A method, device and storage medium for noise reduction processing of speech signals

Abstract

The present application relates to the technical field of speech noise reduction, and discloses a method, device and computer-readable storage medium for noise reduction processing of speech signals. The method comprises: preprocessing a speech signal to be processed to obtain first feature data; processing the first feature data using a preset noise reduction algorithm to filter the stationary noise in the speech signal to be processed and obtain second feature data; inputting the first feature data and the second feature data into a preset noise reduction network to filter the transient noise in the speech signal to be processed and obtain third feature data; determining the speech signal after noise reduction processing according to the first feature data, the second feature data and the third feature data. Through the above method, the preset noise reduction algorithm is used to filter the stationary noise in the speech signal to be processed, and the preset noise reduction network is used to filter the transient noise in the speech signal to be processed, which can combine the advantages of traditional noise reduction and deep learning noise reduction to achieve good noise reduction effect.

Description

Technical Field

[0001] The present application relates to the technical field of speech noise reduction, and in particular to a method and device for noise reduction processing of a speech signal, and a computer-readable storage medium. Background Art

[0002] Speech noise reduction refers to the technology of extracting useful speech signals (or pure speech signals) from noisy speech signals as much as possible and suppressing or reducing noise interference when the speech signals are interfered with or even drowned by various background noises. In order to improve the call quality of mobile phones, headphones and other devices, noise reduction is usually performed on the original data collected by the microphone.

[0003] According to different research ideas, current noise reduction methods mainly include traditional signal processing algorithms and deep learning algorithms. Among them, traditional signal processing algorithms are mostly derived from physical and mathematical principles. Such algorithms cannot eliminate non-steady-state noise, and speech damage is large at low signal-to-noise ratios; while deep learning algorithms use a large amount of speech data or noise data to train the network to learn related features to achieve noise reduction. Such algorithms have the disadvantages of large network structure parameters, large amount of calculation, large delay and poor generalization, and cannot achieve good results with small resources. Summary of the invention

[0004] The main technical problem solved by the present application is to provide a method, device and computer-readable storage medium for noise reduction processing of speech signals, which can solve the problem that the traditional signal processing algorithm and deep learning algorithm in the existing noise reduction methods have poor noise reduction effect when used separately.

[0005] In order to solve the above technical problems, a technical solution adopted by the present application is: to provide a method for noise reduction processing of a speech signal, the method comprising:

[0006] Preprocess the speech signal to be processed to obtain first feature data; use a preset noise reduction algorithm to process the first feature data to filter out stationary noise in the speech signal to be processed and obtain second feature data; input the first feature data and the second feature data into a preset noise reduction network to filter out transient noise in the speech signal to be processed and obtain third feature data; determine the speech signal after noise reduction processing based on the first feature data, the second feature data and the third feature data.

[0007] Optionally, in the above method, preprocessing the speech signal to be processed to obtain first feature data includes:

[0008] The speech signal to be processed is preprocessed to obtain first amplitude data and phase data.

[0009] Optionally, in the above method, a preset noise reduction algorithm is used to process the first feature data to filter stationary noise in the speech signal to be processed, and obtain the second feature data, including:

[0010] The first amplitude data is processed using a preset noise reduction algorithm to filter stationary noise in the speech signal to be processed, and second amplitude data and speech existence probability are obtained.

[0011] Optionally, in the above method, the first feature data and the second feature data are input into a preset noise reduction network to filter transient noise in the speech signal to be processed, and obtain third feature data, including:

[0012] The first amplitude data and the second amplitude data are input into a preset noise reduction network to filter transient noise in the speech signal to be processed and obtain a first time-frequency mask.

[0013] Optionally, in the above method, determining the speech signal after noise reduction processing according to the first feature data, the second feature data and the third feature data includes:

[0014] A speech signal after noise reduction processing is determined according to the first amplitude data, the phase data, the probability of speech existence, and the first time-frequency mask.

[0015] Optionally, in the above method, determining the speech signal after noise reduction processing according to the first amplitude data, the phase data, the speech existence probability, and the first time-frequency mask includes:

[0016] The first time-frequency mask is processed according to the probability of speech existence to obtain a second time-frequency mask; the first amplitude data is processed according to the second time-frequency mask to obtain third amplitude data; and the speech signal after noise reduction processing is determined according to the third amplitude data and phase data.

[0017] Optionally, in the above method, processing the first time-frequency mask according to the speech existence probability to obtain the second time-frequency mask includes:

[0018] Obtain the speech existence probability corresponding to the target frequency band and the speech existence probability corresponding to the effective frequency band; when the speech existence probability corresponding to the target frequency band is less than a first lower limit threshold and the speech existence probability corresponding to the effective frequency band is less than a second lower limit threshold, reduce the first time-frequency mask corresponding to the target frequency band to obtain a second time-frequency mask; or when the speech existence probability corresponding to the target frequency band is greater than a first upper limit threshold or the speech existence probability corresponding to the effective frequency band is greater than a second upper limit threshold, increase the first time-frequency mask corresponding to the target frequency band to obtain a second time-frequency mask.

[0019] Optionally, in the above method, processing the first time-frequency mask according to the speech existence probability to obtain the second time-frequency mask includes:

[0020] Based on formula (1), the first time-frequency mask is weighted and modified using the probability of speech presence to obtain a second time-frequency mask;

[0021]

[0022] Among them, β 1 , β 2 , α 1 , α 2 , α 3 and α 4 are pre-set hyperparameters. α1, α2, α3, and α4 are all floating point numbers in the range of 0-1, and β1 and β2 are both floating point numbers greater than 1. Where N is the number of effective frequency bands after Fourier transform,

[0023] Optionally, in the above method, processing the first amplitude data according to the second time-frequency mask to obtain third amplitude data includes:

[0024] The second time-frequency mask is multiplied by the first amplitude data to obtain third amplitude data.

[0025] Optionally, in the above method, determining the speech signal after noise reduction processing according to the third amplitude data and the phase data includes:

[0026] Determine the speech spectrum data after noise reduction processing according to the third amplitude data and the phase data; perform inverse Fourier transform and overlap-addition on the speech spectrum data to obtain the speech signal after noise reduction processing.

[0027] Optionally, in the above method, before inputting the first feature data and the second feature data into the preset noise reduction network, the method further includes:

[0028] The first feature data and the second feature data are logarithmized to compress the dynamic range of the first feature data and the second feature data.

[0029] Optionally, in the above method, the preset noise reduction network is obtained by training the speech noise reduction network using a training speech set; wherein, training the speech noise reduction network using the training speech set includes:

[0030] A training speech set containing only transient noise and a clean speech set corresponding to the training speech set are obtained; the training speech set is input into a speech denoising network to output a corresponding denoised speech set; and the difference between the clean speech set and the denoised speech set is used to modify the parameters of the speech denoising network.

[0031] In order to solve the above technical problems, another technical solution adopted in the present application is: to provide a noise reduction processing device for a speech signal, the device includes a processor and a memory.

[0032] Specifically, the memory is used to store program instructions, and the processor is used to execute the program instructions to implement the above-mentioned noise reduction processing method for speech signals.

[0033] In order to solve the above technical problems, another technical solution adopted by the present application is: providing a computer-readable storage medium, the computer-readable storage medium storing program instructions, and the program instructions can be executed to implement the above-mentioned noise reduction processing method for speech signals.

[0034] Different from the prior art, the present application provides a method, device and computer-readable storage medium for noise reduction processing of speech signals, the method comprising: preprocessing the speech signal to be processed to obtain first feature data; using a preset noise reduction algorithm to process the first feature data to filter the stationary noise in the speech signal to be processed and obtain second feature data; inputting the first feature data and the second feature data into a preset noise reduction network to filter the transient noise in the speech signal to be processed and obtain third feature data; determining the speech signal after noise reduction processing according to the first feature data, the second feature data and the third feature data. Through the above method, the preset noise reduction algorithm is used to filter the stationary noise in the speech signal to be processed, and the preset noise reduction network is used to filter the transient noise in the speech signal to be processed, which can fully combine the advantages of traditional noise reduction and deep learning noise reduction, and achieve good noise reduction effects on resource-constrained embedded devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0036] Figure 1 It is a flowchart of the first embodiment of the method for noise reduction processing of speech signals provided by the present application;

[0037] Figure 2 It is a flowchart of a second embodiment of a method for noise reduction processing of a speech signal provided by the present application;

[0038] Figure 3 yes Figure 2 Schematic diagram of the process of S24;

[0039] Figure 4 yes Figure 3 Schematic diagram of the process of S241;

[0040] Figure 5 yes Figure 3 Schematic diagram of the process of S243;

[0041] Figure 6 It is a flow chart of training a speech denoising network in an embodiment of a denoising method for speech signals provided in the present application;

[0042] Figure 7 It is a structural schematic diagram of an embodiment of a noise reduction processing device for a speech signal provided by the present application;

[0043] Figure 8 It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It will be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some but not all structures related to the present application are shown in the drawings. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0045] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0046] The steps in the embodiments of the present application do not necessarily have to be processed in the described order of steps. The steps can be selectively rearranged, or the steps in the embodiments can be deleted, or the steps in the embodiments can be added as needed. The step descriptions in the embodiments of the present application are only optional sequence combinations and do not represent all step sequence combinations in the embodiments of the present application. The order of steps in the embodiments cannot be considered as a limitation of the present application.

[0047] The term "and / or" in the embodiments of the present application refers to any and all possible combinations of one or more of the associated enumerated items. It should also be noted that when used in this specification, "include / comprise" specifies the presence of stated features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements and / or components and / or their groups.

[0048] The terms "first", "second", etc. in this application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0049] Among speech denoising methods, traditional denoising methods based on physical models cannot eliminate non-steady-state noise, and speech damage is severe at low signal-to-noise ratios. Deep learning-based denoising methods have the disadvantages of large network structure parameters, large amount of computation, large delay, and poor generalization, and cannot achieve good results with small resources.

[0050] Based on this, the present application proposes a method for denoising speech signals. The method uses a preset denoising algorithm to filter out stationary noise in the speech signal to be processed, and uses a preset denoising network to filter out transient noise in the speech signal to be processed. The method can fully combine the advantages of traditional denoising and deep learning denoising, and achieve good denoising effects on resource-constrained embedded devices.

[0051] See also Figure 1 , Figure 1 1 is a flow chart of a first embodiment of a method for noise reduction processing of a speech signal provided by the present application, the method comprising:

[0052] S11: Preprocess the speech signal to be processed to obtain first feature data.

[0053] Optionally, the speech signal to be processed includes speech and corresponding noise interference, and can be collected by a microphone of an embedded device such as a mobile phone or a headset, which is not specifically limited here.

[0054] Optionally, the method for preprocessing the speech signal to be processed may include operations such as framing, windowing, Fourier transform, and taking absolute value, which are not specifically limited here. In this embodiment, the purpose of preprocessing is to obtain the first feature data.

[0055] S12: Using a preset noise reduction algorithm to process the first feature data to filter out stationary noise in the speech signal to be processed, and obtain second feature data.

[0056] Optionally, traditional single-channel noise reduction methods include general subtraction, Wiener filtering, noise reduction methods based on statistical models, etc., among which the typical representatives are the OMLSA algorithm and the noise reduction algorithm in Webrtc.

[0057] In this embodiment, the OMLSA algorithm is mainly used to process the first feature data to filter the stationary noise in the speech signal to be processed and obtain the second feature data.

[0058] S13: Inputting the first feature data and the second feature data into a preset noise reduction network to filter transient noise in the speech signal to be processed and obtain third feature data.

[0059] Optionally, a preset denoising network is obtained by learning from a large-scale data set. Typical denoising network models include Tasnet, Phasen, Dccrn, DeepXi, etc. Denoising methods based on deep learning can be roughly divided into three categories: spectrum mapping-based methods, time-frequency mask-based methods, and end-to-end methods.

[0060] In this embodiment, the first feature data and the second feature data are input into a preset noise reduction network to filter the transient noise in the speech signal to be processed, and obtain the third feature data.

[0061] Optionally, in some other embodiments, before inputting the first feature data and the second feature data into the preset noise reduction network, the method further includes: taking the logarithm of the first feature data and the second feature data to compress the dynamic range of the first feature data and the second feature data.

[0062] S14: Determine the speech signal after noise reduction processing according to the first feature data, the second feature data and the third feature data.

[0063] Compared with the prior art, this embodiment provides a method for noise reduction processing of a speech signal, the method comprising: preprocessing the speech signal to be processed to obtain first feature data; using a preset noise reduction algorithm to process the first feature data to filter the stationary noise in the speech signal to be processed and obtain second feature data; inputting the first feature data and the second feature data into a preset noise reduction network to filter the transient noise in the speech signal to be processed and obtain third feature data; determining the speech signal after noise reduction processing according to the first feature data, the second feature data and the third feature data. Through the above method, the preset noise reduction algorithm is used to filter the stationary noise in the speech signal to be processed, and the preset noise reduction network is used to filter the transient noise in the speech signal to be processed, which can fully combine the advantages of traditional noise reduction and deep learning noise reduction, and achieve good noise reduction effects on resource-constrained embedded devices.

[0064] See also Figure 2 , Figure 2 1 is a flow chart of a second embodiment of a method for noise reduction processing of a speech signal provided by the present application, the method comprising:

[0065] S21: Preprocess the speech signal to be processed to obtain first amplitude data and phase data.

[0066] Optionally, operations such as framing, windowing, Fourier transform, and absolute value taking are performed on the speech signal to be processed to obtain first amplitude data and phase data.

[0067] S22: Using a preset noise reduction algorithm to process the first amplitude data to filter out stationary noise in the speech signal to be processed, and obtain second amplitude data and speech existence probability.

[0068] Optionally, the first amplitude data is processed using an OMLSA algorithm to filter stationary noise in the speech signal to be processed, and obtain second amplitude data and speech existence probability.

[0069] Optionally, the process of the OMLSA algorithm is: first use the MCRA method to estimate the noise, then estimate the posterior signal-to-noise ratio and the prior signal-to-noise ratio (DD criterion), use the soft decision of the prior signal-to-noise ratio to estimate the prior speech absence probability, and then use the Bayesian criterion to calculate the conditional speech existence probability, and substitute the previously calculated value into the gain function expression to obtain the gain value.

[0070] S23: Inputting the first amplitude data and the second amplitude data into a preset noise reduction network to filter transient noise in the speech signal to be processed and obtain a first time-frequency mask.

[0071] Optionally, the preset noise reduction network in this embodiment adopts a method based on time-frequency masking, the core idea of ​​which is to predict the time-frequency mask by training a deep neural network, which reflects the degree of noise suppression on each time-frequency unit, and then apply the predicted time-frequency mask to the spectrum of the input noisy speech to reconstruct the pure speech signal. Common time-frequency masks include ideal binary mask, ideal ratio mask, phase-sensitive mask, complex ratio mask, etc.

[0072] Optionally, the first amplitude data and the second amplitude data are input into a preset noise reduction network to filter transient noise in the speech signal to be processed and obtain a first time-frequency mask.

[0073] Optionally, in some other embodiments, before inputting the first amplitude data and the second amplitude data into the preset noise reduction network, the method further includes: taking the logarithm of the first amplitude data and the second amplitude data to compress the dynamic range of the first amplitude data and the second amplitude data.

[0074] S24: Determine the speech signal after noise reduction processing according to the first amplitude data, the phase data, the speech existence probability, and the first time-frequency mask.

[0075] See also Figure 3 , Figure 3 yes Figure 2 In the flowchart of S24, S24 may further include:

[0076] S241: Process the first time-frequency mask according to the speech existence probability to obtain a second time-frequency mask.

[0077] Optionally, the first time-frequency mask is processed according to the speech existence probability obtained by processing the first amplitude data with a preset noise reduction algorithm to obtain a second time-frequency mask.

[0078] S242: Process the first amplitude data according to the second time-frequency mask to obtain third amplitude data.

[0079] Optionally, the second time-frequency mask is multiplied by the first amplitude data to obtain third amplitude data.

[0080] S243: Determine the speech signal after noise reduction processing according to the third amplitude data and the phase data.

[0081] Optionally, the third amplitude data and the phase data may be combined to determine a speech signal after noise reduction processing.

[0082] See also Figure 4 , Figure 4 yes Figure 3 In the flowchart of S241, S241 may further include:

[0083] S2411: Obtain the speech existence probability corresponding to the target frequency band and the speech existence probability corresponding to the effective frequency band.

[0084] Optionally, the speech existence probability corresponding to the target frequency band and the speech existence probability corresponding to the effective frequency band can be obtained by processing the first amplitude data according to a preset noise reduction algorithm.

[0085] S2412: When the speech existence probability corresponding to the target frequency band is less than a first lower threshold and the speech existence probability corresponding to the effective frequency band is less than a second lower threshold, reduce the first time-frequency mask corresponding to the target frequency band to obtain a second time-frequency mask.

[0086] Optionally, when the probability of speech existence corresponding to the target frequency band is less than the first lower limit threshold and the probability of speech existence corresponding to the effective frequency band is less than the second lower limit threshold, it means that the probability that the target frequency band is noise is high, and the first time-frequency mask corresponding to the target frequency band should be reduced to obtain the second time-frequency mask.

[0087] Optionally, when the speech existence probability corresponding to the target frequency band is greater than the first upper threshold or the speech existence probability corresponding to the effective frequency band is greater than the second upper threshold, S2413 is executed.

[0088] S2413: When the speech existence probability corresponding to the target frequency band is greater than the first upper threshold or the speech existence probability corresponding to the effective frequency band is greater than the second upper threshold, increase the first time-frequency mask corresponding to the target frequency band to obtain a second time-frequency mask.

[0089] Optionally, when the probability of speech existence corresponding to the target frequency band is greater than the first upper limit threshold or the probability of speech existence corresponding to the effective frequency band is greater than the second upper limit threshold, it means that the probability that the target frequency band is speech is high, and the first time-frequency mask corresponding to the target frequency band should be increased to obtain the second time-frequency mask.

[0090] In an optional embodiment, based on formula (1), the first time-frequency mask is weighted and modified using the probability of speech presence to obtain a second time-frequency mask;

[0091]

[0092] Among them, β 1 , β 2 , α 1 , α 2 , α 3 and α 4 are pre-set hyperparameters. α1, α2, α3, and α4 are all floating point numbers in the range of 0-1, and β1 and β2 are both floating point numbers greater than 1. Where N is the number of effective frequency bands after Fourier transform,

[0093] It can be understood that formula (1) is only one way to process the first time-frequency mask according to the probability of speech existence to obtain the second time-frequency mask. Other functions that meet this characteristic can also achieve this purpose, which is not specifically limited here.

[0094] See also Figure 5 , Figure 5 yes Figure 3 In the flowchart of S243, S243 may further include:

[0095] S2431: Determine the speech spectrum data after noise reduction processing according to the third amplitude data and the phase data.

[0096] Optionally, the third amplitude data and the phase data are combined to determine the speech spectrum data after noise reduction processing.

[0097] S2432: Perform inverse Fourier transform and overlap-addition on the speech spectrum data to obtain a speech signal after noise reduction processing.

[0098] Optionally, the speech spectrum data is subjected to inverse Fourier transform and overlap-addition to obtain a speech signal after noise reduction processing.

[0099] See also Figure 6 , Figure 6 1 is a flow chart of training a speech denoising network in an embodiment of a speech signal denoising processing method provided by the present application, the method comprising:

[0100] S31: Acquire a training speech set containing only transient noise and a clean speech set corresponding to the training speech set.

[0101] Optionally, in order to miniaturize the speech denoising network and achieve better training results, the acquired training speech set contains only transient noise, and a clean speech set containing no noise corresponding to the training speech set is acquired.

[0102] S32: Input the training speech set into the speech denoising network to output a corresponding denoised speech set.

[0103] Optionally, the training speech set is input into the speech denoising network, and after being processed by the speech denoising network, a corresponding denoised speech set can be output.

[0104] Optionally, before the training speech set is input into the speech denoising network, the dynamic range of the training speech set may be compressed, for example, by taking a logarithmic method to compress the dynamic range, so as to obtain a better convergence result.

[0105] S33: Utilizing the difference between the clean speech set and the noise reduction speech set, the parameters of the speech noise reduction network are modified.

[0106] Optionally, there is a difference between the clean speech set and the denoised speech set output by the speech denoising network. The parameters of the speech denoising network can be adjusted according to the difference, and the training process can be repeated to minimize the difference between the clean speech set and the denoised speech set. The speech denoising network obtained at this time is the preset denoising network required by the denoising processing method for speech signals provided in this application.

[0107] See also Figure 7 , Figure 7 4 is a schematic diagram of the structure of an embodiment of a noise reduction processing device for speech signals provided in the present application. The device 400 includes a processor 401 and a memory 402.

[0108] Specifically, the memory 402 is used to store program instructions, and the processor 401 is used to execute the program instructions to implement the method provided by any one or any non-conflicting combination of the above embodiments.

[0109] Optionally, processor 401 is a central processing unit (CPU), which is one of the main devices of an electronic computer and a core component in a computer. Its main function is to interpret computer instructions and process data in computer software. CPU is the core component in a computer responsible for reading instructions, decoding instructions and executing instructions. The central processing unit mainly includes two parts, namely a controller and an arithmetic unit, which also includes a cache memory and a data and control bus that realizes the connection between them. The functions of the central processing unit are mainly to process instructions, perform operations, control time, and process data. In computer architecture, the CPU is a core hardware unit that controls and allocates all hardware resources of the computer (such as memory, input and output units) and performs general operations. The CPU is the computing and control core of the computer. The operations of all software layers in the computer system will eventually be mapped to the operations of the CPU through the instruction set.

[0110] Optionally, the memory 402 is a read-only memory (ROM) or a random access memory (RAM), which is a memory device in a computer system, mainly used to store programs and data. All information in the computer, including input raw data, computer programs, intermediate operation results and final operation results, are stored in the memory. It stores and retrieves information according to the location specified by the controller.

[0111] In an optional embodiment, the noise reduction processing device 400 is a mobile phone. When the user makes a voice call, the voice signal is obtained through the microphone, and the voice signal is subjected to noise reduction processing according to the method provided by any one or any non-conflicting combination of the above embodiments, so that the user's call quality is clearer and a good experience is brought to the user.

[0112] See also Figure 8 , Figure 8 1 is a schematic diagram of a computer-readable storage medium according to an embodiment of the present application, wherein the computer-readable storage medium 500 includes program instructions 501, and the program instructions 501 can be executed to implement the method provided by any one or any non-conflicting combination of the above embodiments. The capacity of the computer-readable storage medium 500 can meet the requirements for storing the program instructions 501.

[0113] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-readable storage media 500 (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by the computer-readable storage medium 500. These computer-readable storage media 500 can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the program instructions 501 executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0115] These computer-readable storage media 500 may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the program instructions 501 stored in the computer-readable storage medium 500 produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0116] These computer-readable storage media 500 can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing program instructions 501 executed on the computer or other programmable device for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0117] The above description is only an implementation method of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made according to the description and drawings of the present application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for noise reduction of speech signals, characterized in that: The method comprises: Preprocessing the speech signal to be processed to obtain first amplitude data and phase data; Processing the first amplitude data using a preset noise reduction algorithm to filter stationary noise in the speech signal to be processed, and obtaining second amplitude data and a probability of speech existence; Inputting the first amplitude data and the second amplitude data into a preset noise reduction network to filter transient noise in the speech signal to be processed and obtain a first time-frequency mask; Processing the first time-frequency mask according to the speech existence probability to obtain a second time-frequency mask; Processing the first amplitude data according to the second time-frequency mask to obtain third amplitude data; A speech signal after noise reduction processing is determined according to the third amplitude data and the phase data.

2. The method according to claim 1, characterized in that The processing of the first time-frequency mask according to the speech existence probability to obtain a second time-frequency mask includes: Obtaining the speech existence probability corresponding to the target frequency band and the speech existence probability corresponding to the effective frequency band; When the speech existence probability corresponding to the target frequency band is less than the first lower threshold and the speech existence probability corresponding to the effective frequency band is less than the second lower threshold, reducing the first time-frequency mask corresponding to the target frequency band to obtain a second time-frequency mask; or When the speech existence probability corresponding to the target frequency band is greater than the first upper threshold or the speech existence probability corresponding to the effective frequency band is greater than the second upper threshold, the first time-frequency mask corresponding to the target frequency band is increased to obtain a second time-frequency mask.

3. The method according to claim 1, characterized in that The processing of the first time-frequency mask according to the speech existence probability to obtain a second time-frequency mask includes: Based on formula (1), weighted modification is performed on the first time-frequency mask using the speech existence probability to obtain a second time-frequency mask; Formula (1): in, , , , , and are pre-set hyperparameters. α1, α2, α3, and α4 are all floating point numbers in the range of 0-1, and β1 and β2 are both floating point numbers greater than 1. , where N is the number of effective frequency bands after Fourier transform, .

4. The method according to claim 1, characterized in that: The processing of the first amplitude data according to the second time-frequency mask to obtain third amplitude data includes: The second time-frequency mask is multiplied by the first amplitude data to obtain third amplitude data.

5. The method according to claim 1, characterized in that The step of determining the speech signal after noise reduction processing according to the third amplitude data and the phase data comprises: Determining speech spectrum data after noise reduction processing according to the third amplitude data and the phase data; The speech spectrum data is subjected to inverse Fourier transform and overlap-addition to obtain a speech signal after noise reduction processing.

6. The method according to claim 1, characterized in that Before inputting the first amplitude data and the second amplitude data into a preset noise reduction network, the method further includes: The first amplitude data and the second amplitude data are logarithmized to compress the dynamic range of the first amplitude data and the second amplitude data.

7. The method according to claim 1, characterized in that The preset noise reduction network is obtained by training the speech noise reduction network using a training speech set; The method of training the speech denoising network using the training speech set includes: Acquire a training speech set containing only transient noise and a clean speech set corresponding to the training speech set; Inputting the training speech set into the speech denoising network to output a corresponding denoised speech set; The parameters of the speech denoising network are modified by using the difference between the clean speech set and the denoised speech set.

8. A noise reduction processing device for a speech signal, characterized in that: The noise reduction processing device comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to execute the program instructions to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: Program instructions are stored, and the program instructions can be executed to implement the method according to any one of claims 1 to 7.

Images