A child-oriented speech emotion recognition model training method

Abstract

This invention relates to the field of children's speech processing technology, specifically a method for training a speech emotion recognition model for children. The method includes: collecting a raw children's speech dataset containing emotion categories and speaker age group labels; preprocessing the dataset to obtain standardized segments through sampling rate unification, silence segment removal, and volume amplitude normalization; extracting multi-dimensional acoustic features to form a raw set; stratifying the dataset by age group; and constructing speaker-independent emotion features within each layer by subtracting the mean of speaker features from feature projections. This feature set is then input into a neural network, trained with an emotion confusion matrix-weighted loss to obtain a preliminary model, and further fine-tuned using artificially induced strong emotionally aroused children's speech. This method improves the accuracy and robustness of children's speech emotion recognition and enhances the ability to distinguish ambiguous emotional speech.

Description

Technical Field

[0001] This invention relates to the field of children's speech processing technology, and in particular to a method for training a speech emotion recognition model for children. Background Technology

[0002] Children's speech emotion recognition faces the dual challenges of physiological developmental stages and individual differences. Current technologies often directly collect children's speech in natural settings, extract acoustic features, and then input them into the model for training. However, this approach fails to fully consider the significant changes in acoustic features across different age groups. For example, emotions from the same year may exhibit different spectral characteristics at different ages due to vocal cord development, and mixing data from different age groups can easily introduce noise. Furthermore, individual differences in children's timbre and pronunciation habits can be mixed into emotional features, making it difficult for conventional feature extraction methods to isolate speaker information, resulting in impure emotional representations. In addition, naturally collected emotional samples often contain ambiguity due to children's subtle expressions, leading to insufficient model ability to distinguish low-intensity or mixed emotions.

[0003] The existing technical solutions have shortcomings. They do not stratify the data by age group, and the overall feature extraction cannot adapt to the stage evolution of children's acoustic features, making the model susceptible to cross-age group differences. They do not specifically eliminate individual speaker differences within the same age group, and residual speaker information in the features reduces the accuracy of emotion recognition. The training samples rely on natural emotional expressions, and high-arousal, strong emotion samples are scarce, resulting in insufficient robustness of the model to ambiguous emotions.

[0004] The current challenge is to address how to stratify data by age group of children's speakers, construct emotional features within each stratum that eliminate speaker differences, and how to leverage artificially induced strong emotion enhancement models for children's speech to overcome the limitations of natural samples in discerning ambiguous emotions. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a training method for a voice emotion recognition model for children.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for training a speech emotion recognition model for children, comprising:

[0007] Collect a raw children's speech dataset containing multiple children as speakers. The raw children's speech dataset contains speech samples with emotion category labels and corresponding speaker age group labels.

[0008] The speech samples of the original children's speech dataset are preprocessed to obtain standardized children's speech segments. The preprocessing includes sampling rate unification, silence segment removal, and volume amplitude normalization.

[0009] Multi-dimensional acoustic features are extracted from the standardized children's speech segments to form an original acoustic feature set;

[0010] The original acoustic feature set is segmented by age based on the speaker's age group label, and speaker-independent emotion feature representations are constructed within each age group. The construction process is achieved by subtracting the speaker's feature mean from the feature projection.

[0011] The speaker-independent emotion feature representation is input into the neural network model for training. During the training process, a loss function weighted by the emotion confusion matrix is ​​introduced to generate a preliminary children's speech emotion recognition model.

[0012] The preliminary children's speech emotion recognition model is adversarially fine-tuned using children's emotionally aroused speech segments, which are children's speech with strong emotional expression generated artificially. The adversarial fine-tuning enhances the model's ability to distinguish speech with ambiguous emotions.

[0013] As a further aspect of the present invention, multi-dimensional acoustic features are extracted from the standardized children's speech segments to form an original acoustic feature set, including:

[0014] The multidimensional acoustic features include fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel frequency cepstral coefficients.

[0015] Each standardized child speech segment is subjected to frame-by-frame windowing processing to obtain a series of short speech frames;

[0016] For each short-time speech frame, perform the following processing:

[0017] The fundamental frequency envelope is calculated using the autocorrelation method, and the fundamental frequency envelope reflects the temporal variation trajectory of speech pitch.

[0018] The harmonic noise ratio is obtained by calculating the ratio of the harmonic component energy to the noise component energy of the speech signal.

[0019] The centroid of the spectrum is obtained by calculating the weighted average frequency of the speech frame spectrum;

[0020] Extract a set of static Mel frequency cepstral coefficients, and calculate the difference and acceleration difference coefficients of the static Mel frequency cepstral coefficients between adjacent speech frames to form the dynamic range of the Mel frequency cepstral coefficients;

[0021] The fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel frequency cepstral coefficients calculated for each short speech frame are arranged in chronological order and combined to form the original acoustic feature set of the child's speech segment.

[0022] As a further aspect of the present invention, the original acoustic feature set is segmented by age based on the speaker's age group label, and speaker-independent emotion feature representations are constructed within each age group, including:

[0023] Based on the speaker's age group label, the original acoustic feature set is divided into a preschool children feature subset, a lower school children feature subset, and an upper school children feature subset;

[0024] For each age stratum feature subset, the acoustic feature mean vector of all speech samples of each speaker in the age stratum is calculated to obtain the personalized feature benchmark of each speaker.

[0025] For each speech sample within an age stratum, the original acoustic feature vector is subtracted from the corresponding speaker's personalized feature baseline to obtain a normalized feature vector.

[0026] Principal component analysis is performed on the normalized eigenvectors, and the top principal components with the highest contribution rates are selected as projection axes to project all normalized eigenvectors onto the subspace.

[0027] The projected feature vector is used as a speaker-independent emotional feature representation, which reduces the influence of acoustic differences among different children under the same emotion.

[0028] As a further aspect of the present invention, the speaker-independent emotion feature representation is input into a neural network model for training. During the training process, a loss function weighted by an emotion confusion matrix is ​​introduced to generate a preliminary child speech emotion recognition model, including:

[0029] Construct a composite neural network containing convolutional layers and long short-term memory network layers as the basic model;

[0030] The speaker-independent emotion features are input into the composite neural network in time series. The convolutional layer is used to extract local acoustic patterns, and the long short-term memory network layer is used to model the temporal evolution dependency of emotion.

[0031] In the early stages of model training, the composite neural network is pre-trained using the standard cross-entropy loss function;

[0032] The performance of the pre-trained model is evaluated on the validation set, and the misclassification ratio of the model among various emotions is calculated to form an emotion confusion matrix, which reflects the emotion category pairs that the model is prone to confuse.

[0033] Based on the element values ​​in the emotion confusion matrix, a higher penalty weight is assigned to each pair of easily confused emotion categories, and a weighted cross-entropy loss function is constructed.

[0034] The pre-trained composite neural network is further trained using the weighted cross-entropy loss function, forcing the model to focus more on distinguishing easily confused emotion categories until the loss function converges, thus obtaining the preliminary children's speech emotion recognition model.

[0035] As a further aspect of the present invention, the adversarial fine-tuning of the preliminary child voice emotion recognition model using child emotionally aroused speech fragments includes:

[0036] Prepare a separate dataset of children's emotionally aroused speech fragments, where each fragment is labeled with a high-confidence emotion tag;

[0037] The preliminary children's speech emotion recognition model is used to predict the children's emotionally aroused speech segment dataset, and segments with model prediction confidence below a threshold are identified as fuzzy emotion speech samples.

[0038] Construct an adversarial discriminator to distinguish whether the input speech features come from the original training data or from the ambiguous emotion speech samples;

[0039] The feature extraction layer of the preliminary children's voice emotion recognition model is connected to the adversarial discriminator to form an adversarial training framework.

[0040] In adversarial training, the parameters of the initial children's voice emotion recognition model are adjusted so that the features it extracts can confuse the adversarial discriminator while maintaining the accuracy of its own emotion classification.

[0041] The adversarial training process enhances the robustness of the preliminary children's speech emotion recognition model to children's speech with indistinct features and ambiguous emotion expression.

[0042] As a further aspect of the present invention, it also includes:

[0043] A framework for transferring children's emotional expression style is constructed, which transfers the standard emotional expression patterns in adult speech to the feature space of children's speech, generating style-enhanced children's emotional speech features to expand the original children's speech dataset;

[0044] During model training, the training weights of the speech samples are automatically adjusted based on the emotional intensity of the speech samples. The emotional intensity is obtained by averaging the scores given by multiple annotators on the vividness of the emotional expression of the speech samples.

[0045] The trained children's speech emotion recognition model is validated on an independent children's speech test set, and the classification decision boundary of the model is calibrated based on the validation results to generate the final deployed children's speech emotion recognition model.

[0046] The construction of the framework for transferring children's emotional expression styles involves transferring standard emotional expression patterns from adult speech to children's speech feature space, specifically including:

[0047] Collect a standard adult speech dataset with emotion labels, and extract the same multi-dimensional acoustic features as children's speech from the standard adult speech dataset to form an adult acoustic feature set;

[0048] Train a cross-domain feature transformation network, the input of which is a speaker-independent emotion feature representation of children's speech, and the output is the transformed features mapped to the adult acoustic feature space;

[0049] Train an emotion classifier to evaluate whether the emotion category of the transformed features is consistent with the original emotion label of the input child speech;

[0050] The goal of jointly training the cross-domain feature transformation network and the emotion classifier is to ensure that the transformed features retain the emotional content of the original child's speech while also having a statistical distribution that approximates the acoustic feature distribution of the corresponding emotion in the adult acoustic feature set.

[0051] Using the trained cross-domain feature transformation network, a portion of the samples in the original children's speech dataset are processed to generate a batch of synthetic children's speech features with more standard emotional acoustic patterns.

[0052] The synthesized children's speech features and their corresponding emotion labels are merged into the original children's speech dataset as new data for subsequent model training.

[0053] As a further aspect of the present invention, the step of automatically adjusting the training weights of the speech samples based on the emotional intensity during model training includes:

[0054] In the original children's speech dataset, an emotion intensity score is assigned to each speech sample. The emotion intensity score is obtained by averaging the scores given by multiple annotators for the vividness of the emotion expression in the speech samples.

[0055] In each iteration of model training, the relative weights of each speech sample in the current batch are calculated based on its emotional intensity score.

[0056] The specific method for calculating the relative weights is to normalize the emotion intensity scores of all voice samples in the current batch, so that the sample with the highest score gets the baseline weight, and the weights of other samples are scaled according to their scores.

[0057] When calculating the model loss, the loss term of each speech sample is multiplied by its corresponding relative weight, so that samples with more vivid emotional expression and more reliable labels contribute more to the gradient update.

[0058] As training progresses, the strategy of dynamically adjusting weight calculations is employed. In the later stages of training, the weight of emotion intensity scores is gradually reduced to balance the model's learning of samples with different levels of expressiveness.

[0059] As a further aspect of the present invention, the calibration of the classification decision boundary of the model based on the verification results includes:

[0060] The trained model was tested using the independent children's speech test set to obtain the model's predicted probability vector and final predicted category for each test sample.

[0061] Statistically analyze the distribution of the predicted probability vectors for samples under each prediction category, especially for samples that were correctly predicted by the model with low confidence or incorrectly predicted with high confidence.

[0062] For each emotion category, a category-specific probability calibration curve is calculated based on its predicted probability distribution. The probability calibration curve is used to map the raw predicted probability output by the model to a calibrated posterior probability.

[0063] Based on the probability calibration curve, the classification decision rule of the model is adjusted, replacing the original decision method of directly selecting the class with the highest probability with a decision method based on the comparison between the calibrated posterior probability and the dynamic threshold.

[0064] The calibrated decision rules and model parameters are solidified to generate the final deployed children's speech emotion recognition model. The model provides a calibrated confidence score when outputting the emotion category.

[0065] As a further aspect of the present invention, it also includes a step of optimizing the attention mechanism of the preliminary children's speech emotion recognition model:

[0066] After the long short-term memory network layer of the composite neural network, an attention calculation layer is added to calculate the contribution weight of acoustic features at different time steps to the final emotion classification.

[0067] The model with an attention computation layer is trained using training data, enabling the model to automatically focus on the segments of speech in which the emotional expression is most prominent.

[0068] The attention weight distribution of the trained model on various emotional speech is analyzed to identify the acoustic feature time regions that play a decisive role in judging specific emotions.

[0069] Based on the analysis results of the attention weight distribution, an auxiliary acoustic feature template is designed, which describes the occurrence pattern of key acoustic events of a specific emotion on the time axis.

[0070] The acoustic feature template is used as an auxiliary supervision signal and introduced into the model's loss function to constrain the model's attention distribution to move closer to the key region described by the acoustic feature template.

[0071] As a further aspect of the present invention, the step of introducing the acoustic feature template as an auxiliary supervision signal into the loss function of the model includes:

[0072] For each target emotion, a corresponding acoustic feature template is defined. The acoustic feature template is a binary sequence that marks the intervals on the speech time axis that are considered crucial for identifying the target emotion.

[0073] The sequence of attention weights distributed along the time axis as output by the attention calculation layer of the computational model;

[0074] Calculate the difference measure between the attention weight distribution sequence and the acoustic feature template sequence of the target emotion;

[0075] Multiply the difference measure by a regularization coefficient to form the attention regularization loss term;

[0076] The attention regularization loss term is added to the original weighted cross-entropy loss function to form the total loss function;

[0077] During model training, by minimizing the total loss function, the model can accurately classify emotions while its attention mechanism can better align with key regions of emotional expression in human knowledge or statistical priors.

[0078] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0079] The original acoustic feature set is stratified based on speaker age group labels. Within each age group, feature projection is subtracted from the speaker's feature mean to construct a speaker-irrelevant emotion feature representation. This technique first divides children's speech into independent sub-layers according to age groups, making the acoustic features within the same layer more consistent in terms of developmental stage influence, reducing the interference of cross-age group differences on emotion representation. Then, within each layer, latent acoustic patterns related to emotion are extracted through feature projection. The mean of each speaker's feature within that layer is then calculated and subtracted to remove irrelevant information such as individual timbre and pronunciation habits, leaving only core emotion-related information in the feature set, thus improving the relevance and purity of emotion representation.

[0080] The initial model is adversarially fine-tuned using artificially induced emotionally aroused speech fragments from children. This technique induces children to produce highly arousing emotional speech in specific contexts, forming a strong emotion sample set to compensate for the scarcity of strong emotion data in naturally collected samples. By inputting these strong emotion samples into the model in an adversarial manner, the model is prompted to pay attention to the subtle differences between strong and ambiguous emotions during training, learn more refined discrimination boundaries, and thus enhance its ability to distinguish between low-intensity, mixed, and other ambiguous emotional speech, improving the model's robustness in complex emotional expression scenarios. Attached Figure Description

[0081] Figure 1 This is a flowchart of a voice emotion recognition model training method for children according to the present invention;

[0082] Figure 2 A flowchart for constructing age group stratification and speaker-independent feature representation;

[0083] Figure 3 A flowchart for adversarial fine-tuning;

[0084] Figure 4 This is a diagram showing the impact of the attenuation coefficient on the final weight.

[0085] Figure 5 Line graph comparing the performance of children's speech emotion recognition models. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0087] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0088] See Figure 1A raw children's speech dataset containing multiple child speakers was collected. Each speech sample in this dataset was labeled with an emotion category and the corresponding speaker's age group. All speech samples in the raw dataset were preprocessed, including standardizing the speech sampling rate to a preset standard, removing silence segments at the beginning and end of the speech, and normalizing the volume amplitude of the speech waveform to obtain standardized children's speech segments. Multi-dimensional acoustic features were extracted from these standardized children's speech segments to form a raw acoustic feature set. Based on the speaker's age group label attached to the speech samples, the raw acoustic feature set was segmented into age groups. Within each segmented age group, speaker-independent emotion feature representations were constructed by projecting features and subtracting the mean of the speaker's features. These representations aimed to reduce the differences in acoustic characteristics among different children for the same emotion. These speaker-independent emotion feature representations were used as input to train a neural network model. During training, a loss function weighted by an emotion confusion matrix was used to guide model optimization, thereby generating a preliminary children's speech emotion recognition model. An adversarial fine-tuning process was performed on a preliminary children's speech emotion recognition model using a set of artificially induced, emotionally aroused speech fragments. This adversarial fine-tuning process aimed to enhance the model's ability to distinguish children's speech fragments with indistinct features and relatively ambiguous emotional expressions.

[0089] See Figure 2In one embodiment of the present invention, the multi-dimensional acoustic features include the fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel-frequency cepstral coefficients. For each standardized child speech segment, frame-by-frame windowing is performed to obtain a series of short-time speech frames. For an example scenario, consider a speech sample from the original child speech dataset, with a duration of 1.5 seconds and a sampling rate of 16kHz. After preprocessing, a standardized child speech segment is obtained. In the frame-by-frame windowing process, a frame length of 25 milliseconds and a frame shift of 10 milliseconds are used to divide the speech segment into a series of short-time speech frames. For each short-time speech frame, the fundamental frequency envelope is calculated using the autocorrelation method, reflecting the temporal trajectory of speech pitch. The ratio of harmonic component energy to noise component energy of the speech signal is calculated to obtain the harmonic noise ratio. The spectral centroid is obtained by calculating the weighted average frequency of the speech frame spectrum. A set of static Mel-frequency cepstral coefficients is extracted, and the difference and acceleration difference coefficients between adjacent speech frames are calculated to constitute the dynamic range of the Mel-frequency cepstral coefficients. The fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel-frequency cepstral coefficients calculated for each short speech frame are arranged in chronological order and combined to form the original acoustic feature set of the children's speech segments. In the data comparison, the original acoustic feature sets of different children's speech segments are consistent in dimensionality, but the feature value distribution varies depending on the speaker's age and emotional state. For example, the fundamental frequency envelope of preschool children's speech usually shows a wider fluctuation range.

[0090] In some embodiments, the original acoustic feature set is stratified by age group based on speaker age group labels, and speaker-independent emotion feature representations are constructed within each age group. According to the speaker age group labels, the original acoustic feature set is divided into a preschool child feature subset, a lower school child feature subset, and an upper school child feature subset. For example, the preschool child feature subset corresponds to children aged 3 to 6, the lower school child feature subset corresponds to children aged 7 to 9, and the upper school child feature subset corresponds to children aged 10 to 12. For each age group feature subset, the mean acoustic feature vector of all speech samples for each speaker within that age group is calculated to obtain a personalized feature benchmark for each speaker. In the example scenario, a preschool child speaker may have 10 speech samples. The average values ​​of the fundamental envelope, harmonic noise ratio, spectral centroid, and Mel-frequency cepstral coefficient dynamic range features of these 10 samples are calculated to form a personalized feature benchmark. For each speech sample within an age stratum, the original acoustic feature vector is subtracted from its corresponding speaker's personalized feature baseline to obtain a normalized feature vector. In data comparison, normalization reduces intra-individual variability among different speech samples from the same child, making the features more focused on emotional differences. Principal component analysis is performed on the normalized feature vectors, selecting the top few principal components with the highest contribution rates as projection axes. All normalized feature vectors are projected onto a subspace. The number of principal components selected can be determined by a cumulative contribution rate threshold, for example, the number of principal components corresponding to a cumulative contribution rate of 95%. The projected feature vectors are used as speaker-independent emotional feature representations. Speaker-independent emotional feature representations weaken the influence of individual acoustic differences among different children under the same emotion. In data comparison, speaker-independent emotional feature representations of different speakers under the same emotion category show higher clustering density in the feature space.

[0091] Optionally, during principal component analysis (PCA), the covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors are solved. The eigenvectors correspond to the principal component directions. PCA can be understood as mapping high-dimensional normalized eigenvectors to a low-dimensional subspace, retaining most of the variance information while removing redundancy. In practical implementation, the mathematical expression of PCA can involve the following formulas:

[0092]

[0093] in: This represents a matrix composed of normalized eigenvectors. This represents the projection matrix consisting of the eigenvectors corresponding to the first k largest eigenvalues. This represents the speaker-independent emotion feature representation matrix after projection. It can be understood that this formula describes the feature projection process, transforming the original features to the principal component space. In some embodiments, the number of age group strata can be adjusted according to the actual data distribution, for example, adding infancy or adolescence strata. Optionally, in the feature normalization step, in addition to subtracting the personalized feature baseline, variance normalization can be introduced to further standardize the feature scale.

[0094] See Figure 3 In one embodiment of the present invention, a composite neural network comprising convolutional layers and long short-term memory (LSM) layers is constructed as the basic model. The convolutional layers may consist of two layers, each using a 3x3 convolutional kernel combined with ReLU activation and pooling operations. The LSM layer may contain 128 hidden units. Speaker-independent emotion feature representations are input into the composite neural network in a time-series manner. In an example scenario, the dimension of the input data is time step × feature dimension, where the feature dimension is the dimension of the speaker-independent emotion feature representation after dimensionality reduction by principal component analysis. The convolutional layers are used to extract local acoustic patterns, such as extracting local correlations from features of adjacent time steps. The LSM layer is used to model the temporal evolution dependency of emotion and capture long-term contextual information in the speech signal. In the early stages of model training, the composite neural network is pre-trained using the standard cross-entropy loss function until the model's classification accuracy on the validation set stabilizes. The performance of the pre-trained model is evaluated on the validation set. The misclassification ratio of the model among various emotions is calculated to form an emotion confusion matrix. The emotion confusion matrix reflects the emotion category pairs that the model is prone to confuse. For example, in a scenario containing four emotions: happiness, sadness, anger, and calmness, the emotion confusion matrix may show that the model frequently misclassifies "sadness" as "calmness" or "anger" as "happiness".

[0095] In some embodiments, a weighted cross-entropy loss function is constructed by assigning a higher penalty weight to each pair of easily confused emotion categories based on the element values ​​in the emotion confusion matrix. The weighted cross-entropy loss function can be calculated using the following formula:

[0096]

[0097] in: This represents the total number of samples in a training batch. This represents the total number of emotion categories. Indicates the first The true sentiment labels of each sample It is an indicator function, when the sample Authentic Labels equals category The value is 1 if it is true, and 0 otherwise. Indicates the model predicts samples Category The probability, This represents the values ​​derived from the sentiment confusion matrix that correspond to the true labels. With category The penalty weights between them. It can be understood that the penalty weights... The value is usually set to be greater than 1, when the category With real labels In the emotion confusion matrix, a higher weight is assigned to a category that exhibits a high misclassification rate. For example, the weight between easily confused "sad" and "calm" categories can be set to 2.0, while the weight for less easily confused categories remains at 1.0. The pre-trained composite neural network is further trained using a weighted cross-entropy loss function, forcing the model to focus more on distinguishing easily confused emotion categories until the loss function converges, resulting in a preliminary children's speech emotion recognition model. In data comparison, the model trained using the weighted cross-entropy loss function shows improved accuracy in distinguishing "sad" and "calm" emotions on the validation set compared to the model trained using the standard cross-entropy loss function.

[0098] Optionally, the preliminary children's speech emotion recognition model can be adversarially fine-tuned using children's emotionally aroused speech fragments. Prepare a separate dataset of children's emotionally aroused speech fragments, where each fragment carries a high-confidence emotion label, such as children's speech fragments expressing strong joy or anger, recorded through professional psychological induction experiments and consistently confirmed by multiple evaluators. Use the preliminary children's speech emotion recognition model to predict on the dataset of children's emotionally aroused speech fragments, identifying fragments whose model prediction confidence is below a threshold as fuzzy emotion speech samples. For example, the model's prediction probability distribution for a fragment of strong joy is flat, with the highest probability value below the set 0.8 threshold. Construct an adversarial discriminator to distinguish whether the input speech features come from the original training data or from fuzzy emotion speech samples. The adversarial discriminator can be a binary classification network containing several fully connected layers. Connect the feature extraction layer of the preliminary children's speech emotion recognition model to the adversarial discriminator to form an adversarial training framework. In this framework, the feature extraction layer attempts to generate features that can confuse the adversarial discriminator, while the adversarial discriminator attempts to correctly determine the source of the features. In the data comparison, the feature distribution from the original training data and the feature distribution from the fuzzy emotional speech samples may differ significantly before adversarial training.

[0099] In some embodiments, during adversarial training, the parameters of the initial children's speech emotion recognition model are adjusted so that the features it extracts can confuse the adversarial discriminator while maintaining its own accuracy in emotion classification. This can be understood as a minimax game process: the initial children's speech emotion recognition model aims to maximize the error rate of the adversarial discriminator, while the adversarial discriminator aims to minimize its own error rate. Through adversarial training, the robustness of the initial children's speech emotion recognition model to children's speech with indistinct features and ambiguous emotional expression is enhanced. In the example scenario, the model, after adversarial fine-tuning, improves the prediction confidence of ambiguous emotional speech samples in the children's emotionally aroused speech segment dataset that originally had low prediction confidence, while maintaining classification accuracy. Optionally, adversarial training can be performed alternately in multiple rounds, with each round updating the parameters of the adversarial discriminator first, and then updating the parameters of the initial children's speech emotion recognition model.

[0100] In one embodiment of the present invention, a framework for transferring children's emotional expression styles is constructed, emotional intensity weights are introduced during training, and decision boundary calibration is performed. The framework aims to transfer standard emotional expression patterns from adult speech to the feature space of children's speech, generating style-enhanced children's emotional speech features for data augmentation. A dataset of standard adult speech labeled with emotions is collected. This dataset may contain speech from multiple adult speakers, labeled with the same emotional categories as children's speech, such as happiness, sadness, anger, surprise, fear, and calmness. Multi-dimensional acoustic features identical to those in children's speech are extracted from the adult standard speech dataset to form an adult acoustic feature set, with the extracted feature dimensions consistent with the original acoustic feature set of children's speech. A cross-domain feature transformation network is trained. The input to the cross-domain feature transformation network is the speaker-independent emotional feature representation of children's speech, and the output is the transformed features mapped to the adult acoustic feature space. The cross-domain feature transformation network can employ an encoder-decoder structure, where the encoder maps children's features to a latent space, and the decoder maps the latent representation to the adult feature space. An emotion classifier is trained to evaluate whether the emotion category of the transformed features matches the original emotion label of the input child speech. The emotion classifier can be a simple fully connected neural network. The cross-domain feature transformation network and the emotion classifier are jointly trained. The goal of joint training is to ensure that the transformed features retain the emotional content of the original child speech while statistically approximating the acoustic feature distribution of the corresponding emotion in the adult acoustic feature set. This can be achieved by minimizing the feature distribution difference loss and the emotion classification loss. The trained cross-domain feature transformation network is used to process a subset of samples from the original child speech dataset to generate a batch of synthetic child speech features with more standard emotional acoustic patterns. The synthetic child speech features and their corresponding emotion labels are then merged into the original child speech dataset as new data for subsequent model training. In one example scenario, the original dataset contains 1000 samples; through style transfer, an additional 500 synthetic samples can be generated, expanding the total data volume to 1500.

[0101] In some embodiments, the training weights of speech samples are automatically adjusted based on their emotional intensity during model training. In the original children's speech dataset, each speech sample is assigned an emotional intensity score. This score is obtained by averaging the scores given independently by multiple annotators to the vividness of the emotional expression of each speech sample. For example, five annotators can rate the vividness of each speech sample on a scale of 1 to 5, and then calculate the average score as the emotional intensity score. In each iteration of model training, the relative weight of each speech sample in the current batch is calculated based on its emotional intensity score. Specifically, the emotional intensity scores of all speech samples in the current batch are normalized, so that the sample with the highest score receives the baseline weight, and the weights of other samples are scaled proportionally to their scores. In data comparison, a happy speech sample with an emotional intensity score of 4.5 has the highest score in the batch, and its relative weight is set to 1.0, while a sad speech sample with an emotional intensity score of 3.0 has a proportionally calculated relative weight of approximately 0.67. When calculating the model loss, the loss term for each speech sample is multiplied by its corresponding relative weight, ensuring that samples with more vivid emotional expression and more reliable labels contribute more to gradient updates. The strategy of dynamically adjusting weight calculations as training progresses is to gradually reduce the weight of emotion intensity scores in the later stages of training in order to balance the model's learning of samples with different expressiveness. For example, the scaling factor of the weights can be linearly reduced from 1.0 to 0.5 in the later stages of the total training rounds.

[0102] Optionally, the trained children's speech emotion recognition model is validated on an independent children's speech test set, and the model's classification decision boundary is calibrated based on the validation results. The trained model is tested using an independent children's speech test set to obtain the model's predicted probability vector and final predicted category for each test sample. This independent children's speech test set is data not used during the training and validation phases. The distribution of the predicted probability vector for each predicted category is statistically analyzed, especially for samples that the model correctly predicted with low confidence or incorrectly predicted with high confidence. For example, the model correctly predicted an "angry" sample with 0.55 confidence but incorrectly predicted a "sad" sample as "calm" with 0.9 confidence. For each emotion category, a category-specific probability calibration curve is calculated based on its predicted probability distribution. The probability calibration curve maps the original predicted probabilities output by the model to calibrated posterior probabilities. Platt scaling or equal-width binning can be used to fit the probability calibration curve. The model's classification decision rules are adjusted based on the probability calibration curve. The original decision-making method of directly selecting the class with the highest probability is replaced by a decision-making method based on a comparison between the calibrated posterior probability and a dynamic threshold. For example, for the "anger" category, a dynamic threshold of 0.6 is set. Only when the calibrated posterior probability of "anger" is greater than 0.6 and is the maximum value is the sample classified as "anger"; otherwise, the next highest probability category is considered. The calibration process aims to ensure that the predicted probabilities output by the model more realistically reflect the actual probability that a sample belongs to a certain category. The calibrated decision rules and model parameters are then solidified into the final deployed children's speech emotion recognition model. The final deployed children's speech emotion recognition model provides a calibrated confidence score when outputting the emotion category. In some embodiments, the dynamic threshold can be determined based on the precision-recall curves of each category on the validation set to achieve a balance between precision and recall.

[0103] In one embodiment of the invention, an emotion intensity score is assigned to each speech sample in the original children's speech dataset. This score is obtained by averaging the scores given independently by multiple annotators to assess the intensity of emotional expression in each speech sample. For example, three annotators can independently score each speech sample on a scale of 1 to 5. The arithmetic mean of the three annotators' scores is then used as the final emotion intensity score for that sample. A strongly expressed "laughing" speech sample might receive an average score of 4.8, while a flat, narrative speech sample might only receive an average score of 2.3. In each iteration of model training, the relative weight of each speech sample within the current batch is calculated based on its emotion intensity score. Specifically, the relative weights are calculated by normalizing the emotion intensity scores of all speech samples within the current batch, ensuring that the sample with the highest score receives the baseline weight, while the weights of other samples are scaled proportionally to their scores. In an example scenario, suppose a training batch contains four voice samples with emotion intensity scores of 4.5, 3.0, 2.5 and 4.0, where the highest score is 4.5 and its baseline weight is set to 1.0. The relative weights of the other samples are obtained by dividing their scores by the highest score of 4.5, as shown in Table 1.

[0104] Table 1: Emotional Intensity Scores and Relative Weights of Voice Samples in Example Training Batches

[0105] Sample number Emotional intensity score Calculated relative weights 1 4.5 1.00 2 3.0 0.67 3 2.5 0.56 4 4.0 0.89

[0106] When calculating the model loss, the loss term for each speech sample is multiplied by its corresponding relative weight. This allows samples with more vivid emotional expressions and more reliable labels to contribute more to gradient updates. This process can be understood as influencing the direction and magnitude of model parameter updates by scaling the loss gradient of each sample. The strategy of dynamically adjusting weight calculations as training progresses, gradually reducing the influence of the emotional intensity score's weight in the later stages of training, aims to balance the model's learning of samples with different levels of expressiveness. One feasible implementation is to calculate the final sample weights... The formula is defined as:

[0107]

[0108] in: Indicates the first The final weights used for each sample when calculating the loss. Indicates the first The emotional intensity score of each sample This represents the maximum emotional intensity score among all samples in the current training batch. It is a decay coefficient between 0 and 1, whose value linearly decreases from 1.0 to 0.5 as the number of training epochs increases. In the data comparison, the model using the emotion intensity weighted training strategy maintains stable classification accuracy for samples with strong emotional expression on the validation set, while improving the recognition performance for samples with mild expression.

[0109] In some embodiments, the model's classification decision boundary is calibrated based on the validation results. The trained model is tested using an independent children's speech test set to obtain the model's predicted probability vector and final predicted category for each test sample. The predicted probability vector is a multi-dimensional vector, where the value of each dimension represents the probability that the model determines the sample belongs to the corresponding emotion category. The distribution of the predicted probability vector for each predicted category is statistically analyzed, particularly for samples that the model correctly predicted with low confidence or incorrectly predicted with high confidence. For example, the model correctly predicted a "calm" sample with 0.52 confidence but incorrectly predicted a "fear" sample as "surprised" with 0.85 confidence. For each emotion category, a category-specific probability calibration curve is calculated based on its predicted probability distribution. The probability calibration curve maps the original predicted probabilities output by the model to calibrated posterior probabilities. This can be achieved using the Platt scaling method, i.e., using a logistic regression model to fit the mapping relationship from the model's original output probabilities to the true posterior probabilities. The model's classification decision rules were adjusted based on the probability calibration curve. The original decision-making method of directly selecting the category with the highest probability was replaced with a comparison decision method based on the calibrated posterior probability and a dynamic threshold. For example, for the "happy" category, a dynamic threshold of 0.7 was set. The model only classifies a sample as "happy" if the calibrated posterior probability of "happy" is greater than 0.7 and is still the highest among all categories. If the calibrated highest posterior probability does not exceed the dynamic threshold of its corresponding category, the model outputs "uncertain" or selects the category with the second highest probability. The introduction of the dynamic threshold adds conservatism to the model's decision-making, helping to reduce the occurrence of high-confidence errors. The calibrated decision rules and model parameters were then solidified into the final deployed children's speech emotion recognition model. The final deployed model provides a calibrated confidence score when outputting the emotion category. In data comparison, after the decision boundary calibration, the proportion of high-confidence error predictions on the independent test set decreased.

[0110] See Figure 4This is a graph showing the effect of the decay coefficient on the final weight. Sample 1 represents a speech sample with an emotion intensity score of 4.5, whose final weight linearly decays from 1.0 to 0.5 with each training round. Sample 2 represents a speech sample with an emotion intensity score of 3.0, whose final weight linearly decays from approximately 0.67 to approximately 0.44 with each training round. Sample 3 represents a speech sample with an emotion intensity score of 2.5, whose final weight linearly decays from approximately 0.56 to approximately 0.37 with each training round. Sample 4 represents a speech sample with an emotion intensity score of 4.0, whose final weight linearly decays from approximately 0.89 to approximately 0.44 with each training round. This visually demonstrates the effectiveness of the dynamic weight decay strategy in this invention. As the number of training rounds increases, the decay coefficient linearly decreases from 1.0 to 0.5, and the final weight of all samples decays synchronously and linearly.

[0111] In one embodiment of the invention, the attention mechanism of the preliminary children's speech emotion recognition model is optimized. An attention calculation layer is added after the long short-term memory network layer of the composite neural network to calculate the contribution weight of acoustic features at different time steps to the final emotion classification. The attention calculation layer can adopt an additive attention mechanism or a dot product attention mechanism, and its output is a weight vector with the same dimension as the input time step, representing the importance of each time step. The model with the attention calculation layer is trained using training data, enabling the model to automatically focus on the segments of speech with the most prominent emotional expression. In an example scenario, the training data contains speaker-independent emotion feature representations and their corresponding emotion labels. The model calculates attention weights through forward propagation, weights and summarizes the time features, and then performs classification. All parameters, including the attention calculation layer, are updated through backpropagation. Analyzing the attention weight distribution of the trained model across various emotional speech patterns identifies acoustic feature time regions that play a decisive role in judging specific emotions. For example, visualization or statistical methods reveal that in speech expressing "anger," the model's attention weights are significantly concentrated in the high-energy and high-fundamental-frequency regions of the middle segment, while in speech expressing "sadness," the attention weights are likely more distributed in the low-energy and fundamental-frequency-decreasing regions at the end of the speech. Based on the analysis of attention weight distribution, auxiliary acoustic feature templates are designed. These templates describe the occurrence patterns of key acoustic events for specific emotions along the time axis. For example, for the emotion of "happiness," the acoustic feature template might be a binary sequence, where the time interval corresponding to laughter or a sudden rise in pitch is marked as 1, and the remaining intervals are marked as 0.

[0112] In some embodiments, acoustic feature templates are introduced as auxiliary supervision signals into the model's loss function. For each target emotion, a corresponding acoustic feature template is defined. The acoustic feature template is a binary sequence that marks the intervals on the speech timeline considered crucial for identifying the target emotion. The sequence length is the same as the number of time steps of the input speech segment. The attention weight distribution sequence output by the model's attention computation layer is calculated on the timeline. This attention weight distribution sequence is a probability distribution vector whose elements sum to 1. A difference measure is calculated between the attention weight distribution sequence and the acoustic feature template sequence of the target emotion. This difference measure can be measured using Kullback-Leibler divergence or mean squared error. The difference measure is multiplied by a regularization coefficient to form the attention regularization loss term, which is then added to the original weighted cross-entropy loss function to form the total loss function. During model training, minimizing the total loss function enables the model to accurately classify emotions while its attention mechanism can better align with key regions of emotional expression in human knowledge or statistical priors. After introducing an attention regularization loss term in the data comparison, the model's attention to the "anger" emotion on the validation set is more concentrated in the time period of key acoustic events labeled by humans.

[0113] In practice, one way to calculate the difference measure is by using the mean squared error, the formula of which is as follows:

[0114]

[0115] in: This represents the attention regularization loss term. Represents the regularization coefficient. This indicates the total number of frames or time steps on the speech timeline. The attention computation layer is in the 1st... The attention weight values ​​output at each time step. The acoustic feature template is represented in the first... The label value at each time step. This can be understood as the regularization coefficient. The task needs to be optimized to balance the emotion classification task and the attention alignment task; for example, a smaller setting can be used in the initial training phase. The value is 0.1, and it is gradually increased as training progresses. In some embodiments, the generation of acoustic feature templates not only relies on manual analysis, but can also automatically extract common temporal patterns with high attention weights from the training data as templates through clustering algorithms.

[0116] See Figure 5This is a line graph comparing the performance of children's speech emotion recognition models. The horizontal axis represents the regularization coefficient, ranging from 0.10 to 0.50, and the vertical axis represents the model accuracy, ranging from 0.70 to 0.95. The line graph contains two independent lines. The first line, marked with a green square, represents the accuracy of the children's speech emotion recognition model with attention regularization. The second line, marked with an orange triangle, represents the accuracy of the children's speech emotion recognition model without attention regularization. It clearly shows the impact of the attention regularization coefficient on model accuracy. Both lines increase with the regularization coefficient, and the accuracy of the model with attention regularization is consistently higher than that of the model without attention regularization, with the difference widening as the regularization coefficient increases.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for training a speech emotion recognition model for children, characterized in that, The method includes: Collect a raw children's speech dataset containing multiple children as speakers. The raw children's speech dataset contains speech samples with emotion category labels and corresponding speaker age group labels. The speech samples of the original children's speech dataset are preprocessed to obtain standardized children's speech segments. The preprocessing includes sampling rate unification, silence segment removal, and volume amplitude normalization. Multi-dimensional acoustic features are extracted from the standardized children's speech segments to form an original acoustic feature set; The original acoustic feature set is segmented by age based on the speaker's age group label, and speaker-independent emotion feature representations are constructed within each age group. The construction process is achieved by subtracting the speaker's feature mean from the feature projection. The speaker-independent emotion feature representation is input into the neural network model for training. During the training process, a loss function weighted by the emotion confusion matrix is ​​introduced to generate a preliminary children's speech emotion recognition model. The preliminary children's speech emotion recognition model is adversarially fine-tuned using children's emotionally aroused speech segments, which are children's speech with strong emotional expression generated artificially. The adversarial fine-tuning enhances the model's ability to distinguish speech with ambiguous emotions.

2. The method for training a speech emotion recognition model for children according to claim 1, characterized in that, Multi-dimensional acoustic features are extracted from the standardized children's speech segments to form an original acoustic feature set, including: The multidimensional acoustic features include fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel frequency cepstral coefficients. Each standardized child speech segment is subjected to frame-by-frame windowing processing to obtain a series of short speech frames; For each short-time speech frame, perform the following processing: The fundamental frequency envelope is calculated using the autocorrelation method, and the fundamental frequency envelope reflects the temporal variation trajectory of speech pitch. The harmonic noise ratio is obtained by calculating the ratio of the harmonic component energy to the noise component energy of the speech signal. The centroid of the spectrum is obtained by calculating the weighted average frequency of the speech frame spectrum; Extract a set of static Mel frequency cepstral coefficients, and calculate the difference and acceleration difference coefficients of the static Mel frequency cepstral coefficients between adjacent speech frames to form the dynamic range of the Mel frequency cepstral coefficients; The fundamental frequency envelope, harmonic noise ratio, spectral centroid, and dynamic range of Mel frequency cepstral coefficients calculated for each short speech frame are arranged in chronological order and combined to form the original acoustic feature set of the child's speech segment.

3. The method for training a speech emotion recognition model for children according to claim 2, characterized in that, The original acoustic feature set is segmented by age group based on the speaker's age group label, and speaker-independent emotion feature representations are constructed within each age group, including: Based on the speaker's age group label, the original acoustic feature set is divided into a preschool children feature subset, a lower school children feature subset, and an upper school children feature subset; For each age stratum feature subset, the acoustic feature mean vector of all speech samples of each speaker in the age stratum is calculated to obtain the personalized feature benchmark of each speaker. For each speech sample within an age stratum, the original acoustic feature vector is subtracted from the corresponding speaker's personalized feature baseline to obtain a normalized feature vector. Principal component analysis is performed on the normalized eigenvectors, and the top principal components with the highest contribution rates are selected as projection axes to project all normalized eigenvectors onto the subspace. The projected feature vector is used as a speaker-independent emotional feature representation, which reduces the influence of acoustic differences among different children under the same emotion.

4. The method for training a speech emotion recognition model for children according to claim 3, characterized in that, The speaker-independent emotion feature representation is input into a neural network model for training. During training, a loss function weighted by an emotion confusion matrix is ​​introduced to generate a preliminary child speech emotion recognition model, including: Construct a composite neural network containing convolutional layers and long short-term memory network layers as the basic model; The speaker-independent emotion features are input into the composite neural network in time series. The convolutional layer is used to extract local acoustic patterns, and the long short-term memory network layer is used to model the temporal evolution dependency of emotion. In the early stages of model training, the composite neural network is pre-trained using the standard cross-entropy loss function; The performance of the pre-trained model is evaluated on the validation set, and the misclassification ratio of the model among various emotions is calculated to form an emotion confusion matrix, which reflects the emotion category pairs that the model is prone to confuse. Based on the element values ​​in the emotion confusion matrix, a higher penalty weight is assigned to each pair of easily confused emotion categories, and a weighted cross-entropy loss function is constructed. The pre-trained composite neural network is further trained using the weighted cross-entropy loss function, forcing the model to focus more on distinguishing easily confused emotion categories until the loss function converges, thus obtaining the preliminary children's speech emotion recognition model.

5. The method for training a speech emotion recognition model for children according to claim 4, characterized in that, The adversarial fine-tuning of the preliminary child voice emotion recognition model using child emotionally aroused speech fragments includes: Prepare a separate dataset of children's emotionally aroused speech fragments, where each fragment is labeled with a high-confidence emotion tag; The preliminary children's speech emotion recognition model is used to predict the children's emotionally aroused speech segment dataset, and segments with model prediction confidence below a threshold are identified as fuzzy emotion speech samples. Construct an adversarial discriminator to distinguish whether the input speech features come from the original training data or from the ambiguous emotion speech samples; The feature extraction layer of the preliminary children's voice emotion recognition model is connected to the adversarial discriminator to form an adversarial training framework. In adversarial training, the parameters of the initial children's voice emotion recognition model are adjusted so that the features it extracts can confuse the adversarial discriminator while maintaining the accuracy of its own emotion classification. The adversarial training process enhances the robustness of the preliminary children's speech emotion recognition model to children's speech with indistinct features and ambiguous emotion expression.

6. The method for training a speech emotion recognition model for children according to claim 5, characterized in that, Also includes: A framework for transferring children's emotional expression style is constructed, which transfers the standard emotional expression patterns in adult speech to the feature space of children's speech, generating style-enhanced children's emotional speech features to expand the original children's speech dataset; During model training, the training weights of the speech samples are automatically adjusted based on the emotional intensity of the speech samples. The emotional intensity is obtained by averaging the scores given by multiple annotators on the vividness of the emotional expression of the speech samples. The trained children's speech emotion recognition model is validated on an independent children's speech test set, and the classification decision boundary of the model is calibrated based on the validation results to generate the final deployed children's speech emotion recognition model. The construction of the framework for transferring children's emotional expression styles involves transferring standard emotional expression patterns from adult speech to children's speech feature space, specifically including: Collect a standard adult speech dataset with emotion labels, and extract the same multi-dimensional acoustic features as children's speech from the standard adult speech dataset to form an adult acoustic feature set; Train a cross-domain feature transformation network, the input of which is a speaker-independent emotional feature representation of children's speech, and the output is a transformed feature mapped to an adult acoustic feature space; Train an emotion classifier to evaluate whether the emotion category of the transformed features is consistent with the original emotion label of the input child speech; The goal of jointly training the cross-domain feature transformation network and the emotion classifier is to ensure that the transformed features retain the emotional content of the original child's speech while also having a statistical distribution that approximates the acoustic feature distribution of the corresponding emotion in the adult acoustic feature set. Using the trained cross-domain feature transformation network, a portion of the samples in the original children's speech dataset are processed to generate a batch of synthetic children's speech features with more standard emotional acoustic patterns. The synthesized children's speech features and their corresponding emotion labels are merged into the original children's speech dataset as new data for subsequent model training.

7. The method for training a speech emotion recognition model for children according to claim 6, characterized in that, The automatic adjustment of training weights based on the emotional intensity of speech samples during model training includes: In the original children's speech dataset, an emotion intensity score is assigned to each speech sample. The emotion intensity score is obtained by averaging the scores given by multiple annotators for the vividness of the emotion expression in the speech sample. In each iteration of model training, the relative weights of each speech sample in the current batch are calculated based on its emotional intensity score. The specific method for calculating the relative weights is to normalize the emotion intensity scores of all voice samples in the current batch, so that the sample with the highest score gets the baseline weight, and the weights of other samples are scaled according to their scores. When calculating the model loss, the loss term of each speech sample is multiplied by its corresponding relative weight, so that samples with more vivid emotional expression and more reliable labels contribute more to the gradient update. As training progresses, the strategy of dynamically adjusting weight calculations is employed. In the later stages of training, the weight of emotion intensity scores is gradually reduced to balance the model's learning of samples with different levels of expressiveness.

8. The method for training a speech emotion recognition model for children according to claim 7, characterized in that, The calibration of the model's classification decision boundary based on the validation results includes: The trained model was tested using the independent children's speech test set to obtain the model's predicted probability vector and final predicted category for each test sample. Statistically analyze the distribution of the predicted probability vectors for samples under each prediction category, especially for samples that were correctly predicted by the model with low confidence or incorrectly predicted with high confidence. For each emotion category, a category-specific probability calibration curve is calculated based on its predicted probability distribution. The probability calibration curve is used to map the raw predicted probability output by the model to a calibrated posterior probability. Based on the probability calibration curve, the classification decision rule of the model is adjusted, replacing the original decision method of directly selecting the class with the highest probability with a decision method based on the comparison between the calibrated posterior probability and the dynamic threshold. The calibrated decision rules and model parameters are solidified to generate the final deployed children's speech emotion recognition model. The model provides a calibrated confidence score when outputting the emotion category.

9. The method for training a speech emotion recognition model for children according to claim 8, characterized in that, It also includes a step of optimizing the attention mechanism of the preliminary children's speech emotion recognition model: After the long short-term memory network layer of the composite neural network, an attention calculation layer is added to calculate the contribution weight of acoustic features at different time steps to the final emotion classification. The model with an attention computation layer is trained using training data, enabling the model to automatically focus on the segments of speech where the emotional expression is most prominent. The attention weight distribution of the trained model on various emotional speech is analyzed to identify the acoustic feature time regions that play a decisive role in judging specific emotions. Based on the analysis results of the attention weight distribution, an auxiliary acoustic feature template is designed, which describes the occurrence pattern of key acoustic events of a specific emotion on the time axis. The acoustic feature template is used as an auxiliary supervision signal and introduced into the model's loss function to constrain the model's attention distribution to move closer to the key region described by the acoustic feature template.

10. A method for training a speech emotion recognition model for children according to claim 9, characterized in that, The step of introducing the acoustic feature template as an auxiliary supervision signal into the model's loss function includes: For each target emotion, a corresponding acoustic feature template is defined. The acoustic feature template is a binary sequence that marks the intervals on the speech time axis that are considered crucial for identifying the target emotion. The sequence of attention weights distributed along the time axis as output by the attention calculation layer of the computational model; Calculate the difference measure between the attention weight distribution sequence and the acoustic feature template sequence of the target emotion; Multiply the difference measure by a regularization coefficient to form the attention regularization loss term; The attention regularization loss term is added to the original weighted cross-entropy loss function to form the total loss function; During model training, by minimizing the total loss function, the model can accurately classify emotions while its attention mechanism can better align with key regions of emotional expression in human knowledge or statistical priors.

Images