A method and apparatus for training a speech denoising model

Abstract

This invention discloses a method and apparatus for training a speech denoising model. The method involves acquiring speech samples to be trained and inputting the training samples and RNNt-1 hidden state values ​​into the speech model. Forward propagation is performed on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech result and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state values ​​are then reset. A weighted error result is calculated using the MSN with a weight mask. The speech model is then backpropagated based on the weighted error result to update the speech model weights, resulting in the updated output speech model weights. Forward propagation is then performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. This method can solve the overfitting problem in real-world applications without changing the model structure and training data, improving the denoising effect, enhancing the model's ability to suppress sudden noise, and thus improving the overall speech denoising performance.

Description

Technical Field

[0001] This invention relates to the field of speech processing technology, and in particular to a method and apparatus for training a speech noise reduction model. Background Technology

[0002] In order to efficiently utilize GPUs to train speech denoising models, speech data is first reconstructed into independent samples with uniform format by slicing. Then, the independent samples are randomly sampled to form training samples, which are used as inputs for model training. At the same time, MSE is usually used as a loss function to evaluate the error between the current output and the expected output of the model, and the weights in the speech denoising model are updated based on this error.

[0003] However, in existing model training schemes, on the one hand, independent samples and training samples are usually of equal size, and the duration of independent samples is extremely limited due to the size of GPU memory. Typically, the duration of independent samples is set to 10s, 15s, or 20s. This can lead to overfitting of the RNN units in the model when dealing with speech data that is much longer than 20s in real-world application scenarios, resulting in abnormal noise reduction effects. On the other hand, the MSE loss function evaluates the model output error with equal weight, which does not match the subjective perception of the human ear. Specifically, the human ear has different sensitivities to speech degradation at different volumes, while the MSE provides the same error feedback at different volumes. In addition, the human ear is sensitive to sudden noise, while the MSE is not sensitive to sudden noise error feedback because sudden noise accounts for a small proportion of independent samples. Therefore, existing model training schemes have limitations, and the trained models cannot converge to the optimal solution perceived by the human ear. Summary of the Invention

[0004] This invention provides a speech denoising model training method and apparatus, which can solve the overfitting problem of the model in real application scenarios without changing the model structure and training data, improve the denoising effect of the model, enhance the model's ability to suppress sudden noise, and improve the speech denoising effect.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a speech denoising model training method and apparatus, comprising:

[0006] Acquire the speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples.

[0007] Forward propagation of the training samples and RNNt-1 hidden state values ​​yields the actual speech results and the updated RNNt-1 hidden state values.

[0008] The updated RNNt-1 hidden state value is reset by clearing it to zero using probability p;

[0009] The error between the actual speech result and the expected result is calculated using the MSN with weight masking. The speech model is then backpropagated based on the weighted error result to update the speech model weights and obtain the updated speech model weights. Then, forward propagation is performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model.

[0010] In this embodiment, training samples are acquired and input into the speech model along with the training samples and RNNt-1 hidden state values. Forward propagation is performed on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech result and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state values ​​are then reset to zero using probability p to obtain the reset RNNt-1 hidden state values. The error between the actual speech result and the expected result is calculated using the MSN with weight mask to obtain the weighted error result. The speech model is then backpropagated based on the weighted error result to update the speech model weights and output the updated speech model weights. Forward propagation is then performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. By using this method, the overfitting problem that occurs in real-world application scenarios can be solved without changing the model structure and training data, improving the model's noise reduction effect, enhancing the model's ability to suppress sudden noise, and improving the speech noise reduction effect.

[0011] As a preferred approach, the data size of the RNNt-1 hidden state value is obtained based on the input data of the GRU and the number of frames after windowing and framing of independent samples.

[0012] If the input data size for GRU is N×L×Hin, then the size of the corresponding RNNt-1 hidden state value is N×L×Hout.

[0013] Where N is the number of independent samples in the training samples, L is the number of frames after windowing and framing of the independent samples, Hin is the input feature dimension of GRU, and Hout is the output feature dimension of GRU.

[0014] By implementing the embodiments of the present invention, the duration of independent samples is effectively increased by adding RNNt-1 hidden state values ​​as inputs to the model training phase, based on the original model input.

[0015] As a preferred approach, the weighted error result is obtained by calculating the error between the actual speech result and the expected result using the weighted MSN (Maximum Speech Number) with a weight mask, specifically as follows:

[0016] MSN is used to calculate the error results for each frame between the actual speech result and the expected result.

[0017] Each independent sample data is processed by frame-by-frame windowing, and the energy of each frame signal of each independent sample is calculated by short-time Fourier transform. The energy of each frame signal is then normalized to obtain the normalized energy of each frame.

[0018] The weight mapping curve is determined based on the set energy threshold and weight threshold, and the energy of each frame in the independent sample is mapped according to the weight mapping curve to obtain the weight corresponding to each frame energy.

[0019] The weights corresponding to the energy of each frame are normalized to obtain the MSE weight mask. The normalization formula is as follows:

[0020]

[0021] Among them, Weig This is represented by the weight corresponding to the energy of each frame;

[0022] The weighted error result is obtained by multiplying the MSE weight mask by the error result of each frame.

[0023] In this embodiment, the independent sample data is processed by frame-by-frame windowing, and the energy of each frame signal of each independent sample is calculated by short-time Fourier transform. The energy of each frame signal is normalized to obtain the normalized energy of each frame. The weight mapping curve is determined according to the set energy threshold and weight threshold. After mapping the energy of each frame in the independent sample to obtain the weight corresponding to each frame energy according to the weight mapping curve, the weight corresponding to each frame energy is normalized to obtain the MSE weight mask. Then, the MSE weight mask is multiplied by the error result of each frame of the actual speech result and the expected result calculated by MSN to obtain the weighted error result. By applying the weight mask to adjust the MSE result before backpropagation, the error before backpropagation can be closer to the actual subjective perception of the human ear. This makes the noise reduction effect of the model after convergence closer to the subjective listening experience of the human ear, thus improving the noise reduction effect of the model.

[0024] As a preferred approach, the energy of each frame of signal is normalized to obtain the normalized energy of each frame, specifically as follows:

[0025] The energy of each frame of signal for each independent sample is normalized using the maximum energy value of each independent sample. The normalization formula is as follows:

[0026]

[0027] Among them, Energy norm Energy represents the normalized signal energy of each frame. max This represents the maximum energy of an independent sample.

[0028] As a preferred option, the formula for the weight mapping curve is:

[0029]

[0030] Among them, Energy t Weigt represents the energy threshold for each frame. t Represents the weight threshold for each frame, Energy norm This represents the normalized signal energy of each frame.

[0031] As a preferred solution, to address the same technical problem, embodiments of the present invention also provide a speech denoising model training device, including a sample acquisition module, a forward propagation module, a reset module, and a weighted error result calculation module.

[0032] The sample acquisition module is used to acquire speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples.

[0033] The forward propagation module is used to perform forward propagation on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech results and the updated RNNt-1 hidden state values;

[0034] The reset module is used to clear and reset the updated RNNt-1 hidden state value using probability p to obtain the reset RNNt-1 hidden state value;

[0035] The weighted error calculation module is used to calculate the error between the actual speech result and the expected result using the MSN with weight masking to obtain the weighted error result. This allows the speech model to backpropagate based on the weighted error result and update the speech model weights to obtain the updated speech model weights. Then, based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples, forward propagation is performed until the speech model converges to obtain the target speech model.

[0036] As a preferred embodiment, the weighted error result calculation module includes an MSN calculation unit, an energy calculation unit, a weight calculation unit, a weight mask calculation unit, and a weighted error result calculation unit.

[0037] The MSN calculation unit is used to calculate the error result for each frame of the actual speech result and the expected result using MSN;

[0038] The energy calculation unit is used to perform frame-by-frame windowing processing on each independent sample data, and calculate the energy of each frame signal of each independent sample through short-time Fourier transform, and normalize the energy of each frame signal to obtain the normalized energy of each frame.

[0039] The weight calculation unit is used to determine the weight mapping curve based on the set energy threshold and weight threshold, and to map the energy of each frame in the independent sample according to the weight mapping curve to obtain the weight corresponding to each frame energy.

[0040] The weight mask calculation unit is used to normalize the weights corresponding to the energy of each frame to obtain the MSE weight mask;

[0041] The weighted error result calculation unit is used to multiply the MSE weight mask by the error result of each frame to obtain the weighted error result.

[0042] As a preferred solution, in order to solve the same technical problem, embodiments of the present invention also provide a speech denoising model training device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the speech denoising model training method as shown in the embodiments of the present invention.

[0043] As a preferred solution, in order to solve the same technical problem, embodiments of the present invention also provide a storage medium storing a computer program, which, when executed by a processor, implements the speech denoising model training method as shown in the embodiments of the present invention.

[0044] This invention acquires training samples and inputs the training samples and RNNt-1 hidden state values ​​into a speech model. Forward propagation is performed on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech result and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state values ​​are then reset using probability p to obtain the reset RNNt-1 hidden state values. The error between the actual speech result and the expected result is calculated using a weighted error masked MSN to obtain a weighted error result. The speech model then performs backpropagation based on the weighted error result and updates the speech model weights to obtain the updated output speech model weights. Forward propagation is then performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. This method can solve the overfitting problem in real-world applications without changing the model structure and training data, improving the model's noise reduction effect, enhancing its ability to suppress sudden noise, and improving overall speech noise reduction performance. Attached Figure Description

[0045] Figure 1 : A schematic flowchart of an embodiment of the speech denoising model training method provided by the present invention;

[0046] Figure 2 : A schematic diagram of the training process of an embodiment of the speech denoising model training method provided by the present invention;

[0047] Figure 3 : A schematic diagram of the device structure of one embodiment of the speech denoising model training device provided by the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] Please refer to Figure 1 This invention provides a speech denoising model training method, which includes steps 101 to 104, each step of which is detailed below:

[0051] Step 101: Obtain the speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples.

[0052] In this embodiment, at the initial training time, the speech data is reconstructed into independent samples with a uniform format by slicing. The independent samples are then randomly sampled to form training samples as inputs for model training. At the same time, the RNN hidden state values ​​are used as supplementary inputs to the model. Then, the training speech sample data and the RNN hidden state values ​​are input together into the model for forward propagation.

[0053] Step 102: Perform forward propagation on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech results and the updated RNNt-1 hidden state values.

[0054] Optionally, the size of the RNNt-1 hidden state values ​​is determined based on the input data of the GRU and the number of frames after windowing and framing of independent samples;

[0055] The input data size for GRU is N×L×H. in Then the corresponding RNN t-1 The size of the hidden state value is N×L×H out ;

[0056] Where N is the number of independent samples in the training samples, L is the number of frames after windowing and framing of the independent samples, and H in H represents the input feature dimension corresponding to GRU. ou t represents the output feature dimension of the GRU.

[0057] In this embodiment, the training speech sample data and the RNN hidden state values ​​are used together as model input data and fed into the model for forward propagation to obtain the model output and the updated GRUt hidden state values.

[0058] As an example of this embodiment, taking the GRU in the hidden state value of an RNN as an example, the input data size corresponding to the GRU during the forward propagation of the training samples is N×L×H. in The size of the corresponding RNNt-1 hidden state is N×L×H. out Where N is the number of independent samples in the training samples, L is the number of frames after windowing and framing of the independent samples, and H in H represents the input feature dimension corresponding to GRU. ou t represents the output feature dimension of the GRU.

[0059] Step 103: Use probability p to reset the updated RNNt-1 hidden state value to zero to obtain the reset RNNt-1 hidden state value.

[0060] In this embodiment, as Figure 2As shown, the training speech sample data and the RNN hidden state values ​​are input into the model as model input data for forward propagation to obtain the model output and the updated GRUt hidden state values. Then, the updated RNNt-1 hidden state values ​​are reset to zero using probability p to obtain the reset RNNt-1 hidden state values, which are used as the GRUt-1 hidden state values ​​for the next forward propagation.

[0061] As an example of this embodiment, the GRUt hidden state value is reset to 0 according to probability p, which is preferably 0.1 in this invention. The GRUt hidden state value after the probabilistic reset is retained as the GRUt-1 hidden state value for the next forward propagation.

[0062] Step 104: Calculate the error between the actual speech result and the expected result using the weighted MSN. Then, the speech model performs backpropagation based on the weighted error result and updates the speech model weights to obtain the updated speech model weights. Finally, perform forward propagation based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model.

[0063] Optionally, a weighted error result can be obtained by using the weighted MSN to calculate the error between the actual speech result and the expected result, specifically as follows:

[0064] The error results for each frame between the actual speech result and the expected result are calculated using MSN;

[0065] Each independent sample data is processed by frame-by-frame windowing, and the energy of each frame signal of each independent sample is calculated by short-time Fourier transform. The energy of each frame signal is then normalized to obtain the normalized energy of each frame.

[0066] The weight mapping curve is determined based on the set energy threshold and weight threshold, and the energy of each frame in the independent sample is mapped according to the weight mapping curve to obtain the weight corresponding to each frame energy.

[0067] The weights corresponding to the energy of each frame are normalized to obtain the MSE weight mask. The normalization formula is as follows:

[0068]

[0069] Among them, Weig This is represented by the weight corresponding to the energy of each frame;

[0070] The weighted error result is obtained by multiplying the MSE weight mask by the error result of each frame.

[0071] Optionally, the formula for the weight mapping curve is:

[0072]

[0073] Among them, Energy t Weigt represents the energy threshold for each frame. t Represents the weight threshold for each frame, Energy norm This represents the normalized signal energy of each frame.

[0074] In this embodiment, after obtaining the current model output, the MSE is typically used as a loss function to evaluate the error between the current model output and the expected output, and the weights in the speech denoising model are updated based on this error.

[0075] First, the error between the current output and the expected output of the model is calculated using the MSE loss function to obtain the raw MSE result. Then, the independent samples are processed by frame-by-frame windowing, and the energy of each frame signal is calculated by short-time Fourier transform. The energy of each frame of the independent samples is normalized according to the maximum energy of the independent samples, expressed in dB. The specific formula is as follows:

[0076]

[0077] Then, based on the set energy threshold... th and weight threshold th Determine the weight mapping curve, and map the energy of each frame in the independent sample according to the weight mapping curve to obtain its corresponding weight, that is, the unnormalized MSE weight mask corresponding to the independent sample. The relationship of the weight mapping curve is as follows:

[0078]

[0079] Among them, Energy th Weight represents the energy threshold for each frame. th Represents the weight threshold for each frame, Energy norm This represents the normalized signal energy of each frame.

[0080] It should be noted that, in this invention, Energyth is preferably 40dB, and the weight threshold is Weight th In this invention, 0.1 is preferred.

[0081] Subsequently, the unnormalized MSE weight mask is normalized to obtain the final MSE weight mask. The normalization formula is as follows:

[0082]

[0083] in, This is represented by the weight corresponding to the energy in each frame;

[0084] Finally, the MSE weight mask is multiplied by the original MSE result of each frame of data to obtain the final weighted MSE result, and the model weights are updated by backpropagation based on this MSE result.

[0085] Example 2

[0086] Accordingly, see Figure 3 , Figure 3 This is a schematic diagram of a speech denoising model training device provided by the present invention. As shown in the figure, the speech denoising model training device includes a sample acquisition module 301, a forward propagation module 302, a reset module 303, and a weighted error result calculation module 304. The specific units of each module are as follows:

[0087] The sample acquisition module 301 is used to acquire the speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples.

[0088] Forward propagation module 302 is used to perform forward propagation on training samples and RNNt-1 hidden state values ​​to obtain actual speech results and updated RNNt-1 hidden state values;

[0089] The reset module 303 is used to reset the updated RNNt-1 hidden state value to zero using probability p to obtain the reset RNNt-1 hidden state value;

[0090] The weighted error result calculation module 304 is used to calculate the error between the actual speech result and the expected result using the MSN with weight masking to obtain the weighted error result. This allows the speech model to backpropagate based on the weighted error result and update the speech model weights to obtain the updated speech model weights. Then, based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples, forward propagation is performed until the speech model converges to obtain the target speech model.

[0091] Optionally, the weighted error result calculation module 304 includes an MSN calculation unit 3041, an energy calculation unit 3042, a weight calculation unit 3043, a weight mask calculation unit 3044, and a weighted error result calculation unit 3045.

[0092] MSN calculation unit 3041 is used to calculate the error result of each frame of the actual speech result and the expected result using MSN;

[0093] The energy calculation unit 3042 is used to perform frame-by-frame windowing processing on each independent sample data, and calculate the energy of each frame signal of each independent sample through short-time Fourier transform, and normalize the energy of each frame signal to obtain the normalized energy of each frame.

[0094] The weight calculation unit 3043 is used to determine the weight mapping curve according to the set energy threshold and weight threshold, and to map the energy of each frame in the independent sample according to the weight mapping curve to obtain the weight corresponding to each frame energy.

[0095] The weight mask calculation unit 3044 is used to normalize the weights corresponding to the energy of each frame to obtain the MSE weight mask;

[0096] The weighted error result calculation unit 3045 is used to multiply the MSE weight mask by the error result of each frame to obtain the weighted error result.

[0097] For a more detailed explanation of the working principle and process of this embodiment, please refer to the relevant description in Embodiment 1.

[0098] By acquiring training samples and inputting the training samples and RNNt-1 hidden state values ​​into the speech model, forward propagation is performed on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech result and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state values ​​are then reset to zero using probability p to obtain the reset RNNt-1 hidden state values. The error between the actual speech result and the expected result is calculated using the MSN with weight mask to obtain the weighted error result. The speech model is then backpropagated based on the weighted error result to update the speech model weights and output the updated speech model weights. Forward propagation is then performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. By using this method, the overfitting problem that occurs in real-world application scenarios can be solved without changing the model structure and training data, improving the model's noise reduction effect, enhancing the model's ability to suppress sudden noise, and improving the speech noise reduction effect.

[0099] Example 3

[0100] Accordingly, the speech denoising model training device provided by the present invention includes: a processor and a memory.

[0101] The memory stores a computer program, which is configured to be executed by a processor, which performs operations corresponding to the speech denoising model training method shown in Embodiment 1 of this application.

[0102] Example 4

[0103] Accordingly, the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the speech denoising model training method as shown in Embodiment 1.

[0104] Compared to existing technologies, this invention acquires training samples and inputs the training samples and RNNt-1 hidden state values ​​into a speech model. Forward propagation is performed on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech result and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state values ​​are then reset using probability p to obtain the reset RNNt-1 hidden state values. The error between the actual speech result and the expected result is calculated using a weighted error masked MSN to obtain a weighted error result. The speech model then performs backpropagation based on the weighted error result and updates the speech model weights to obtain the updated output speech model weights. Forward propagation is then performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. This method can solve the overfitting problem in real-world applications without changing the model structure and training data, improving the model's noise reduction effect, enhancing its ability to suppress sudden noise, and improving overall speech noise reduction performance.

[0105] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for training a speech denoising model, characterized in that, include: Acquire the speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples. The training samples and RNNt-1 hidden state values ​​are forward propagated to obtain the actual speech results and the updated RNNt-1 hidden state values. The updated RNNt-1 hidden state value is reset by clearing it to zero using probability p; The error between the actual speech result and the expected result is calculated using the MSN with a weighted mask. The speech model is then backpropagated based on the weighted error result to update the speech model weights and obtain the updated speech model weights. Then, forward propagation is performed based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples until the speech model converges to obtain the target speech model. The weight mask is determined based on the energy of each frame signal of each independent sample.

2. The speech denoising model training method as described in claim 1, characterized in that, The data size of the RNNt-1 hidden state value is obtained based on the input data of the GRU and the number of frames after windowing and framing of independent samples. The input data size for GRU is N×L×H. in The corresponding RNN t-1 The size of the hidden state value is N×L×H out ; Where N is the number of independent samples in the training samples, L is the number of frames after windowing and framing of the independent samples, and H in H represents the input feature dimension corresponding to GRU. ou t represents the output feature dimension of the GRU.

3. The speech denoising model training method as described in claim 1, characterized in that, The weighted error result is obtained by calculating the error between the actual speech result and the expected result using the weighted masked MSN, specifically as follows: The error results for each frame between the actual speech result and the expected result are calculated using MSN; The independent sample data is processed by frame-by-frame windowing, and the energy of each frame signal of each independent sample is calculated by short-time Fourier transform. The energy of each frame signal is then normalized to obtain the normalized energy of each frame. The weight mapping curve is determined based on the set energy threshold and weight threshold, and the energy of each frame in the independent sample is mapped according to the weight mapping curve to obtain the weight corresponding to each frame energy. The weights corresponding to the energy of each frame are normalized to obtain the MSE weight mask, where the normalization formula is: in, This is represented by the weight corresponding to the energy in each frame; The weighted error result is obtained by multiplying the MSE weight mask by the error result of each frame.

4. The speech denoising model training method as described in claim 3, characterized in that, The process of normalizing the energy of each frame signal to obtain the normalized energy of each frame is specifically as follows: The energy of each frame of signal from each independent sample is normalized using the maximum energy value of each independent sample. The normalization formula is as follows: in, This represents the normalized signal energy of each frame. This represents the signal energy of each frame. This represents the maximum energy of an independent sample.

5. The speech denoising model training method as described in claim 3, characterized in that, The formula for the weight mapping curve is: in, This represents the energy threshold for each frame. This represents the weight threshold for each frame. This represents the normalized signal energy of each frame.

6. A speech denoising model training device, characterized in that, It includes a sample acquisition module, a forward propagation module, a reset module, and a weighted error result calculation module; The sample acquisition module is used to acquire the speech samples to be trained and input the training samples and RNNt-1 hidden state values ​​into the speech model. The training samples include multiple independent samples. The forward propagation module is used to perform forward propagation on the training samples and RNNt-1 hidden state values ​​to obtain the actual speech results and the updated RNNt-1 hidden state values; The reset module is used to clear and reset the updated RNNt-1 hidden state value using probability p to obtain the reset RNNt-1 hidden state value; The weighted error result calculation module is used to calculate the error between the actual speech result and the expected result using the MSN with weight mask to obtain the weighted error result. This allows the speech model to backpropagate based on the weighted error result and update the speech model weights to obtain the updated speech model weights. Then, based on the updated speech model weights, the reset RNNt-1 hidden state values, and the training samples, forward propagation is performed until the speech model converges to obtain the target speech model. The weight mask is determined based on the energy of each frame signal of each independent sample.

7. The speech denoising model training device as described in claim 6, characterized in that, The weighted error result calculation module includes an MSN calculation unit, an energy calculation unit, a weight calculation unit, a weight mask calculation unit, and a weighted error result calculation unit. The MSN calculation unit is used to calculate the error result of each frame of the actual speech result and the expected result using MSN; The energy calculation unit is used to perform frame-by-frame windowing processing on each of the independent sample data, and calculate the energy of each frame signal of each independent sample through short-time Fourier transform, and normalize the energy of each frame signal to obtain the normalized energy of each frame. The weight calculation unit is used to determine the weight mapping curve according to the set energy threshold and weight threshold, and to map the energy of each frame in the independent sample according to the weight mapping curve to obtain the weight corresponding to each frame energy. The weight mask calculation unit is used to normalize the weights corresponding to the energy of each frame to obtain the MSE weight mask. The weighted error result calculation unit is used to multiply the MSE weight mask by the error result of each frame to obtain the weighted error result.

8. A speech denoising model training device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the speech denoising model training method as described in any one of claims 1 to 5.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the speech denoising model training method as described in any one of claims 1 to 5.

Images