Voice Denoising Method, Device, Electronic Device and Storage Medium

Abstract

An embodiment of the present application provides a method, device, electronic device, and storage medium for voice denoising. The method includes: performing frame division processing on the acquired original noisy voice signal to obtain a plurality of voice frames; obtaining first feature information and second feature information of a first voice frame, where the first voice frame is any one of the plurality of voice frames, the first feature information includes the envelope spectrum feature of the first voice frame, and the second feature information includes feature information obtained according to the pitch feature of the first voice frame; obtaining a target frequency point gain corresponding to the first voice frame according to the first feature information and the second feature information; performing denoising processing on the first voice frame based on the target frequency point gain to obtain a target voice signal. This method can effectively remove the noise signal in the voice signal without damaging the voice.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of speech processing technology, and more particularly, to a speech denoising method, apparatus, electronic device, and computer-readable storage medium. Background Art

[0002] During a voice call, the voice signal is usually contaminated by surrounding noise signals, resulting in the receiving party being unable to hear the pure voice signal, thus greatly hindering the smoothness and comfort of communication. Therefore, how to conveniently and effectively remove the noise signal from the voice signal has attracted more and more attention.

[0003] In the related art, generally, the voice signal and the noise signal are distinguished based on their different characteristics, and then the noise component in the noisy voice signal is estimated and removed to denoise the voice. For example, the traditional voice denoising method based on voice activity detection (VAD) estimates the noise component in the non-speech segment and then deletes it from the noisy voice to achieve the effect of voice denoising.

[0004] Currently, when performing voice denoising processing based on deep learning technology, it is usually based on a large network model to estimate the pure voice signal from the noisy voice signal. However, this type of method usually requires voice denoising processing based on a large network model. After shrinking this type of network model to fit a mobile terminal, performing voice denoising processing based on the shrunk network model often leads to a sharp decline in performance and may even cause problems such as damaging the voice. Summary of the Invention

[0005] An object of the present disclosure is to provide a new technical solution for removing the noise signal from the voice signal to solve the problem that the existing voice denoising method may have a performance decline or even damage to the voice.

[0006] In a first aspect of the present disclosure, a voice denoising method is provided, and the method includes:

[0007] Performing frame division processing on the obtained original noisy voice signal to obtain a plurality of voice frames;

[0008] Obtaining first feature information and second feature information of a first voice frame, where the first voice frame is any one of the plurality of voice frames, the first feature information includes the envelope spectrum feature of the first voice frame, and the second feature information includes the feature information obtained according to the pitch feature of the first voice frame;

[0009] Obtaining a target frequency point gain corresponding to the first voice frame according to the first feature information and the second feature information;

[0010] Perform denoising processing on the first speech frame based on the target frequency point gain to obtain a target speech signal.

[0011] In a second aspect of the present disclosure, a speech denoising device is further provided, including:

[0012] A frame segmentation processing module, configured to perform frame segmentation processing on the acquired original noisy speech signal to obtain a plurality of speech frames;

[0013] A feature extraction module, configured to obtain first feature information and second feature information of a first speech frame, where the first speech frame is any one of the plurality of speech frames, the first feature information includes the envelope spectrum feature of the first speech frame, and the second feature information includes the feature information obtained according to the pitch feature of the first speech frame;

[0014] A gain prediction module, configured to obtain a target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information;

[0015] A denoising processing module, configured to perform denoising processing on the first speech frame based on the target frequency point gain to obtain a target speech signal.

[0016] In a third aspect of the present disclosure, an electronic device is further provided, including:

[0017] A memory, configured to store executable instructions;

[0018] A processor, configured to run the electronic device to execute the method according to the first aspect of the present disclosure under the control of the executable instructions.

[0019] In a fourth aspect of the present disclosure, a computer-readable storage medium is further provided. The computer-readable storage medium stores a computer program that can be read and executed by a computer. The computer program is configured to execute the method according to the first aspect of the present disclosure when being read and run by the computer.

[0020] One beneficial effect of the embodiments of the present disclosure is that, according to the embodiments of the present disclosure, by performing frame segmentation on the original noisy speech signal, and respectively obtaining the first feature information and the second feature information of each speech frame, then, based on the first feature information and the second feature information, obtaining the target frequency point gain corresponding to the first speech frame, and denoising the first speech frame based on the target frequency point gain, the target speech signal can be obtained. In the embodiments of the present disclosure, by constructing the first feature information based on the envelope spectrum feature of the speech frame and constructing the second feature information based on the pitch feature of the speech frame, using the envelope gain estimated from the first feature information as the guiding information, it is possible to take into account the detailed estimation of the speech signal while effectively reducing the size of the network model, so as to achieve the effect of effectively removing the noise signal without damaging the speech. Therefore, denoising the speech signal generated during the user's call process based on this speech denoising method can provide a more comfortable call communication experience.

[0021] Other features and advantages of the present specification will become clear through the following detailed description of the exemplary embodiments of the present specification with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings incorporated in the specification and constituting a part of the specification illustrate the embodiments of the present specification, and together with the description thereof are used to explain the principles of the present specification.

[0023] Figure 1 is a schematic flowchart of a speech denoising method provided by an embodiment of the present disclosure.

[0024] Figure 2 is a schematic processing diagram for obtaining the second feature information provided by an embodiment of the present disclosure.

[0025] Figure 3 is a schematic processing diagram for obtaining the target frequency point gain provided by an embodiment of the present disclosure.

[0026] Figure 4 is a schematic block diagram of a speech denoising device provided by an embodiment of the present disclosure.

[0027] Figure 5 is a schematic hardware structure diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] Now, various exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. It should be noted that: unless otherwise specifically stated, the relative arrangements, numerical expressions, and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0029] The following description of at least one exemplary embodiment is merely illustrative and is in no way a limitation on the present invention, its application, or its use.

[0030] Technologies, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such technologies, methods, and devices should be considered as part of the specification.

[0031] In all examples shown and discussed herein, any specific values should be construed as merely exemplary and not as a limitation. Thus, other examples of the exemplary embodiments may have different values.

[0032] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0033] <Method Embodiment>

[0034] To solve the problem that when performing speech denoising, it is necessary to rely on a large network model structure, and in the case of reducing the network model, there may be a performance degradation or even damage to the speech, the embodiments of the present disclosure provide a speech denoising method. Please refer to Figure 1 , which is a schematic flowchart of the speech denoising method provided by the embodiments of the present disclosure. This method can be implemented by an electronic device, which can be a terminal device, for example, a mobile phone, a tablet computer, etc.; of course, this electronic device can also be a server, for example, a blade server, a rack server, etc., and no special limitation is made here.

[0035] As Figure 1 described, the method of this embodiment may include the following steps S1100 - S1400, which are described in detail below.

[0036] Step S1100, perform frame splitting on the acquired original noisy speech signal to obtain a plurality of speech frames.

[0037] Specifically, after the original noisy speech signal is acquired, since the original noisy speech signal is usually a non - stationary audio data stream, generally, it is necessary to perform frame splitting on the original noisy speech signal to obtain speech frames corresponding to the original noisy speech signal and composed of a plurality of quasi - stationary signals, and then perform denoising processing on the plurality of quasi - stationary speech frames to obtain the target speech signal.

[0038] In specific implementation, the original noisy speech signal can be framed in an overlapping segmented manner so that there is a smooth transition between frames and its continuity is maintained. Among them, in the overlapping segmented manner, the overlapping part between the previous frame and the next frame is called the frame shift. In specific implementation, the speech signal can be framed by using a weighted method with a movable finite-length window. Generally, the frame length of a frame can be taken as 20 - 50 ms. The window functions that can be used include but are not limited to any one of the following window functions: Rectangle window, Hanning window, Hamming window, and Blackman window. There is no special limitation on how to perform the framing process here.

[0039] Step S1200, obtain the first feature information and the second feature information of the first speech frame, where the first speech frame is any one of the multiple speech frames, the first feature information includes the envelope spectrum feature of the first speech frame, and the second feature information includes the feature information obtained according to the pitch feature of the first speech frame.

[0040] After the original noisy speech signal is framed in step S1100, the target speech signal with the noise signal removed can be obtained by performing denoising processing on the obtained multiple speech frames respectively. In the embodiment of the present disclosure, the first speech frame can be any one of the multiple speech frames.

[0041] Specifically, the envelope spectrum feature of the first speech frame can be extracted as the first feature information based on any envelope spectrum feature extraction method. For example, the spectral data of the first speech frame can be band-divided according to the critical band (Bark), that is, the Bark band, by using the human auditory model, and the audio energy of the masking band and its corresponding first-order difference data and second-order difference data are calculated. Then, the envelope spectrum feature of the first speech frame is obtained based on the first-order difference data and the second-order difference data as the first feature information.

[0042] It should be noted that since the division of the Bark band is much smaller than the complete spectral bandwidth, calculating the envelope gain of the speech frame based on the envelope spectrum feature obtained by the Bark band division can greatly reduce the network size for predicting the envelope enhancement estimation part, thus making it more suitable for applying this method in terminal devices. In the embodiment of the present disclosure, the Bark band can be divided into 24 bands. Of course, in specific implementation, it can be not limited to this number of divided bands.

[0043] In the embodiment of the present disclosure, the second feature information can be the feature information obtained according to the pitch (fundamental tone) feature of the first speech frame.

[0044] Please refer to Figure 2, which is a schematic processing diagram for obtaining second feature information provided by an embodiment of the present disclosure. As Figure 2 shown, in one embodiment, obtaining the second feature information of the first speech frame includes: extracting the pitch feature of the first speech frame; performing Fourier transform processing on the pitch feature, and using the energy spectrum of the processed pitch feature as the first sub-feature information; and obtaining the second feature information according to the first sub-feature information.

[0045] In specific implementation, the pitch feature of the first speech frame can be obtained based on methods such as the cepstrum method, the short-time autocorrelation method, the short-time average magnitude difference method, the LPC method, etc., and the detailed processing process is not described here.

[0046] Please continue to refer to Figure 2 , in one embodiment, obtaining the second feature information of the first speech frame further includes: performing a preset harmonic reconstruction process on the first speech frame to obtain a first harmonic reconstruction speech signal; performing Fourier transform processing on the first harmonic reconstruction speech signal to obtain second sub-feature information; and performing a preset process on the first harmonic reconstruction speech signal and performing Fourier transform processing on the first harmonic reconstruction speech signal after the preset process to obtain third sub-feature information; then, obtaining the second feature information according to the first sub-feature information obtained based on the pitch feature, the second sub-feature information, and the third sub-feature information.

[0047] As Figure 2 shown, performing the preset process on the first harmonic reconstruction speech signal includes: setting the signal corresponding to the amplitude less than a preset value in the first harmonic reconstruction speech signal to the preset value.

[0048] In this embodiment, the preset value can be 0, that is, the first harmonic reconstruction speech signal obtained after harmonic reconstruction processing can be subjected to this preset process based on the function mapping relationship: Max(x, 0), where x is the signal amplitude in the first harmonic reconstruction speech signal. Of course, the first harmonic reconstruction speech signal can also be subjected to the preset process based on other function mapping relationships. For example, it can also be processed based on the Min() function or a combination of these two types of functions, and no special limitation is made here.

[0049] Specifically, in the embodiment of the present disclosure, the second feature information used for detail enhancement processing of the first speech frame may include the feature information obtained based on the pitch feature of the first speech frame, and may also include the feature information obtained by Fourier transform processing after harmonic reconstruction processing of the first speech frame. Among them, how to perform harmonic reconstruction processing and Fourier transform processing on the speech signal is not described here.

[0050] Step S1300, obtain the target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information.

[0051] After the first feature information and the second feature information of the first speech frame are extracted based on the above steps, by inputting the first feature information and the second feature information into a pre-trained target gain prediction model, the target frequency point gain corresponding to the first speech frame can be estimated, and then the noise signal in the first speech frame can be removed based on the target frequency point gain.

[0052] Please refer to Figure 3 , which is a schematic diagram of the process for obtaining the target frequency point gain provided by an embodiment of the present disclosure. As Figure 3 shown, in the embodiment of the present disclosure, the target gain prediction model may include an envelope enhancement sub-model and a detail enhancement sub-model. Among them, the envelope enhancement sub-model is used to predict the envelope gain of the first speech frame and output the gain, that is, the IRM_bar gain; the detail enhancement sub-model is used to perform detail enhancement on the intervals between the spectra obtained by Fourier transform in the second feature information and output the frequency point gain after the enhancement process, that is, IRM_bin; where IRM_bin can also be referred to as frequency bin.

[0053] It should be noted that in specific implementation, the envelope enhancement sub-model may be a network model based on the structure of a recurrent neural network model (RNN, Recurrent Neural Network), and the detail enhancement sub-model may be a network model based on the structure of a gated convolutional network model (GatedCNN). Of course, only examples are given here. In specific implementation, the two sub-models may also be constructed based on other model structures according to needs.

[0054] As Figure 3 shown, in one embodiment, the obtaining of the target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information includes: inputting the first feature information into the envelope enhancement sub-model of the target gain prediction model, and obtaining the first critical band gain by predicting the envelope gain in the first feature information; inputting the second feature information into the detail enhancement sub-model of the target gain prediction model, and obtaining the first frequency point gain to be processed by predicting the gain at the corresponding frequency points in the second feature information; performing upsampling processing on the first critical band gain, and adding the first critical band gain after the upsampling processing to the first frequency point data to be processed to obtain the target frequency point gain, where the upsampling processing is used to perform a dimensionality increase process on the first critical band gain to increase the dimension of the first critical band gain to be consistent with the dimension of the first frequency point gain to be processed.

[0055] The preset critical band data can be the band data obtained after dividing the Bark band into 24 bands as described in the above description.

[0056] Specifically, in this embodiment, the envelope gain of the first speech frame can be predicted based on the envelope enhancement sub-model with a smaller network model structure, that is, the IRM_bar gain is used as the guiding information. By upsampling the IRM_bar gain to match the shape of the first frequency point gain to be processed output by the detail enhancement sub-model, that is, the IRM_bin gain, and then adding the two to obtain the target frequency point gain finally predicted. Since the detail enhancement sub-model can output the interval of the magnitude spectrum, that is, the gain of gain, based on each frequency point in the spectral domain, that is, bin, therefore, the IRM_bin gain output by the detail enhancement sub-model contains both detail components and envelope components. Based on the IRM_bar gain output by the envelope enhancement sub-model as the guiding information of the IRM_bin gain, the role of detail estimation of the detail enhancement part is given, and at the same time, the requirement for the fitting ability of the network model can be reduced, so that under the premise of a smaller network model structure, the role of detail estimation of the detail enhancement part can be realized, and the problem of insufficient details that may exist when only using the envelope enhancement part can be avoided.

[0057] It should be noted that in specific implementation, when adding the first critical band gain after upsampling processing, that is, the IRM_bar gain and the first frequency point data to be processed, that is, the IRM_bin gain, it can be directly adding the two; or, it can also be weighted summing the two according to the weight values obtained by pre-testing and corresponding to the two respectively; or, it can also be obtaining the target frequency point gain by first directly splicing and then mapping them, and there is no special limitation on how to add the two here.

[0058] Step S1400, denoising the first speech frame based on the target frequency point gain to obtain a target speech signal.

[0059] After predicting the target frequency point gain corresponding to the first speech frame based on the target gain prediction model in step S1300 above, the speech signal can be restored based on the target frequency point gain to obtain a target speech signal.

[0060] In one embodiment, the denoising process of the first speech frame based on the target frequency point gain to obtain the target speech signal includes: multiplying the target frequency point gain by the speech spectrum of the first speech frame to obtain the amplitude spectrum data of the noise-removed signal; multiplying the amplitude spectrum data by the phase information of the first speech frame to obtain the first result data; performing an inverse Fourier transform on the first result data to obtain the first denoised speech signal corresponding to the first speech frame; and obtaining the target speech signal according to the first denoised speech signal.

[0061] Specifically, after predicting the target frequency point gain corresponding to each speech frame in the multiple speech frames obtained by frame-dividing the original noisy speech signal based on the above steps, the target frequency point gain can be regarded as a masking threshold. By multiplying the target frequency point gain by the speech spectrum of the corresponding speech frame, the amplitude spectrum data of the noise-removed signal is obtained. Then, the phase in the original speech signal with noise is used as the final phase information, which is multiplied by the amplitude spectrum data, and the multiplication result is subjected to an inverse Fourier transform to obtain the denoised speech signal corresponding to the current speech frame. By performing the above processing on each speech frame one by one, the target speech signal corresponding to the original noisy speech signal and with the noise signal removed can be obtained.

[0062] In the above description, the first feature information and the second feature information are input into the target gain prediction model to predict and obtain a target frequency point gain, and then the original noisy speech signal is denoised by denoising the first speech frame according to the target frequency point gain to obtain the target speech signal. It should be noted that in specific implementation, the target gain prediction model may not necessarily perform an addition process on the first critical band gain and the first frequency point to be processed obtained based on the first feature information and the second feature information, but can directly output the first critical band gain and the first frequency point to be processed. After the electronic device obtains the first critical band gain and the first frequency point to be processed, by multiplying the first critical band gain and the first frequency point to be processed by the speech spectrum of the first speech frame respectively, and performing an addition process on the two multiplication results respectively to obtain the first denoised speech signal corresponding to the first speech frame, and then obtaining the target speech signal according to the first denoised speech signal. This embodiment does not make special limitations on how to apply the first critical band gain and the first frequency point to be processed to denoise the speech frame.

[0063] In summary, the voice denoising method provided by the embodiments of the present disclosure performs frame segmentation on the original noisy voice signal, and respectively obtains the first feature information and the second feature information of each voice frame. Then, based on the first feature information and the second feature information, the target frequency point gain corresponding to the first voice frame is obtained, and the first voice frame is denoised based on the target frequency point gain to obtain the target voice signal. In the embodiments of the present disclosure, by constructing the first feature information based on the envelope spectrum feature of the voice frame and constructing the second feature information based on the pitch feature of the voice frame, and using the envelope gain estimated from the first feature information as the guiding information, it is possible to take into account the detailed estimation of the voice signal on the premise of effectively reducing the size of the network model, so as to achieve the effect of effectively removing the noise signal without damaging the voice. Therefore, denoising the voice signal generated during the user's call based on this voice denoising method can provide a more comfortable call communication experience.

[0064] <Device Embodiment>

[0065] Corresponding to the above method embodiment, in this embodiment, a voice denoising device is further provided, as Figure 4 described, the device 400 may include a frame segmentation processing module 410, a feature extraction module 420, a gain prediction module 430, and a denoising processing module 440.

[0066] The frame segmentation processing module 410 is configured to perform frame segmentation on the acquired original noisy voice signal to obtain a plurality of voice frames; the feature extraction module 420 is configured to obtain the first feature information and the second feature information of the first voice frame, where the first voice frame is any one of the plurality of voice frames, the first feature information includes the envelope spectrum feature of the first voice frame, and the second feature information includes the feature information obtained according to the pitch feature of the first voice frame; the gain prediction module 430 is configured to obtain the target frequency point gain corresponding to the first voice frame according to the first feature information and the second feature information; the denoising processing module 440 is configured to perform denoising processing on the first voice frame based on the target frequency point gain to obtain a target voice signal.

[0067] <Device Embodiment>

[0068] In this embodiment, an electronic device is further provided, as Figure 5 described, the electronic device 500 may further include a processor 520 and a memory 510, the memory 510 is configured to store executable instructions; the processor 520 is configured to operate the electronic device according to the control of the instructions to execute the method according to any embodiment of the present disclosure.

[0069] <Computer-readable Storage Medium Embodiment>

[0070] This embodiment provides a computer-readable storage medium. An executable command is stored in the storage medium. When the executable command is executed by a processor, the method described in any method embodiment of this specification is executed.

[0071] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this specification. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of an instruction, and the module, the segment of the program, or the part of the instruction contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions marked in the block may occur in an order different from that marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that implementation by hardware, implementation by software, and implementation by the combination of software and hardware are equivalent.

[0072] The embodiments of this specification have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein. The scope of this application is defined by the appended claims.

Claims

1. A voice denoising method, characterized in that, Including: Performing frame segmentation on the acquired original noisy speech signal to obtain a plurality of speech frames; Obtaining first feature information and second feature information of a first speech frame, where the first speech frame is any one of the plurality of speech frames, the first feature information includes the envelope spectrum feature of the first speech frame, and the second feature information includes the feature information obtained based on the pitch feature of the first speech frame; Obtaining a target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information; Performing denoising processing on the first speech frame based on the target frequency point gain to obtain a target speech signal; Wherein, the obtaining the target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information includes: Inputting the first feature information into an envelope enhancement sub-model of a target gain prediction model, and obtaining a first critical band gain by predicting the envelope gain in the first feature information; Inputting the second feature information into a detail enhancement sub-model of the target gain prediction model, and obtaining a first frequency point gain to be processed by predicting the gain at the corresponding frequency points in the second feature information; Performing up-sampling processing on the first critical band gain, and performing summation processing on the up-sampled first critical band gain and the first frequency point data to be processed to obtain the target frequency point gain, where the up-sampling processing is used to perform dimension elevation processing on the first critical band gain to elevate the dimension of the first critical band gain to be consistent with the dimension of the first frequency point gain to be processed.

2. The method according to claim 1, wherein The envelope enhancement sub-model includes a network model based on a recurrent neural network model structure, and the detail enhancement sub-model includes a network model based on a gated convolutional network model structure.

3. The method according to claim 1, wherein The performing denoising processing on the first speech frame based on the target frequency point gain to obtain a target speech signal includes: Multiplying the target frequency point gain by the speech spectrum of the first speech frame to obtain amplitude spectrum data of the noise-removed signal; Multiplying the amplitude spectrum data by the phase information of the first speech frame to obtain first result data; Performing inverse Fourier transform processing on the first result data to obtain a first denoised speech signal corresponding to the first speech frame; Obtaining the target speech signal according to the first denoised speech signal.

4. The method according to claim 1, characterized in that The obtaining the second feature information of the first speech frame includes: Extracting the pitch feature of the first speech frame; Performing Fourier transform processing on the pitch feature, and using the energy spectrum of the processed pitch feature as first sub-feature information; Obtaining the second feature information according to the first sub-feature information.

5. The method according to claim 4, wherein The obtaining the second feature information of the first speech frame further includes: Performing a preset harmonic reconstruction process on the first speech frame to obtain a first harmonic reconstruction speech signal; Performing Fourier transform processing on the first harmonic reconstruction speech signal to obtain second sub-feature information; Perform preset processing on the first harmonic reconstructed speech signal, and perform Fourier transform processing on the first harmonic reconstructed speech signal after the preset processing to obtain third sub-feature information; Obtain the second feature information according to the first sub-feature information, the second sub-feature information, and the third sub-feature information.

6. The method according to claim 5, wherein The performing preset processing on the first harmonic reconstructed speech signal includes: According to a preset function mapping relationship, set the signals in the first harmonic reconstructed speech signal whose corresponding amplitudes are less than a preset value to the preset value.

7. A voice denoising device, characterized in that, It includes: A frame division processing module, configured to perform frame division processing on the acquired original noisy speech signal to obtain a plurality of speech frames; A feature extraction module, configured to obtain first feature information and second feature information of a first speech frame, where the first speech frame is any one of the plurality of speech frames, the first feature information includes the envelope spectrum feature of the first speech frame, and the second feature information includes the feature information obtained according to the pitch feature of the first speech frame; A gain prediction module, configured to obtain the target frequency point gain corresponding to the first speech frame according to the first feature information and the second feature information; A denoising processing module, configured to perform denoising processing on the first speech frame based on the target frequency point gain to obtain a target speech signal; Wherein, the gain prediction module is specifically configured to: Input the first feature information into an envelope enhancement sub-model of a target gain prediction model, and obtain a first critical band gain by predicting the envelope gain in the first feature information; Input the second feature information into a detail enhancement sub-model of the target gain prediction model, and obtain a first to-be-processed frequency point gain by predicting the gain at the corresponding frequency points in the second feature information; Perform upsampling processing on the first critical band gain, and perform summation processing on the first critical band gain after the upsampling processing and the first to-be-processed frequency point data to obtain the target frequency point gain, where the upsampling processing is used to perform dimension elevation processing on the first critical band gain to elevate the dimension of the first critical band gain to be consistent with the dimension of the first to-be-processed frequency point gain.

8. An electronic device, characterized in that, It includes: A memory, configured to store executable instructions; A processor, configured to control the operation of the electronic device according to the instructions to execute the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be read and executed by a computer. The computer program is configured to execute the method according to any one of claims 1-6 when being read and run by the computer.

Images