A single-channel speech enhancement method, system and storage medium based on neural network sub-band modeling

Abstract

The present invention provides a single-channel speech enhancement method, system and storage medium for neural network sub-band modeling. The single-channel speech enhancement method includes the following steps: Step 1: Collecting noisy speech signals and sending them to a logarithmic power spectrum extraction module and a Bark cepstral coefficient extraction module; Step 2: Using a logarithmic power spectrum extraction module and a Bark cepstral coefficient extraction module to receive the noisy speech signal of step 1, and then using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to extract features of the noisy speech signal, and finally sending the extracted features to a frequency band feature division module; Step 3: Using a frequency band feature division module to receive the features extracted in step 2, and then using the frequency band feature division module to allocate sub-band features to the extracted features. The beneficial effects of the present invention are: the present invention performs separate neural network modeling on each sub-band of the speech signal, reduces the task difficulty of the neural network, and reduces the parameters of the model.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a single-channel speech enhancement method, system and storage medium for neural network sub-band modeling. Background Art

[0002] At present, a series of voice electronic products on the market, such as communication products and human-computer interaction products, are affected by various noise interferences. Noisy background noise not only affects the quality of communication between people, but also brings great challenges to human-computer interaction. For example, for voice interactive electronic devices such as smart speakers, smart TVs, and car-mounted equipment, voice recognition is an indispensable technology, and the accuracy of voice recognition in a quiet environment can fully meet the requirements of people's use. However, when there is a certain amount of background noise, the recognition accuracy of the machine will be greatly affected. Therefore, it is necessary to use voice enhancement technology to perform noise reduction processing on voice signals, reduce the impact of interference noise, improve the quality of voice, and enable the machine to achieve a high recognition accuracy in a complex acoustic environment. In addition, for some voice products that have strict requirements on noise reduction and delay, such as hearing aids, intercoms, and ear monitors, the voice enhancement algorithm must not only ensure a good noise reduction effect, but also have the characteristics of low computational complexity and low delay. Summary of the invention

[0003] The present invention provides a single-channel speech enhancement method based on neural network sub-band modeling, comprising the following steps:

[0004] Step 1: Collect noisy speech signals and send them to the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module;

[0005] Step 2: Using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to receive the noisy speech signal in step 1, and then using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to extract features from the noisy speech signal, and finally sending the extracted features to the frequency band feature division module;

[0006] Step 3: The frequency band feature division module is used to receive the features extracted in step 2, and then the frequency band feature division module is used to allocate the extracted features to sub-band features, and the features on each sub-band are input to the corresponding neural network mapping module to estimate the prior signal-to-noise ratio, and finally the estimated prior signal-to-noise ratios on all sub-bands are combined and sent to the full-band Wiener filter module;

[0007] Step 4: Use the full-band Wiener filter module to receive and process the estimated prior signal-to-noise ratio on all sub-bands in step 3 to obtain an enhanced speech signal.

[0008] As a further improvement of the present invention, in step 2, the logarithmic power spectrum feature extraction module extracts features from the noisy speech signal and further comprises performing the following steps:

[0009] The first step: pre-processing the speech signal x(n) collected by the microphone by frame division and windowing;

[0010] The second step: perform fast Fourier transform to obtain the signal spectrum and obtain the power spectrum S in the frequency domain 2 (k);

[0011] The third step: perform natural logarithm operation;

[0012] Step 4: Compress the power spectrum in the logarithmic domain to obtain the extracted logarithmic power spectrum feature Y log (k), as shown in the following formula (1):

[0013] Y log (k) = ln(S 2 (k)),k=1,2,...,N (1)

[0014] Wherein, k is the frequency index, N is the frame length of a frame of speech signal, and in the single-channel speech enhancement method, a sampling rate of 16 kHz is adopted, a frame length of 16 ms per frame, a frame shift of 8 ms, and N is 129.

[0015] As a further improvement of the present invention, in step 2, the Bark cepstral coefficient extraction module extracts features from the noisy speech signal further comprising performing the following steps:

[0016] Step S1: pre-processing the input speech signal x(n) by frame division and windowing;

[0017] Step S2: Perform fast Fourier transform to transform the data from time domain to frequency domain;

[0018] Step S3: Calculate the frequency domain power spectrum S 2 (k);

[0019] Step S4: The calculated frequency domain power spectrum S 2 (k) After passing through the Bark filter, the filtered energy spectrum is calculated, as shown in formula (2):

[0020]

[0021] Where b is the order index of the Bark energy spectrum, B is the number of Bark filters, which is 24 here. Each filter corresponds to a frequency band in the Bark domain. The transfer function expression of the Bark frequency filter is shown in the following formula (3):

[0022]

[0023] Step S5: Take the logarithm of the Bark energy spectrum of each frame and perform discrete cosine transform (DCT), as shown in formula (4), to obtain the Bark cepstral coefficient feature,

[0024]

[0025] Among them, Y bark (n) is the extracted BFCC feature, n is the frequency band index of the feature, and the dimension of the feature is consistent with the number of Bark filters, which is 24 dimensions.

[0026] As a further improvement of the present invention, in step 3, the frequency band feature division module further includes sequentially performing the following steps:

[0027] Subband division step: The frequency domain range of 0-8000 Hz is divided into 8 subbands, and the indexes of the features on different subbands are given according to the different numbers of LPS features and BFCC features corresponding to each subband;

[0028] Feature splicing step: The LPS and BFCC features on each sub-band are spliced ​​together and sent to their respective neural network mapping modules for prior signal-to-noise ratio estimation.

[0029] As a further improvement of the present invention, in step 3, the neural network mapping module includes 5 neural layers, wherein the first layer and the last layer are both feedforward neural network layers, the middle three layers are GRU neural layers, and the feedforward neural network layer performs weighted summation in a fully connected manner and performs nonlinear activation, as shown in the following formula:

[0030] h=g(W·X+b) (5)

[0031] Among them, W and b are the weight and bias of the neuron respectively, h represents the output of the feedforward neural network layer, X is the input of the feedforward neural network layer, and g(·) represents the nonlinear activation operation. The first layer of the feedforward neural network layer here uses the ReLU activation function, and the last layer of the feedforward neural network layer needs to estimate the prior signal-to-noise ratio, so no activation operation is performed, only linear weighted summation is performed.

[0032] As a further improvement of the present invention, the memory update mechanism in the GRU layer of the neural network mapping module is specifically as follows:

[0033] The GRU unit takes the feature x of the current frame as input t The output h of the previous frame retained before t-1 Combined, processed by the update gate and reset gate, an output h of the current frame is generated t , and the iteration continues in this way. The calculation formulas for each gate and output are as follows:

[0034] r t =σ(W r ·[h t-1 ,x t ]) (6)

[0035] z t =σ(W z ·[h t-1 ,x t ]) (7)

[0036]

[0037] where σ(·) and tanh(·) represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively. t represents the output of the current frame update gate, z t Represents the output of the forget gate for the current frame.

[0038] As a further improvement of the present invention, in step 3, the priori signal-to-noise ratio values ​​on each sub-band estimated by the neural network mapping module are combined to obtain a 129-dimensional output.

[0039] As a further improvement of the present invention, in step 4, the full-band Wiener filtering module further includes performing the following steps:

[0040] Step Y1: Calculate the gain function for filtering, which is expressed as follows:

[0041]

[0042] in, is the prior signal-to-noise ratio value output by the neural network mapping module;

[0043] Step Y2: Filter the input noisy speech using the estimated gain function, and finally perform inverse Fourier transform to obtain the denoised speech signal The formula is as follows:

[0044]

[0045] Formula (11) is the frequency domain filtering process of Wiener filtering, where S(k) is the spectrum of the input noisy speech signal, and N is the number of frequency points in each frame, which is 129 here. For the enhanced speech signal spectrum, the inverse Fourier transform in formula (12) is performed to obtain the final time domain signal output

[0046] The present invention also discloses a single-channel speech enhancement system based on neural network sub-band modeling, comprising: a memory, a processor and a computer program stored on the memory, wherein the computer program is configured to implement the steps of the single-channel speech enhancement method described in the present invention when called by the processor.

[0047] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the single-channel speech enhancement method of the present invention when called by a processor.

[0048] The beneficial effects of the present invention are: 1. The single-channel speech enhancement method of the present invention performs separate neural network modeling on each sub-band of the speech signal, which reduces the task difficulty of the neural network, reduces the parameters of the model, and achieves lower algorithm complexity; 2. The single-channel speech enhancement method of the present invention adopts a neural network model to estimate the prior signal-to-noise ratio of the signal, and then combines it with the traditional filtering method to perform noise reduction, which effectively improves the generalization ability of the neural network noise reduction algorithm; 3. The single-channel speech enhancement method of the present invention has a higher mapping accuracy for the neural network model trained separately for each sub-band, and can achieve better speech noise reduction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a principle block diagram of the single-channel speech enhancement method of the present invention;

[0050] Figure 2 It is a block diagram of the principle of extracting logarithmic power features of the single-channel speech enhancement method of the present invention;

[0051] Figure 3 It is a block diagram of the extraction principle of the BFCC feature of the single-channel speech enhancement method of the present invention;

[0052] Figure 4 It is a structural diagram of each sub-band of the neural network mapping module of the single-channel speech enhancement method of the present invention;

[0053] Figure 5 It is a block diagram of the memory update principle in the GRU layer of the single-channel speech enhancement method of the present invention. DETAILED DESCRIPTION

[0054] like Figure 1As shown, the present invention discloses a single-channel speech enhancement method based on neural network subband modeling. The single-channel speech enhancement method is based on a neural network model, uses log power spectrum features (LPS) and Bark-frequency cepstral coefficients (BFCC) to estimate the prior signal-to-noise ratio of the target speech, and combines it with the Wiener filtering method to achieve a good compromise between noise reduction effect and computational complexity. The single-channel speech enhancement method includes executing the following steps:

[0055] Step 1: A single microphone collects a noisy speech signal and sends it to a logarithmic power spectrum extraction module and a Bark cepstral coefficient extraction module;

[0056] Step 2: Using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to receive the noisy speech signal in step 1, and then using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to extract features from the noisy speech signal, and finally sending the extracted features to the frequency band feature division module;

[0057] Step 3: The frequency band feature division module is used to receive the features extracted in step 2, and then the frequency band feature division module is used to allocate sub-band features to the two sets of extracted features, and the features on each sub-band are input to the corresponding neural network mapping module to estimate the prior signal-to-noise ratio, and finally the estimated prior signal-to-noise ratios on all sub-bands are combined and sent to the full-band Wiener filter module;

[0058] Step 4: Use the full-band Wiener filter module to receive and process the estimated prior signal-to-noise ratio on all sub-bands in step 3 to obtain an enhanced speech signal.

[0059] In the single-channel speech enhancement method of the present invention, we selected 4800 sentences (24 males and 24 females, each speaking 100 sentences) from the Aishell Chinese data set [1] as the pure speech data of the training set, and then randomly mixed them with 100 different noise types selected from the Freesound website [2]. The mixed signal-to-noise ratio conforms to the uniform distribution in the interval range of [-5, 20], and the total noisy training data length is about 100 hours. Subsequently, the BFCC features and logarithmic power spectrum features of each sub-band are extracted, and their corresponding ideal prior signal-to-noise ratio values ​​are constructed. Then, the back propagation algorithm is used to train each neural network, and 10% of all training data is divided as a validation set. When the loss on the training set and the validation set is minimized, the model is saved, and the neural network mapping model corresponding to different sub-bands is obtained. The above is the processing flow of the entire single-channel speech enhancement method and the training process of the neural network model. Next, each key module will be introduced in detail.

[0060] like Figure 2 As shown, in step 2, the logarithmic power spectrum feature extraction module is used to extract the frequency domain logarithmic power features of the speech signal, and the logarithmic power spectrum feature extraction module performs feature extraction on the noisy speech signal and further comprises performing the following steps:

[0061] The first step: pre-processing the speech signal x(n) collected by the microphone by frame division and windowing;

[0062] The second step: perform fast Fourier transform (FFT) to obtain the signal spectrum and obtain the power spectrum S in the frequency domain 2 (k);

[0063] The third step: perform natural logarithm operation;

[0064] Step 4: Compress the power spectrum in the logarithmic domain to obtain the extracted logarithmic power spectrum feature Y log (k), as shown in the following formula (1):

[0065] Y log (k) = ln(S 2 (k)),k=1,2,...,N (1)

[0066] Wherein, k is the frequency index, N is the frame length of a frame of speech signal, and in the single-channel speech enhancement method, a sampling rate of 16 kHz is adopted, a frame length of 16 ms per frame, and a frame shift of 8 ms, so N is 129.

[0067] like Figure 3 As shown, in step 2, the Bark cepstral coefficient feature extraction module performs feature extraction in the frequency domain using the Bark scale, simulating the masking effect of the human auditory system on sound, and making full use of the characteristic that the low-frequency resolution of the human ear for sound signals is higher than its high-frequency resolution, thereby extracting a spectrum feature that is very close to human subjective perception. The Bark cepstral coefficient extraction module performs feature extraction on the noisy speech signal and further includes performing the following steps:

[0068] Step S1: pre-processing the input speech signal x(n) by frame division and windowing;

[0069] Step S2: Perform fast Fourier transform to transform the data from the time domain to the frequency domain;

[0070] Step S3: Calculate the frequency domain power spectrum S 2 (k);

[0071] Step S4: The calculated frequency domain power spectrum S 2 (k) After passing through the Bark filter, the filtered energy spectrum is calculated, as shown in formula (2):

[0072]

[0073] Where b is the order index of the Bark energy spectrum, B is the number of Bark filters, which is 24 here. Each filter corresponds to a frequency band in the Bark domain. The transfer function expression of the Bark frequency filter is shown in the following formula (3):

[0074]

[0075] Step S5: Take the logarithm of the Bark energy spectrum of each frame and perform discrete cosine transform (DCT), as shown in formula (4), to obtain the Bark cepstral coefficient feature,

[0076]

[0077] Among them, Y bark (n) is the extracted BFCC feature, n is the frequency band index of the feature, and the dimension of the feature is consistent with the number of Bark filters, which is 24 dimensions.

[0078] In step 3, the frequency band feature division module divides the extracted Bark cepstral coefficient features and logarithmic power spectrum features of each frame signal into sub-bands, and each sub-band only contains the BFCC features and LPS features within its frequency range, as shown in Table 1.

[0079] Table 1 Feature allocation of frequency domain subbands

[0080]

[0081] The frequency band feature division module further includes sequentially performing the following steps:

[0082] Subband division step: The frequency domain range of 0-8000Hz is divided into 8 subbands, and considering that most speech signals are concentrated in the low-frequency range, the low-frequency subbands are divided more finely. In addition, according to the different numbers of LPS features and BFCC features corresponding to each subband, the indexes of the features on different subbands are given, as shown in Table 1;

[0083] Feature splicing step: The LPS and BFCC features on each sub-band are spliced ​​together and sent to their respective neural network mapping modules for prior signal-to-noise ratio estimation.

[0084] In step 3, the neural network mapping module models each sub-band feature and customizes a dedicated noise reduction model for different sub-bands. Considering the time-series correlation characteristics of the speech signal, a model with the ability to map the prior signal-to-noise ratio is constructed based on the gated recurrent unit (GRU) in the neural network mapping module.

[0085] like Figure 4 As shown in the figure, after the allocation of the sub-band division module, the characteristics of each sub-band are input into the designed neural network structure to perform a priori signal-to-noise ratio The neural network mapping module includes 5 neural layers, wherein the first and last layers are feedforward neural network layers, and the middle three layers are GRU neural layers. The feedforward neural network layers perform weighted summation in a fully connected manner and perform nonlinear activation, as shown in the following formula (5):

[0086] h=g(W·X+b) (5)

[0087] Among them, W and b are the weight and bias of the neuron respectively, h represents the output of the feedforward neural network layer, X is the input of the feedforward neural network layer, and g(·) represents the nonlinear activation operation. The first layer of the feedforward neural network layer here uses the ReLU activation function, and the last layer of the feedforward neural network layer needs to estimate the prior signal-to-noise ratio, so no activation operation is performed, only linear weighted summation is performed.

[0088] like Figure 5 As shown, the memory update mechanism in the GRU layer of the neural network mapping module is as follows:

[0089] The GRU unit takes the feature x of the current frame as input t The output h of the previous frame retained before t-1 Combined, processed by the update gate and reset gate, an output h of the current frame is generated t , and the iteration continues in this way. The calculation formulas for each gate and output are as follows:

[0090] r t =σ(W r ·[h t-1 ,x t ]) (6)

[0091] z t =σ(W z ·[h t-1 ,x t ]) (7)

[0092]

[0093] where σ(·) and tanh(·) represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively. t represents the output of the current frame update gate, z t Represents the output of the forget gate of the current frame.

[0094] In addition, since the number of features on each sub-band is different, although the neural network structure on each sub-band is the same, considering the different difficulty of the task in each sub-band, the number of neurons in the neural network model corresponding to each sub-band is different, as shown in Table 2 below.

[0095] Table 2 Neuron configuration of different sub-band neural network modules

[0096]

[0097] In step 3, the priori signal-to-noise ratio values ​​on each sub-band estimated by the neural network mapping module are combined to obtain a 129-dimensional output.

[0098] In step 4, the full-band Wiener filtering module further includes executing the following steps: Step Y1: Calculate a gain function for filtering, expressed as shown in the following formula (10):

[0099]

[0100] in, is the prior signal-to-noise ratio value output by the neural network mapping module;

[0101] Step Y2: Use the estimated gain function to filter the input noisy speech, and finally perform inverse Fourier transform to obtain the denoised speech signal. The formula is as follows:

[0102]

[0103] Formula (11) is the frequency domain filtering process of Wiener filtering, where S(k) is the spectrum of the input noisy speech signal, and N is the number of frequency points in each frame, which is 129 here. For the enhanced speech signal spectrum, the inverse Fourier transform as in formula (12) is performed to obtain the final time domain signal output

[0104] The beneficial effects of the present invention are: 1. The single-channel speech enhancement method of the present invention performs separate neural network modeling on each sub-band of the speech signal, which reduces the task difficulty of the neural network, reduces the parameters of the model, and achieves lower algorithm complexity; 2. The single-channel speech enhancement method of the present invention adopts a neural network model to estimate the prior signal-to-noise ratio of the signal, and then combines it with the traditional filtering method to perform noise reduction, which effectively improves the generalization ability of the neural network noise reduction algorithm; 3. The single-channel speech enhancement method of the present invention has a higher mapping accuracy for the neural network model trained separately for each sub-band, and can achieve better speech noise reduction effect.

[0105] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A single-channel speech enhancement method based on neural network sub-band modeling, characterized in that: This involves performing the following steps: Step 1: Collect noisy speech signals and send them to the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module; Step 2: Using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to receive the noisy speech signal in step 1, and then using the logarithmic power spectrum extraction module and the Bark cepstral coefficient extraction module to extract features from the noisy speech signal, and finally sending the extracted features to the frequency band feature division module; Step 3: The frequency band feature division module is used to receive the features extracted in step 2, and then the frequency band feature division module is used to allocate the extracted features to sub-band features, and the features on each sub-band are input to the corresponding neural network mapping module to estimate the prior signal-to-noise ratio, and finally the estimated prior signal-to-noise ratios on all sub-bands are combined and sent to the full-band Wiener filter module; Step 4: Using a full-band Wiener filter module to receive and process the estimated priori signal-to-noise ratios on all sub-bands in step 3, an enhanced speech signal is obtained; In step 2, the logarithmic power spectrum feature extraction module extracts features from the noisy speech signal and further comprises performing the following steps: The first step: pre-processing the speech signal x(n) collected by the microphone by frame division and windowing; The second step: perform fast Fourier transform to obtain the signal spectrum and obtain the power spectrum S in the frequency domain 2 (k); The third step: perform natural logarithm operation; Step 4: Compress the power spectrum in the logarithmic domain to obtain the extracted logarithmic power spectrum feature Y log (k), as shown in the following formula (1): Y log (k)=ln(S 2 (k)),k=1,2,...,N(1) Wherein, k is the frequency index, N is the frame length of a frame of speech signal, in this single-channel speech enhancement method, a sampling rate of 16kHz is adopted, a frame length of 16ms per frame, a frame shift of 8ms, and N is 129; In step 2, the Bark cepstral coefficient extraction module extracts features from the noisy speech signal and further includes performing the following steps: Step S1: pre-processing the input speech signal x(n) by frame division and windowing; Step S2: Perform fast Fourier transform to transform the data from the time domain to the frequency domain; Step S3: Calculate the frequency domain power spectrum S 2 (k); Step S4: The calculated frequency domain power spectrum S 2 (k) After passing through the Bark filter, the filtered energy spectrum is calculated, as shown in formula (2): Where b is the order index of the Bark energy spectrum, B is the number of Bark filters, which is 24 here. Each filter corresponds to a frequency band in the Bark domain. The transfer function expression of the Bark frequency filter is shown in the following formula (3): Step S5: Take the logarithm of the Bark energy spectrum of each frame and perform discrete cosine transform (DCT), as shown in formula (4), to obtain the Bark cepstral coefficient feature, Among them, Y bark (n) is the extracted BFCC feature, n is the frequency band index of the feature, and the dimension of the feature is consistent with the number of Bark filters, which is 24 dimensions; In step 3, the neural network mapping module includes 5 neural layers, wherein the first layer and the last layer are both feedforward neural network layers, and the middle three layers are GRU neural layers. The feedforward neural network layer performs weighted summation in a fully connected manner and performs nonlinear activation, as shown in the following formula (5): h=g(WX+b) (5) Wherein, W and b are the weight and bias of the neuron respectively, h represents the output of the feedforward neural network layer, X is the input of the feedforward neural network layer, g() represents the nonlinear activation operation, the first layer of the feedforward neural network layer here adopts the ReLU activation function, and the last layer of the feedforward neural network layer needs to estimate the prior signal-to-noise ratio, so no activation operation is performed, only linear weighted summation is performed; the memory update mechanism in the GRU layer of the neural network mapping module is as follows: The GRU unit takes the feature x of the current frame as input t The output h of the previous frame retained before t-1 Combined, processed by the update gate and reset gate, an output h of the current frame is generated t , and the iteration continues in this way. The calculation formulas for each gate and output are as follows: r t =σ(W r ·[h t-1 ,x t ])(6) z t =σ(W z ·[h t-1 ,x t ]) (7) where σ(·) and tanh(·) represent the Sigmoid activation function and the hyperbolic tangent activation function, respectively. t represents the output of the current frame update gate, z t Represents the output of the forget gate for the current frame.

2. The single-channel speech enhancement method according to claim 1, characterized in that: In step 3, the frequency band feature division module further includes sequentially performing the following steps: Subband division step: The frequency domain range of 0-8000 Hz is divided into 8 subbands, and the indexes of the features on different subbands are given according to the different numbers of LPS features and BFCC features corresponding to each subband; Feature splicing step: The LPS and BFCC features on each sub-band are spliced ​​together and sent to their respective neural network mapping modules for prior signal-to-noise ratio estimation.

3. The single-channel speech enhancement method according to claim 1, characterized in that: In step 3, the priori signal-to-noise ratio values ​​on each sub-band estimated by the neural network mapping module are combined to obtain a 129-dimensional output.

4. The single-channel speech enhancement method according to claim 3, characterized in that: In step 4, the full-band Wiener filtering module further comprises performing the following steps: Step Y1: Calculate the gain function for filtering, which is expressed as follows: in, is the prior signal-to-noise ratio value output by the neural network mapping module; Step Y2: Filter the input noisy speech using the estimated gain function, and finally perform inverse Fourier transform to obtain the denoised speech signal The formula is as follows: Formula (11) is the frequency domain filtering process of Wiener filtering, where S(k) is the spectrum of the input noisy speech signal, and N is the number of frequency points in each frame, which is 129 here. For the enhanced speech signal spectrum, the inverse Fourier transform in formula (12) is performed to obtain the final time domain signal output 5. A single-channel speech enhancement system based on neural network sub-band modeling, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the computer program is configured to implement the steps of the single-channel speech enhancement method according to any one of claims 1 to 4 when called by the processor.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the single-channel speech enhancement method according to any one of claims 1 to 4 when called by a processor.

Images