A method, device and electronic device for reducing speech noise

Abstract

The present invention discloses a speech denoising method, device and electronic device, and relates to the technical field of speech signal processing. The method comprises: obtaining an amplitude spectrum of a source speech signal; partitioning the audio based on a posteriori signal-to-noise ratio value and a preset threshold range to obtain an audio zone, smoothing the filter parameters corresponding to all audio zones to obtain filter parameters, using a Wiener filter to perform denoising on the amplitude spectrum to obtain a denoised amplitude spectrum; using a Griffin‑Lim phase recovery algorithm based on an alternating direction multiplier method, an augmented Lagrangian function and Gaussian iteration to perform phase estimation on the denoised amplitude spectrum to obtain an estimated phase, and integrating the denoised amplitude spectrum and the estimated phase to obtain a denoised speech signal. The present invention improves the convergence of the Griffin‑Lim phase recovery algorithm and reduces the number of iterations, thereby reducing the algorithm complexity and reducing computing resources; its Wiener filter realizes reducing the speech loss degree in the denoising algorithm while reducing the noise residue.

Description

Technical Field

[0001] The present invention relates to the technical field of speech signal processing, and in particular to a speech noise reduction method, device and electronic equipment. Background Art

[0002] In traditional speech denoising algorithms, the amplitude spectrum in the speech signal is often processed to obtain the amplitude spectrum of the denoised speech signal, and the denoised speech is obtained based on the amplitude spectrum of the denoised speech signal and the phase of the noisy speech before denoising. In the above-mentioned speech denoising algorithm, only the amplitude spectrum is processed, while the phase information is ignored, resulting in distortion of the speech signal. Especially when the signal-to-noise ratio of the noisy speech signal is low, only processing the amplitude spectrum will cause distortion of the speech signal, while accurate estimation and correction of the phase information can help reduce this distortion, making the denoised speech more natural and clear.

[0003] Therefore, the Griffin-Lim algorithm is usually used to implement the phase reconstruction method. By iteratively estimating the phase spectrum from the amplitude spectrum, a natural speech signal is gradually reconstructed. However, the Griffin-Lim algorithm has disadvantages such as slow convergence. A large number of iterations are required to reconstruct a relatively satisfactory speech signal. The algorithm is highly complex and requires large computing resources.

[0004] In summary, how to reduce the algorithm complexity of the phase recovery method and reduce the amount of calculation of the phase recovery method are technical problems that need to be solved. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a speech noise reduction method, device, electronic device and storage medium.

[0006] Technical solution:

[0007] According to one aspect of the present disclosure, there is provided a method for reducing speech noise, comprising:

[0008] Obtaining a source speech signal;

[0009] Performing short-time Fourier transform on the source speech signal after frame division and windowing to obtain the amplitude spectrum and phase information of the source speech signal;

[0010] Partitioning the audio based on a posteriori signal-to-noise ratio value and a preset threshold range to obtain a plurality of audio regions, smoothing the filter parameters corresponding to all the audio regions to obtain updated filter parameters, and performing denoising on the amplitude spectrum using a Wiener filter of the updated filter parameters to obtain a denoised amplitude spectrum;

[0011] An improved Griffin-Lim phase recovery algorithm is used to perform phase estimation on the amplitude spectrum after noise reduction to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the amplitude spectrum after noise reduction, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm;

[0012] Integrate the noise-reduced amplitude spectrum and the estimated phase, perform inverse Fourier transform, and obtain a noise-reduced time domain signal;

[0013] The noise-reduced time domain signals are overlapped and added to obtain noise-reduced speech signals.

[0014] According to another aspect of the present disclosure, there is provided a speech noise reduction device, comprising:

[0015] An acquisition module is used to acquire a source speech signal;

[0016] A short-time Fourier transform module is used to perform short-time Fourier transform on the source speech signal after performing frame division and windowing to obtain an amplitude spectrum of the source speech signal;

[0017] A filter module, partitioning the audio based on a posterior signal-to-noise ratio value and a preset threshold range to obtain multiple audio areas, smoothing the filter parameters corresponding to all the audio areas to obtain updated filter parameters, and using the Wiener filter of the updated filter parameters to perform denoising on the amplitude spectrum to obtain a denoised amplitude spectrum;

[0018] A phase estimation module, which uses an improved Griffin-Lim phase recovery algorithm to perform phase estimation on the amplitude spectrum after noise reduction to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the amplitude spectrum after noise reduction, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm;

[0019] An inverse Fourier transform module integrates the noise-reduced amplitude spectrum and the estimated phase, performs an inverse Fourier transform, and obtains a noise-reduced time domain signal;

[0020] A time domain module after noise reduction is obtained, and the time domain signal after noise reduction is overlapped and added to obtain a noise-reduced speech signal.

[0021] According to another aspect of the present disclosure, there is provided an electronic device, including:

[0022] at least one processor; and

[0023] a memory communicatively connected to the at least one processor; wherein,

[0024] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the aforementioned aspect.

[0025] According to yet another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to execute the method described in the aforementioned aspect.

[0026] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, the method described in the aforementioned aspect is implemented.

[0027] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description.

[0028] Beneficial effects: The phase correlation problem is converted into a minimum optimization problem through the alternating direction multiplier method (ADMM), which improves the objective function in the Griffin-Lim algorithm, avoids the problem of non-existence of the solution, and converts the minimum optimization problem into a two-variable optimization problem, thereby simplifying the amount of calculation for each iteration and speeding up the overall solution speed. Subsequently, by introducing the augmented Lagrangian function and using Gaussian iteration to solve the two-variable optimization problem, the convergence of the algorithm can be improved, the number of iterations can be reduced, and the algorithm complexity and computing resources can be reduced.

[0029] At the same time, the audio is partitioned by the posterior signal-to-noise ratio value and the preset threshold range to obtain multiple audio areas, the filter parameters corresponding to all the audio areas are smoothed to obtain updated filter parameters, and the amplitude spectrum is denoised using the Wiener filter with the updated filter parameters to obtain the denoised amplitude spectrum, thereby avoiding the problem that the energy at the tail of the speech is small and is easily misjudged as noise, thereby achieving the reduction of speech loss in the noise reduction algorithm while reducing noise residue. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of a speech noise reduction method according to embodiment 1 of the present invention;

[0031] Figure 2 is a flow chart of an improved Wiener filter of a speech noise reduction method according to embodiment 1 of the present invention;

[0032] Figure 3 It is a flow chart of an improved Griffin-Lim phase recovery algorithm of a speech noise reduction method in embodiment 1 of the present invention;

[0033] Figure 4 is a schematic diagram of the principle of noise reduction processing in a scenario of embodiment 1 of the present invention;

[0034] Figure 5 It is a structural schematic diagram of a speech noise reduction device according to embodiment 2 of the present invention. DETAILED DESCRIPTION

[0035] In order to make the technical solution of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] Figure 1 A flowchart of a speech noise reduction method provided in Example 1 of the present disclosure.

[0038] The speech noise reduction method provided in the embodiment of the present disclosure is performed by a speech noise reduction device. The speech noise reduction device can be an electronic device, or can be configured in an electronic device to integrate the amplitude spectrum after noise reduction and the estimated phase, suppress the noise in the source speech, and obtain noise-reduced speech. The embodiment of the present disclosure is described by taking the speech noise reduction device configured in an electronic device as an example.

[0039] Among them, the electronic device can be any stationary or mobile computing device capable of data processing, such as mobile computing devices such as laptops, smart phones, wearable devices, or stationary computing devices such as desktop computers, or servers, or other types of computing devices, etc., and the present disclosure does not limit this.

[0040] like Figure 1 As shown, the speech noise reduction method comprises:

[0041] Step S101, obtaining a source voice signal;

[0042] Step S102, performing short-time Fourier transform (STFT) on the source speech signal after frame division and windowing, so as to obtain an amplitude spectrum of the source speech signal;

[0043] In an exemplary embodiment, a Hanning window may be used to perform frame windowing and short-time Fourier transform on the input noisy speech signal;

[0044] Step S103, partitioning the audio based on the a posteriori signal-to-noise ratio value and a preset threshold range to obtain multiple audio areas, smoothing the filter parameters corresponding to all the audio areas to obtain updated filter parameters, and using the Wiener filter of the updated filter parameters to perform denoising on the amplitude spectrum to obtain a denoised amplitude spectrum;

[0045] It should be noted that, in order to solve the problem that the energy of the voice tail is weak, the noise reduction estimator identifies it as noise and over-suppresses it, resulting in voice distortion, the present disclosure adopts step S103 to improve the Wiener filter, wherein the expression of the Wiener filter is shown in the following formula:

[0046] ; (1)

[0047] Among them, Gain is used to represent the Wiener filter, Indicates The a priori signal-to-noise ratio of the frame is calculated using the decision-directed method. are the Wiener filter parameters.

[0048] In an exemplary embodiment, it can be seen from the above formula (1) that the larger the Wiener filter parameter is, the easier it is to generate more speech distortion and less residual noise. Conversely, the smaller the Wiener filter parameter is, the easier it is to generate less speech distortion and more residual noise. Based on this, the present disclosure divides the speech area, weak sound area and noise area by threshold range based on the parameters obtained by averaging the posterior signal-to-noise ratio of all frequency points;

[0049] Specifically, the step S103 of the present disclosure partitions the audio based on the a posteriori signal-to-noise ratio value and the preset threshold range to obtain multiple audio regions, and smoothes the filter parameters of all audio regions to obtain updated filter parameters;

[0050] Figure 2 is a flow chart of an improved Wiener filter of a speech noise reduction method according to embodiment 1 of the present invention, such as Figure 2 As shown, improving the Wiener filter by updating the filter parameters specifically includes:

[0051] Step S1031, averaging the a posteriori signal-to-noise ratio values ​​of all frequency points to obtain an average parameter corresponding to each frequency point;

[0052] Step S1032, setting multiple threshold ranges, and dividing all frequency points into a speech area, a weak sound area, and a noise area by using average parameters and threshold ranges;

[0053] Wherein, the step S1032 specifically includes:

[0054] Set the minimum and maximum partition thresholds;

[0055] When the average parameter of the frequency point is less than the minimum division threshold, dividing the frequency point into the noise area;

[0056] When the average parameter of the frequency point is greater than the maximum division threshold, dividing the frequency point into the speech area;

[0057] When the average parameter of the frequency point is less than or equal to the maximum division threshold, and greater than or equal to the minimum division threshold, the frequency point is divided into the weak sound area;

[0058] As shown in the following formula:

[0059] ; (2)

[0060] Among them, Pro is used to indicate the area where the voice exists, and th is used to indicate the average parameter of the frequency point. , K is used to represent the total number of frequency points, Used to represent the amplitude spectrum of noisy speech, Used to represent the amplitude spectrum of the estimated noise signal;

[0061] 0 is used to represent the noise area, 1 is used to represent the speech area, and 2 is used to represent the weak sound area;

[0062] th min Used to indicate the minimum partition threshold, i.e. the noise area threshold;

[0063] th max Used to indicate the maximum division threshold, that is, the speech area threshold.

[0064] Step S1033, obtaining filter parameters corresponding to the speech area, weak sound area and noise area, and performing smoothing processing on all filter parameters to obtain processed filter parameters.

[0065] Wherein, the step S1033 specifically includes:

[0066] The filter parameters corresponding to each of the audio regions are obtained, as shown in the following formula:

[0067] ; (3)

[0068] in, , and is a constant, satisfying , Used to represent filter parameters;

[0069] In order to prevent the rapid change of the gain function and thus generate musical noise, all filter parameters are smoothed to obtain the processed filter parameters, as shown in the following formula:

[0070] ; (4)

[0071] Among them, β is used to represent the smoothing factor, 0<β<1.

[0072] Step S104, using an improved Griffin-Lim phase recovery algorithm to perform phase estimation on the denoised amplitude spectrum to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the denoised amplitude spectrum, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method (ADMM), converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm;

[0073] In an exemplary embodiment, Figure 3 is a flowchart of an improved Griffin-Lim phase recovery algorithm of a speech noise reduction method in embodiment 1 of the present invention, wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm in step S104 is as follows: Figure 3 As shown, specifically including:

[0074] Step S1041, obtaining the phase correlation problem between the complex spectrum with phase information and the amplitude spectrum after noise reduction, specifically includes:

[0075] According to the given denoised amplitude spectrum Find the corresponding phase information and calculate the noise-reduced amplitude spectrum and the corresponding phase information form a complex spectrum X with a phase, and the complex spectrum must meet the preset conditions, which are shown in the following formula:

[0076] ; (5)

[0077] in, , , , represents the short-time Fourier transform, represents the image of the short-time Fourier transform;

[0078] It should be noted that for a given amplitude spectrum , and there is no phase information, it is necessary to find appropriate phase information based on the amplitude spectrum A to form a complex spectrum X with phase. The complex spectrum must meet the preset conditions, as shown in the above formula.

[0079] Among them, the short-time Fourier transform formula is shown in the following formula:

[0080] ; (6)

[0081] in, is used to represent the short-time Fourier transform, τ is used to represent the signal time variable, , k is used to indicate the frequency subscript, , Used to represent the complex spectrum after short-time Fourier transform, , L is used to represent the length of a frame signal, l is used to represent the subscript of a frame time domain signal, and x is used to represent the speech signal. , Used to represent vector transpose, Used to indicate the signal length. For representing the field of real numbers, Used to indicate frame shift time, is used to represent the analysis window, j is used to represent, 3 units, Used to represent the amplitude spectrum, Used to indicate phase;

[0082] The inverse Fourier transform formula corresponding to the short-time Fourier transform formula is shown in the following formula:

[0083] ; (7)

[0084] Step S1042, in order to solve the above phase correlation problem, that is, the formula (5) corresponding to the preset condition, the present disclosure converts the above phase correlation problem into a minimum optimization problem based on the alternating direction multiplier method, in order to prevent , resulting in the non-existence of a solution, and the minimum optimization problem is shown in the following formula:

[0085] ; (8)

[0086] in,

[0087] ; (9)

[0088] ; (10)

[0089] ; (11)

[0090] Among them, X is used to represent the complex spectrum with phase information, λ is used to represent the penalty factor, represents the Hadamard product, is the L2 norm of the matrix, called the Frobenius norm, which can also be expressed as , express norm, represents the short-time Fourier transform, represents inverse Fourier transform;

[0091] Step S1043, in order to simplify the amount of calculation for each iteration and speed up the overall solution speed, the minimum optimization problem is transformed into a two-variable optimization problem, as shown in the following formula:

[0092] ; (12)

[0093] ; (13)

[0094] Step S1044, using the augmented Lagrangian function to solve the two variable optimization problem, the augmented Lagrangian function is shown in the following formula:

[0095] ; (14)

[0096] in , For penalty items;

[0097] It should be noted that by introducing the penalty term, the augmented Lagrangian function can improve the convergence of the original Lagrangian method, and it is relatively insensitive to the choice of initial values ​​and can obtain a better solution in a larger range;

[0098] Step S1045, for the multivariate function (14), in order to improve convergence, an iterative method can be used to solve it, in which only one variable is updated in each step, and the other two variables are fixed, and then the idea of ​​the Gaussian iteration method is used to calculate When using the already calculated replace Calculate, calculate When using the already calculated Replace Calculate, specifically:

[0099] For Iterations:

[0100] Step 1: ; (15)

[0101] Step2: ; (16)

[0102] Step 3: ; (17)

[0103] Step S1046, through variable substitution method, let , substituting into the augmented Lagrangian function, we can get the following formula:

[0104] ; (18)

[0105] The corresponding iterative process becomes:

[0106] Step 1: ; (19)

[0107] Step2: ; (20)

[0108] Step 3: ;(twenty one)

[0109] Step S1047, performing a solution to obtain an improved Griffin-Lim phase recovery algorithm; specifically comprising:

[0110] For the minimum optimization problem in Step 1 and Step 2, the minimum value is obtained by taking the derivative of the function and setting the derivative to 0, as follows:

[0111] For Step 1:

[0112] make ,but

[0113] ;(twenty two)

[0114] ;(twenty three)

[0115] in , The Elements , the calculation formula is

[0116] ;(twenty four)

[0117] Will , then formula (22) can be converted to:

[0118] ;(twenty four)

[0119] in,

[0120] ; (25)

[0121] For Step 2:

[0122] ; (26)

[0123] ; (27)

[0124] ; (28)

[0125] In an exemplary embodiment, the expression of the improved Griffin-Lim phase recovery algorithm can be obtained by using the above formula (21), formula (25) and formula (28), as shown in the following formula:

[0126] ; (29)

[0127] ; (30)

[0128] ; (31)

[0129] Among them, X (0) =A,ρ>0,λ>0,Z (0) =X (0) ,ω (0) =0,m=0,1,…M;

[0130] X (m+1) It is used to represent the complex spectrum with phase information of the m+1th iteration, A is used to represent the amplitude spectrum without phase information after speech denoising, m represents the iteration subscript, ρ, λ, ω, Z are auxiliary variables, is the soft threshold operator, Z (m+1) ,ω (m+1) represents the value of the m+1th iteration, Used to express After performing inverse Fourier transform, short-time Fourier transform is performed to obtain the value.

[0131] As a specific embodiment, the improved Griffin-Lim phase recovery algorithm process is:

[0132] enter: ,

[0133] Initialization parameters:

[0134] Output:

[0135] for m=0,1,...,M

[0136]

[0137]

[0138]

[0139] end

[0140] Step S105, integrating the denoised amplitude spectrum and the estimated phase, performing an inverse Fourier transform (ISTFT), and obtaining a denoised time domain signal;

[0141] Step S106: performing overlap-addition on the denoised time domain signal to obtain a denoised speech signal.

[0142] In order to more clearly illustrate the above noise reduction process, an example is now given in conjunction with the model structure diagram.

[0143] refer to Figure 4 , Figure 4 This is a schematic diagram of the principle of noise reduction processing in a scene, such as Figure 4 As shown, a source speech signal is obtained, and the source speech signal is framed, windowed, and short-time Fourier transformed in succession to obtain an amplitude spectrum of the source speech signal, and then the amplitude spectrum is input into an improved Wiener filter to obtain a denoised amplitude spectrum, and then an improved Griffin-Lim phase recovery algorithm is used to perform phase estimation on the denoised amplitude spectrum to obtain a corresponding estimated phase, and finally the denoised amplitude spectrum and the estimated phase are integrated, and an inverse Fourier transform is performed to obtain a denoised time domain signal, and the denoised time domain signal is overlapped and added to obtain a denoised speech signal;

[0144] The step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the denoised amplitude spectrum, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm;

[0145] Among them, the step of improving the Wiener filter includes: partitioning the audio based on the posterior signal-to-noise ratio value and the preset threshold range to obtain multiple audio areas, and smoothing the filter parameters corresponding to all audio areas to obtain updated filter parameters.

[0146] In daily acoustic environments, speech is usually interfered by noise, so the source speech needs to be denoised. The present invention integrates the amplitude spectrum and estimated phase after denoising to perform denoising on the source speech, thereby obtaining denoised speech; wherein, the denoised speech is the speech after the noise is suppressed in the source speech, which can clearly characterize the speech information of the source speech.

[0147] It should be noted that the alternating direction method of multipliers (ADMM) is a method for solving separable convex optimization problems, which has the advantages of fast processing speed and good convergence performance.

[0148] It should be noted that the above phase correlation problem is the objective function in the Griffin-Lim algorithm.

[0149] In summary, the phase correlation problem is converted into a minimum optimization problem through the alternating direction multiplier method (ADMM), which realizes the improvement of the objective function in the Griffin-Lim algorithm, avoids the problem of non-existence of the solution, and converts the minimum optimization problem into a two-variable optimization problem, thereby simplifying the amount of calculation for each iteration and accelerating the overall solution speed. Subsequently, by introducing the augmented Lagrangian function and using Gaussian iteration to solve the two-variable optimization problem, the convergence of the algorithm can be improved and the number of iterations can be reduced, thereby reducing the complexity of the algorithm and further reducing computing resources. It is applied to the speech denoising algorithm to improve the intelligibility of the denoised speech.

[0150] At the same time, the audio is partitioned by the posterior signal-to-noise ratio value and the preset threshold range to obtain multiple audio areas, the filter parameters corresponding to all the audio areas are smoothed to obtain updated filter parameters, and the amplitude spectrum is denoised using the Wiener filter with the updated filter parameters to obtain the denoised amplitude spectrum, thereby avoiding the problem that the energy at the tail of the speech is small and is easily misjudged as noise, thereby achieving the reduction of speech loss in the noise reduction algorithm while reducing noise residue.

[0151] Example 2

[0152] Combine the following Figure 5 , the speech noise reduction device provided by the present invention is described.

[0153] Figure 5 It is a structural schematic diagram of a speech noise reduction device provided according to the second embodiment of the present disclosure.

[0154] like Figure 5 As shown, a speech noise reduction device comprises:

[0155] An acquisition module is used to acquire a source speech signal;

[0156] A short-time Fourier transform module is used to perform short-time Fourier transform on the source speech signal after performing frame division and windowing to obtain an amplitude spectrum of the source speech signal;

[0157] A filter module, partitioning the audio based on a posterior signal-to-noise ratio value and a preset threshold range to obtain multiple audio areas, smoothing the filter parameters corresponding to all the audio areas to obtain updated filter parameters, and using the Wiener filter of the updated filter parameters to perform denoising on the amplitude spectrum to obtain a denoised amplitude spectrum;

[0158] A phase estimation module, which uses an improved Griffin-Lim phase recovery algorithm to perform phase estimation on the amplitude spectrum after noise reduction to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the amplitude spectrum after noise reduction, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm;

[0159] An inverse Fourier transform module integrates the noise-reduced amplitude spectrum and the estimated phase, performs an inverse Fourier transform, and obtains a noise-reduced time domain signal;

[0160] The noise reduction module performs overlapping addition on the noise-reduced time domain signals to obtain a noise-reduced speech signal.

[0161] The specific implementation of the speech noise reduction device provided in the embodiment of the present disclosure is consistent with the speech noise reduction method provided in Example 1, and will not be described in detail here.

[0162] In the above-mentioned embodiment 2, the phase correlation problem is converted into a minimum optimization problem by using the alternating direction multiplier method (ADMM), so as to improve the objective function in the Griffin-Lim algorithm, avoid the problem of non-existence of the solution, and convert the minimum optimization problem into a two-variable optimization problem, thereby simplifying the amount of calculation for each iteration and accelerating the overall solution speed. Subsequently, by introducing the augmented Lagrangian function and using Gaussian iteration to solve the two-variable optimization problem, the convergence of the algorithm can be improved and the number of iterations can be reduced, thereby reducing the complexity of the algorithm, and further reducing computing resources, and applying it to the speech denoising algorithm to improve the intelligibility of the denoised speech;

[0163] At the same time, the audio is partitioned by the posterior signal-to-noise ratio value and the preset threshold range to obtain multiple audio areas, the filter parameters corresponding to all the audio areas are smoothed to obtain updated filter parameters, and the amplitude spectrum is denoised using the Wiener filter with the updated filter parameters to obtain the denoised amplitude spectrum, thereby avoiding the problem that the energy at the tail of the speech is small and is easily misjudged as noise, thereby achieving the reduction of speech loss in the noise reduction algorithm while reducing noise residue.

[0164] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0165] An electronic device includes:

[0166] at least one processor; and

[0167] a memory communicatively connected to the at least one processor; wherein,

[0168] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the speech noise reduction method described in the aforementioned aspect.

[0169] Among them, a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the speech noise reduction method described in the aforementioned aspect.

[0170] A computer program product includes a computer program, and when the computer program is executed by a processor, the speech noise reduction method described in the above aspect is implemented.

[0171] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0172] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0173] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0174] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0175] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0176] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A speech noise reduction method, characterized in that: include: Acquire a source speech signal; Performing short-time Fourier transform on the source speech signal after frame division and windowing to obtain an amplitude spectrum of the source speech signal; Partitioning the audio based on the a posteriori signal-to-noise ratio value and a preset threshold range to obtain a plurality of audio regions, smoothing the filter parameters corresponding to all the audio regions to obtain updated filter parameters, and performing denoising on the amplitude spectrum using a Wiener filter of the updated filter parameters to obtain a denoised amplitude spectrum; An improved Griffin-Lim phase recovery algorithm is used to perform phase estimation on the amplitude spectrum after noise reduction to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the amplitude spectrum after noise reduction, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm; Integrate the noise-reduced amplitude spectrum and the estimated phase, perform inverse Fourier transform, and obtain a noise-reduced time domain signal; Overlapping and adding the noise-reduced time domain signals to obtain a noise-reduced speech signal; The expression of the improved Griffin-Lim phase recovery algorithm is as shown in the following formula: ； ； ； Among them, X (0) =A, ρ>0, λ>0, Z (0) =X (0) Oh, oh (0) =0,m=0,1,…M; X (m+1) It is used to represent the complex spectrum with phase information of the m+1th iteration, A is used to represent the amplitude spectrum without phase information after speech denoising, m represents the iteration subscript, ρ, λ, ω, Z are auxiliary variables, is the soft threshold operator, Z (m+1) ,ω (m+1) represents the corresponding m+1th iteration value, Used to express After performing inverse Fourier transform, perform short-time Fourier transform to obtain the value; The step of obtaining the improved Griffin-Lim phase recovery algorithm specifically includes: The phase correlation problem between the complex spectrum with phase information and the denoised amplitude spectrum is obtained, specifically including: finding the corresponding phase information according to the given denoised amplitude spectrum, and forming a complex spectrum with phase according to the denoised amplitude spectrum and the corresponding phase information, and the complex spectrum needs to meet the preset conditions, as shown in the following formula: ； Among them, A is used to represent the amplitude spectrum after noise reduction, , X is used to represent the complex spectrum, , , , It is used to represent the complex spectrum after short-time Fourier transform, and τ is used to represent the signal time variable. , k is used to indicate the frequency subscript, represents the short-time Fourier transform, It represents the image of short-time Fourier transform; st is the abbreviation of subject to, which is used to indicate being subject to or satisfying preset conditions; represents a set of complex matrices, Used to indicate plural; Among them, based on the alternating direction multiplier method, the above phase correlation problem is converted into a minimum optimization problem, and the minimum optimization problem is shown in the following formula: ； in, ； ； ； Among them, X is used to represent the complex spectrum with phase information, λ is used to represent the penalty factor, represents the Hadamard product, is the L2 norm of the matrix, express norm, represents the short-time Fourier transform, represents inverse Fourier transform; The minimum optimization problem is transformed into a two-variable optimization problem, as shown in the following formula: ； Among them, Z is used to represent the auxiliary variable; The two variable optimization problem is solved using the augmented Lagrangian function, which is shown in the following formula: ； in , For penalty items; In Used to represent vector transpose; The augmented Lagrangian function is solved by using Gaussian iteration method and variable substitution method to obtain an improved Griffin-Lim phase retrieval algorithm; Among them, the variable substitution method is used, The augmented Lagrangian function is improved to obtain the improved augmented Lagrangian function as shown in the following formula: ； The Gaussian iteration method is used to solve the improved augmented Lagrangian function. The calculation process is: Step1: ; Step2: ; Step3: 。 2. The speech noise reduction method according to claim 1, characterized in that: The audio is partitioned based on the a posteriori signal-to-noise ratio value and the preset threshold range to obtain multiple audio areas, and the filter parameters corresponding to all the audio areas are smoothed to obtain updated filter parameters; Specifically include: The posterior signal-to-noise ratio values ​​of all frequencies are averaged to obtain the average parameter corresponding to each frequency; Setting multiple threshold ranges, dividing all frequency points into a speech area, a weak sound area and a noise area by using an average parameter and the threshold ranges; Filter parameters corresponding to the speech area, the weak sound area and the noise area are obtained, and all the filter parameters are smoothed to obtain processed filter parameters.

3. The speech noise reduction method according to claim 2, characterized in that: The multiple threshold ranges are set, and the frequency points are divided into a speech area, a weak sound area and a noise area by using an average parameter and a threshold range; Specifically include: Set the minimum and maximum partition thresholds; When the average parameter of the frequency point is less than the minimum division threshold, dividing the frequency point into the noise area; When the average parameter of the frequency point is greater than the maximum division threshold, dividing the frequency point into the speech area; When the average parameter of the frequency point is less than or equal to the maximum division threshold, and greater than or equal to the minimum division threshold, the frequency point is divided into the weak sound area; As shown in the following formula: ； Among them, Pro is used to indicate the type of area where voice exists, 0 is used to indicate the noise area, 1 is used to indicate the voice area, 2 is used to indicate the weak sound area, and th is used to indicate the average parameter of the frequency point. , K is used to represent the total number of frequency points, Used to represent the amplitude spectrum of noisy speech, Used to represent the amplitude spectrum of the estimated noise signal, th min Used to indicate the minimum partition threshold, th max Used to indicate the maximum partition threshold.

4. The speech noise reduction method according to claim 3, characterized in that: The obtaining of the filter parameters corresponding to the speech area, the weak sound area and the noise area, and smoothing all the filter parameters to obtain the processed filter parameters specifically includes: The filter parameters corresponding to the speech area, the weak sound area and the noise area are obtained, as shown in the following formula: ； in, , and is a constant, satisfying , Used to represent filter parameters; All the filter parameters are smoothed to obtain the processed filter parameters, as shown in the following formula: ； Among them, β is used to represent the smoothing factor, 0<β<1.

5. A speech noise reduction device, characterized in that: include: An acquisition module is used to acquire a source speech signal; A short-time Fourier transform module is used to perform short-time Fourier transform on the source speech signal after performing frame division and windowing to obtain an amplitude spectrum of the source speech signal; A filter module, partitioning the audio based on a posterior signal-to-noise ratio value and a preset threshold range to obtain a plurality of audio regions, smoothing the filter parameters corresponding to all the audio regions to obtain updated filter parameters, and performing denoising on the amplitude spectrum using a Wiener filter of the updated filter parameters to obtain a denoised amplitude spectrum; A phase estimation module, which uses an improved Griffin-Lim phase recovery algorithm to perform phase estimation on the amplitude spectrum after noise reduction to obtain a corresponding estimated phase; wherein the step of obtaining the improved Griffin-Lim phase recovery algorithm includes: obtaining a phase correlation problem between a complex spectrum with phase information and the amplitude spectrum after noise reduction, converting the phase correlation problem into a minimum optimization problem based on an alternating direction multiplier method, converting the minimum optimization problem into a two-variable optimization problem, and solving the two-variable optimization problem based on an augmented Lagrangian function and Gaussian iteration to obtain an improved Griffin-Lim phase recovery algorithm; An inverse Fourier transform module integrates the noise-reduced amplitude spectrum and the estimated phase, performs an inverse Fourier transform, and obtains a noise-reduced time domain signal; Obtain a denoised time domain module, and perform overlapping addition on the denoised time domain signal to obtain a denoised speech signal; The expression of the improved Griffin-Lim phase recovery algorithm is as shown in the following formula: ； ； ； Among them, X (0) =A, ρ>0, λ>0, Z (0) =X (0) Oh, oh (0) =0,m=0,1,…M; X (m+1) It is used to represent the complex spectrum with phase information of the m+1th iteration, A is used to represent the amplitude spectrum without phase information after speech denoising, m represents the iteration subscript, ρ, λ, ω, Z are auxiliary variables, is the soft threshold operator, Z (m+1) ,ω (m+1) represents the corresponding m+1th iteration value, Used to express After performing inverse Fourier transform, perform short-time Fourier transform to obtain the value; The step of obtaining the improved Griffin-Lim phase recovery algorithm specifically includes: The phase correlation problem between the complex spectrum with phase information and the denoised amplitude spectrum is obtained, specifically including: finding the corresponding phase information according to the given denoised amplitude spectrum, and forming a complex spectrum with phase according to the denoised amplitude spectrum and the corresponding phase information, and the complex spectrum needs to meet the preset conditions, as shown in the following formula: ； Among them, A is used to represent the amplitude spectrum after noise reduction, , X is used to represent the complex spectrum, , , , It is used to represent the complex spectrum after short-time Fourier transform, and τ is used to represent the signal time variable. , k is used to indicate the frequency subscript, represents the short-time Fourier transform, It represents the image of short-time Fourier transform; st is the abbreviation of subject to, which is used to indicate being subject to or satisfying preset conditions; represents a set of complex matrices, Used to indicate plural; Among them, based on the alternating direction multiplier method, the above phase correlation problem is converted into a minimum optimization problem, and the minimum optimization problem is shown in the following formula: ； in, ； ； ； Among them, X is used to represent the complex spectrum with phase information, λ is used to represent the penalty factor, represents the Hadamard product, is the L2 norm of the matrix, express norm, represents the short-time Fourier transform, represents inverse Fourier transform; The minimum optimization problem is transformed into a two-variable optimization problem, as shown in the following formula: ； Among them, Z is used to represent the auxiliary variable; The two variable optimization problem is solved using the augmented Lagrangian function, which is shown in the following formula: ； in , For penalty items; In Used to represent vector transpose; The augmented Lagrangian function is solved by using Gaussian iteration method and variable substitution method to obtain an improved Griffin-Lim phase retrieval algorithm; Among them, the variable substitution method is used, The augmented Lagrangian function is improved to obtain the improved augmented Lagrangian function as shown in the following formula: ； The Gaussian iteration method is used to solve the improved augmented Lagrangian function. The calculation process is: Step1: ; Step2: ; Step3: 。 6. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 4.

8. A computer program product, characterized in that The invention comprises a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 4.

Images