Small sample speech emotion recognition method and device based on meta metric learning

Abstract

The application provides a small sample speech emotion recognition method and device based on meta metric learning. The method proposes an adaptive oversampling data imbalance processing method based on density distribution evaluation on one hand, dynamically processes the imbalance problem between different categories of data, and proposes a small sample learning method based on metric learning on the other hand, and combines the training mode of meta learning, can not only perform emotion recognition on a small sample data set, but also get rid of the pre-training process of the model on a large database. Compared with the traditional small sample learning method, the method has high recognition accuracy and good realizability.

Description

Technical Field

[0001] This invention relates to the field of emotion recognition technology, and in particular to a few-sample speech emotion recognition method and apparatus based on metametric learning. Background Technology

[0002] Speech emotion recognition methods are mainly divided into two parts: those based on emotion feature extraction and those based on spectrogram classification. Current research widely combines both methods with algorithms from the field of deep learning to achieve better recognition results. Patent application CN114664289A, entitled "A Speech Emotion Recognition Method Based on Convolutional Neural Networks," discloses a speech emotion recognition method based on convolutional neural networks. This method uses spectrograms as input and employs deep learning technology for speech emotion recognition. Patent application CN111429947B, entitled "A Speech Emotion Recognition Method Based on Multi-Level Residual Convolutional Neural Networks," constructs a multi-level residual convolutional neural network and inputs the spectrogram into it for training. Patent application CN114566188A, entitled "A Speech Emotion Recognition Method Based on Spectrograms," inputs the spectrogram into a deep convolutional neural network for deep emotion feature learning and, based on this, captures the contextual relationships between different speech segments through a long short-term memory network, ultimately achieving the recognition of the emotion in the input speech.

[0003] Current speech emotion recognition technologies mostly utilize large-scale datasets for model training, achieving significant results and high recognition accuracy. However, this relies heavily on sufficient labeled data, which is difficult to obtain in practice. When data volume is insufficient, the accuracy of the recognition model drops significantly. In small sample environments, the model's emotion recognition accuracy is relatively low. When the sample data is limited, overfitting is prone to occur when training traditional speech emotion recognition models, often leading to a decrease in emotion recognition accuracy and failing to achieve the desired results.

[0004] Current speech emotion recognition models generally do not consider the problem of data imbalance. This leads to an overemphasis on the features of the majority class samples during recognition, neglecting the learning of features from both the majority and minority classes. This results in inter-class differences in recognition rates. Furthermore, real-world speech emotion data typically exhibits dynamic characteristics, with data distribution changing over time, potentially leading to shifts in the majority and minority classes. However, most current imbalanced data classification models are static models; once trained and determined, their parameters and structure are not dynamically adjusted, preventing them from learning features from newly added data in real time. Summary of the Invention

[0005] To address the overfitting issues that arise when training traditional speech emotion recognition models with limited sample data, and the problem of inter-class differences in recognition rates due to neglecting data imbalance, this invention proposes a few-shot speech emotion recognition method and apparatus based on meta-metric learning. This method employs an adaptive oversampling data imbalance handling method based on density distribution evaluation to dynamically address imbalances between different data categories. Furthermore, it combines a few-shot learning method based on metric learning with a meta-learning training approach. This not only enables emotion recognition on small-sample datasets but also eliminates the need for pre-training models on large-scale databases. Compared to traditional few-shot learning methods, this method offers higher recognition accuracy and better feasibility.

[0006] According to one aspect of the present invention, a few-shot speech emotion recognition method based on meta-metric learning is provided, comprising the following steps:

[0007] S1: Obtain a small sample speech emotion dataset and perform data processing, including imbalanced data processing, speech signal preprocessing and spectrogram generation, to obtain a spectrogram dataset;

[0008] S2: Based on the meta-learning standard, the spectrogram dataset is split into the support set S and query set Q used in the meta-training phase, and the support set S' and query set Q' used in the meta-testing phase;

[0009] S3: Build a few-sample speech emotion recognition model based on metric learning;

[0010] S4: The small sample speech emotion recognition model is trained and tested in two stages: meta-training and meta-testing. The meta-training stage trains hyperparameters, and the meta-testing stage trains parameters. After training is completed, the trained model is obtained.

[0011] S5: Perform speech emotion recognition using the trained model to obtain the speech emotion recognition result.

[0012] Preferably, step S1 includes:

[0013] S11: Obtain a small sample speech emotion dataset;

[0014] S12: Perform imbalanced data processing on the small sample voice emotion dataset;

[0015] S13: After completing the imbalanced data processing, preprocessing operations are performed on the speech signal, including sampling, quantization, pre-emphasis, and windowing, to generate the original sentiment spectrogram;

[0016] S14: Select a fixed step size to crop the original sentiment spectrogram to obtain the spectrogram dataset.

[0017] Preferably, the steps of the unbalanced data processing include:

[0018] S121: Define a potentially imbalanced speech emotion dataset as D = {x1, x2, ..., x...} n Let n be the number of sentiment categories in the dataset. Then, the n-1 categories are the minority classes, and the remaining 1 category is the majority class with the largest number of data points. Let the set of the majority classes be Dq = {x1, x2, ..., xn}. q}, that is, the number of majority classes is q, and for n-1 minority classes, the number of samples in any minority class is denoted as p;

[0019] S122: Iterate through steps S123-S126 until the data is balanced, thus completing the unbalanced data processing process;

[0020] S123: Calculate the imbalance index i. When i < 1, the data is considered imbalanced. The formula for calculating i is:

[0021]

[0022] Where λ is a filtering parameter used to filter out minor data imbalance problems that have little impact on subsequent sentiment classification tasks, and its value ranges from [p / q, 1].

[0023] S124: Calculate the total number of samples required to synthesize for this minority class.

[0024] N = α(qp), α ∈ (0, 1],

[0025] Where α is an adjustment parameter used to set the required degree of adjustment. When α = 1, it means that the data needs to be adjusted until it is completely balanced.

[0026] S125: Calculate x for each minority class sample according to the Euclidean distance formula. i k nearest neighbors

[0027]

[0028] Where t represents the feature dimension of speech emotion, x i Representing the minority class sample, y j Let x represent the majority class sample, and let x be the denoted x. i The number of majority class samples among these k nearest neighbors is s. i ;

[0029] S126: Calculate the minority class sample x i The density distribution ρ of the majority class samples within the range of k nearest neighbors. i And after normalization, we get

[0030] ρ i =s i / k

[0031]

[0032] S127: Based on the normalized density distribution value, calculate the number N of new samples that need to be synthesized from this sample. i

[0033]

[0034] Return to step S122.

[0035] Preferably, step S2 includes:

[0036] S21: Split the spectrogram dataset into an original training set and an original test set;

[0037] S22: Select N classes of samples from the original test set. These N classes of samples contain a large number of samples.

[0038] S23: Randomly select k+x samples from each of these N classes of samples. K of these samples are placed into the support set S' used in the meta-testing phase, and the other x samples are placed into the query set Q' used in the meta-testing phase.

[0039] S24: Following the same method as steps S22-S23, obtain the support set S and query set Q used in the meta-training phase.

[0040] Preferably, step S3 includes:

[0041] Set the embedded function f θ The expression is as follows:

[0042]

[0043] Among them, Z l+1 Z represents the convolution output of the (l+1)th layer. l w represents the output of the convolution at layer l. l+1 Let b be the parameter matrix of the (l+1)th layer, and b be the bias.

[0044] Make the support set as

[0045] S={(x1,y1),(x2,y2),...,(x n ,y n )}, i=1,2,...,n,

[0046] Where, x i It is the i-th eigenvector, y iThis is the corresponding tag information;

[0047] After being projected into the embedding space, the prototype of each class is obtained using the following formula.

[0048]

[0049] Where n represents the true class of the sample, z n S represents the prototype of class n. n f represents a set of classes. θ Indicates an embedded function;

[0050] For each sample in the query set, calculate the probability of belonging to each class as follows:

[0051]

[0052] Where n' represents the predicted sample class, n represents the true class of the sample, x is the feature vector of the sample point, y is the corresponding label information, and d represents the Euclidean distance;

[0053] Therefore, the loss function is defined as follows:

[0054] J θ =-log(p θ (y=n'|x)).

[0055] Preferably, in step S3, an adaptive motion estimation algorithm is selected to optimize the parameters.

[0056] Preferably, step S3 further includes: setting the initial hyperparameters and initial model parameters of the few-sample speech emotion recognition model.

[0057] Preferably, step S4 includes:

[0058] The initial hyperparameters set in step S3 are trained through the meta-training phase. After the best hyperparameters are trained, the meta-testing phase is entered, which trains the model parameters to obtain the optimal model parameters.

[0059] According to another aspect of the present invention, the present invention also provides a few-shot speech emotion recognition device based on meta-metric learning, comprising the following modules:

[0060] The data processing module is used to acquire and process small sample speech emotion datasets, including imbalanced data processing, speech signal preprocessing, and spectrogram generation, to obtain spectrogram datasets.

[0061] The data splitting module is used to split the spectrogram dataset into the support set S and query set Q used in the meta-training phase, and the support set S' and query set Q' used in the meta-testing phase, according to the meta-learning standard.

[0062] The network construction module is used to build a few-sample speech emotion recognition model based on metric learning.

[0063] The network training module is used to train and test the small sample speech emotion recognition model in two stages: meta-training and meta-testing. The meta-training stage trains hyperparameters, and the meta-testing stage trains parameters. After training is completed, the trained model is obtained.

[0064] The emotion recognition module is used to perform speech emotion recognition through the trained model and obtain the speech emotion recognition result.

[0065] The technical solution provided by this invention has the following beneficial effects:

[0066] (1) It performs speech emotion recognition based on small samples, avoiding the difficulty of obtaining large-scale and labeled data;

[0067] (2) This method does not require large-scale datasets for model pre-training, and is easier to implement and deploy than traditional few-sample emotion recognition systems;

[0068] (3) The speech emotion recognition method based on metametric learning considers the problem of classifying unknown samples at the theoretical level. This method can better handle the emotion recognition problem and achieve higher emotion recognition accuracy.

[0069] (4) This method can effectively improve the overfitting phenomenon caused by insufficient sample size in deep neural networks and achieve higher emotion recognition accuracy.

[0070] (5) The proposed data imbalance processing method takes into account the feature distribution of the majority class samples, adaptively generates minority class samples, dynamically processes the imbalance problem between different classes of data, and can overcome the problem of overlapping sample features caused by direct interpolation and oversampling of minority class samples. Attached Figure Description

[0071] The specific effects of the present invention will be further explained below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0072] Figure 1 This is a flowchart of the few-sample speech emotion recognition method based on metametric learning in an embodiment of the present invention;

[0073] Figure 2 This is a system framework diagram of the few-sample speech emotion recognition method based on metametric learning in this embodiment of the invention;

[0074] Figure 3 This is a flowchart illustrating the specific preprocessing steps in an embodiment of the present invention.

[0075] Figure 4 This refers to the division of the training set and the test set in meta-learning in this embodiment of the invention;

[0076] Figure 5 This is a prototype network diagram in an embodiment of the present invention;

[0077] Figure 6 This is a structural diagram of a few-sample speech emotion recognition device based on metametric learning in an embodiment of the present invention. Detailed Implementation

[0078] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0079] Example 1:

[0080] refer to Figure 1 and Figure 2 This invention provides a few-sample speech emotion recognition method based on meta-metric learning, which specifically includes the following steps:

[0081] S1: Obtain a small sample speech emotion dataset and perform data processing, including imbalanced data processing, speech signal preprocessing and spectrogram generation, to obtain a spectrogram dataset;

[0082] The processing flow of step S1 is as follows: Figure 3 As shown, we first select a commonly used small-sample speech emotion dataset and then perform data balancing on the speech segments within it to ensure that the amount of data for each emotion category is roughly the same, thus avoiding data imbalance problems.

[0083] This imbalanced data processing method employs an adaptive oversampling imbalance handling approach based on density distribution assessment. The specific steps are as follows:

[0084] S121: Define a potentially imbalanced speech emotion dataset as D = {x1, x2, ..., x...} n Let n be the number of sentiment categories in the dataset. Then, the n-1 categories are the minority classes, and the remaining 1 category is the majority class with the largest number of data points. Let the set of the majority classes be Dq = {x1, x2, ..., xn}. q}, that is, the number of majority classes is q, and for n-1 minority classes, the number of samples in any minority class is denoted as p;

[0085] S122: Iterate through steps S123-S126 until the data is balanced, thus completing the unbalanced data processing process;

[0086] S123: Calculate the imbalance index i. When i < 1, the data is considered imbalanced. The formula for calculating i is:

[0087]

[0088] Where λ is a filtering parameter used to filter out minor data imbalance problems that have little impact on subsequent sentiment classification tasks, and its value ranges from [p / q, 1].

[0089] S124: Calculate the total number of samples required to synthesize for this minority class.

[0090] N = α(qp), α ∈ (0, 1],

[0091] Where α is an adjustment parameter used to set the required degree of adjustment. When α = 1, it means that the data needs to be adjusted until it is completely balanced.

[0092] S125: Calculate x for each minority class sample according to the Euclidean distance formula. i k nearest neighbors

[0093]

[0094] Where t represents the feature dimension of speech emotion, x i Representing the minority class sample, y j Let x represent the majority class sample, and let x be the denoted x. i The number of majority class samples among these k nearest neighbors is s. i ;

[0095] S126: Calculate the minority class sample x i The density distribution ρ of the majority class samples within the range of k nearest neighbors. i And after normalization, we get

[0096] ρ i =s i / k

[0097]

[0098] S127: Based on the normalized density distribution value, calculate the number N of new samples that need to be synthesized from this sample. i

[0099]

[0100] Return to step S122.

[0101] After the imbalanced data processing is completed, the speech signal is subjected to necessary data preprocessing operations, including sampling, quantization, pre-emphasis, and windowing. Then, the original sentiment spectrogram is plotted. At this time, the spectrogram is a long spectrogram. Next, the spectrogram is cropped with a fixed step size to achieve the purpose of data augmentation, thereby obtaining the spectrogram dataset.

[0102] S2: Based on the standards of meta-learning, the spectrogram dataset is further split;

[0103] The training process of meta-learning is called meta-learning training, or simply "meta-training." Similarly, its testing process is called meta-learning testing, or simply "meta-testing." In meta-learning, both the training set and the test set are further split into a support set and a query set. The support set and test set in the training set are referred to as S and Q, respectively, while the support set and query set in the test set are referred to as S' and Q', respectively. This division is as follows: Figure 4 As shown.

[0104] (1) Select N types of samples from the original test set. These N types of samples contain a large number of samples.

[0105] (2) Randomly select k+x samples from each of the N classes of samples (x means any number can be selected, but k+x cannot exceed the total number of samples in each class), and put k of the samples into the support set S', and the other x samples into the query set Q'.

[0106] (3) Following the steps in (2), obtain the support set S and query set Q during the meta-training process.

[0107] You can select x classes of samples from the original training set, where x can be arbitrarily specified. Here, we specify that it should still be N. Alternatively, you can choose not to select k+x samples from each class. Here, we specify that the number of samples selected should still be consistent with the number selected from the test set.

[0108] Based on the above, the spectrogram dataset obtained in step one is converted into four parts: the support set S and query set Q used in the meta-training stage, and the support set S' and query set Q' used in the meta-testing stage.

[0109] S3: Build a few-sample speech emotion recognition model based on metric learning;

[0110] A few-shot speech emotion recognition model is built based on metric learning. Specifically, the model is implemented using a prototype network. The prototype network maps sample data from each category to a space and extracts their "mean" to represent the prototype of that category.

[0111] The few-shot learning algorithm based on prototype networks is a combination of metric learning and meta-learning. Its basic idea is to transform the input into a feature vector through a neural network and project it into an embedding space. In this embedding space, features with similar distances are grouped together, and their average value is taken as the center of this class, which is called the "prototype". The framework diagram of the prototype network is shown below. Figure 5 As shown in the figure. S H×W×C R represents the sample set. D Denotes the embedding space, f θ This represents an embedding function constructed from a neural network.

[0112] The core of the prototype network is the embedding function f, which is composed of neural networks. θ First, select a suitable network model to form the embedding function. Taking a convolutional neural network as an example, then set the initial hyperparameters and select an optimization algorithm. Generally, the Adaptive Motion Estimation (Adam) algorithm is selected to optimize the parameters, and the loss function is set.

[0113] Specifically, set the embedding function f θ The expression is as follows:

[0114]

[0115] Among them, Z l+1 Z represents the convolution output of the (l+1)th layer. l w represents the output of the convolution at layer l. l+1 Let b be the parameter matrix of the (l+1)th layer, and b be the bias.

[0116] Make the support set as

[0117] S={(x1,y1),(x2,y2),...,(x n ,y n )}, i=1,2,...,n,

[0118] Where, x i It is the i-th eigenvector, y i This is the corresponding tag information;

[0119] After being projected into the embedding space, the prototype of each class is obtained using the following formula.

[0120]

[0121] Where n represents the true class of the sample, z n S represents the prototype of class n. n f represents a set of classes. θ Indicates an embedded function;

[0122] For each sample in the query set, calculate the probability of belonging to each class as follows:

[0123]

[0124] Where n' represents the predicted sample class, n represents the true class of the sample, x is the feature vector of the sample point, y is the corresponding label information, and d represents the Euclidean distance;

[0125] Therefore, the loss function is defined as follows:

[0126] J θ =-log(p θ (y=n'|x)).

[0127] S4: The small sample speech emotion recognition model is trained and tested in two stages: meta-training and meta-testing. The meta-training stage trains hyperparameters, and the meta-testing stage trains parameters. After training is completed, the trained model is obtained.

[0128] S5: Perform speech emotion recognition using the trained model to obtain the speech emotion recognition result.

[0129] Example 2:

[0130] refer to Figure 6 In this embodiment, a few-sample speech emotion recognition device based on meta-metric learning is also provided, including the following modules:

[0131] Data processing module 01 is used to acquire a small sample speech emotion dataset and perform data processing, including imbalanced data processing, speech signal preprocessing and spectrogram generation, to obtain a spectrogram dataset;

[0132] The data splitting module 02 is used to split the spectrogram dataset into the support set S and query set Q used in the meta-training stage, and the support set S' and query set Q' used in the meta-testing stage, according to the meta-learning standard.

[0133] Network building module 03 is used to build a few-sample speech emotion recognition model based on metric learning;

[0134] The network training module 04 is used to train and test the small sample speech emotion recognition model in two stages: meta-training and meta-testing. The meta-training stage trains hyperparameters, and the meta-testing stage trains parameters. After training is completed, the trained model is obtained.

[0135] The emotion recognition module 05 is used to perform speech emotion recognition through the trained model and obtain the speech emotion recognition result.

[0136] 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 system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0137] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0138] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A small sample speech emotion recognition method based on meta metric learning, characterized in that, The method comprises the following steps: S1: obtaining a small sample voice emotion dataset and performing data processing, including unbalanced data processing, voice signal preprocessing and spectrogram generation, to obtain a spectrogram dataset; S2: according to the standard of meta-learning, the spectrogram dataset is split into a support set S and a query set Q used in the meta-training stage, and a support set S' and a query set Q' used in the meta-testing stage; S3: building a small sample voice emotion recognition model based on metric learning; S4: training and testing the small sample voice emotion recognition model in two stages of meta-training and meta-testing, wherein the meta-training stage trains hyperparameters, and the meta-testing stage trains parameters, and after the training is completed, a trained model is obtained; S5: performing voice emotion recognition through the trained model to obtain a voice emotion recognition result; The step of unbalanced data processing comprises: S121: define the speech emotion dataset that may exist data imbalance as D = {x1, x2,..., x n}, define the number of emotion categories contained in the dataset as n, then n-1 categories are minority classes, and the remaining 1 category is the majority class with the largest data volume, and the set of majority classes is denoted as Dq = {x1, x2,..., x q}, that is, the number of majority classes is q, and for n-1 minority classes, the number of samples of any minority class is denoted as p; S122: performing steps S123-S126 iteratively until the data is balanced, that is, the unbalanced data processing process is completed; S123: calculating an unbalanced index i, when i < 1, it is considered that there is data imbalance, and the calculation formula of i is Wherein, λ is a filtering parameter, used to filter slight data imbalance problems that have little effect on subsequent emotion classification tasks, and the value range is [p / q, 1]; S124: calculating the total number of samples required to be synthesized for the minority class, N = α (q-p), α ∈ (0, 1], Wherein, α is an adjustment parameter, used to set the required adjustment degree, when α = 1, it is required to adjust to complete data balance; S125: According to the Euclidean distance calculation formula, find the k nearest neighbors of each minority class sample x i ​ wherein t represents the feature dimension of the speech emotion, x i represents the minority class sample, y j represents the majority class sample, and x i the number of the majority class samples in the k neighboring samples of x i ; S126: Calculate the density distribution p of the majority class samples in the range of k neighboring samples of the minority class sample x i i ​​ ρ i = s i / k S127: According to the normalized density distribution value, the number N of new samples to be synthesized for the sample is obtained i Return to step S122; Step S2 comprises: S21: splitting the spectrogram dataset into an original training set and an original test set; S22: selecting N class samples from the original test set, the selected N class samples contain a plurality of samples; S23: randomly selecting k+x samples from each of the N class samples, wherein k samples are put into the support set S' used in the meta-testing stage, and the other x samples are put into the query set Q' used in the meta-testing stage; S24: obtaining the support set S and the query set Q used in the meta-training stage according to the same method of steps S22-S23.

2. The meta-metric learning based small sample speech emotion recognition method according to claim 1, characterized in that, Step S1 comprises: S11: obtaining a small sample voice emotion dataset; S12: performing unbalanced data processing on the small sample voice emotion dataset; S13: after completing the unbalanced data processing, performing preprocessing operations on the voice signal, including sampling, quantization, pre-emphasis, windowing, and generating original emotion spectrograms; S14: selecting a fixed step to crop the original emotion spectrograms to obtain a spectrogram dataset.

3. The meta-metric learning based few-shot speech emotion recognition method according to claim 1, wherein, Step S3 comprises: Setting the embedding function f θ is expressed as follows: wherein Z l+1 represents the convolution output of the l+1th layer, Z l represents the convolution output of the lth layer, w l+1 is the parameter matrix of the l+1th layer, and b is the bias quantity. Let the support set be S={(x1,y1),(x2,y2),...,(x n ,y n )},i=1,2,...,n, wherein x i is the i-th feature vector, y i is the corresponding label information; After projection into the embedding space, the prototype of each class is calculated according to the following formula where n denotes the true class of the sample, z n denotes a prototype of class n, S n denotes a set of classes, f θ denotes an embedding function; For each sample of the query set, the probability of belonging to each class is calculated as follows: Wherein, n' represents the predicted sample category, n represents the true category of the sample, x is the feature vector of the sample point, y is the corresponding label information, and d represents the Euclidean distance; Therefore, the loss function is defined as J θ = -log(p θ (y = n'|x)).

4. The meta-metric learning based small sample speech emotion recognition method according to claim 1, characterized in that, In step S3, an adaptive motion estimation algorithm is selected to optimize the parameters.

5. The method of claim 1, wherein the meta-metric learning based small sample speech emotion recognition method is characterized by, The step S3 further comprises: setting initial hyperparameters and initial model parameters of the small sample speech emotion recognition model.

6. The method of claim 5, wherein the meta-metric learning-based small sample speech emotion recognition method is characterized by, The step S4 comprises: The initial hyperparameters set in the step S3 are trained through a meta-training phase, and after the best hyperparameters are trained, a meta-testing phase is entered, i.e., model parameters are trained to obtain optimal model parameters.

7. A meta-metric learning based few-shot speech emotion recognition device for implementing the method of any one of claims 1-6. The system comprises the following modules: A data processing module is configured to acquire a small sample speech emotion dataset and perform data processing, including unbalanced data processing, speech signal preprocessing and spectrogram generation, to obtain a spectrogram dataset; A data splitting module is configured to split the spectrogram dataset into a support set S and a query set Q used in a meta-training phase and a support set S' and a query set Q' used in a meta-testing phase according to a meta-learning standard; A network building module is configured to build a small sample speech emotion recognition model based on metric learning; A network training module is configured to train and test the small sample speech emotion recognition model in two phases of meta-training and meta-testing, wherein the meta-training phase trains hyperparameters and the meta-testing phase trains parameters, and after the training is completed, a trained model is obtained; An emotion recognition module is configured to perform speech emotion recognition through the trained model to obtain a speech emotion recognition result.

Images