless microphone network that does not require a fusion center, and has gradually expanded from a restricted network topology to an arbitrary network topology. In the existing distributed signal processing technology, the method that requires a fusion center usually sends the data of each node to the fusion center, which will cause a large computing load to the fusion center, and if the fusion center is damaged, the entire network will stop working. When the wireless microphone network does not have a fusion center, some methods require the connection topology between nodes to be fully connected or tree-type. Although these methods have achieved very good results in simulation experiments, they can only be applied to specific topologies, so they cannot be well applied in practice. This requires the development of more robust distributed speech enhancement methods that can be applied to arbitrary network topologies.

[0004] In the prior art, a distributed noise elimination technology based on generalized eigenvalue decomposition is proposed. This technology uses some prior knowledge of the sound source or array geometry and further expands the distributed adaptive node-specific signal estimation algorithm. Although this technology further expands the existing distributed algorithm based on generalized eigenvalue decomposition and can make the final output result of distributed noise elimination approximately the same as the effect achieved by the data processing center, this technology requires some prior information to be known before it can be implemented.

[0005] In the prior art, a robust distributed multi-channel Wiener filter noise elimination technology is proposed. This technology calculates the signal-to-noise ratio of the signal received by each node, thereby combining the compressed signal sent by its neighbors at the node with the maximum input signal-to-noise ratio to achieve speech enhancement, and finally sends the obtained enhanced signal to all other nodes. Although this technology implements a topology-independent distributed noise elimination algorithm, it must first prune the tree topology of any network topology, and then compare the input signal-to-noise ratio in the tree topology to achieve distributed noise elimination. However, the above operation slows down the running speed of the distributed noise elimination technology.

[0006] Most of the existing distributed speech noise elimination technologies are limited to specific network topologies and achieve certain distributed speech noise elimination effects in these specific networks. Although some technologies can be applied to any network topology, their implementation methods usually assume that some prior information that cannot be obtained in practice is known. These methods cannot perform very well in practical applications. The present invention provides a weighted parameterized distributed noise elimination technical solution without being constrained by network topology and prior information. It is a distributed noise elimination system that allows all nodes to simultaneously complete noise elimination in combination with neighbor node information. Summary of the invention

[0007] In view of the problems existing in the prior art, the present invention discloses a weighted parameterized distributed noise elimination method, comprising the following steps:

[0008] Perform frame-by-frame windowing on the E-dimensional signal received by J nodes in the wireless microphone network, and perform discrete Fourier transform on each frame signal to obtain a discrete spectrum signal;

[0009] Performing data compression processing on the discrete spectrum signal;

[0010] The compressed data of each node is sent to all its neighboring nodes, and each node performs weighted summation on the compressed data from all neighboring nodes to obtain a sum signal;

[0011] List the summed signal below the noisy speech signal vector of the node to construct local signal data;

[0012] Based on the characteristics of the speech signal and the logarithmic spectrum distance, the discrete spectrum of each microphone signal is respectively subjected to speech activity detection, thereby obtaining a speech activity detection result;

[0013] updating the correlation matrix of the constructed local signal based on the voice activity detection result;

[0014] Update the local filter based on the correlation matrix of each node;

[0015] The local signal is de-noised at each node to obtain the enhanced speech signal at each node;

[0016] The enhanced speech signal is subjected to inverse discrete Fourier transform to obtain the time domain output speech signal of the current frame, and the time domain output speech signal of each frame is overlapped and added to obtain the final output signal.

[0017] Furthermore, the compressed data of each node is sent to all its neighboring nodes. At each node, the compressed data from all neighboring nodes are weighted summed, and the coefficient matrix M is defined as:

[0018]

[0019] where n j Represents the degree of a node, that is, the number of neighbors of the node excluding itself. Represents the neighbor set of the node, and the connection matrix L between the nodes is defined as:

[0020]

[0021] The compressed data of all nodes are stacked and represented as a vector:

[0022]

[0023] in[·] T Represents the transpose of a vector or matrix. Based on the above symbolic representation, each node sums the compressed data sent by all its neighboring nodes and the signal is represented as:

[0024]

[0025] Where ⊙ represents the Hadamard product.

[0026] Furthermore, the following expression is used when constructing the local signal data:

[0027]

[0028] In each iterative calculation, the summed signal and the node noisy speech signal vector required to construct the signal will be updated, so the constructed signal changes with the number of iterations.

[0029] Furthermore, the following expression is used to update the local filter:

[0030]

[0031] or

[0032]

[0033] in(·) -1 represents the inverse operation of the matrix, tr{·} represents the trace of the matrix, is a selection vector, in which only the n0th element is 1, and the other elements are all 0. β is a trading parameter, which weighs the relationship between speech distortion and residual noise. That is, the larger β is, the greater the speech distortion is and the smaller the residual noise is. After each node updates the above formula, it is necessary to update the filter corresponding to the noisy signal received by the node:

[0034]

[0035] where α i ∈(0,1] is a decreasing sequence.

[0036] A weighted parameterized distributed noise cancellation system, comprising:

[0037] The discrete Fourier transform module performs frame-by-frame windowing on the E-dimensional signal received by J nodes in the wireless microphone network, and performs discrete Fourier transform on each frame signal to obtain a discrete spectrum signal;

[0038] A data compression module compresses the discrete frequency spectrum of the node signal obtained by the discrete Fourier transform module;

[0039] The data fusion module sends the compressed data of each node to all its neighboring nodes and performs weighted summation of the compressed data of all neighboring nodes of each node;

[0040] A data construction module, which lists the summed signal below the discrete spectrum signal to construct local signal data;

[0041] The voice activity detection module performs voice activity detection on the discrete spectrum of each microphone signal based on the characteristics of the voice signal and the logarithmic spectrum distance, thereby obtaining a voice activity detection result;

[0042] A correlation matrix updating module, which updates the correlation matrix of the constructed local signal based on the voice activity detection result;

[0043] A filter updating module receives the correlation matrix estimate of each node updated by the correlation matrix updating module, and updates the local filter at each node simultaneously;

[0044] A noise elimination module, based on the updated local filter, performs noise elimination on the local signal data at each node to obtain an enhanced speech signal at each node;

[0045] The inverse discrete Fourier transform module receives the enhanced speech signal sent by the noise elimination module, performs inverse discrete Fourier transform on the signal to obtain the time domain output speech signal of the current frame, and overlaps and adds each frame of the time domain output speech signal to obtain the final output signal.

[0046] Due to the adoption of the above technical solution, the present invention provides a weighted parameterized distributed noise elimination method and system. The present invention is a weighted parameterized distributed noise elimination technology that can be applied to any network topology connection. It achieves the construction of local node signals by weighted summing of compressed signals sent by all neighboring nodes at each node, and enables all nodes to update filters at the same time, so that all nodes can obtain node-specific enhanced speech signals in each iteration. This application proposes a weighted parameterized distributed noise elimination system, and defines the algorithm as WP-DANSE-I / II in this application. This technology achieves two enhancement effects and can be applied to any connected network topology. At the same time, each node can obtain different speech enhancement signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 A flow chart of a weighted parameterized distributed noise elimination method of the present invention;

[0049] Figure 2 It is a block diagram of the weighted parameterized distributed noise cancellation principle of the system of the present invention;

[0050] Figure 3 is a schematic diagram of an unconstrained wireless microphone network in the present invention;

[0051] Figure 4 ] is the PESQ value after speech noise elimination in each method under different input signal-to-noise ratios in the embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the technical solutions and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention:

[0053] like Figure 2A weighted parameterized distributed noise elimination system shown includes a discrete Fourier transform (DFT) module 1, a data compression module 2, a data fusion module 3, a data construction module 4, a voice activity detection module 5, a correlation matrix update module 6, a filter update module 7, a noise elimination module 8, and an inverse discrete Fourier transform (IDFT) module 9.

[0054] like Figure 1 As shown, the function of the discrete Fourier transform module 1 is to perform frame windowing processing on the E-dimensional signal received by J nodes in the wireless microphone network, and then perform discrete Fourier transform on each frame signal to obtain a discrete spectrum signal.

[0055] As a preferred method, the working principle of the discrete Fourier transform module 1 is: in the wireless microphone network, there are a total of J linear microphone arrays, and each array j consists of E j microphones, one of which is also called a node. First, the signals y of each node j are j,e (n) Perform frame division and windowing processing respectively, and then perform DFT on each frame signal, where y j,e (n) represents the noisy speech signal of the eth channel received by the jth node. In the verification of this application, the sampling frequency fs of the speech signal is 16kHz, the window function adopts the Hanning window, the frame shift size is 50%, and the length of each frame data is set to M=256 points, where the expression of the Hanning window is as follows:

[0056] ω(m)=0.5-0.5cos(2πm / M), m=0,1,...,M-1 (1)

[0057] By using the Hanning window expression, the windowed signal can be expressed as:

[0058] y' j,e (m) = y j,e (n)ω(m) (2)

[0059] Then, DFT is performed on the windowed signal of the noisy speech received by each microphone, and the discrete spectrum is expressed as:

[0060]

[0061] Wherein, k represents the frequency index, and l represents the frame index.

[0062] The signals Y of each node j,e (k,l) are stacked, and the stacked vector form is expressed as follows:

[0063]

[0064] For the convenience of subsequent representation, the indices k and l will be omitted. In addition, y j =x j +v j , where x j is the speech signal part, v j is the noise signal part.

[0065] Furthermore, the function of the data compression module 2 is to perform data compression on the discrete frequency spectrum of the node signal obtained by the discrete Fourier transform module 1 .

[0066] As a preferred method, the working principle of the data compression module 2 is to compress the discrete spectrum of the node signal obtained by the discrete Fourier transform module 1. Since the subsequent calculation method is iterative calculation, the subsequent variable representation will be represented by the superscript i to represent the iteration index. The data compression process is expressed as follows:

[0067]

[0068] in is the scalar data after compression, is the compression vector and also the local filter of the node, (·) H Indicates the conjugate transpose of a vector or matrix. In addition, the compressed vector needs to be initialized in practical applications. In the present application, the compressed vector elements are initialized to unit interval random numbers that obey uniform distribution during verification.

[0069] The function of the data fusion module 3 is that after the data compression module 2 sends the data compressed by each node to all its neighboring nodes, each node will perform weighted summation of the compressed data from all neighboring nodes.

[0070] As a preferred method, the working principle of the data fusion module 3 is that after the data compression module 2 sends the data compressed by each node to all its neighboring nodes, the data fusion module 3 performs weighted summation of the compressed data from all neighboring nodes at each node. The coefficient matrix M is defined as:

[0071]

[0072] where n j Represents the degree of a node, that is, the number of neighbors of the node excluding itself. Represents the neighbor set of the node. In addition, the connection matrix L between nodes is defined as:

[0073]

[0074] The compressed data of all nodes are stacked and represented as a vector:

[0075]

[0076] in[·] T Represents the transpose of a vector or matrix. Based on the above symbolic representation, each node sums the compressed data sent by all its neighbors and the signal can be expressed as:

[0077]

[0078] Where ⊙ represents the Hadamard product.

[0079] The function of the data construction module 4 is to receive the weighted sum of the compressed signals obtained by the data fusion module 3 at each node, and to construct the local signal of the node in combination with the noisy speech signal obtained by the discrete Fourier transform module 1 at each node.

[0080] As a preferred method, the working principle of the data construction module 4 is to receive the weighted sum of the compressed signals obtained by the data fusion module 3 at each node, and to construct the local signal of the node in combination with the noisy speech signal vector obtained by the discrete Fourier transform module 1 at each node. The specific method is: the summed signal is listed below the noisy speech signal vector of the node, so as to construct the local signal data. The constructed signal is expressed as:

[0081]

[0082] In each iterative calculation, since the signals from the discrete Fourier transform module 1 and the data fusion module 3 will change, the constructed signal will also change with the change of the number of iterations.

[0083] The function of the voice activity detection module 5 is to use the characteristic that most of the first second of speech is a non-speech segment, and combine the logarithmic spectrum distance to detect the voice activity of each discrete spectrum of the microphone signal obtained by the discrete Fourier transform module 1, so as to obtain the voice activity detection result.

[0084] As a preferred method, the working principle of the voice activity detection module 5 is to detect the voice activity of each discrete spectrum of the microphone signal obtained by the discrete Fourier transform module 1. Under normal circumstances, the first second of the audio signal received by the microphone does not contain voice. Using this feature and combining the processing process of frame division and windowing, the number of the initial non-voice frames of the microphone received signal is set as NIS frames, where NIS = fs / (50%×M)-1 = 124. Then, the noise average spectrum estimated by this NIS frame is:

[0085]

[0086] The specific operation represented by formula (11) is to sum and average all the frequency points corresponding to each frame signal, and then get the estimate of the noise logarithmic spectrum as follows:

[0087]

[0088] Where |·| represents a modulo operation. Then, the logarithmic spectrum of each frame signal is calculated:

[0089]

[0090] From equations (12) and (13), we can get the logarithmic spectrum distance between each frame signal and the noise signal. The logarithmic spectrum distance formula is as follows:

[0091]

[0092] In summary, the judgment method of voice activity detection can be obtained: first, set a non-speech segment counter, the initial value of which can be set to 125, and set a log spectrum distance threshold to 3. Then calculate the log spectrum distance d between each frame signal and the noise frame spec , judge d spec Is it less than the log spectrum distance threshold? If so, the frame is a speechless frame, and the speechless segment counter is incremented by 1. If not, the frame is a speechless frame, and the speechless segment counter must be reset to zero regardless of its value. Finally, it should be noted that if the value of the speechless segment counter before it is reset to zero is less than the minimum speechless length, then all the speechless frames after the speechless segment counter was reset to zero last time and before it is reset to zero this time are considered speechless frames. Here, the minimum speechless length is set to 10.

[0093] In order to reduce speech distortion during verification, this application considers each microphone signal to be a noise frame only when the speech activity detection result is a noise frame. Otherwise, it is considered to be a speech frame.

[0094] The function of the correlation matrix updating module 6 is to update the correlation matrix of the local signal constructed by the data constructing module 4 according to the voice activity detection result obtained by the voice activity detection module 5 .

[0095] As a preferred method, the working principle of the correlation matrix updating module 6 is to update the correlation matrix of the local signal constructed by the data construction module 4 according to the voice activity detection result obtained by the voice activity detection module 5. When the detection result shows that it is a voice frame, the noisy voice signal correlation matrix of each node is updated as follows:

[0096]

[0097] Where α is the forgetting factor used in the above recursive estimation, which is set to 0.997 in the verification of this application.j =x j +v j The local signal constructed by the data construction module 4 can be expressed in a similar way as:

[0098]

[0099] When the detection result of the voice activity detection module 5 shows that it is a noise frame, the noise signal correlation matrix of each node is updated as follows:

[0100]

[0101] Usually, the speech signal part and the noise signal part are uncorrelated, so the correlation matrix estimation method of the speech signal part is:

[0102]

[0103] The function of the filter updating module 7 is to receive the correlation matrix estimate of each node updated by the correlation matrix updating module 6, and to update the local filter at each node simultaneously.

[0104] As a preferred mode, the working principle of the filter updating module 7 is to receive the correlation matrix estimate of each node updated by the correlation matrix updating module 6, and update the local filter at each node simultaneously:

[0105]

[0106] or:

[0107]

[0108] in(·) -1 represents the inverse operation of the matrix, tr{·} represents the trace of the matrix, is a selection vector, in which only the n0th element is 1, and the other elements are all 0. β is a trading parameter, and its size can weigh the relationship between speech distortion and residual noise, that is, the larger β is, the greater the speech distortion is, the smaller the residual noise is, and vice versa. This application sets β to 20 during verification. After each node updates formula (19) or formula (20), it needs to update the filter corresponding to the noisy signal received by the node:

[0109]

[0110] where α i ∈(0,1] is a decreasing sequence. In this application, when verifying, let α i =1 / i.

[0111] The function of the noise elimination module 8 is to use the filter obtained by the filter updating module 7 to perform noise elimination on the local signal constructed by the data construction module 4 at each node, so as to obtain the enhanced speech signal of each node.

[0112] As a preferred method, the working principle of the noise elimination module 8 is to use the filter obtained by the filter update module 7 to eliminate the noise of the local signal constructed by the data construction module 4 at each node, and the enhanced speech signal of each node is expressed as:

[0113]

[0114] The function of the inverse discrete Fourier transform module 9 is to receive the enhanced speech signal sent by the noise elimination module 8, perform inverse discrete Fourier transform on the signal to obtain the time domain output speech signal of the current frame, and overlap and add each frame of the time domain output speech signal to obtain the final output signal.

[0115] As a preferred method, the working principle of the discrete Fourier inverse transform module 9 is to transform the enhanced speech signal obtained by the noise elimination module 8 at each node Perform IDFT to obtain the time domain enhanced speech signal output by the current frame. The IDFT formula is as follows:

[0116]

[0117] Here, i and l represent the same meaning, that is, the i-th iteration is equivalent to the l-th frame, so when they appear at the same time below, the iteration index i is omitted.

[0118] Because the present application performs frame windowing processing on each signal in module 1, and the frame shift is 50%, when the first frame output voice signal is obtained, an overlap-add operation is performed with the second frame output voice signal, and the overlap portion accounts for 50%. The specific formula is as follows:

[0119]

[0120] Where [·] is a rounding operation, and [a] represents the largest integer not exceeding the number a.

[0121] The present invention discloses a weighted parameterized distributed noise cancellation system. To verify the effectiveness of the method of the present invention, the present application uses the Imgae model to simulate a room of 8m×8m×3m in size. Eight microphone arrays are randomly distributed in the room, each of which contains 5 microphones and the microphone spacing is 2 cm. By gradually increasing the communication distance between the arrays until any array can reach all other arrays through a collection of some links, the wireless connection topology between the arrays is completed. The network formed by this method is called an unconstrained wireless microphone network. Figure 3 The schematic diagram of the network is shown, which gives the coordinates of the speaker and the connection topology between the arrays. In addition, the height of the speaker is set to 1.7 meters, the height of all arrays is set to 1 meter, and the speed of sound propagation in the room is set to 340m / s. The speech signal used by the speaker comes from the TIMIT database [ https: / / download.csdn.net / download / sdhyfxh / 4086482 ] is a pure speech signal randomly selected from it, with a time length of 6 seconds. The background noise is white noise, and the sampling frequency of speech and noise is 16kHz.

[0122] At this time, the weighted parameterized distributed adaptive node-specific signal estimation (WP-DANSE-I / II) method proposed in this application is used to eliminate the noise of the array noisy speech signal in the entire wireless microphone network, where the WP-DANSE-I method corresponds to equation (19) and the WP-DANSE-II method corresponds to equation (20). At the same time, the methods in references [1] and [2] are used to eliminate speech noise in this experiment. Figure 4 A performance comparison of perceptual evaluation of speech quality (PESQ) for different input signal-to-noise ratios under white noise is given, and the results show that both methods proposed in this application have achieved good performance, especially the WP-DANSE-II method, which has achieved performance improvement compared to noisy speech signals regardless of the input signal-to-noise ratio.

[0123] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A weighted parameterized distributed noise elimination method, characterized in that include: Perform frame-by-frame windowing on the E-dimensional signal received by J nodes in the wireless microphone network, and perform discrete Fourier transform on each frame signal to obtain a discrete spectrum signal; Performing data compression processing on the discrete spectrum signal; The compressed data of each node is sent to all its neighboring nodes, and each node performs weighted summation on the compressed data from all neighboring nodes to obtain a sum signal; List the summed signal below the noisy speech signal vector of the node to construct local signal data; Based on the characteristics of the speech signal and the logarithmic spectrum distance, the discrete spectrum of each microphone signal is respectively subjected to speech activity detection, thereby obtaining a speech activity detection result; updating the correlation matrix of the constructed local signal based on the voice activity detection result; Update the local filter based on the correlation matrix of each node; The local signal is de-noised at each node to obtain the enhanced speech signal at each node; The enhanced speech signal is subjected to inverse discrete Fourier transform to obtain the time domain output speech signal of the current frame, and the time domain output speech signal of each frame is overlapped and added to obtain the final output signal.

2. The weighted parameterized distributed noise elimination method according to claim 1, characterized in that: The compressed data of each node is sent to all its neighboring nodes. At each node, the compressed data from all neighboring nodes are weighted summed. The coefficient matrix M of the wireless microphone network is defined as: Where n j Represents the degree of a node, that is, the number of neighbors of the node excluding itself. Represents the neighbor set of the node, and the connection matrix L between the nodes is defined as: The compressed data of all nodes are stacked and represented as a vector: in[·] T Represents the transpose of a vector or matrix. Based on the above symbolic representation, each node sums the compressed data sent by all its neighboring nodes and the signal is represented as: Where ⊙ represents the Hadamard product.

3. The weighted parameterized distributed noise elimination method according to claim 1, characterized in that: The following expression is used to construct local signal data: In each iterative calculation, the summed signal and the node noisy speech signal vector required to construct the signal will be updated, so the constructed signal changes with the number of iterations.

4. The weighted parameterized distributed noise elimination method according to claim 1, characterized in that: The following expression is used to update the local filter: or in(·) -1 represents the inverse operation of the matrix, tr{·} represents the trace of the matrix, is a selection vector, in which only the n0th element is 1, and the other elements are all 0. β is a trading parameter, which weighs the relationship between speech distortion and residual noise. That is, the larger β is, the greater the speech distortion is and the smaller the residual noise is. After each node updates the above formula, it is necessary to update the filter corresponding to the noisy signal received by the node: where α i ∈(0,1] is a decreasing sequence.

5. A weighted parameterized distributed noise cancellation system, characterized in that include: The discrete Fourier transform module performs frame-by-frame windowing on the E-dimensional signal received by J nodes in the wireless microphone network, and performs discrete Fourier transform on each frame signal to obtain a discrete spectrum signal; A data compression module compresses the discrete frequency spectrum of the node signal obtained by the discrete Fourier transform module; The data fusion module sends the compressed data of each node to all its neighboring nodes and performs weighted summation of the compressed data of all neighboring nodes of each node; A data construction module, which lists the summed signal below the discrete spectrum signal to construct local signal data; The voice activity detection module performs voice activity detection on the discrete spectrum of each microphone signal based on the characteristics of the voice signal and the logarithmic spectrum distance, thereby obtaining a voice activity detection result; A correlation matrix updating module, which updates the correlation matrix of the constructed local signal based on the voice activity detection result; A filter updating module receives the correlation matrix estimate of each node updated by the correlation matrix updating module, and updates the local filter at each node simultaneously; A noise elimination module, based on the updated local filter, performs noise elimination on the local signal data at each node to obtain an enhanced speech signal at each node; The inverse discrete Fourier transform module receives the enhanced speech signal sent by the noise elimination module, performs inverse discrete Fourier transform on the signal to obtain the time domain output speech signal of the current frame, and overlaps and adds each frame of the time domain output speech signal to obtain the final output signal.

Images