Real-time adaptive beam forming method and device based on maximum SNR (Signal to Noise Ratio)

Abstract

The invention provides a real-time self-adaptive beam forming method and device based on the maximum SNR, and the method comprises the following steps: carrying out the windowing preprocessing of time domain audio signals collected by M microphones, and carrying out the FFT conversion, and obtaining an M * F-dimensional frequency domain signal matrix X; updating an M * M-dimensional covariance matrix Rnow (f) of each frequency point by adopting an exponential weighted moving average mechanism based on the frequency domain signal matrix X; the inverse matrix Rnow-1 (f) of the covariance matrix Rnow (f) is subjected to recursive updating by adopting a Sherman-Morrison-Woodbury matrix inversion lemma, and the inverse matrix Rnow-1 (f) of the covariance matrix Rnow (f) is subjected to recursive updating by adopting the Sherman-Morrison-Woodbury matrix inversion lemma; calculating the SNR of each microphone channel in parallel, screening an optimal reference microphone and extracting a corresponding beam forming weight, and for each frequency point f, performing beam forming processing on the frequency domain signal X (f) by applying the weight to obtain an output frequency spectrum Y (f); and constructing a gain compensation mechanism based on the optimal reference microphone, performing frequency-point-by-frequency gain compensation on the output spectrum Y (f), performing IFFT (Inverse Fast Fourier Transform) to restore the output spectrum Y (f) into a time domain signal, and outputting enhanced voice. Therefore, the technical support of current voice interaction is adapted.

Description

Technical Field

[0001] This invention relates to the fields of voice technology, artificial intelligence and quantum computing, and in particular to a real-time adaptive beamforming method and apparatus based on maximum SNR. Background Technology

[0002] With the rapid development of smart terminals, remote conferencing, autonomous driving, and other fields, the demand for reliable voice interaction is increasing. Voice signal processing technology, as the core support for voice interaction, directly determines the interactive experience. Among these technologies, beamforming has become a research hotspot due to its ability to enhance voice through spatial filtering.

[0003] Therefore, how to develop a real-time adaptive beamforming method and device based on maximum SNR to adapt to the current technical support for voice interaction has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The present invention aims to solve these problems and provide a real-time adaptive beamforming method and apparatus based on maximum SNR to meet the technical support of current voice interaction.

[0005] This invention provides a real-time adaptive beamforming method based on maximum SNR, comprising the following steps:

[0006] After windowing preprocessing of the time-domain audio signals acquired by M microphones, an FFT transformation is performed to obtain an M×F dimensional frequency domain signal matrix X.

[0007] Based on the frequency domain signal matrix X, the M×M covariance matrix R at each frequency point is updated using an exponentially weighted moving average mechanism. now (f);

[0008] The covariance matrix R is obtained by applying the Sherman-Morrison-Woodbury matrix inversion lemma. now The inverse matrix R of (f) now -1 (f) Perform recursive updates;

[0009] Parallel calculation of the SNR of each microphone channel, selection of the optimal reference microphone, and extraction of the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f);

[0010] A gain compensation mechanism is constructed based on the optimal reference microphone. After frequency-by-frequency gain compensation is performed on the output spectrum Y(f), an IFFT transform is performed to restore it to the time domain signal, and the enhanced speech is output.

[0011] in,

[0012] The steps described involve performing windowing preprocessing on the time-domain audio signals acquired by the M microphones, followed by FFT transformation to obtain an M×F dimensional frequency-domain signal matrix X; which includes...

[0013] The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; wherein, the sampling rate of the time-domain signal is 16kHz, the frame length is 10ms, M is the number of microphones in the microphone array, which is 8, and F is the total number of frequency points;

[0014] The aforementioned step, based on the frequency domain signal matrix X, uses an exponentially weighted moving average mechanism to update the M×M dimension covariance matrix R at each frequency point. now (f); includes

[0015] The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is:

[0016] ,

[0017] in, Let X(f) be the covariance matrix of the frequency point f in the previous frame, and let X(f) be the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.95.

[0018] in,

[0019] The aforementioned step employs the Sherman-Morrison-Woodbury matrix inversion lemma to apply the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; including...

[0020] During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is:

[0021] R now -1 (f)=inv(R now (f))

[0022] in, This represents the matrix inversion operation;

[0023] The formula for calculating the scalar value k is:

[0024] in, α is the inverse of the covariance matrix of the frequency point f corresponding to the previous frame, and α is the forgetting factor defined in the above steps;

[0025] Current frame inverse matrix R now -1 The formula for calculating (f) is:

[0026]

[0027] This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

[0028] in,

[0029] The steps described above involve parallel calculation of the SNR of each microphone channel, selection of the optimal reference microphone, and extraction of the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f); which includes,

[0030] The specific process for selecting the optimal beamformer for the SNR includes:

[0031] For the m-th microphone, construct an M×1 dimensional steering vector pointing to that microphone. The m-th element is 1, and the rest are 0, which corresponds to the beamformer weights. The calculation formula is: Where m ranges from 1 to M, Guide vector The conjugate transpose of;

[0032] The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows:

[0033] in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power.

[0034] Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight for the current frame. The output spectrum is obtained by applying this weight. .

[0035] in,

[0036] The steps described include: constructing a gain compensation mechanism based on the optimal reference microphone; performing frequency-point gain compensation on the output spectrum Y(f); performing an IFFT transform to restore the time-domain signal; and outputting enhanced speech.

[0037] Construct the reference microphone selection vector s ref The vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions;

[0038] The formula for calculating the pointwise gain factor g(f) in the frequency domain is:

[0039] in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone;

[0040] The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is:

[0041] =Y(f) / max(g(f),ε)

[0042] Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero;

[0043] right Perform an IFFT transform to convert the frequency domain signal back to the time domain, outputting a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

[0044] According to another aspect of the present invention, the present invention also provides a real-time adaptive beamforming apparatus based on maximum SNR, comprising:

[0045] The preprocessing unit is configured to perform windowing preprocessing on the time-domain audio signals collected by the M microphones, and then perform FFT transformation to obtain an M×F-dimensional frequency domain signal matrix X.

[0046] The moving average processing unit is configured to update the M×M covariance matrix R of each frequency point using an exponentially weighted moving average mechanism based on the frequency domain signal matrix X. now (f);

[0047] The recursive update unit is configured to apply the Sherman-Morrison-Woodbury matrix inversion lemma to the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates;

[0048] The beamforming processing unit is configured to calculate the SNR of each microphone channel in parallel, select the optimal reference microphone, and extract the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f);

[0049] The output unit is configured to construct a gain compensation mechanism based on the optimal reference microphone, perform frequency-by-frequency gain compensation on the output spectrum Y(f), perform IFFT transformation to restore it to the time domain signal, and output enhanced speech.

[0050] in,

[0051] The preprocessing unit includes

[0052] The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; wherein, the sampling rate of the time-domain signal is 16kHz, the frame length is 10ms, M is the number of microphones in the microphone array, which is 8, and F is the total number of frequency points;

[0053] The moving average processing unit includes

[0054] The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is:

[0055]

[0056] in, Let X(f) be the covariance matrix of the frequency point f in the previous frame, and let X(f) be the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.95.

[0057] in,

[0058] The recursive update unit includes

[0059] During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is:

[0060] R now -1 (f)=inv(R now (f))

[0061] in, This represents the matrix inversion operation;

[0062] The formula for calculating the scalar value k is:

[0063] in, α is the inverse of the covariance matrix of the frequency point f corresponding to the previous frame, and α is the forgetting factor defined in the above steps;

[0064] Current frame inverse matrix R now -1 The formula for calculating (f) is:

[0065]

[0066] This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

[0067] in,

[0068] The beamforming processing unit includes...

[0069] The specific process for selecting the optimal beamformer for the SNR includes:

[0070] For the m-th microphone, construct an M×1 dimensional steering vector pointing to that microphone. The m-th element is 1, and the rest are 0, which corresponds to the beamformer weights. The calculation formula is:

[0071] Where m ranges from 1 to M, Guide vector The conjugate transpose of;

[0072] The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows:

[0073]

[0074] in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power.

[0075] Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight for the current frame. The output spectrum is obtained by applying this weight. .

[0076] in,

[0077] The output unit includes,

[0078] Construct the reference microphone selection vector s refThe vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions;

[0079] The formula for calculating the pointwise gain factor g(f) in the frequency domain is:

[0080] in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone;

[0081] The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is:

[0082] =Y(f) / max(g(f),ε)

[0083] Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero;

[0084] right Perform an IFFT transform to convert the frequency domain signal back to the time domain, outputting a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

[0085] This invention relates to a real-time adaptive beamforming method based on maximum SNR, comprising: performing windowing preprocessing on time-domain audio signals acquired by M microphones, followed by FFT transformation to obtain an M×F dimensional frequency-domain signal matrix X; and updating the M×M dimensional covariance matrix R at each frequency point based on the frequency-domain signal matrix X using an exponentially weighted moving average mechanism. now (f); The covariance matrix R is obtained by applying the Sherman-Morrison-Woodbury matrix inversion lemma. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; calculate the SNR of each microphone channel in parallel, select the optimal reference microphone, and extract the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f); a gain compensation mechanism is constructed based on the optimal reference microphone, and after frequency-by-frequency gain compensation is performed on the output spectrum Y(f), an IFFT transform is performed to restore it to the time domain signal, and the enhanced speech is output.

[0086] This method effectively addresses the uncertainty of sound source location, thereby improving the robustness of speech enhancement. The technology acquires speech signals through a microphone array, converts the time-domain speech signal to the frequency domain for processing, and optimizes the beamformer weights by dynamically updating the covariance and inverse matrices of the speech signal spectrum. Without requiring prior DOA information, the optimal microphone channel can be selected based on SNR evaluation. Gain compensation is then applied to the spectrum using the optimal beamformer, ultimately converting the frequency-domain speech to the time-domain speech, resulting in high-fidelity enhanced speech output.

[0087] The above description is merely an overview of this solution. In order to better understand the technical means of this invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this invention more apparent and understandable, specific embodiments of this invention are described below. Attached Figure Description

[0088] To more clearly illustrate this technical solution, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this technology. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0089] Figure 1 This is a schematic flowchart illustrating an embodiment of the real-time adaptive beamforming method based on maximum SNR according to the present invention.

[0090] Figure 2 This invention schematically illustrates a flowchart of Embodiment 2 of a real-time adaptive beamforming method based on maximum SNR;

[0091] Figure 3 This invention schematically illustrates a structural block diagram of a third embodiment of a real-time adaptive beamforming device based on maximum SNR;

[0092] Figure 4 This invention schematically illustrates a structural block diagram of a fourth embodiment of a real-time adaptive beamforming device based on maximum SNR. Detailed Implementation

[0093] The embodiments of the present invention will be described in detail below, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0094] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0095] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to that other component.

[0096] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0097] The technical terms involved in this invention are defined as follows:

[0098] Beamforming: By weighting and synthesizing the multi-channel signals received by the microphone array, a beam pointing towards a specific sound source is formed, while suppressing interference noise outside the beam, thus achieving spatial filtering.

[0099] Covariance matrix: A matrix that describes the correlation between multi-channel signals. Its elements are the cross-correlation coefficients of different channel signals and are the core basis for beamformer weight calculation.

[0100] SNR (Signal-to-Noise Ratio): The ratio of speech signal power to noise signal power. It is a core indicator for measuring speech quality. The higher the value, the less noise interference the speech is affected by.

[0101] DOA (Direction of Arrival): Spatial angle information of the sound source signal arriving at the microphone array. Traditional adaptive beamforming technology needs to use this to determine the beam direction.

[0102] Sherman-Morrison-Woodbury matrix inversion lemma: an efficient method for inverting rank-corrected matrices, which transforms high-complexity matrix inversion operations into low-order operations, suitable for real-time signal processing scenarios.

[0103] This invention belongs to the field of speech signal processing technology, specifically relating to a real-time adaptive beamforming method based on maximum SNR. This method can effectively address the uncertainty of sound source location, thereby improving the robustness of speech enhancement. The technology acquires speech signals through a microphone array, converts the time-domain speech signal to the frequency domain for processing, and optimizes the beamformer weights by dynamically updating the covariance matrix and inverse matrix of the speech signal spectrum. It does not require prior sound source DOA information; the optimal microphone channel can be selected based on SNR evaluation. Gain compensation is applied to the spectrum using the optimal beamformer, and finally, the frequency-domain speech is converted back to the time-domain speech, resulting in high-fidelity enhanced speech output.

[0104] Currently, mainstream beamforming technologies are mainly divided into two categories, and the specific technical solutions and sources are as follows:

[0105] Fixed beamforming technology, with DS (Delay Summation) beamforming as its core, involves pre-setting the geometric parameters of the microphone array (such as array shape and element spacing), calculating the delay compensation for each microphone channel based on the preset direction of the target sound source, and directly superimposing the signals from each channel after delay alignment. The core feature of this technology is that the weighting coefficients are fixed and do not adjust with environmental changes. (Source: *Microphone Array Signal Processing* (National Defense Industry Press, 2015, Section 3.3); Patent CN102680946A (Ultra-wideband beamformer and implementation method based on differential operators))

[0106] Adaptive beamforming technology, with MVDR (Minimum Variance Distortionless Response) beamforming as a typical example, employs the following technical solution: First, it acquires real-time directional information of the sound source through a DOA estimation module (such as the MUSIC algorithm or ESPRIT algorithm), and constructs a steering vector based on this information. Then, it acquires multi-channel signals in real time and calculates the signal covariance matrix. Finally, it calculates the beamformer weights based on the steering vector and the covariance matrix to achieve speech enhancement of the target sound source. The core feature of this technology is that the weight coefficients are dynamically updated with the DOA information. Sources: Patent CN117857976A (Noise Reduction Method, Terminal, and Storage Medium Based on MVDR Beamforming); Patent CN119943078A (An MVDR Beamforming Method and Apparatus).

[0107] Beamforming, a core method for speech enhancement, essentially achieves spatial filtering by weighting the multi-channel signals received by a microphone array. However, current mainstream beamforming technologies all have significant limitations in practical engineering applications, as detailed below:

[0108] Fixed beamforming technology directs the beam in a fixed direction by pre-setting weighting coefficients for the microphone array. Although it has a simple structure and low computational complexity (approximately...), it is effective in this field. , (The number of microphones) is limited, but it cannot adapt to dynamic noise environments. When interference signals are incident from the main lobe of the beam, the suppression effect drops sharply; in scenarios such as multi-person conferences and mobile terminals where the location of the sound source changes dynamically, the fixed beam is difficult to track in real time, resulting in excessive fluctuations in the amplitude of the voice signal, which seriously affects the coherence of the speech.

[0109] Adaptive beamforming technologies such as MVDR improve environmental adaptability by updating weights in real time, but still have many shortcomings: First, they suffer from poor computational stability. The calculation of the inverse covariance matrix often uses a direct inversion method, which can easily lead to singular values ​​in the matrix when there are sudden changes in the signal (such as sudden noise), causing fluctuations in the weight calculation results. Second, beam pointing depends on prior DOA information. When the DOA information has errors or even large errors, the main lobe of the beam cannot accurately point to the sound source, resulting in a decrease in SNR. Third, there is a lack of a quantitative evaluation mechanism for beamformer performance, making it impossible to adaptively select the optimal filtering parameters in complex sound fields. Fourth, some technologies neglect the gain compensation stage, which can easily lead to distortion of the amplitude of the speech signal after beamforming, affecting the auditory experience.

[0110] like Figure 1 As shown in Embodiment 1 of this patent, a real-time adaptive beamforming method based on maximum SNR includes the following steps:

[0111] After windowing preprocessing of the time-domain audio signals acquired by M microphones, an FFT transformation is performed to obtain an M×F dimensional frequency domain signal matrix X.

[0112] Based on the frequency domain signal matrix X, the M×M covariance matrix R at each frequency point is updated using an exponentially weighted moving average mechanism. now (f);

[0113] The covariance matrix R is obtained by applying the Sherman-Morrison-Woodbury matrix inversion lemma. now The inverse matrix R of (f) now -1 (f) Perform recursive updates;

[0114] Parallel calculation of the SNR of each microphone channel, selection of the optimal reference microphone, and extraction of the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f);

[0115] A gain compensation mechanism is constructed based on the optimal reference microphone. After frequency-by-frequency gain compensation is performed on the output spectrum Y(f), an IFFT transform is performed to restore it to the time domain signal, and the enhanced speech is output.

[0116] in,

[0117] The steps described involve performing windowing preprocessing on the time-domain audio signals acquired by the M microphones, followed by FFT transformation to obtain an M×F dimensional frequency-domain signal matrix X; which includes...

[0118] The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; wherein, the sampling rate of the time-domain signal is 16kHz, the frame length is 10ms, M is the number of microphones in the microphone array, which is 8, and F is the total number of frequency points;

[0119] The aforementioned step, based on the frequency domain signal matrix X, uses an exponentially weighted moving average mechanism to update the M×M dimension covariance matrix R at each frequency point. now (f); includes

[0120] The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is:

[0121]

[0122] in, Let X(f) be the covariance matrix of the frequency point f in the previous frame, and let X(f) be the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.95.

[0123] in,

[0124] The aforementioned step employs the Sherman-Morrison-Woodbury matrix inversion lemma to apply the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; including...

[0125] During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is:

[0126] R now -1 (f)=inv(R now (f))

[0127] in, This represents the matrix inversion operation;

[0128] The formula for calculating the scalar value k is:

[0129] in, α is the inverse of the covariance matrix of the frequency point f corresponding to the previous frame, and α is the forgetting factor defined in the above steps;

[0130] Current frame inverse matrix R now -1 The formula for calculating (f) is:

[0131]

[0132] This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

[0133] in,

[0134] The steps described above involve parallel calculation of the SNR of each microphone channel, selection of the optimal reference microphone, and extraction of the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f); which includes,

[0135] The specific process for selecting the optimal beamformer for the SNR includes:

[0136] For the m-th microphone, construct an M×1 dimensional steering vector pointing to that microphone. The m-th element is 1, and the rest are 0, which corresponds to the beamformer weights. The calculation formula is:

[0137] Where m ranges from 1 to M, Guide vector The conjugate transpose of;

[0138] The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows:

[0139]

[0140] in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power.

[0141] Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight for the current frame. The output spectrum is obtained by applying this weight. .

[0142] in,

[0143] The steps described include: constructing a gain compensation mechanism based on the optimal reference microphone; performing frequency-point gain compensation on the output spectrum Y(f); performing an IFFT transform to restore the time-domain signal; and outputting enhanced speech.

[0144] Construct the reference microphone selection vector s ref The vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions;

[0145] The formula for calculating the pointwise gain factor g(f) in the frequency domain is:

[0146] in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone;

[0147] The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is:

[0148] =Y(f) / max(g(f),ε)

[0149] Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero;

[0150] right Perform an IFFT transform to convert the frequency domain signal back to the time domain, outputting a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

[0151] To illustrate the specific implementation process of this embodiment one in detail, the following example uses an 8-channel microphone array embedded terminal real-time voice enhancement embodiment.

[0152] This embodiment is applied to an embedded remote conferencing terminal. The terminal is equipped with an 8-channel equidistant linear microphone array with an array element spacing of 5cm. The core processor is an ARM Cortex-A7 with a main frequency of 1GHz and an FPU floating-point arithmetic unit. It needs to achieve real-time voice enhancement in multi-person dynamic speaking scenarios, with a single frame processing time of ≤10ms, an output voice SNR of ≥15dB, and no obvious amplitude distortion.

[0153] The microphone array has 8 microphones (M=8), a sampling rate of 16kHz, a speech duration of 10ms per frame, and 160 sampling points per frame.

[0154] The Hanning window has a length of 160 points, the FFT transform has 256 points (F=256 frequency points), the forgetting factor α=0.9, and the minimum value ε= ;

[0155] A statistical method for speech still segments is used, and silent segments are identified through a VAD module. The initial noise covariance matrix is ​​calculated by accumulating 50 frames of silent segment signals. .

[0156] The specific process includes: the terminal's eight microphones simultaneously acquire time-domain speech signals, with each signal being a discrete time-domain sequence of 160 points. Hanning windowing is applied to each time-domain signal to suppress spectral leakage; subsequently, a 256-point FFT transform is performed on the windowed signal to convert each time-domain signal into a frequency-domain signal with 256 frequency points, ultimately yielding an 8×256-dimensional frequency-domain signal matrix X.

[0157] For each frequency point f (f=0,1,...,255), based on the 8×1 dimensional signal vector X(f) of the corresponding frequency point in the frequency domain signal matrix X, the covariance matrix is ​​updated according to the following formula:

[0158] in, Let f be the covariance matrix of the frequency point f in the previous frame. Let X(f) be the conjugate transpose of X(f). This update mechanism preserves the statistical characteristics of historical signals while also responding quickly to changes in sound source location and environmental noise fluctuations in a conference setting.

[0159] The inverse matrix is ​​recursively updated using the Sherman-Morrison-Woodbury matrix inversion lemma. The specific process is as follows:

[0160] During the processing of the first frame, the inverse matrix R is obtained by direct inversion. now -1 (f)=inv(R now (f));

[0161] Calculate scalar values ,in This is the inverse matrix of the previous frame;

[0162] Inverse matrix update: .

[0163] This method reduces the complexity of matrix inversion from O(8³)=512 to O(8²)=64, significantly improving computational efficiency.

[0164] Construct guide vectors for each of the eight microphones. (M×1 dimensional, the m-th element is 1, and the rest are 0), calculate the beamforming weights for each microphone. ;

[0165] SNR is calculated using only 64 frequency points (f=0~63) corresponding to 0~2000Hz, and the formula is as follows: The numerator is the sum of the speech signal power and the denominator is the sum of the noise signal power;

[0166] Compare the SNR values ​​of the 8 microphones and select the weight vector corresponding to the maximum value as the optimal weight. The output spectrum is obtained by applying this weight. .

[0167] Construct the reference microphone selection vector s ref (8×1 dimension, the optimal microphone index position is 1, and the rest are 0);

[0168] Calculate the gain factor This reflects the gain characteristics of the beamformer for the optimal microphone signal; This avoids calculation errors caused by a denominator of zero. right A 256-point IFFT transform is performed to convert the signal to the time domain. Inter-frame distortion is eliminated by overlapping and adding, and finally, one channel of enhanced speech is output.

[0169] In a meeting scenario with three people speaking dynamically, the terminal has an average processing time of 4.2ms per frame, which meets the real-time requirements. The output voice SNR is 12~18dB higher than the original signal, the voice amplitude fluctuation is ≤3dB, and the clarity and naturalness of the sound are significantly better than traditional MVDR beamforming technology, with no obvious distortion or stuttering.

[0170] The above example illustrates real-time voice enhancement in an embedded terminal with an 8-channel microphone array. Real-time voice enhancement in an embedded terminal with a 4-channel microphone array can also be achieved, and the specific implementation can be found in the above example. It will not be repeated here.

[0171] like Figure 2 As shown, this invention also provides a real-time adaptive beamforming method for maximizing SNR in next-generation quantum computing. This method is theoretically achievable. If planned for release, it can be protected through priority rights; if not planned for release, it can be disclosed to prevent others from applying for it. Detailed technical implementation methods are as follows.

[0172] Includes the following steps: Step 1: Time-domain to frequency-domain signal quantization conversion; specifically, after preprocessing the time-domain audio signals acquired by M microphones using Hanning windowing, the traditional FFT transform is replaced by quantum Fourier transform (QFT) to convert each time-domain signal into a quantum state frequency-domain signal with F frequency points, constructing an M×F dimensional quantum frequency-domain signal matrix. The number of qubits is configured according to the number of frequency points F, satisfying 2^q≥F (q is the number of qubits). Step two, dynamic update of the quantum state speech covariance matrix; specifically, based on the quantum frequency domain signal matrix. The quantum superposition property is used to calculate the M×M dimension covariance matrix quantum state at each frequency point in parallel. The update is performed using a quantum exponential weighted moving average mechanism, and the update formula is as follows:

[0173] in, The quantum state of the covariance matrix of the frequency point f corresponding to the previous frame. Let f be the M×1 dimensional quantum frequency domain signal vector of the current frame frequency point. Represents the quantum tensor product. The dual vector representing the quantum state (corresponding to the classical conjugate transpose), α is the forgetting factor, with a value ranging from 0.8 to 0.95; Step 3: Efficient update of the inverse quantum covariance matrix; specifically, based on the quantum version of the Sherman-Morrison-Woodbury matrix inversion lemma, the quantum phase estimation (QPE) technique and quantum multipliers are used to recursively update the quantum state of the inverse matrix. This avoids the high complexity of classical matrix inversion and further reduces the computational complexity from O(M²) to O(logM). Step four: Optimal beamformer selection based on maximum SNR in the quantum domain; specifically, leveraging the advantages of quantum parallel computing, simultaneously weighting the quantum states of the M microphone channels. The solution is performed to enhance the quantum state probability amplitude of high SNR channels using quantum amplitude amplification technology, and the optimal quantum state beamforming weights corresponding to the reference microphone are selected. The quantum frequency domain signal matrix |X> is beamformed using this weight to obtain the quantum state output spectrum. ; Step 5, Quantum gain compensation and signal restoration; specifically, constructing the quantum reference microphone selection vector. The quantum state of the frequency domain pointwise gain factor is calculated through quantum inner product operation. ,right Performing quantum gain compensation yields , )( For the quantum state minimum, take the value of Finally, the quantum state frequency domain signal is restored to the time domain classical signal through inverse quantum Fourier transform (IQFT) to output enhanced speech.

[0174] This embodiment is applied to an ultra-high precision remote quantum conferencing system. The system is equipped with a 16-channel distributed microphone array with an array element spacing of 3cm. The core processing unit is a superconducting quantum computer with 64 qubits and a quantum gate fidelity of ≥99.5%. It needs to cope with large-scale multi-person interactive scenarios, such as ≤10 people speaking dynamically at the same time. The single frame processing time is required to be ≤10ms, and the output voice SNR is ≥20dB. At the same time, it meets the requirements of low latency and high robustness.

[0175] The microphone array consists of 16 microphones (M=16), with a sampling rate of 16kHz, a frame length of 10ms, and 160 sampling points per frame. In the quantum processing parameters, the Hanning window length is 160 points, and the quantum Fourier transform (QFT) is configured with 8 qubits. Forgetting factor α = 0.92, quantum state minimum Quantum gate operation accuracy ≤ 0.1%; Quantum noise covariance matrix estimation employs a quantum minimum tracking algorithm, which tracks the quantum state minimum of the signal power at each frequency point through quantum superposition states in parallel, dynamically updating the quantum state of the noise covariance matrix. Quantum measurement error ≤0.5%.

[0176] Sixteen microphones simultaneously acquire time-domain speech signals. Each signal is windowed using a Hanning window and then converted into a 16-bit digital signal by an analog-to-digital converter (ADC). A quantum encoder then maps the classical digital signal to a quantum state (using ground-state encoding, where each sample value corresponds to the phase or amplitude of a qubit). A quantum Fourier transform circuit is constructed based on eight qubits to perform a QFT on each quantized time-domain signal. Utilizing quantum parallelism, all sampling points are processed simultaneously, converting each time-domain signal into a 256-frequency quantum state frequency-domain signal, ultimately forming a 16×256-dimensional quantum frequency-domain signal matrix. The quantum circuit depth of the QFT transform is O(q²) = 64 (q = 8), which is significantly lower than the O(FlogF) complexity of the traditional FFT.

[0177] For each frequency point f (f=0~255), a quantum tensor product circuit is constructed. The quantum state is used to simulate the outer product operation of the classical covariance matrix. An exponentially weighted moving average is implemented using quantum adders and quantum multipliers to convert the quantum state of the previous frame's covariance matrix. A linear combination with the current frame's quantum outer product state yields... The quantum superposition property allows the calculation of the covariance matrix at all frequencies to be performed in parallel, improving efficiency by a factor of F (F=256) compared to classical serial computation.

[0178] Based on the quantum version of the Sherman-Morrison-Woodbury lemma, the specific process is as follows: During the processing of the first frame, quantum phase estimation (QPE) is used to... Eigenvalue estimation is performed, and |R| is obtained by combining it with a quantum inverse transform circuit. now -1 (f)>; Calculating scalar quantum states using quantum multipliers Parallel multiplication of multiple quantum states is achieved through quantum interference effects; The update formula is executed by quantum subtractor and quantum divider circuits. Taking advantage of the parallel operation characteristics of quantum gates, all elements of the inverse matrix are updated simultaneously. The computational delay depends only on the operation time of the quantum gates (≤100ns).

[0179] Simultaneously constructing the steering vector quantum states of 16 microphones using quantum parallel circuits. (M×1 dimensional, the quantum state of the m-th element is) The rest are ), using quantum inner product circuits to compute 16-way weighted quantum states in parallel ; The system utilizes quantum superposition to simultaneously process 64 frequency points corresponding to 0~2000Hz. A quantum power calculation circuit is used to solve for the power quantum states of the speech signal and the noise signal in parallel. Based on a quantum divider, 16 SNR quantum states are obtained. ; A quantum amplitude amplification algorithm is used to amplify the probability amplitude of quantum states corresponding to high SNR to close to 1, and the optimal weighted quantum state is obtained through quantum measurement. Applying this weight yields .

[0180] Constructing a quantum reference microphone selection vector (16×1 dimensional, the optimal microphone index position quantum state is) The rest are ); Calculation via quantum inner product circuit Accurate gain characterization is achieved by utilizing the coherent properties of quantum states; Executed via quantum divider circuit This avoids the risk of quantum state collapse when the denominator is zero; To | The inverse quantum Fourier transform (IQFT) is performed to convert the quantum state frequency domain signal into a classical time domain signal. After being read by the quantum measurement module, the inter-frame distortion is eliminated by the overlap addition method, and one channel of enhanced speech is output.

[0181] In a remote conference scenario with 10 people speaking dynamically, this quantum computing solution achieves an average processing time of only 1.2ms per frame, a 71.4% speedup compared to the classical solution (4.2ms). Quantum parallel computing and quantum amplitude amplification technologies improve SNR calculation accuracy by 30%, resulting in an output speech SNR that is 22-28dB higher than the original signal, with speech amplitude fluctuations ≤1.5dB. The anti-interference characteristics of quantum states effectively resist quantum disturbances caused by environmental noise, maintaining 99.8% signal stability even in complex electromagnetic interference environments. The clarity and naturalness of the sound are significantly improved compared to the classical solution, fully meeting the requirements of ultra-high precision voice interaction scenarios.

[0182] like Figure 3 As shown in Embodiment 3, according to another aspect of this patent, this patent also provides a real-time adaptive beamforming device based on maximum SNR, comprising: The preprocessing unit is configured to perform windowing preprocessing on the time-domain audio signals collected by the M microphones, and then perform FFT transformation to obtain an M×F-dimensional frequency domain signal matrix X. The moving average processing unit is configured to update the M×M covariance matrix R of each frequency point using an exponentially weighted moving average mechanism based on the frequency domain signal matrix X. now (f); The recursive update unit is configured to apply the Sherman-Morrison-Woodbury matrix inversion lemma to the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; The beamforming processing unit is configured to calculate the SNR of each microphone channel in parallel, select the optimal reference microphone, and extract the corresponding beamforming weights. For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f); The output unit is configured to construct a gain compensation mechanism based on the optimal reference microphone, perform frequency-by-frequency gain compensation on the output spectrum Y(f), perform IFFT transformation to restore it to the time domain signal, and output enhanced speech.

[0183] in, The preprocessing unit includes The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; wherein, the sampling rate of the time-domain signal is 16kHz, the frame length is 10ms, M is the number of microphones in the microphone array, which is 8, and F is the total number of frequency points; The moving average processing unit includes The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is:

[0184] in, Let X(f) be the covariance matrix of the frequency point f in the previous frame, and let X(f) be the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.95.

[0185] in, The recursive update unit includes During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is: R now -1 (f)=inv(R now (f)) in, This represents the matrix inversion operation; The formula for calculating the scalar value k is:

[0186] in, α is the inverse of the covariance matrix of the frequency point f corresponding to the previous frame, and α is the forgetting factor defined in the above steps; Current frame inverse matrix R now -1 The formula for calculating (f) is:

[0187] This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

[0188] in, The beamforming processing unit includes... The specific process for selecting the optimal beamformer for the SNR includes: For the m-th microphone, construct an M×1 dimensional steering vector pointing to that microphone. The m-th element is 1, and the rest are 0, which corresponds to the beamformer weights. The calculation formula is: Where m ranges from 1 to M, Guide vector The conjugate transpose of; The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows:

[0189] in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power. Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight for the current frame. The output spectrum is obtained by applying this weight. .

[0190] in, The output unit includes, Construct the reference microphone selection vector s ref The vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions; The formula for calculating the pointwise gain factor g(f) in the frequency domain is:

[0191] in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone; The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is:

[0192] Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero; right Perform an IFFT transform to convert the frequency domain signal back to the time domain, outputting a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

[0193] like Figure 4 As shown in Embodiment 4, this quantum computing maximum SNR real-time adaptive beamforming device includes... The quantization conversion unit is configured for time-domain to frequency-domain signal quantization conversion. Specifically, after preprocessing the time-domain audio signals acquired by M microphones using Hanning windowing, it replaces the traditional FFT with quantum Fourier transform (QFT) to convert each time-domain signal into a quantum state frequency-domain signal with F frequency points, constructing an M×F dimensional quantum frequency-domain signal matrix. The number of qubits is configured according to the number of frequency points F, satisfying 2^q≥F (q is the number of qubits). The dynamic update unit is configured to dynamically update the quantum-state speech covariance matrix; specifically, it is based on the quantum frequency domain signal matrix. The quantum superposition property is used to calculate the M×M dimension covariance matrix quantum state at each frequency point in parallel. The update is performed using a quantum exponential weighted moving average mechanism, and the update formula is as follows:

[0194] in, The quantum state of the covariance matrix of the frequency point f corresponding to the previous frame. Let f be the M×1 dimensional quantum frequency domain signal vector of the current frame frequency point. Represents the quantum tensor product. The dual vector representing the quantum state (corresponding to the classical conjugate transpose), α is the forgetting factor, with a value ranging from 0.8 to 0.95; The inverse matrix update unit is configured for efficient updating of the quantum covariance matrix inverse; specifically, it recursively updates the inverse matrix quantum state |R| using quantum phase estimation (QPE) techniques and quantum multipliers, based on the quantum version of the Sherman-Morrison-Woodbury matrix inversion lemma. now -1 (f)>, avoids the high complexity of classical matrix inversion, and further reduces the computational complexity from O(M²) to O(logM); The selection unit is configured to perform optimal beamformer selection in the quantum domain based on the maximum SNR; specifically, it leverages the advantages of quantum parallel computing to simultaneously weight the quantum states of the M microphone channels. The solution is performed to enhance the quantum state probability amplitude of high SNR channels using quantum amplitude amplification technology, and the optimal quantum state beamforming weights corresponding to the reference microphone are selected. Applying this weight to the quantum frequency domain signal matrix Beamforming processing is performed to obtain the quantum state output spectrum. ; The compensation and restoration unit is configured for quantum gain compensation and signal restoration; specifically, it constructs a quantum reference microphone selection vector. The quantum state of the frequency domain pointwise gain factor is calculated through quantum inner product operation. right Performing quantum gain compensation yields , ( For the quantum state minimum, take the value of Finally, the quantum state frequency domain signal is restored to the time domain classical signal through inverse quantum Fourier transform (IQFT) to output enhanced speech.

[0195] For details of the embodiments of the apparatus claims, please refer to the method embodiments section. To avoid repetition, they will not be repeated here.

[0196] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time adaptive beamforming method based on maximum SNR, comprising the following steps: After windowing preprocessing of the time-domain audio signals acquired by M microphones, an FFT transformation is performed to obtain an M×F dimensional frequency domain signal matrix X. Based on the frequency domain signal matrix X, an exponential weighted moving average mechanism is used to update the MxM dimensional covariance matrix R of each frequency point now (f); Sherman-Morrison-Woodbury matrix inversion lemma to the covariance matrix R now the inverse matrix R of (f) now -1 (f) recursively updating; Parallelly compute SNR of each microphone channel, screen optimal reference microphone and extract corresponding beamforming weight W opt (f), for each frequency point f, apply the weight to beamforming process of frequency domain signal X(f) to obtain output spectrum Y(f); A gain compensation mechanism is constructed based on the optimal reference microphone. After frequency-by-frequency gain compensation is performed on the output spectrum Y(f), an IFFT transform is performed to restore it to the time domain signal, and the enhanced speech is output.

2. The real-time adaptive beamforming method based on maximum SNR as described in claim 1, characterized in that, The steps described involve performing windowing preprocessing on the time-domain audio signals acquired by the M microphones, followed by FFT transformation to obtain an M×F dimensional frequency-domain signal matrix X; which includes... The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; The aforementioned step, based on the frequency domain signal matrix X, uses an exponentially weighted moving average mechanism to update the M×M dimension covariance matrix R at each frequency point. now (f); includes The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is: ， Among them, R pre X(f) is the covariance matrix of the frequency point f in the previous frame, and X(f) is the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.

95.

3. The real-time adaptive beamforming method based on maximum SNR as described in claim 2, characterized in that, The aforementioned step employs the Sherman-Morrison-Woodbury matrix inversion lemma to apply the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; including... During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is: R now -1 (f)=inv(R now (f)) Where inv(.) represents the matrix inversion operation; The formula for calculating the scalar value k is: ， Among them, R pre -1 (f) is the inverse of the covariance matrix of the frequency point f in the previous frame, and α is the forgetting factor defined in the above steps; Current frame inverse matrix R now -1 The formula for calculating (f) is: ， This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

4. The real-time adaptive beamforming method based on maximum SNR as described in claim 3, characterized in that, The steps described above involve parallel calculation of the SNR of each microphone channel, selection of the optimal reference microphone, and extraction of the corresponding beamforming weight W. opt For each frequency point f, the weight is applied to perform beamforming processing on the frequency domain signal X(f) to obtain the output spectrum Y(f); include, The specific process for selecting the optimal beamformer for the SNR includes: For the m-th microphone, construct an M×1 dimensional steering vector s pointing to that microphone. m The m-th element is 1, and the rest are 0, corresponding to the beamformer weight w. m The calculation formula is: , Where m ranges from 1 to M, Guide vector The conjugate transpose of; The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows: ， in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power. Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight W for the current frame. opt (f) The output spectrum is obtained by applying this weight. .

5. The real-time adaptive beamforming method based on maximum SNR as described in claim 4, characterized in that, The steps described above involve constructing a gain compensation mechanism based on the optimal reference microphone, performing frequency-point gain compensation on the output spectrum Y(f), and then performing an IFFT transform to restore it to a time-domain signal, thereby outputting enhanced speech. Including, Construct the reference microphone selection vector s ref The vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions; The formula for calculating the pointwise gain factor g(f) in the frequency domain is: ， in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone; The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is: ， Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero; right Perform an IFFT transform to convert the frequency domain signal back to the time domain and output a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

6. A real-time adaptive beamforming device based on maximum SNR, characterized in that, include: The preprocessing unit is configured to perform windowing preprocessing on the time-domain audio signals collected by the M microphones, and then perform FFT transformation to obtain an M×F-dimensional frequency domain signal matrix X. The moving average processing unit is configured to update the M×M covariance matrix R of each frequency point using an exponentially weighted moving average mechanism based on the frequency domain signal matrix X. now (f); The recursive update unit is configured to apply the Sherman-Morrison-Woodbury matrix inversion lemma to the covariance matrix R. now The inverse matrix R of (f) now -1 (f) Perform recursive updates; The beamforming processing unit is configured to calculate the SNR of each microphone channel in parallel, select the optimal reference microphone, and extract the corresponding beamforming weight W. opt (f), for each frequency point f, the weight is applied to the frequency domain signal X(f) to perform beamforming processing to obtain the output spectrum Y(f); The output unit is configured to construct a gain compensation mechanism based on the optimal reference microphone, perform frequency-by-frequency gain compensation on the output spectrum Y(f), perform IFFT transformation to restore it to the time domain signal, and output enhanced speech.

7. A real-time adaptive beamforming device based on maximum SNR as described in claim 6, characterized in that, The preprocessing unit includes The windowing preprocessing uses Hanning window to suppress spectral leakage, and the FFT transform converts each time-domain signal into a frequency-domain signal with F frequency points; The moving average processing unit includes The exponentially weighted moving average mechanism balances the retention of historical signal features with the response speed to environmental changes through a forgetting factor α, and the covariance matrix R... now The update formula for (f) is: ， Among them, R pre X(f) is the covariance matrix of the frequency point f in the previous frame, and X(f) is the M×1 dimensional frequency domain signal vector of the frequency point f in the current frame. This represents the conjugate transpose operation, where α is the forgetting factor, with a value ranging from 0.8 to 0.

95.

8. The real-time adaptive beamforming device based on maximum SNR as described in claim 7, characterized in that, The recursive update unit includes During the processing of the first frame, the inverse matrix R now -1 (f) Obtained by directly inverting the formula, the formula is: R now -1 (f)=inv(R now (f)), in, This represents the matrix inversion operation; The formula for calculating the scalar value k is: ， in, α is the inverse of the covariance matrix of the frequency point f corresponding to the previous frame, and α is the forgetting factor defined in the above steps; Current frame inverse matrix R now -1 The formula for calculating (f) is: ， This update method reduces the complexity of matrix inversion from O(M³) to O(M²), meeting the real-time processing requirements of embedded devices.

9. A real-time adaptive beamforming device based on maximum SNR as described in claim 8, characterized in that, The beamforming processing unit includes... The specific process for selecting the optimal beamformer for the SNR includes: For the m-th microphone, construct an M×1 dimensional steering vector pointing to that microphone. The m-th element is 1, and the rest are 0, corresponding to the beamformer weight w. m The calculation formula is: ， Where m ranges from 1 to M, Guide vector The conjugate transpose of; The SNR is calculated by selecting the frequency points corresponding to 0~2000Hz. The calculation formula is as follows: ， in, Let f be the noise covariance matrix at frequency point f, which is estimated by minimum tracking or statistical methods for speech still segments. The numerator is the sum of the speech signal power after beamforming, and the denominator is the sum of the noise signal power. Compare the SNR calculation results of the M microphones, and select the weight vector corresponding to the maximum value as the beamforming weight w for the current frame. opt (f) The output spectrum is obtained by applying this weight. 。 10. A real-time adaptive beamforming device based on maximum SNR as described in claim 9, characterized in that, The output unit includes, Construct the reference microphone selection vector s ref The vector takes a value of 1 at the optimal microphone index position and a value of 0 at other positions; The formula for calculating the pointwise gain factor g(f) in the frequency domain is: ,in, For the conjugate transpose of the optimal beamforming weight vector, s ref Select a vector for the reference microphone; The output spectrum Y(f) is corrected at each frequency point to obtain the final output spectrum. The formula is: ， Where ε is the local minimum value, and takes the value of This is used to avoid cases where the denominator is zero; right Perform an IFFT transform to convert the frequency domain signal back to the time domain and output a high-fidelity enhanced speech signal. The amplitude characteristics of this signal are consistent with the original signal acquired by the optimal reference microphone, with no amplitude distortion.

Images