Method and device for training machine learning model and computer readable storage medium

Abstract

The invention provides a method for training a machine learning model. In the method, sampling is performed from a first training set according to a plurality of weights to train a machine learning model in a first iteration. The first training set includes a plurality of levels corresponding to a plurality of weights. In addition, after the first iteration, a plurality of first efficiencies of the machine learning model at a plurality of levels are evaluated using the validation set; updating a weight corresponding to each level according to the first efficiencies; and sampling from the first training set again according to the updated weights so as to train the machine learning model in a second iteration after the first iteration. In addition, the invention also provides a device using the method and a computer readable storage medium.

Description

Technical Field

[0001] The present disclosure relates to a machine learning technology, and more particularly to a method, device, and computer-readable storage medium for dynamically adjusting a training set to train a machine learning model. Background Art

[0002] In the field of speech emotion recognition (SER), noise is a major challenge. Noise can confuse the system and reduce the recognition accuracy. In the real world, to achieve a reliable and accurate speech emotion recognition system, it must be able to handle a variety of noisy environments.

[0003] There are two main solutions to this problem. The first strategy is to remove noise first and then perform recognition on clean speech data. The second strategy is to train directly on noisy speech data and try to make the system model still be able to accurately recognize speech emotions in the presence of noise.

[0004] In recent years, data augmentation technology has been widely used in the field of speech emotion recognition to enhance the robustness of the model to noise. These enhancement methods usually predefine the noise level in the training data and then perform static training based on these predefined levels.

[0005] However, static training requires a prior understanding of the importance of different noise levels in model training in order to reasonably assign weights to different noise levels. However, since too many factors such as the application environment and the distribution of training data may affect the importance, such prior knowledge is unlikely to be obtained, resulting in static training methods causing the model to perform poorly at certain noise levels, which in turn affects the overall performance of the system. Summary of the invention

[0006] In view of this, the present disclosure proposes a method, device and computer program product for training a machine learning model, which can dynamically adjust the training set during the training process to improve the accuracy of the machine learning model.

[0007] A first aspect of the present disclosure proposes a computer-implemented method for training a machine learning model, comprising: sampling from a first training set according to multiple weights to train the machine learning model in a first iteration, wherein the first training set includes multiple levels corresponding to the multiple weights; evaluating the first performance of the machine learning model at each level using a validation set after the first iteration; updating the weight corresponding to each level according to each first performance; and re-sampling from the first training set according to the updated weights to train the machine learning model in a second iteration after the first iteration.

[0008] In some embodiments of the first aspect, the method further includes evaluating a second performance of the machine learning model at each level using a validation set after the second iteration; and updating a weight corresponding to each level according to each second performance.

[0009] In some embodiments of the first aspect, the updated weight corresponding to each level is negatively correlated with the first performance of the machine learning model at each level.

[0010] In some embodiments of the first aspect, each weight is not less than a preset minimum value.

[0011] In some embodiments of the first aspect, the method further comprises mixing the noisy data into the noise-free training set to generate the first training set.

[0012] In some embodiments of the first aspect, sampling from a first training set according to multiple weights to train a machine learning model in a first iteration includes: sampling at multiple levels of the first training set according to multiple weights to obtain a first sample training set; merging the first sample training set and the noise-free training set to generate a second training set; and using the second training set to train the machine learning model in the first iteration.

[0013] In some embodiments of the first aspect, the first training set comprises an ordered dataset.

[0014] In some embodiments of the first aspect, the method further includes dividing the first training set into multiple levels according to a distortion index.

[0015] In some embodiments of the first aspect, the machine learning model includes a speech recognition model.

[0016] In some embodiments of the first aspect, dividing the first training set into multiple levels according to the distortion index includes dividing the first training set into multiple levels according to at least one of perceptual evaluation of speech quality (PESQ), short-time objective intelligibility (STOI), and frequency-weighted signal-to-noise ratio segment (fwSNRseq).

[0017] A second aspect of the present disclosure provides an electronic device, comprising a memory and a processor. The memory is used to store at least one instruction. The processor is coupled to the memory. When the processor executes the at least one instruction, the electronic device is used to execute the above-mentioned computer-implemented method.

[0018] A third aspect of the present disclosure provides a computer-readable storage medium storing a program, wherein the program includes instructions, which, when executed by one or more processors of an electronic device, cause the electronic device to execute the above-mentioned computer-implemented method. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0020] Figure 1 Detailed description of the invention is a flowchart illustrating a method for training a machine learning model according to an embodiment of the present invention.

[0021] Figure 2 is a schematic diagram showing a method for training a machine learning model according to an embodiment of the present disclosure.

[0022] Figure 3 is a schematic diagram showing the generation of an ordered first training set according to an embodiment of the present disclosure.

[0023] Figure 4 is a schematic block diagram showing a computing system according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0025] In this disclosure, the term "plurality" means two or more, unless otherwise specified. In this disclosure, the term "and / or" describes the association relationship of associated objects, covering any one of the listed objects and all possible combinations. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0026] In the present disclosure, unless otherwise specified, the terms "first", "second", etc. are used to distinguish similar objects without intending to limit their positional relationship, temporal relationship, or importance relationship. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in ways other than those illustrated or described herein.

[0027] In addition, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or elements is not necessarily limited to those steps or elements explicitly listed, but may include other steps or elements not explicitly listed or inherent to such process, method, system, product, or apparatus.

[0028] The present disclosure proposes a training method that can improve the accuracy of a machine learning model. It should be noted that in multiple embodiments of the present disclosure, the training method is exemplarily described by means of a speech emotion recognition (SER) model, but the present disclosure does not limit the machine learning model to which the training method is applicable. Technical personnel in the field of the present disclosure can apply the training method to the machine learning model to be trained based on the technical concepts introduced in these embodiments.

[0029] Figure 1 is a flowchart illustrating a method for training a machine learning model according to an embodiment of the present disclosure; Figure 2 is a schematic diagram showing a method for training a machine learning model according to an embodiment of the present disclosure.

[0030] Please refer to Figure 1 and Figure 2 In step S110, a first training set TR is provided. 1 , and the first training set TR 1 Including multiple levels L 1 ~L 5 (For example but not limited to, 5 levels), each level L 1 ~L 5 It will correspond to a weight.

[0031] In some embodiments, the first training set TR 1 For example, an ordinal data set includes multiple ordinal data. Specifically, due to the data type, the first training set TR 1 The data in can be divided into multiple levels L 1 ~L 5 For example, the first training set TR 1This may include signal data with noise, such as, but not limited to, audio signal data with background noise, music, or speech.

[0032] In some embodiments, the trained machine learning model M is, for example but not limited to, a speech emotion recognition model.

[0033] It should be noted that, as long as it can be quantified or clustered into multiple levels or categories, the present disclosure does not limit the data type of the first training set. In some embodiments, the first training set may be a category data set, which includes multiple category data.

[0034] In some embodiments, the first training set TR 1 For example, the training set is expanded by mixing multiple levels of noisy data with the noise-free training set.

[0035] Figure 3 is a schematic diagram showing the generation of an ordered first training set according to an embodiment of the present disclosure.

[0036] Please refer to Figure 3 , noise-free training set X clean For example, a speech emotion dataset is included. For example, the speech emotion dataset may be from MSP-Podcast corpus, Multimodal Emotion Lines Dataset (MELD) or other speech emotion datasets, and the present disclosure is not limited thereto.

[0037] On the other hand, the noise data X noise For example, the noise data X includes at least one of noise, music and speech. noise It may be from the MUSAN (Music, Speech, And Noise) dataset, but the present disclosure is not limited thereto.

[0038] In some embodiments, the noise-free training set X clean For example, with different levels of noise data X noise Mix to get the first training set TR 1 After mixing, the first training set TR 1 For example, the amount of data is a noise-free data set X clean multiples (for example, but not limited to, 30 times) of the amount of data.

[0039] In some embodiments, the first training set TR 1For example, it is divided into K levels (for example, but not limited to, K=5) according to the distortion index. The distortion index includes, for example, a combination of one or more of Perceptual Evaluation of Speech Quality (PESQ), Short-Time Objective Intelligibility (STOI), and Frequency-Weighted Signal-to-Noise Ratio Segmental (fwSNRseq), but the present disclosure does not limit the type of distortion index.

[0040] In some embodiments, the first training set TR 1 The method of dividing into K levels is, for example, through uniform splitting or Gaussian mixture model (GMM), but the present disclosure is not limited thereto. 1 The number of data in each level is related to the clustering method and the first training set TR 1 is related to the data distribution.

[0041] exist Figure 2 In the embodiment, the first training set TR 1 For example, it is divided into 5 levels based on short-term objective comprehension. 1 ~L 5 .

[0042] In addition, in step S110, each level L 1 ~L 5 For example, an initial weight w will be given 11 ~w 15 In some embodiments, the initial weights may be predefined by the model developer. For example, each level L 1 ~L 5 The initial weight w 11 ~w 15 can be the same (for example, w 11 =w 12 =w 13 =w 14 =w 15 =0.2). However, the present disclosure is not limited thereto. Each level L 1 ~L 5 The initial weight w 11 ~w 15 It may also be set differently based on prior knowledge.

[0043] In some embodiments, in order to ensure the robustness of subsequent training, a very small weight (for example but not limited to 0.05 or 0.1) may be set, and the weight corresponding to each level will not be less than the very small weight.

[0044] In step S120, the first training set TR is selected according to a plurality of weights. 1 Multiple levels of L 1 ~L 5 Specifically, in the first training set TR 1 In each level, data with a quantity positively correlated to the corresponding weight will be sampled to obtain the first sample training set.

[0045] In some embodiments, the total number of data in the first sample training set is the same as that in the noise-free training set X clean However, the present disclosure is not limited thereto, and those skilled in the art to which the present disclosure pertains can adjust the total number of data in the first sample training set according to their needs.

[0046] Please refer to Figure 2 In some embodiments, in the nth iteration (n is a positive integer) of training the machine learning model M, according to a plurality of weights w n1 ~w n5 From the first training set TR 1 Multiple levels of L 1 ~L 5 Sampling is done to obtain the first sample training set {X n1 ,X n2 ,X n3 ,X n4 ,X n5}. In detail, from the first training set TR 1 The first level of L 1 Sample multiple sample data X n1 ; From the first training set TR 1 The second level of L 2 Sample multiple sample data X n2 ; From the first training set TR 1 The third level of L 3 Sample multiple sample data X n3 ; From the first training set TR 1 The fourth level L 4 Sample multiple sample data X n4 ; and from the first training set TR 1 The fifth level L 5 Sample multiple sample data X n5 , where the sample data X n1 ~X n5 The number of samples is positively correlated with the weight w n1~w n5 .

[0047] Please go back Figure 1 In step S130, the machine learning model M is trained for one iteration according to the sampling results. Specifically, according to the multiple weights w n1 ~w n5 In the first data set TR 1 Multiple levels of L 1 ~L 5 After a sampling is performed in , the machine learning model M is trained based on the sampling result for at least one iteration (for example, m iterations, where m is a positive integer).

[0048] Please refer to Figure 2 In some embodiments, according to a plurality of weights w n1 ~w n5 From the first training set TR 1 Multiple levels of L 1 ~L 5 Sampling is done to obtain the first sample training set {X n1 ,X n2 ,X n3 ,X n4 ,X n5}, based on the first sample training set {X n1 ,X n2 ,X n3 ,X n4 ,X n5}Train the nth iteration of the machine learning model M.

[0049] In some embodiments, the first sample training set {X n1 ,X n2 ,X n3 ,X n4 ,X n5} will be compared with the noise-free data set X clean Merge to generate the second training set TR 2 In the nth iteration to the n+m-1th iteration, the second training set TR 2 Will be used to train the machine learning model M.

[0050] In some embodiments, m is equal to 1, that is, each iteration will resample to train the machine learning model M in the next iteration. However, the present disclosure is not limited to this. In some embodiments, m can also be determined by the model developer.

[0051] Please go back Figure 1 In step S140, the validation set is used to evaluate the machine learning model M at each level L 1 ~L 5The efficacy of 1 ~p 5 .

[0052] Specifically, since the data type in the validation set is the same as that in the first training set TR 1 And the second training set TR 2 The data type of the validation set can also be based on the first training set TR 1 The same classification standard is divided into multiple levels 1 ~L 5 In this way, the validation set can be used to evaluate the machine learning model M for each level L 1 ~L 5 The effectiveness of the data 1 ~p 5 .

[0053] In some embodiments, the validation set is obtained by mixing multiple levels of noise data X noise The verification set is generated with a noise-free verification set, but the disclosure does not limit the generation method of the verification set.

[0054] Please refer to Figure 2 In some embodiments, after evaluating the machine learning model M using the validation set, the machine learning model M for each level L can be obtained. 1 ~L 5 The effectiveness of the data 1 ~p 5 .from Figure 2 It can be seen that the machine learning model M targets each level L of distortion from low to high 1 ~L 5 The effectiveness of the data 1 ~p 5 Gradually lower.

[0055] In some embodiments, the performance indicator used to evaluate the performance is, for example, the F1 score, but the present disclosure is not limited thereto. In other embodiments, other performance indicators such as accuracy, precision, or recall may also be used.

[0056] In some embodiments, when according to a plurality of weights w n1 ~w n5 In the first data set TR 1 Multiple levels of L 1 ~L 5 After a sampling is performed in , the machine learning model M is trained for m iterations based on the sampling result, and after the machine learning model M is trained in the n+m-1th iteration, the validation set is used to evaluate the machine learning model M at each level L 1 ~L5 The efficacy of 1 ~p 5 , then proceed to step S150.

[0057] In some embodiments, m is equal to 1. n1 ~w n5 In the first data set TR 1 Multiple levels of L 1 ~L 5 After a sampling is performed in , the machine learning model M is trained for one iteration based on the sampling result, and after the machine learning model M is trained in this iteration, the validation set is used to evaluate the machine learning model M at each level L 1 ~L 5 The efficacy of 1 ~p 5 , then proceed to step S150.

[0058] In some embodiments, when according to a plurality of weights w n1 ~w n5 In the first data set TR 1 Multiple levels of L 1 ~L 5 After a sampling is performed in , the machine learning model M is trained for m iterations based on the sampling result, and after each iteration of training the machine learning model M, the validation set is used to evaluate the machine learning model M at each level L 1 ~L 5 The efficacy of 1 ~p 5 , and after the machine learning model M is trained and evaluated at the n+m-1th iteration, step S150 is entered. In this case, m can be, for example, the performance p evaluated after each iteration. 1 ~p 5 to decide.

[0059] For example, when each level L 1 ~L 5 The corresponding performance p 1 ~p 5 When the consistency is too low, it means that the training of the data at a specific level may need to be strengthened, so step S150 is entered to update the weights and perform subsequent resampling.

[0060] For example, when each level L 1 ~L 5 The corresponding performance p 1 ~p 5 The ratio of each level L before the previous sampling 1 ~L 5When the corresponding performance ratios are close, it means that the effect or significance of updating weights and resampling may not be significant, so training can continue to the next iteration.

[0061] For example, when each level L 1 ~L 5 The corresponding performance p 1 ~p 5 When all the preset thresholds are reached, the training can be ended.

[0062] In other words, the number of iterations m to train the machine learning model M based on a sampling result can be calculated based on the number of levels L evaluated after each iteration. 1 ~L 5 The corresponding performance p 1 ~p 5 To dynamically determine, but the present disclosure is not limited thereto.

[0063] Please go back Figure 1 In step S150, according to each level L 1 ~L 5 The efficacy of 1 ~p 5 Update each level L 1 ~L 5 The corresponding weight is then returned to step S120 for re-sampling.

[0064] Specifically, since the sampling will be returned to step S120 after the weights are updated, the updated weights can determine the distribution of training data for subsequent training of the machine learning model M. When the machine learning model M has a low performance for a specific level of data, it means that a higher proportion of the data at the specific level may be required for the machine learning model M to learn. Therefore, according to the performance p evaluated in step S140, 1 ~p 5 To update each level L 1 ~L 5 The corresponding weights can dynamically adjust the training data distribution during the training process.

[0065] In some embodiments, in order to ensure the robustness of subsequent training, a very small weight (for example but not limited to 0.05 or 0.1) may be set, and the weight corresponding to each level after the update will not be less than the very small weight.

[0066] In some embodiments, multiple levels L 1 ~L 5 The sum of the corresponding multiple weights is 1, for example.

[0067] In some embodiments, after updating, each level L 1 ~L 5 The corresponding weight w(n+m)1 ~w (n+m)5 For example, each level L evaluated in step S140 1 ~L 5 The corresponding performance p 1 ~p 5 Negatively correlated.

[0068] In some embodiments, the updated weight w (n+m)1 ~w (n+m)5 For example, it is positively correlated with the efficacy p 1 ~p 5 The reciprocal of .

[0069] Please refer to Figure 2 In some embodiments, in step 140, for example, the noise-free validation set is also used to evaluate the standard performance std of the machine learning model M for the noise-free validation set. In order to increase the weight corresponding to the level with lower performance, the standard performance std of the noise-free validation set is calculated and the standard performance std of each level L is calculated. 1 ~L 5 The effectiveness of the data 1 ~p 5 The difference between (n+m-1)1 ~g (n+m-1)5 , and make the updated weight w (n+m)1 ~w (n+m)5 Positively correlated with the difference g (n+m-1)1 ~g (n+m-1)5 .

[0070] In some embodiments, m is equal to 1. After the machine learning model M is trained and the performance is evaluated at the nth iteration, the standard performance std of the noise-free validation set is calculated and the standard performance std of each level L is calculated. 1 ~L 5 The effectiveness of the data 1 ~p 5 The difference between n1 ~g n5 , and make the updated weight w (n+1)1 ~w (n+1)5 Positively correlated with the difference g n1 ~g n5 .

[0071] Please go back Figure 1 After updating the weights, return to step S120 to resample and train the machine learning model M again for at least one iteration according to the sampling results in step S130. And so on until the training is completed.

[0072] In some embodiments, the training completion condition is as described in the previous paragraph, when the performance p1-p5 corresponding to each level L1-L5 reaches a preset threshold, the training can be terminated. However, it should be noted that the present disclosure does not limit the training completion condition, and those skilled in the art can implement it according to their own requirements for the model.

[0073] Table 1 below compares the accuracy of the method for training a machine learning model introduced in the embodiment of the present disclosure (represented by "this method" in the table) and the traditional training method after training a speech emotion recognition model, and the evaluation index used is, for example, the weighted F1 score. In the experiment in Table 1, the first training set is divided into multiple levels using short-term objective intelligibility as the distortion index, and the weight is updated in each iteration using M equal to 1. Table 1

[0074] It can be seen from Table 1 that, whether it is the MSP-Podcast or MELD dataset, compared with the traditional training method, the method of training the machine learning model according to the embodiment of the present disclosure to train the speech emotion recognition model has better performance in almost every degree of distortion data.

[0075] Figure 4 is a schematic block diagram showing a computing system according to an embodiment of the present disclosure.

[0076] Please refer to Figure 4 , the method for training a machine learning model described herein is a computer-implemented method, which can be implemented on a computing system 400 having a variety of hardware components. In some embodiments, the computing system 400 can be implemented in the form of an electronic device, which includes, but is not limited to, one or more of the following components: a processor (Central Processing Unit, CPU) 410, a graphics processor (Graphics Processing Unit, GPU) 420, an input / output component 430, a network component 440, and a memory 450, and the above components can communicate and transmit data via the system bus 460. However, the present disclosure does not limit the specific model, quantity, and configuration of each component, and those skilled in the art can adjust, select, or add or reduce components according to the specific needs and operating environment when implementing the present disclosure.

[0077] In some embodiments, the main computing core within the computing system 400 is one or more processors 410. The processor 410 may be responsible for running the main computing process and related control logic of algorithms such as deep learning. In some embodiments, the processor 410 is used to execute processing instructions (i.e., machine executable instructions) stored in a non-volatile computer-readable medium (e.g., storage device 470).

[0078] In some embodiments, in order to improve the computing efficiency of deep learning, the computing system 400 may also include one or more image processors 420 designed for large-scale parallel computing. The image processor 420 can effectively improve the computing power of the system when performing deep learning training and reasoning.

[0079] In some embodiments, the computing system 400 may include a variety of input / output components 430 for receiving user input and displaying system output. For example, the input / output components 430 may include a keyboard, a mouse, a touch pad, a display screen, a speaker, and other types of sensing devices.

[0080] In some embodiments, the computing system 400 may also include a network element 440 for network communication. For example, the network element 440 may include a network interface card for wired or wireless network connection, or a communication module for 3G, 4G, 5G or other wireless communication technologies.

[0081] In some embodiments, the computing system 400 includes one or more memories 450, such as volatile memory elements such as random access memory (RAM). The memory 450 stores parameters of the deep learning model and other data and programs used to run algorithms such as deep learning.

[0082] In addition, the computing system 400 may further include one or more of the following elements: a storage device 470 , a power management element 480 , and other elements 490 (eg, other hardware elements).

[0083] In some embodiments, the computing system 400 may include one or more storage devices 470, such as non-volatile memory devices such as hard disks (HDD) or solid state drives (SSD). These storage devices 470 can be used to store information such as program code, training data, model parameters, etc. of deep learning software. In addition, the storage device 470 can also be used to store intermediate results and final outputs of deep learning algorithms.

[0084] In some embodiments, the computing system 400 may include one or more power management components 480 for providing power to various hardware components of the computing system 400 and managing their power consumption. The power management components 480 may include batteries, power converters, and other power management devices.

[0085] In some embodiments, the computing system 400 may further include other components 490 , such as a cooling fan, a heat sink, and other various control and monitoring devices, but the present disclosure is not limited thereto.

[0086] In addition, the embodiments of the present invention may also be implemented as one or more computer program products, including a computer program having one or more instructions. Specifically, a computer program (also referred to as a program, software, script, or code) may be presented in any form of programming language, and the computer program may be deployed in any form. During the operation of the computing system 400 (e.g., an electronic device), the instructions or a portion thereof may also reside completely or at least partially in the processor 410 so that the processor 410 performs the method described herein. In addition, the embodiments of the present invention may also be implemented as a computer-readable storage medium storing the above-mentioned computer program.

[0087] In summary, the method, device, and computer-readable storage medium for training a machine learning model proposed in the embodiments of the present disclosure dynamically adjust the data distribution of the training set according to the evaluation results of the validation set during the training of the machine learning model. In this way, the robustness of the trained machine learning model can be improved, so that it can perform well for input data with different levels of noise.

[0088] Although this paper includes many details, these details should not be interpreted as limitations to the scope of the claimed invention, but are interpreted as descriptions of features peculiar to specific embodiments. Some features described in the context of different embodiments herein may also be combined in a single embodiment. On the contrary, the various features described in the context of a single embodiment may also be separated or implemented in multiple embodiments with any suitable sub-combination. In addition, although features may be described as acting with some combinations and even initially claimed above, one or more features from the claimed combination may be removed from the combination in some cases, and the claimed combination may be directed to a variation of a sub-combination or a sub-combination. Similarly, although operations are depicted in a particular order in the drawings, this should not be understood as requiring to perform such operations in the particular order shown or in a sequential order, or to perform all described operations to achieve desired results.

[0089] Only a few implementations and examples are described, and other implementations, enhancements, and variations may be made based on what is described and illustrated in this disclosure.

Claims

1. A computer-implemented method for training a machine learning model, characterized in that: include: Sampling from a first training set according to a plurality of weights to train the machine learning model in a first iteration, wherein the first training set includes a plurality of levels corresponding to the plurality of weights; After the first iteration, evaluating a plurality of first performances of the machine learning model at the plurality of levels using a validation set; updating the plurality of weights corresponding to the plurality of levels according to the plurality of first capabilities; and Re-sampling from the first training set according to the updated multiple weights to train the machine learning model in a second iteration, wherein the second iteration occurs after the first iteration.

2. The computer-implemented method of claim 1, wherein: Also includes: After the second iteration, evaluating a plurality of second performances of the machine learning model at the plurality of levels using the validation set; as well as The weights corresponding to the levels are updated according to the second capabilities.

3. The computer-implemented method of claim 1, wherein: The multiple weights corresponding to the updated multiple levels are negatively correlated with the multiple first performances of the machine learning model at the multiple levels.

4. The computer-implemented method of claim 3, wherein: Each of the plurality of weights is not less than a preset minimum value.

5. The computer-implemented method of claim 1, wherein: Also includes: The noisy data is mixed into the noise-free training set to generate the first training set.

6. The computer-implemented method of claim 5, wherein: Sampling from the first training set according to the plurality of weights to train the machine learning model in the first iteration comprises: Sampling respectively in the multiple levels of the first training set according to the multiple weights to obtain a first sample training set; Combining the first sample training set and the noise-free training set to generate a second training set; and The machine learning model is trained in the first iteration using the second training set.

7. The computer-implemented method of claim 1, wherein: The first training set includes an ordered data set.

8. The computer-implemented method of claim 1, wherein: Also includes: The first training set is divided into the multiple levels according to the distortion index.

9. The computer-implemented method of claim 8, wherein: The machine learning model includes a speech recognition model.

10. The computer-implemented method of claim 9, wherein: According to the distortion index, dividing the first training set into the multiple levels comprises: The first training set is divided into the plurality of levels according to at least one of a perceptual assessment of speech quality, a short-term objective intelligibility, and a frequency-weighted signal-to-noise ratio segment.

11. An electronic device, characterized in that: include: A memory for storing at least one instruction; as well as A processor is coupled to the memory, wherein when the processor executes the at least one instruction, the electronic device is used to perform the computer-implemented method as claimed in any one of claims 1 to 10.

12. A computer-readable storage medium storing a program, wherein the program includes instructions, characterized in that: When the instructions are executed by one or more processors of an electronic device, the instructions cause the electronic device to perform the computer-implemented method of any one of claims 1 to 10.

Images