Method and apparatus for noise reduction of voice data

Abstract

The application discloses a noise reduction method and device for voice data. The method comprises the following steps: obtaining a voice data sample, wherein the voice sample data comprises multiple frames of voice data mixed with noise; obtaining label information corresponding to the voice data sample, wherein the label information is used for marking pure voice features, noise voice features and voice activity features in the voice data sample; determining a preset neural network model; training the preset neural network model by using the voice data sample and the label information, so that the preset neural network model learns the corresponding relationship between the pure voice features and the noise voice features, and the corresponding relationship between the noise voice features and the voice activity features, and obtains a voice processing model. In the model training process, the voice data without noise and the sample mixed with noise are used for training, and the voice activity features in the sample are combined for training, so that the noise reduction performance is better under the condition of the same model calculation amount.

Description

[0001] This invention is a divisional application, with the national application number 202210225936.9, the application date being March 7, 2022, and the invention title being "Training method for speech processing model, noise reduction method and apparatus for speech data". Technical Field

[0002] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for denoising voice data. Background Technology

[0003] Deep learning-based denoising methods typically take noisy speech and its transformation features as input and output clean speech and its transformation features as output. Achieving good denoising performance usually requires complex network models, which consume significant computational and storage resources. This reduces the flexibility of the network model and makes it unsuitable for use on resource-constrained devices. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, or at least partially solve the above-mentioned technical problems, this application provides a method and apparatus for denoising voice data.

[0005] According to one aspect of the embodiments of this application, a method for training a speech processing model is provided, comprising:

[0006] Acquire speech data samples, wherein the speech sample data includes multiple frames of speech data obtained after noise mixing;

[0007] Obtain the label information corresponding to the speech data sample, wherein the label information is used to mark the clean speech features, noisy speech features and speech activity features in the speech data sample;

[0008] Determine the preset neural network model;

[0009] The preset neural network model is trained using the speech data samples and the label information, so that the preset neural network model learns the correspondence between the clean speech features and the noisy speech features, as well as between the noisy speech features and the speech activity features, to obtain a speech processing model.

[0010] Furthermore, the acquisition of voice data samples includes:

[0011] Acquire initial speech data, wherein the initial speech data is speech data carrying clean speech features and speech activity features;

[0012] Determine the initial noise characteristics of various types and the preset signal-to-noise ratio;

[0013] The initial speech data and the initial noise features are mixed according to the preset signal-to-noise ratio to obtain the speech data sample.

[0014] Furthermore, the preset neural network model includes: a convolutional neural network, a first recurrent neural network, a second recurrent neural network, a first deep neural network, and a second deep neural network;

[0015] The convolutional neural network is connected to the first recurrent neural network and the second recurrent neural network respectively. The first recurrent neural network is connected to the first deep neural network, and the second recurrent neural network is connected to the second deep neural network. The convolutional neural network is connected to the first recurrent neural network through a connection unit, and the second recurrent neural network is also connected to the first recurrent neural network through the connection unit.

[0016] Furthermore, the preset neural network model is trained using the speech data samples and the label information, so that the preset neural network model learns the correspondence between the clean speech features and the noisy speech features, as well as between the noisy speech features and the speech activity features, to obtain a speech processing model, including:

[0017] The speech data sample is input into the preset neural network model, so that the convolutional neural network in the preset neural network model extracts the first feature of the speech data sample. The first feature is input into the connection unit and the second recurrent neural network respectively. The second recurrent neural network detects the first speech activity feature and the first noise speech feature in the first feature, and extracts the first speech activity feature. The first speech activity feature is input into the connection unit and the second deep neural network respectively. The connection unit concatenates the first feature and the first speech activity feature to obtain the second feature. The second feature is input into the first recurrent neural network, and the first recurrent neural network detects the second speech activity feature and the second noise speech feature in the second feature, extracts the second speech activity feature, and inputs the second speech activity feature into the first deep neural network for feature superposition to obtain the third feature. At the same time, the target value output by the second deep neural network is determined.

[0018] When the target value is used to indicate that the speech activity feature meets the preset feature, the initial speech data corresponding to the speech data sample is determined, and the fourth feature corresponding to the initial speech data is extracted;

[0019] The loss function value is calculated based on the third feature and the fourth feature. If the loss function value is less than a preset threshold, the preset neural network model is determined as the speech processing model.

[0020] According to another aspect of the embodiments of this application, a method for denoising voice data is also provided, comprising:

[0021] Acquire the raw speech data to be processed;

[0022] The raw speech data is input into a pre-trained speech processing model so that the speech processing model extracts features from the raw speech data and outputs target values ​​and target speech features based on the features.

[0023] When the target value is used to indicate that the target speech feature meets the preset feature, target speech data is generated based on the target speech feature, wherein the preset feature is the feature of speech data that does not carry noise.

[0024] Furthermore, the step of inputting the target speech data into a pre-trained speech processing model, so that the speech processing model outputs target values ​​and target speech features, includes:

[0025] The raw speech data is input into the speech processing model, so that the convolutional neural network in the speech processing model extracts the raw features of the raw speech data. The raw features are then input into the connection unit and the second recurrent neural network. The second recurrent neural network detects the raw speech activity features and raw noise speech features in the first raw features, and extracts the raw speech activity features. The raw speech activity features are then input into the connection unit and the second deep neural network. The connection unit concatenates the raw features and raw speech activity features to obtain concatenated features. The concatenated features are then input into the first recurrent neural network, which detects the target speech activity features and target noise speech features in the concatenated features, extracts the target speech activity features, and inputs the target speech activity features into the first deep neural network for feature superposition to obtain the target speech features. At the same time, the target value output by the second deep neural network is determined.

[0026] According to another aspect of the embodiments of this application, a training apparatus for a speech processing model is also provided, comprising:

[0027] The first acquisition module is used to acquire voice data samples, wherein the voice sample data includes multiple frames of voice data obtained after noise mixing;

[0028] The second acquisition module is used to acquire the label information corresponding to the speech data sample, wherein the label information is used to mark the clean speech features, noisy speech features and speech activity features in the speech data sample;

[0029] The determination module is used to determine the preset neural network model;

[0030] The training module is used to train the preset neural network model using the speech data samples and the label information, so that the preset neural network model learns the correspondence between the clean speech features and the noisy speech features, as well as between the noisy speech features and the speech activity features, to obtain a speech processing model.

[0031] According to another aspect of the embodiments of this application, a noise reduction device for voice data is also provided, comprising:

[0032] The acquisition module is used to acquire the raw voice data to be processed;

[0033] An extraction module is used to input the raw speech data into a pre-trained speech processing model so that the speech processing model extracts the features of the raw speech data and outputs a target value and target speech features based on the features;

[0034] The processing module is configured to generate target speech data based on the target speech features when the target value indicates that the target speech features satisfy preset features, wherein the preset features are features of speech data that do not carry noise.

[0035] According to another aspect of the embodiments of this application, a storage medium is also provided, the storage medium including a stored program that executes the above steps when the program is run.

[0036] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; wherein: the memory is used to store computer programs; and the processor is used to execute the steps in the above method by running the programs stored in the memory.

[0037] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps in the above-described method.

[0038] Compared with the prior art, the technical solution provided in this application has the following advantages: In the model training process, this application uses noiseless speech data and samples mixed with noise, and combines the speech activity features in the samples for training, which can achieve better noise reduction performance under the same model computational load. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a training method for a speech processing model provided in an embodiment of this application;

[0042] Figure 2 This is a schematic diagram of the structure of the preset neural network model provided in the embodiments of this application;

[0043] Figure 3 A flowchart illustrating a method for denoising voice data, provided in another embodiment of this application;

[0044] Figure 4 A block diagram of a training device for a speech processing model provided in an embodiment of this application;

[0045] Figure 5 A block diagram of a voice data noise reduction device provided in another embodiment of this application;

[0046] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another similar entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0049] This application provides a method for training a speech processing model, a method for denoising speech data, and an apparatus. The methods provided in this invention can be applied to any electronic device as needed, such as servers, terminals, or other electronic devices. No specific limitations are made here, and for ease of description, they will be referred to as electronic devices.

[0050] According to one aspect of the embodiments of this application, a method embodiment for training a speech processing model is provided. Figure 1 A flowchart illustrating a training method for a speech processing model provided in this application embodiment is shown below. Figure 1 As shown, the method includes:

[0051] Step S11: Obtain speech data samples, wherein the speech sample data includes multiple frames of speech data obtained after noise mixing.

[0052] In this embodiment of the application, step S11, obtaining voice data samples, includes the following steps A1-A3:

[0053] Step A1: Obtain initial speech data, wherein the initial speech data is speech data carrying clean speech features and speech activity features.

[0054] The method provided in this application is applied to a voice processing device, which may be a smartphone, computer, etc. The voice processing device can control a voice acquisition device to collect initial voice data, which may be a recorder, voice recorder, etc. It is understood that the initial voice data is noise-free voice data (i.e., clean voice data). The initial voice data carries clean voice features and voice activity features; the voice activity features are used to characterize the sounds present in the initial voice data, etc.

[0055] Step A2: Determine the initial noise characteristics of various types and the preset signal-to-noise ratio.

[0056] In this embodiment of the application, in order to construct a speech data sample carrying noise, it is also necessary to determine various types of noise data, which may include: white noise, additive noise, and multiplicative noise. The preset signal-to-noise ratio may be a pre-set noise mixing ratio.

[0057] Step A3: Mix the initial speech data and initial noise features according to the preset signal-to-noise ratio to obtain speech data samples.

[0058] In this embodiment of the application, the speech processing device can mix the initial speech data and noise data according to a preset signal comparison to obtain a speech data sample. The speech data sample obtained at this time is the speech data obtained after noise mixing of multiple frames of clean noise features.

[0059] Step S12: Obtain the label information corresponding to the speech data sample, wherein the label information is used to mark the clean speech features, noisy speech features and speech activity features in the speech data sample.

[0060] In this embodiment, the labeling information of the speech data samples is used to mark the speech activity features and noisy speech features in the speech data samples. The speech data samples can be speech codes, and the speech activity features can be the encoded features of the speech codes, etc.

[0061] Step S13: Determine the preset neural network model.

[0062] In the embodiments of this application, Figure 2 This is a schematic diagram of the structure of the preset neural network model provided in the embodiments of this application, such as... Figure 2 As shown, the preset neural network model includes: Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN1), Recurrent Neural Network (RNN2), Deep Neural Network (DNN1), and Deep Neural Network (DNN2).

[0063] A convolutional neural network (CNN) connects a first recurrent neural network (RNN1) and a second recurrent neural network (RNN2) to each other. The first recurrent neural network (RNN1) is connected to a first deep neural network (DNN1), and the second recurrent neural network (RNN2) is connected to a second deep neural network (DNN2). The convolutional neural network is connected to the first recurrent neural network (RNN1) through a connection unit (Cat), and the second recurrent neural network (RNN2) is also connected to the first recurrent neural network (RNN1) through a connection unit (Cat).

[0064] It should be noted that CNN is used for deep feature extraction. It extracts deep speech features through a multi-layer CNN network. In this embodiment, a 7-layer CNN network is used, with Batch Norm added between each layer and PReLU activation function used.

[0065] RNN2: RNN2 includes LSTM (Long Short-Term Memory) units (or GRU units) to extract speech activity features. In this application embodiment, two layers of LSTM units are used to extract speech activity features.

[0066] Cat: The connection unit concatenates the features output by the CNN with the speech activity features to synthesize a single feature.

[0067] RNN1 and DNN1: The noise reduction branch outputs mask information. The activation function of DNN1 is Sigmoid, which is multiplied by the original noisy features to obtain the estimated clean features. Then, the clean speech is obtained through inverse transformation and overlapping addition operations.

[0068] DNN2: Uses a 1-layer fully connected network to identify speech activity features and obtain target values. The target value is 0 or 1, where 0 indicates that the current speech activity feature does not match the preset activity feature, and 1 indicates that the current speech activity feature matches the preset activity feature.

[0069] Step S14: Train the preset neural network model using speech data samples and label information so that the preset neural network model learns the correspondence between clean speech features and noisy speech features, as well as between noisy speech features and speech activity features, to obtain a speech processing model.

[0070] In this embodiment, step S14 involves training a preset neural network model using speech data samples and label information to enable the preset neural network model to learn the correspondence between clean speech features and noisy speech features, as well as between noisy speech features and speech activity features, thereby obtaining a speech processing model. This includes the following steps B1-B3:

[0071] Step B1: Input the speech data sample into a preset neural network model so that the convolutional neural network in the preset neural network model extracts the first feature of the speech data sample. The first feature is then input into the connection unit and the second recurrent neural network. The second recurrent neural network detects the first speech activity feature and the first noise speech feature in the first feature and extracts the first speech activity feature. The first speech activity feature is then input into the connection unit and the second deep neural network. The connection unit concatenates the first feature and the first speech activity feature to obtain the second feature. The second feature is then input into the first recurrent neural network, which detects the second speech activity feature and the second noise speech feature in the second feature and extracts the second speech activity feature. The second speech activity feature is then input into the first deep neural network for feature superposition to obtain the third feature. At the same time, the target value output by the second recurrent neural network is determined.

[0072] It should be noted that the preset neural network model includes two branches. The first branch is used to extract speech activity features and includes a second recurrent neural network and a second deep neural network. The second branch is used to denoise the speech data and includes connection units, a first recurrent neural network, and a first deep neural network. Both the first and second recurrent neural networks are used to detect speech activity features and noisy speech features in the features.

[0073] Step B2: When the target value is used to indicate that the speech activity feature meets the preset feature, determine the initial speech data corresponding to the speech data sample, and extract the fourth feature corresponding to the initial speech data.

[0074] Step B3: Calculate the loss function value based on the third and fourth features. If the loss function value is less than a preset threshold, determine the preset neural network model as the speech processing model.

[0075] In this embodiment, the calculation process of the loss function is as follows:

[0076]

[0077] In the embodiments of this application, s target For cross-entropy, and e noise For characteristic changes, < and > denote the dot product of vectors. s is the fourth feature, s is the third feature, and SI-SNR is the loss function value.

[0078] This application uses noise-free speech data and noise-mixed samples during model training, and combines this with speech activity features from the samples for training, which improves the noise reduction effect of speech under the same model computational requirements. Furthermore, this application employs different loss functions for different branches during model training: SI-SNR is used as the loss function for the first branch, and cross-entropy is used as the loss function for the second branch. This multi-task learning approach yields the final speech processing model, resulting in better noise reduction performance.

[0079] Figure 3 A flowchart of a method for denoising speech data provided in an embodiment of this application is shown below. Figure 3 As shown, the method may include the following steps:

[0080] Step S21: Obtain the raw speech data to be processed.

[0081] Step S22: Input the raw speech data into the pre-trained speech processing model so that the speech processing model can extract the feature information of the raw speech data and output the target value and target speech features based on the feature information.

[0082] In this embodiment of the application, step S22, inputting the original speech data into a pre-trained speech processing model so that the speech processing model outputs target values ​​and target speech features, includes:

[0083] The raw speech data is input into the speech processing model, where the convolutional neural network extracts the raw features of the raw speech data. The raw features are then input into a connection unit and a second recurrent neural network. The second recurrent neural network detects the raw speech activity features and raw noise speech features in the first raw features and extracts the raw speech activity features. The raw speech activity features are then input into the connection unit and the second deep neural network. The connection unit concatenates the raw features and raw speech activity features to obtain concatenated features. The concatenated features are then input into the first recurrent neural network, where the first recurrent neural network detects the target speech activity features and target noise speech features in the concatenated features and extracts the target speech activity features. The target speech activity features are then input into the first deep neural network for feature superposition to obtain the target speech features. At the same time, the target value output by the second deep neural network is determined.

[0084] It should be noted that, in this embodiment, the speech processing model has two branches after the convolutional neural network. The first branch is used to extract speech activity features and includes a second recurrent neural network and a second deep neural network. The second branch is used to denoise the speech data and includes a connection unit, a first recurrent neural network, and a first deep neural network.

[0085] Specifically, the convolutional neural network (CNN) is used to extract the raw features from the original speech data. The CNN then inputs these raw features into the connection unit and the second recurrent neural network (RNN). The second RNN extracts the raw speech activity features from the first set of features and inputs these features into the second deep neural network. The second deep neural network then determines the authenticity of the raw speech activity features and obtains the target value. Simultaneously, the second RNN also provides the raw speech activity features to the connection unit.

[0086] The connection unit concatenates the original speech activity features with the first feature to obtain the concatenated features. Then, the concatenated features are denoised by the first recurrent neural network and the first deep neural network to obtain the target speech features.

[0087] Step S23: When the target value is used to represent that the target speech features meet the preset features, target speech data is generated based on the target speech features. The preset features are the features of speech data that do not carry noise.

[0088] In this embodiment of the application, when the target value is 1, the target value is determined to indicate that the target speech features meet the preset features, and at this time the target speech data can be generated according to the target speech features.

[0089] This application's embodiments employ a two-branch approach to process speech data. Firstly, the presence of speech activity features in the original speech data can be directly determined using the target value; if no speech activity features are found, no processing is performed, thus improving the efficiency of the speech processing model. Secondly, the branch extracting speech activity features uses a simpler model structure with far fewer parameters than the noise reduction branch, reducing computational load and storage resource consumption.

[0090] Figure 4 This is a block diagram of a speech processing model training device provided in an embodiment of this application. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 4 As shown, the device includes:

[0091] The first acquisition module 41 is used to acquire speech data samples, wherein the speech sample data includes multiple frames of speech data obtained after noise mixing.

[0092] The second acquisition module 42 is used to acquire the label information corresponding to the speech data sample, wherein the label information is used to mark the clean speech features, noisy speech features and speech activity features in the speech data sample.

[0093] Module 43 is used to determine the preset neural network model.

[0094] Training module 44 is used to train the preset neural network model using the speech data samples and the label information, so that the preset neural network model learns the correspondence between the clean speech features and the noisy speech features, as well as the noisy speech features and the speech activity features, to obtain a speech processing model.

[0095] In this embodiment of the application, the first acquisition module 41 is used to acquire initial speech data, wherein the initial speech data is speech data carrying clean speech features and speech activity features; determine multiple types of initial noise features and a preset signal-to-noise ratio; and mix the initial speech data and the initial noise features according to the preset signal-to-noise ratio to obtain the speech data sample.

[0096] In this embodiment, the preset neural network model includes: a convolutional neural network, a first recurrent neural network, a second recurrent neural network, a first deep neural network, and a second deep neural network; the convolutional neural network is connected to the first recurrent neural network and the second recurrent neural network respectively, the first recurrent neural network is connected to the first deep neural network, and the second recurrent neural network is connected to the second deep neural network. The convolutional neural network is connected to the first recurrent neural network through a connection unit, and the second recurrent neural network is also connected to the first recurrent neural network through a connection unit.

[0097] In this embodiment, the training module 44 is used to input speech data samples into a preset neural network model, so that the convolutional neural network in the preset neural network model extracts a first feature of the speech data sample, and inputs the first feature into a connection unit and a second recurrent neural network respectively. The second recurrent neural network detects a first speech activity feature and a first noise speech feature in the first feature, and extracts the first speech activity feature. The first speech activity feature is input into the connection unit and the second deep neural network respectively. The connection unit concatenates the first feature and the first speech activity feature to obtain a second feature. The second feature is input into the first recurrent neural network, and the first recurrent neural network detects the second speech activity feature and the second noise speech feature in the second feature, extracts the second speech activity feature, and inputs the second speech activity feature into the first deep neural network for feature superposition to obtain a third feature. At the same time, the target value output by the second deep neural network is determined.

[0098] Figure 5 This is a block diagram of a voice data noise reduction device provided in an embodiment of this application. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 5 As shown, the device includes:

[0099] The acquisition module 51 is used to acquire the raw voice data to be processed.

[0100] Extraction module 52 is used to input the raw speech data into a pre-trained speech processing model so that the speech processing model can extract the features of the raw speech data and output the target value and target speech features based on the features.

[0101] The processing module 53 is used to generate target speech data based on the target speech features when the target value is used to represent the target speech features that meet the preset features. The preset features are the features of speech data that do not carry noise.

[0102] This application also provides an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 1501, a communication interface 1502, a memory 1503, and a communication bus 1504, wherein the processor 1501, the communication interface 1502, and the memory 1503 communicate with each other through the communication bus 1504.

[0103] Memory 1503 is used to store computer programs;

[0104] When the processor 1501 executes the computer program stored in the memory 1503, it implements the steps of the above embodiments.

[0105] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0106] The communication interface is used for communication between the aforementioned terminal and other devices.

[0107] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0108] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0109] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the training method of any of the speech processing models described in the above embodiments.

[0110] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the training method of any of the speech processing models described in the above embodiments.

[0111] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive).

[0112] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

[0113] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for denoising speech data, characterized in that, include: Acquire the raw speech data to be processed; The raw speech data is input into a pre-trained speech processing model so that the speech processing model can extract the feature information of the raw speech data and output the target value and target speech features based on the feature information. When the target value is used to indicate that the target speech feature satisfies a preset feature, target speech data is generated based on the target speech feature, wherein the preset feature is a feature of speech data that does not carry noise; The step of inputting the raw speech data into a pre-trained speech processing model, so that the speech processing model extracts feature information from the raw speech data, and outputs target values ​​and target speech features based on the feature information, includes: The raw speech data is input into the speech processing model, whereby the convolutional neural network in the speech processing model extracts the raw features of the raw speech data. The raw features are then input into a connection unit and a second recurrent neural network. The second recurrent neural network detects the raw speech activity features and raw noise speech features in the raw features and extracts the raw speech activity features. The raw speech activity features are then input into the connection unit and the second deep neural network. The connection unit concatenates the raw features and raw speech activity features to obtain concatenated features. The concatenated features are then input into the first recurrent neural network, where the first recurrent neural network detects the target speech activity features and target noise speech features in the concatenated features and extracts the target speech activity features. The target speech activity features are then input into the first deep neural network for feature superposition to obtain the target speech features. Simultaneously, the target value output by the second deep neural network is determined. The speech activity features are used to characterize the sounds present in the initial speech data.

2. A noise reduction device for voice data, employing the method described in claim 1, characterized in that, include: The acquisition module is used to acquire the raw voice data to be processed; An extraction module is used to input the raw speech data into a pre-trained speech processing model so that the speech processing model extracts the features of the raw speech data and outputs a target value and target speech features based on the features; The processing module is configured to generate target speech data based on the target speech features when the target value indicates that the target speech features satisfy preset features, wherein the preset features are features of speech data that do not carry noise.

3. A storage medium, characterized in that, The storage medium includes a stored program, wherein the program executes the steps of the method described in claim 1 when it is run.

4. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other through the communication bus; wherein: Memory, used to store computer programs; A processor for executing the steps of the method of claim 1 by running a program stored in memory.

Images