Data expansion methods, apparatus, electronic devices and readable storage media

Abstract

This application discloses a data augmentation method, apparatus, electronic device, and readable storage medium, belonging to the field of data processing technology. The method includes: acquiring second acoustic feature data matching first acoustic feature data based on a first model; determining target acoustic feature data from the second acoustic feature data based on a second model; and recombining the target acoustic feature data to obtain augmented acoustic feature data corresponding to the target acoustic feature data.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, specifically relating to a data expansion method, apparatus, electronic device, and readable storage medium. Background Technology

[0002] Currently, the training of voice recognition models usually depends heavily on the quality and quantity of training acoustic feature data.

[0003] However, as the types of inputs to the aforementioned sound recognition model increase, the quality of the training acoustic feature data corresponding to the sound recognition model deteriorates, and the gap between the number of any type of training acoustic feature data and the number of other training acoustic feature data becomes larger and larger. This leads to a higher probability of overfitting when learning a certain type of training acoustic feature data. Summary of the Invention

[0004] The purpose of this application is to provide a data augmentation method, apparatus, electronic device, and readable storage medium that can solve the problem of a high probability of overfitting when learning a certain type of training acoustic feature data.

[0005] In a first aspect, embodiments of this application provide a data augmentation method, which includes: acquiring second acoustic feature data that matches first acoustic feature data based on a first model; determining target acoustic feature data from the second acoustic feature data based on a second model; and recombining the target acoustic feature data to obtain augmented acoustic feature data corresponding to the target acoustic feature data.

[0006] Secondly, embodiments of this application provide a data augmentation device, which includes an acquisition module, a determination module, and a processing module; the acquisition module is used to acquire second acoustic feature data that matches the first acoustic feature data based on a first model; the determination module is used to determine target acoustic feature data from the second acoustic feature data based on a second model; and the processing module is used to reorganize the target acoustic feature data to obtain augmented acoustic feature data corresponding to the target acoustic feature data.

[0007] Thirdly, embodiments of this application provide an electronic device including a processor and a memory, wherein the memory stores programs or instructions executable on the processor, and the programs or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0009] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0010] In a sixth aspect, embodiments of this application provide a computer program product stored in a storage medium, which is executed by at least one processor to implement the method described in the first aspect.

[0011] In this embodiment, based on a first model, second acoustic feature data matching the first acoustic feature data can be obtained; and based on a second model, target acoustic feature data can be determined from the second acoustic feature data; and the target acoustic feature data can be reorganized to obtain expanded acoustic feature data corresponding to the target acoustic feature data. Through this scheme, target acoustic feature data in the second acoustic feature data matching the first acoustic feature data can be obtained sequentially based on the first and second models, thereby improving the quality of the second acoustic feature data; and the expansion of the target acoustic feature data can be achieved through the reorganization process, thus reducing the probability of overfitting in the learning of the target acoustic feature data. Attached Figure Description

[0012] Figure 1 This is a flowchart of the data augmentation method provided in the embodiments of this application;

[0013] Figure 2 This is one of the architectural schematic diagrams of the first sub-model in some embodiments of this application;

[0014] Figure 3 This is the second schematic diagram of the architecture of the first sub-model in some embodiments of this application;

[0015] Figure 4 This is a schematic diagram of the architecture of the second sub-model in some embodiments of this application;

[0016] Figure 5 This is one of the schematic diagrams of the data augmentation method provided in the embodiments of this application;

[0017] Figure 6a This is a second schematic diagram of the data augmentation method provided in the embodiments of this application;

[0018] Figure 6b This is the third schematic diagram of the data augmentation method provided in the embodiments of this application;

[0019] Figure 7 This is a schematic diagram of the data expansion device provided in the embodiments of this application;

[0020] Figure 8This is a schematic diagram of the electronic device provided in the embodiments of this application;

[0021] Figure 9 This is a hardware schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0023] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0024] The following section will first explain some of the terms or terms used in the specification and claims of this application.

[0025] Acoustic feature data refers to the data of physical quantities that can represent the acoustic characteristics of speech, and is a general term for the acoustic performance of various sound elements. For example, acoustic feature data can be the formant frequency, formant intensity or bandwidth representing timbre, or the duration, fundamental frequency, average speech power, etc. representing the prosodic characteristics of speech.

[0026] Generative adversarial models consist of a generative model and a discriminative model. The generative model is responsible for capturing the distribution of the sample data; the discriminative model is typically a binary classifier that distinguishes between real data and generated samples.

[0027] Voice recognition model: An algorithmic model that can recognize voice as a predefined class.

[0028] The data expansion method, apparatus, electronic device, and readable storage medium provided in this application will be described in detail below with reference to the accompanying drawings and through specific embodiments and application scenarios.

[0029] Sound contains a wealth of information about everyday life and physical events. In cognitive science research, auditory cognition is often considered the second most important sensory system after visual cognition. Clearly, auditory cognition, as a crucial way of perceiving the environment, has undeniable research value and development potential. Sound recognition aims to identify sounds as predefined classes. The significance of sound recognition research is evident in many fields, such as urban noise monitoring, fire early warning systems, and health status detection.

[0030] Currently, the training of sound recognition models typically relies heavily on the quality and quantity of training acoustic feature data. However, as the types of inputs to these sound recognition models increase, the quality of the corresponding training acoustic feature data deteriorates, and the gap between the quantity of any one type of training acoustic feature data and the quantity of other types widens. This increases the probability that the sound recognition model will overfit to a particular type of training acoustic feature data.

[0031] The data augmentation method provided in this application can be applied to scenarios involving the training of sound recognition models.

[0032] In this embodiment, a second acoustic feature data matching the training acoustic feature data of the sound recognition model (e.g., the first model in this embodiment) can be obtained based on a generative adversarial model (e.g., the first acoustic feature data in this embodiment). Based on the sound recognition model (e.g., the second model in this embodiment), target acoustic feature data (e.g., dog bark data or cat meow data) is determined from the second acoustic feature data. The target acoustic feature data is then recombined to obtain expanded acoustic feature data corresponding to the target acoustic feature data. This scheme allows for the sequential acquisition of target acoustic feature data from the second acoustic feature data matching the training acoustic feature data, based on the generative adversarial model and the sound recognition model, thereby improving the quality of the training acoustic feature data corresponding to the sound recognition model. Furthermore, the expansion of the target acoustic feature data can be achieved through recombining the target acoustic feature data, thus reducing the probability of overfitting during learning of the target acoustic feature data. This enables the sound recognition model to learn more spectral knowledge and improve its generalization ability.

[0033] It should be noted that the data expansion method provided in this application can be executed by a data expansion device, an electronic device, or a functional module within an electronic device. Some embodiments of this application use an electronic device executing the data expansion method as an example to illustrate the data expansion method provided in this application.

[0034] Figure 1 A flowchart of the data augmentation method provided in an embodiment of this application is shown. Figure 1 As shown, the data augmentation method provided in this application embodiment may include the following steps 101 to 103.

[0035] Step 101: The electronic device acquires second acoustic feature data that matches the first acoustic feature data based on the first model.

[0036] In some embodiments of this application, the first model described above can be used to generate the second acoustic feature data described above.

[0037] In some embodiments of this application, the first model described above can be the generative adversarial model described above.

[0038] In some embodiments of this application, the second model described above can be used to identify the audio type corresponding to the acoustic feature data.

[0039] In some embodiments of this application, the above-mentioned audio type may include any type such as dog barking, airplane sound, or flowing water sound.

[0040] In some embodiments of this application, the second model described above may be the aforementioned sound recognition model.

[0041] In some embodiments of this application, the first acoustic feature data mentioned above may be the trained acoustic feature data corresponding to the second model mentioned above.

[0042] In some embodiments of this application, the electronic device can extract the audio length and sampling rate information corresponding to the trained audio samples corresponding to the second model, and perform preprocessing such as frame segmentation and windowing on the extracted audio length and sampling rate information in sequence. Then, the preprocessed audio length and sampling rate information are processed by Fourier transform and a filter for extracting acoustic features in sequence to obtain the first acoustic feature data.

[0043] In some embodiments of this application, the dimension of the first acoustic feature data can be any dimension such as 500×128 or 600×256; for example, the dimension of the first acoustic feature data is 500×128, of which the time domain dimension is 500 and the frequency domain dimension is 128.

[0044] In some embodiments of this application, the dimension of the second acoustic feature data is the same as the dimension of the first acoustic feature data.

[0045] In some embodiments of this application, the matching of the second acoustic feature data with the first acoustic feature data can be understood as: the matching degree between the second acoustic feature data and the first acoustic feature data is greater than or equal to the matching degree threshold.

[0046] In some embodiments of this application, the matching threshold can be preset by the system or set by the user according to actual usage needs.

[0047] The specific methods for electronic devices to acquire the aforementioned second acoustic feature data are explained in detail below.

[0048] In some embodiments of this application, the first model may include a first sub-model and a second sub-model; then step 101 may be implemented by steps 101a to 101c as described below.

[0049] Step 101a: The electronic device inputs the randomly generated Q groups of noise data into the first sub-model in groups to obtain the third acoustic feature data of the Q groups.

[0050] Among them, the above Q groups of noise data correspond one-to-one with the above Q groups of third acoustic feature data. The dimension of each group of noise data in the Q groups of noise data is the same as the dimension of the above first acoustic feature data, and Q is a positive integer.

[0051] In some embodiments of this application, the electronic device may first input one set of noise data from the Q groups of noise data into the first sub-model to obtain a set of third acoustic feature data corresponding to the set of noise data; then the electronic device may input another set of noise data from the Q groups of noise data into the first sub-model to obtain a set of third acoustic feature data corresponding to the other set of noise data; and so on, the electronic device may obtain the Q groups of third acoustic feature data that correspond one-to-one with the Q groups of noise data.

[0052] In some embodiments of this application, the dimension of each group of noise data is the same as the dimension of the first acoustic feature data, so that the dimension of each group of third acoustic feature data in the obtained Q groups of third acoustic feature data is the same as the dimension of the first acoustic feature data, thereby ensuring that the obtained third acoustic feature data has a large degree of matching with the first acoustic feature data.

[0053] In some embodiments of this application, the specific value of Q can be any positive integer preset by the system, or it can be any positive integer set by the user according to actual usage needs.

[0054] In some embodiments of this application, the first sub-model described above can be the generative model in the generative adversarial model described above.

[0055] In some embodiments of this application, the above-described generative model can be a generator model Gmodel obtained through training.

[0056] For example, such as Figure 2As shown, the main architecture of the above generative model consists of an encoder network 21 and a decoder network 22. The encoder network 21 is used to reduce the size of the features, increase the number of channels, and increase the receptive field of the convolutional kernel on the feature map, thereby further iterating the information to be generated; the decoder network 22 is used to increase the feature size, reduce the number of channels, and gradually restore the size of the original feature map through the learned information, thereby completing the generation of the new feature map.

[0057] The encoder network 21 consists of multiple encoding blocks (i.e., Encoder_block) and multiple decoding blocks (i.e., Decoder_block).

[0058] The structure of the above-mentioned coded block is as follows: Figure 3 As shown, frequency domain downsampling is performed first, followed by time domain downsampling, and finally, the encoded information from both is integrated using a channel attention mechanism. The encoding of the frequency and time domains is mainly achieved through two methods: one is downsampling followed by convolution, which helps to increase the receptive field during the encoding process; the other is convolution followed by pooling, which helps to sparsify the model and improve its convergence speed and performance. The outputs of the two methods are then superimposed, as detailed in the following steps:

[0059] First, the 500×128 dimensional noise data is used as the input feature map x. Frequency domain downsampling + convolution and convolution + frequency domain pooling operations are performed respectively to indirectly expand the receptive field of the convolution kernel in the frequency domain. The outputs of the two branches are superimposed to obtain x2. Second, the feature map x2 is subjected to frequency domain downsampling + convolution and convolution + temporal pooling operations respectively to indirectly expand the receptive field of the convolution kernel in the temporal domain. The outputs of the two branches are superimposed to obtain x3. Then, the result is compressed into a one-dimensional vector x3 through global pooling and the weights of each channel are recalibrated through a fully connected layer and a sigmoid function to obtain x4. Finally, the convolutioned x2 and the weight vector x4 are multiplied to obtain the output feature map x5 to realize the inter-channel attention mechanism. The algorithm flow is shown in the following formula (1):

[0060]

[0061] Among them, D[·] 1×r This indicates frequency domain downsampling operations, where 1×r represents the downsampling size; D[·] r×1 This represents the temporal downsampling operation; [w1,w2,w3,w4] represents the weight matrix of the four convolutional layers, and [b1,b2,b3,b4] represents the bias values ​​of the four convolutional layers; F sq(·) is the squeezing activation function, which compresses the feature map of size H×W×C (H is the time domain dimension, W is the frequency domain dimension, and C is the number of channels) into a one-dimensional vector; σ[·] represents the sigmoid function, f(·) represents the linear rectified activation function, p(·) represents the pooling operation, and w5 represents the weight of the fully connected layer.

[0062] Thus, the encoder network 21 described above uses two downsampling methods and adds them together, which can help the model learn the backpropagation loss better and continuously increase the similarity between the generated features and the original features; and the channel attention mechanism can help align the correspondence between the feature maps before and after each downsampling.

[0063] The structure of the above decoding block is implemented by simple upsampling and convolution operations, which can reduce the number of parameters and help the model converge quickly. The convolutional layers in the structure of this decoding block adopt a two-dimensional convolutional neural network, with the number of convolutional kernels set to 256, 128, 64, and 32 respectively. Before each convolution operation, a 2*2 upsampling kernel is used to progressively increase the feature size.

[0064] Step 101b: The electronic device inputs the first acoustic feature data and a set of third acoustic feature data from the Q group into the second sub-model in sequence to obtain the probability score corresponding to each set of third acoustic feature data.

[0065] Each probability score in the probability scores corresponding to the third acoustic feature data in the Q group is used to indicate the degree of similarity between the corresponding third acoustic feature data and the first acoustic feature data.

[0066] In some embodiments of this application, the value range of the above probability score can be [0, 1].

[0067] In some embodiments of this application, if the probability score corresponding to a set of third acoustic feature data is close to 0 or 1, it can be determined that the similarity between the set of third acoustic feature data and the first acoustic feature data is small (i.e., the difference is large, and the set of third acoustic feature data is unusable); conversely, it can be determined that the similarity between the set of third acoustic feature data and the first acoustic feature data is large (i.e., the difference is small, and the set of third acoustic feature data is usable).

[0068] In some embodiments of this application, the second sub-model described above can be the discriminative model in the generative adversarial model described above.

[0069] In some embodiments of this application, the above-mentioned discriminant model can be a discriminator network D_model obtained through training.

[0070] For example, such as Figure 4As shown, the discriminator network 41 employs a Convolutional Neural Network (CNN) model, specifically a two-dimensional CNN, with the number of convolutional kernels set to 16, 32, 64, and 128 respectively. The electronic device can input the acoustic feature data generated by the first sub-model into the discriminator network 41 for training, and output the probability score (i.e., the aforementioned probability score) of whether the acoustic feature data generated by the first sub-model is real data.

[0071] In some embodiments of this application, the electronic device may first input the first acoustic feature data and one set of third acoustic feature data from the Q groups into the second sub-model to obtain the probability score corresponding to the set of third acoustic feature data; then input the first acoustic feature data and another set of third acoustic feature data from the Q groups into the second sub-model to obtain the probability score corresponding to the other set of third acoustic feature data; and so on, so that the electronic device can obtain the probability score corresponding to each set of third acoustic feature data.

[0072] It should be noted that, in actual implementation, the electronic device can first generate the Q sets of third acoustic feature data through the first sub-model, and after the generation is completed, input the first acoustic feature data and one set of third acoustic feature data from the Q sets of third acoustic feature data into the second sub-model in sequence; or, the electronic device can first generate one set of third acoustic feature data from the Q sets of third acoustic feature data through the first sub-model, input the first acoustic feature data and the set of third acoustic feature data into the second sub-model, then generate another set of third acoustic feature data from the Q sets of third acoustic feature data through the first sub-model, and input the first acoustic feature data and the other set of third acoustic feature data into the second sub-model, and so on.

[0073] Step 101c: The electronic device calculates the credibility of each group of third acoustic feature data based on the probability score corresponding to each group of third acoustic feature data, and saves K groups of third acoustic feature data with credibility greater than or equal to the second threshold as second acoustic feature data.

[0074] Where K is a positive integer less than or equal to Q.

[0075] In some embodiments of this application, when the probability score corresponding to a group of third acoustic feature data in the above-mentioned Q group approaches a certain value a between 0 and 1, the electronic device can calculate the credibility of the group of third acoustic feature data based on the probability score.

[0076] Specifically, assume that the probability score corresponding to the above set of third acoustic feature data is p. QThen, the confidence level T of the third acoustic feature data can be calculated using the following formula (2):

[0077] T = a / |p Q -a|; (2)

[0078] In some embodiments of this application, the electronic device may set the second threshold R = a × N (10 ≤ N ≤ 20), where N is the stringency coefficient of the screening. If T ≤ R, the above set of third acoustic feature data is determined to be unreliable data, and there is no need to save the set of third acoustic feature data. If T > R, the set of third acoustic feature data is determined to be reliable data, and the set of third acoustic feature data is saved as the above second acoustic feature data.

[0079] It should be noted that, in actual implementation, the second threshold can be any value set by the user according to actual usage needs, and this application embodiment does not limit it.

[0080] In this embodiment of the application, since the electronic device can expand the first acoustic feature data based on the first sub-model and select the second acoustic feature data with higher credibility from the expanded acoustic feature data based on the second sub-model, the quantity and quality of the expanded acoustic feature data can be improved.

[0081] Step 102: The electronic device determines the target acoustic feature data from the second acoustic feature data based on the second model.

[0082] In some embodiments of this application, the second acoustic feature data may include acoustic feature data of multiple audio types, and the target acoustic feature data may be acoustic feature data corresponding to the same audio type.

[0083] In some embodiments of this application, the target acoustic feature data described above may be the acoustic feature data that the second model needs to learn.

[0084] In some embodiments of this application, step 102 can be specifically implemented by steps 102a and 102b as described below.

[0085] Step 102a: The electronic device scores each acoustic feature data in the second acoustic feature data using the second model.

[0086] For the specific method of scoring each of the above acoustic feature data by electronic devices, please refer to the specific description in the relevant technology. To avoid repetition, it will not be elaborated here.

[0087] Step 102b: The electronic device identifies the acoustic feature data in the second acoustic feature data whose corresponding score is greater than or equal to the first threshold as the target acoustic feature data.

[0088] In some embodiments of this application, the first threshold can be any numerical value that can indicate the audio type corresponding to the target acoustic feature data.

[0089] In some embodiments of this application, the first threshold can be one of a plurality of thresholds, each of which can be used to determine acoustic feature data of an audio type.

[0090] The data augmentation method provided in the embodiments of this application will be described exemplarily below.

[0091] For example, suppose that the score corresponding to one of the acoustic feature data in the second acoustic feature data is A, and the first threshold is B; then, if A < B, the acoustic feature data is determined not to be the target acoustic feature data and is discarded; if A ≥ B, the acoustic feature data is determined to be the target acoustic feature data and is retained.

[0092] In this embodiment of the application, since the electronic device can determine the acoustic feature data whose corresponding score is greater than or equal to the first threshold in the second acoustic feature data as the target acoustic feature data, the accuracy of screening acoustic feature data can be improved.

[0093] Step 103: The electronic device reconstructs the target acoustic feature data to obtain the extended acoustic feature data corresponding to the target acoustic feature data.

[0094] In some embodiments of this application, the aforementioned extended acoustic feature data can be used as extended data for the aforementioned target acoustic feature data and learned together by the aforementioned second model.

[0095] The following section provides a detailed explanation of the specific methods used by electronic devices to reconstruct the acoustic feature data of the aforementioned targets.

[0096] In some embodiments of this application, step 103 can be specifically implemented by steps 103a to 103c as described below.

[0097] Step 103a: The electronic device cuts the target acoustic feature data into M data segments.

[0098] Where M is an integer greater than or equal to 2.

[0099] In some embodiments of this application, the electronic device can divide the aforementioned target acoustic feature data into the aforementioned M data segments based on the time dimension.

[0100] In some embodiments of this application, any two data segments among the M data segments can have the same dimension.

[0101] For example, if the dimension of the target acoustic feature data is 500×128, then the dimension of each data segment in the above M data segments is 500 / M×128, that is, any two data segments in the M data segments have the same dimension.

[0102] In some embodiments of this application, step 103a above can be specifically implemented by step 103a1 below.

[0103] Step 103a1: The electronic device cuts the target acoustic feature data into M data segments according to the minimum cutting time, maximum cutting time and cutting step size corresponding to the target acoustic feature data.

[0104] In some embodiments of this application, the electronic device may first determine the specific value of M based on the minimum cutting time, the maximum cutting time, and the cutting step size, and then cut the target acoustic feature data into the M data segments.

[0105] For example, assume the minimum cutting time mentioned above is l min The maximum cutting time mentioned above is l max The above cutting step size is l hop Then the electronic device can determine the above M through the following steps (a) and (b):

[0106] (a) From l min Initially, the length increases by l each time. hop until it equals l max So far, the target acoustic feature data is cut sequentially and the scores are tested;

[0107] (b) For each cut, retain the highest score of the segment and the current cut length to obtain (l max -l min ) / l hop The maximum score values ​​are compared, and the cutting length corresponding to the maximum score value is M as mentioned above.

[0108] In this embodiment, since the electronic device can divide the target acoustic feature data into M data segments according to the minimum cutting time, maximum cutting time and cutting step size corresponding to the target acoustic feature data, it can adaptively determine the optimal segment value for different target acoustic feature data, thereby improving the flexibility of determining the segment value of acoustic feature data.

[0109] Step 103b: The electronic device stores the M data segments in a set of M score intervals based on the highest score of the acoustic feature data contained in each of the M data segments.

[0110] Among them, the above M score interval sets correspond one-to-one with the above M data segments.

[0111] In some embodiments of this application, the one-to-one correspondence between the above-mentioned M score interval sets and the above-mentioned M data segments can be understood as: in one of the score interval sets of the M score interval sets, one of the data segments of the M data segments is stored.

[0112] The following is an exemplary description of the specific method by which an electronic device stores the aforementioned M data segments in the aforementioned M score interval sets.

[0113] For example, suppose we statistically analyze the above M data segments T j The highest score S of acoustic feature data in (j = 1, 2, ..., M) j (j = 1, 2, ..., M), corresponding to M score interval sets U. j (j = 1, 2, ..., M); then the electronic device can store the M data segments in the set of M score intervals in the following way:

[0114] (1) When 0 ≤ S j When T < (1 / M), j Place it in U1;

[0115] (2) When (1 / M)≤S j When T < (2 / M), j Place it in U2;

[0116] ...

[0117] (M) When [(M-1) / M]≤S j When <1, T j Placed in U M .

[0118] It should be noted that, at this point, the aforementioned set U of M score intervals... j The amount of acoustic feature data contained in them can vary.

[0119] Step 103c: The electronic device extracts N sets of acoustic feature data from the M score interval sets, and splices each set of acoustic feature data based on the time dimension to obtain expanded acoustic feature data.

[0120] Among them, any one set of acoustic feature data in the above N sets includes: an acoustic feature data randomly selected from each of the at least two score interval sets, wherein the at least two score interval sets are the score interval sets among the above M score interval sets, and N is a positive integer.

[0121] For example, suppose the above at least two score interval sets are score interval set a and score interval set b, and score interval set a stores acoustic feature data 1 and acoustic feature data 2, and score interval set b stores acoustic feature data 3 and acoustic feature data 4; then one set of acoustic feature data in the above N sets of acoustic feature data may include: acoustic feature data 1 randomly selected from score interval set a, and acoustic feature data 4 randomly selected from score interval set b.

[0122] In some embodiments of this application, the electronic device may employ a segment extractor to extract the above-mentioned N sets of acoustic feature data from the above-mentioned M score interval sets.

[0123] The following example illustrates the specific method for obtaining the expanded acoustic feature data using the above-mentioned fragment extractor in an electronic device, taking the above-mentioned at least two score interval sets as the above-mentioned M score interval sets.

[0124] For example, the electronic device can first shuffle the data order in the aforementioned M score interval sets to ensure that the subsequently extracted segments do not come from the same feature map, increasing feature diversity. Then, the electronic device can set a segment extractor W, which contains M elements (0, 1, 2, ..., M-1), representing the index values ​​of the M score interval sets. Next, the electronic device can shuffle the element positions within the segment extractor W and extract an acoustic feature data from the corresponding score interval set according to the index in the segment extractor W. Then, the extracted M acoustic feature data (dimension 500 / M×128) are concatenated along the time dimension to obtain an augmented acoustic feature data (dimension restored to 500×128). This process can be repeated to obtain all the aforementioned augmented acoustic feature data. This ensures that the obtained augmented acoustic feature data has the same size as the target acoustic feature data, and that the newly combined features cover all data segments.

[0125] In some embodiments of this application, the electronic device can specify the content of the target feature in the required reconstructed feature by setting the elements inside the above-mentioned fragment extractor W (the higher the score of the data fragment, the greater the content of the target feature); for example, when the values ​​of the above M elements are all set to M-1, the data fragments inside the obtained feature all come from the set of the highest score intervals, so the content of the target feature is greater.

[0126] In some embodiments of this application, the electronic device can adaptively adjust the values ​​inside the segment extractor W according to the learning rate of the second model; the specific implementation process is as follows:

[0127] (1) Map the learning rate variation range of the second model above to the interval [M×(M-1) / 2, M×(M-1)], where M×(M-1) / 2 corresponds to the minimum acceptable sum of values ​​in the fragment extractor W, and M×(M-1) corresponds to the maximum sum of values ​​in the fragment extractor W.

[0128] (2) When the value of the above learning rate mapping is y, set the sum of the values ​​in the above fragment extractor W to [3M×(M-1) / 2]-y, so as to ensure that the learning rate is inversely proportional to the sum of the values ​​in the fragment extractor W after being mapped proportionally.

[0129] (3) When the second model above adopts a fixed learning rate, the sum of the values ​​in the above fragment extractor W is fixed at M×(M-1) / 2 to ensure the balance of the extracted data.

[0130] For example, when M=5, the usage process of the above fragment extractor is as follows: Figure 5 As shown, the shuffled extraction order of the segment extractor is B, E, D, A, C. Thus, the segment extractor sequentially extracts one acoustic feature data from interval (i.e., the set of score intervals) B, interval E, interval D, interval A, and interval C, and then splices the extracted acoustic feature data to obtain the acoustic feature data used for training (i.e., the expanded acoustic feature data).

[0131] In some embodiments of this application, when the second model has a large learning rate, it can train extended acoustic features composed of combinations of various score segments to help the second model improve its discriminative power and avoid overfitting or getting trapped in local optima; when the second model has a small learning rate, it trains extended acoustic features composed of combinations of higher score segments to help the second model continuously learn features and gradually approach the model's optimal point.

[0132] In this embodiment, since the electronic device can store the M data segments in sets according to the highest score of the acoustic feature data contained in the M data segments obtained by cutting the target acoustic feature data, and extract the acoustic feature data from them and splice them based on the time dimension to obtain the above-mentioned expanded acoustic feature data, the amount of expanded acoustic feature data is greatly increased.

[0133] In some embodiments of this application, before step 103b above, the data augmentation method provided in the embodiments of this application may further include step 103d or step 103e below.

[0134] Step 103d: If the M data segments include P consecutive first data segments, the electronic device increases the highest score of the acoustic feature data in the second data segment among the M data segments by a first value.

[0135] Wherein, the highest score of the acoustic feature data in each of the P first data segments is greater than or equal to the third threshold, and the second data segment includes: a data segment adjacent to and preceding the P first data segments, and a data segment adjacent to and following the P first data segments, where P is an integer less than M and greater than or equal to 2.

[0136] In some embodiments of this application, the first value can be any value such as 0.1, 0.2 or 0.3, and can be determined according to actual usage requirements. This application does not limit the specific value.

[0137] In some embodiments of this application, the aforementioned third threshold may be preset by the system or may be set by the user according to actual usage needs.

[0138] In some embodiments of this application, the specific value of P can be determined based on the corresponding target acoustic feature data, and different target acoustic feature data can correspond to different P values.

[0139] In some embodiments of this application, the third threshold corresponding to the data segments cut from acoustic feature data of different audio types may be the same or different.

[0140] The data expansion method provided in the embodiments of this application will be described exemplarily below with reference to the accompanying drawings.

[0141] For example, such as Figure 6a As shown, the P segments are the aforementioned P first data segments. The electronic device can increase the highest score of the acoustic feature data in the segment before and after the P segments (i.e., the second data segment) by 0.1 (i.e., the first value) to ensure the smoothness of the scores of the segments when the continuous sound appears and when it ends.

[0142] Step 103e: If the highest score of the acoustic feature data in the third data segment among the M data segments is less than the third threshold, and the highest score of the acoustic feature data in the fourth data segment among the M data segments is greater than or equal to the third threshold, then the electronic device will increase the highest score of the acoustic feature data in the third data segment by a second value.

[0143] The fourth data segment includes: a data segment adjacent to and preceding the third data segment, and a data segment adjacent to and following the third data segment.

[0144] In some embodiments of this application, the second value can be any value such as 0.1, 0.2 or 0.3, and can be determined according to actual usage requirements. This application does not limit the specific value.

[0145] In some embodiments of this application, the second value may be the same as or different from the first value.

[0146] The data expansion method provided in the embodiments of this application will be described exemplarily below with reference to the accompanying drawings.

[0147] For example, such as Figure 6b As shown, the highest score of the acoustic feature data in data segment 61 (i.e., the fourth data segment) and the highest score of the acoustic feature data in data segment 62 (i.e., the fourth data segment) both exceed the threshold (i.e., the third threshold), while the highest score of the acoustic feature data in data segment 63 (i.e., the third data segment) does not exceed the threshold. Therefore, the electronic device can increase the highest score of the acoustic feature data in data segment 63 by 0.1 (i.e., the second value) to ensure that the score of the noise segment is compensated when noise occurs in the middle of the continuous sound.

[0148] In this embodiment, since the electronic device can use different scoring compensation strategies to compensate for the highest score of the acoustic feature data in the second / third data segment, the smoothness of the scores between the segmented data segments can be ensured, so as to ensure that the scores of the data placed in different score intervals are more realistic.

[0149] It should be noted that if the electronic device performs step 103d or step 103e, the electronic device can store the M data segments in the set of M score intervals based on the highest score of the acoustic feature data contained in each of the data segments after compensation.

[0150] In the data augmentation method provided in this application embodiment, target acoustic feature data in the second acoustic feature data that matches the first acoustic feature data can be obtained sequentially based on the first model and the second model to improve the quality of the second acoustic feature data; and the target acoustic feature data can be augmented by reorganizing the target acoustic feature data; thereby reducing the probability of overfitting when learning the target acoustic feature data.

[0151] The data expansion method provided in this application can be executed by a data expansion device. This application uses an example of a data expansion device executing the data expansion method to illustrate the data expansion device provided in this application.

[0152] Combination Figure 7This application provides a data augmentation device 70, which may include an acquisition module 71, a determination module 72, and a processing module 73. The acquisition module 71 is used to acquire second acoustic feature data that matches the first acoustic feature data based on a first model. The determination module 72 is used to determine target acoustic feature data from the second acoustic feature data based on a second model. The processing module 73 is used to reconstruct the target acoustic feature data to obtain augmented acoustic feature data corresponding to the target acoustic feature data.

[0153] In some embodiments of this application, the determination module 72 can be specifically used to score each acoustic feature data in the second acoustic feature data using the second model; and to determine the acoustic feature data in the second acoustic feature data whose corresponding score is greater than or equal to the first threshold as the target acoustic feature data.

[0154] In some embodiments of this application, the processing module 73 can be specifically used to cut the target acoustic feature data into M data segments, where M is an integer greater than or equal to 2; and according to the highest score of the acoustic feature data contained in each of the M data segments, store the M data segments in M ​​score interval sets, where the M score interval sets correspond one-to-one with the M data segments; and extract N sets of acoustic feature data from the M score interval sets, and splice each set of acoustic feature data based on the time dimension to obtain the expanded acoustic feature data, wherein any set of acoustic feature data in the N sets of acoustic feature data includes: an acoustic feature data randomly extracted from each of at least two score interval sets, where the at least two score interval sets are score interval sets in the M score interval sets, and N is a positive integer.

[0155] In some embodiments of this application, the processing module 73 described above can be specifically used to cut the target acoustic feature data into the above M data segments according to the minimum cutting time, maximum cutting time, and cutting step size corresponding to the target acoustic feature data.

[0156] In some embodiments of this application, the first model may include a first sub-model and a second sub-model. Specifically, the acquisition module 71 can be used to input Q groups of randomly generated noise data into the first sub-model to obtain Q groups of third acoustic feature data. Each group of noise data corresponds one-to-one with the first acoustic feature data, and the dimension of each group of noise data is the same as the dimension of the first acoustic feature data, where Q is a positive integer. The first acoustic feature data is then sequentially input into the second sub-model along with one group of third acoustic feature data from the Q groups to obtain a probability score corresponding to each group of third acoustic feature data. Each probability score indicates the similarity between the corresponding group of third acoustic feature data and the first acoustic feature data. The confidence level of each group of third acoustic feature data is calculated based on its probability score, and K groups of third acoustic feature data with confidence levels greater than or equal to a second threshold are saved as the second acoustic feature data, where K is a positive integer less than or equal to Q.

[0157] In the data augmentation device provided in this application embodiment, target acoustic feature data in the second acoustic feature data that matches the first acoustic feature data can be obtained sequentially based on the first model and the second model to improve the quality of the second acoustic feature data; and the target acoustic feature data can be augmented by reorganizing the target acoustic feature data; thereby reducing the probability of overfitting when learning the target acoustic feature data.

[0158] The data expansion device in this application embodiment can be an electronic device or a component within an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. It can also be a server, network attached storage (NAS), personal computer (PC), television set (TV), ATM, or self-service machine, etc. This application embodiment does not specifically limit the device.

[0159] The data expansion device in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0160] The data expansion device provided in this application embodiment can achieve... Figure 1 To avoid repetition, the various processes implemented in the method embodiment shown in Figure 6 will not be described again here.

[0161] like Figure 8 As shown, this application embodiment also provides an electronic device 800, including a processor 801 and a memory 802. The memory 802 stores a program or instructions that can run on the processor 801. When the program or instructions are executed by the processor 801, they implement the various steps of the data expansion method embodiment described above and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0162] It should be noted that the electronic devices in the embodiments of this application include mobile electronic devices and non-mobile electronic devices.

[0163] Figure 9 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.

[0164] The electronic device 1000 includes, but is not limited to, components such as: radio frequency unit 1001, network module 1002, audio output unit 1003, input unit 1004, sensor 1005, display unit 1006, user input unit 1007, interface unit 1008, memory 1009, and processor 1010.

[0165] Those skilled in the art will understand that the electronic device 1000 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 1010 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 9 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0166] The processor 1010 can be used to acquire second acoustic feature data that matches the first acoustic feature data based on the first model; and to determine target acoustic feature data from the second acoustic feature data based on the second model; and to reorganize the target acoustic feature data to obtain extended acoustic feature data corresponding to the target acoustic feature data.

[0167] In some embodiments of this application, the processor 1010 can be specifically used to score each acoustic feature data in the second acoustic feature data using the second model; and to determine the acoustic feature data in the second acoustic feature data whose corresponding score is greater than or equal to the first threshold as the target acoustic feature data.

[0168] In some embodiments of this application, the processor 1010 can be specifically used to cut the target acoustic feature data into M data segments, where M is an integer greater than or equal to 2; and according to the highest score of the acoustic feature data contained in each of the M data segments, store the M data segments in M ​​score interval sets, where the M score interval sets correspond one-to-one with the M data segments; and extract N sets of acoustic feature data from the M score interval sets, and splice each set of acoustic feature data based on the time dimension to obtain the expanded acoustic feature data, wherein any set of acoustic feature data in the N sets of acoustic feature data includes: an acoustic feature data randomly extracted from each of at least two score interval sets, where the at least two score interval sets are score interval sets in the M score interval sets, and N is a positive integer.

[0169] In some embodiments of this application, the processor 1010 can be specifically used to cut the target acoustic feature data into the M data segments according to the minimum cutting time, maximum cutting time and cutting step size corresponding to the target acoustic feature data.

[0170] In some embodiments of this application, the first model may include a first sub-model and a second sub-model. Specifically, the processor 1010 may input Q groups of randomly generated noise data into the first sub-model to obtain Q groups of third acoustic feature data, where each group of noise data corresponds one-to-one with the first acoustic feature data, and the dimension of each group of noise data is the same as the dimension of the first acoustic feature data, where Q is a positive integer; and input the first acoustic feature data sequentially with one group of third acoustic feature data from the Q groups of third acoustic feature data into the second sub-model to obtain a probability score corresponding to each group of third acoustic feature data, where each probability score indicates the similarity between the corresponding group of third acoustic feature data and the first acoustic feature data; and calculate the confidence level of each group of third acoustic feature data based on the probability score corresponding to each group of third acoustic feature data, and save K groups of third acoustic feature data with confidence levels greater than or equal to a second threshold as the second acoustic feature data, where K is a positive integer less than or equal to Q.

[0171] In the electronic device provided in this application embodiment, target acoustic feature data in second acoustic feature data that matches the first acoustic feature data can be obtained sequentially based on the first model and the second model to improve the quality of the second acoustic feature data; and the target acoustic feature data can be expanded by reorganizing the target acoustic feature data; thereby reducing the probability of overfitting when learning the target acoustic feature data.

[0172] It should be understood that, in this embodiment, the input unit 1004 may include a graphics processing unit (GPU) 10041 and a microphone 10042. The GPU 10041 processes image data of still images or videos obtained by an image capture device (such as a camera) in video capture mode or image capture mode. The display unit 1006 may include a display panel 10061, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like. The user input unit 1007 includes a touch panel 10071 and at least one of other input devices 10072. The touch panel 10071 is also called a touch screen. The touch panel 10071 may include a touch detection device and a touch controller. Other input devices 10072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, and joysticks, which will not be described in detail here.

[0173] The memory 1009 can be used to store software programs and various data. The memory 1009 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 1009 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 1009 in this embodiment includes, but is not limited to, these and any other suitable types of memory.

[0174] The processor 1010 may include one or more processing units; optionally, the processor 1010 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into the processor 1010.

[0175] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described data expansion method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0176] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0177] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described data expansion method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0178] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0179] This application provides a computer program product that is stored in a storage medium and executed by at least one processor to implement the various processes of the data expansion method embodiments described above, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0180] It should be noted that, in this document, 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 that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0181] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0182] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A data augmentation method, characterized in that, The method includes: Based on the first model, obtain second acoustic feature data that matches the first acoustic feature data; Based on the second model, target acoustic feature data are determined from the second acoustic feature data; The target acoustic feature data is reconstructed to obtain the expanded acoustic feature data corresponding to the target acoustic feature data; The reorganization process of the target acoustic feature data to obtain the expanded acoustic feature data corresponding to the target acoustic feature data includes: The target acoustic feature data is divided into M data segments, where M is an integer greater than or equal to 2; Based on the highest score of the acoustic feature data contained in each of the M data segments, the M data segments are stored in M ​​score interval sets, and the M score interval sets correspond one-to-one with the M data segments; N sets of acoustic feature data are extracted from the M score interval sets, and each set of acoustic feature data is concatenated based on the time dimension to obtain the expanded acoustic feature data. Any set of acoustic feature data in the N sets of acoustic feature data includes: one acoustic feature data randomly extracted from each of the at least two score interval sets, where the at least two score interval sets are the score interval sets in the M score interval sets, and N is a positive integer.

2. The method according to claim 1, characterized in that, The step of determining the target acoustic feature data from the second acoustic feature data based on the second model includes: The second model is used to score each acoustic feature data in the second acoustic feature data; The acoustic feature data in the second acoustic feature data whose corresponding score is greater than or equal to the first threshold is determined as the target acoustic feature data.

3. The method according to claim 1, characterized in that, The step of dividing the target acoustic feature data into M data segments includes: Based on the minimum cutting time, maximum cutting time, and cutting step size corresponding to the target acoustic feature data, the target acoustic feature data is cut into the M data segments.

4. The method according to claim 1, characterized in that, The first model includes a first sub-model and a second sub-model; The step of obtaining second acoustic feature data that matches the first acoustic feature data based on the first model includes: The Q groups of randomly generated noise data are input into the first sub-model respectively to obtain the Q groups of third acoustic feature data. The Q groups of noise data correspond one-to-one with the Q groups of third acoustic feature data. The dimension of each group of noise data is the same as the dimension of the first acoustic feature data. Q is a positive integer. The first acoustic feature data is sequentially input into the second sub-model along with a set of third acoustic feature data from the Q group to obtain a probability score corresponding to each set of third acoustic feature data. Each probability score is used to indicate the degree of similarity between the corresponding set of third acoustic feature data and the first acoustic feature data. The credibility of each group of third acoustic feature data is calculated based on the probability score corresponding to each group of third acoustic feature data, and K groups of third acoustic feature data with credibility greater than or equal to the second threshold are saved as the second acoustic feature data, where K is a positive integer less than or equal to Q.

5. A data expansion device, characterized in that, The device includes an acquisition module, a determination module, and a processing module; The acquisition module is used to acquire second acoustic feature data that matches the first acoustic feature data based on the first model; The determining module is used to determine target acoustic feature data from the second acoustic feature data based on the second model; The processing module is used to reorganize the target acoustic feature data to obtain the expanded acoustic feature data corresponding to the target acoustic feature data; The processing module is specifically used to cut the target acoustic feature data into M data segments, where M is an integer greater than or equal to 2; and to store the M data segments in M ​​score interval sets according to the highest score of the acoustic feature data contained in each of the M data segments, wherein the M score interval sets correspond one-to-one with the M data segments; and to extract N sets of acoustic feature data from the M score interval sets, and to splice each set of acoustic feature data based on the time dimension to obtain the expanded acoustic feature data, wherein any set of acoustic feature data in the N sets of acoustic feature data includes: an acoustic feature data randomly extracted from each of at least two score interval sets, wherein the at least two score interval sets are score interval sets among the M score interval sets, and N is a positive integer.

6. The apparatus according to claim 5, characterized in that, The determining module is specifically used to score each acoustic feature data in the second acoustic feature data using the second model; and to determine the acoustic feature data in the second acoustic feature data whose corresponding score is greater than or equal to the first threshold as the target acoustic feature data.

7. The apparatus according to claim 5, characterized in that, The processing module is specifically used to cut the target acoustic feature data into the M data segments according to the minimum cutting time, maximum cutting time, and cutting step size corresponding to the target acoustic feature data.

8. The apparatus according to claim 5, characterized in that, The first model includes a first sub-model and a second sub-model; The acquisition module is specifically used to input Q groups of randomly generated noise data into the first sub-model to obtain Q groups of third acoustic feature data. The Q groups of noise data correspond one-to-one with the Q groups of third acoustic feature data. The dimension of each group of noise data is the same as the dimension of the first acoustic feature data, and Q is a positive integer. The first acoustic feature data is then sequentially input into the second sub-model along with one group of third acoustic feature data from the Q groups of third acoustic feature data to obtain a probability score corresponding to each group of third acoustic feature data. Each probability score is used to indicate the similarity between the corresponding group of third acoustic feature data and the first acoustic feature data. The module calculates the credibility of each group of third acoustic feature data based on the probability score corresponding to each group of third acoustic feature data, and saves K groups of third acoustic feature data with credibility greater than or equal to a second threshold as the second acoustic feature data, where K is a positive integer less than or equal to Q.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the data augmentation method as described in any one of claims 1-4.

Images