Emotion recognition model training method and system and emotion recognition method

Abstract

The invention relates to the signal processing technology, and discloses a training method and system of an emotion recognition model and an emotion recognition method.The method comprises the steps that noise adding and enhancement processing is carried out on an obtained voice sample, an enhanced voice signal is generated, and the voice sample is a sample with an emotion label; inputting the enhanced voice signal into a deep learning model to obtain voice coding features; decoupling the voice coding features to separate voice features and noise features; performing emotion classification on the voice features through the deep learning model, and calculating emotion classification loss; performing noise classification on the noise features, and calculating auxiliary task loss; and performing iterative updating on the deep learning model based on the emotion classification loss and the auxiliary task loss to obtain an emotion recognition model. And the robustness of the model is improved through training, so that more accurate emotion classification is realized.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a training method, system, and emotion recognition method for an emotion recognition model. Background Art

[0002] In recent years, with the rapid development of speech processing technology, real-time speech emotion recognition has become a key research area in the field of artificial intelligence. However, in practical applications, speech signals are often interfered with by various types of environmental noise, which poses a significant challenge to the performance of emotion recognition systems. Background noise particularly impacts speech signal quality in typical application scenarios such as car cockpits and family gatherings. Specifically, high-frequency noise often obscures the fundamental frequency variation characteristics associated with emotion, while low-frequency noise can interfere with the accurate extraction of short-term energy features. These noise interferences not only reduce the accuracy of emotion feature extraction but also directly impact the overall performance of emotion classification. Therefore, achieving accurate speech emotion recognition in complex noisy environments has become a key technical challenge that needs to be addressed urgently. Summary of the Invention

[0003] In view of this, the embodiments of the present application provide a training method, system, and emotion recognition method for an emotion recognition model, which can effectively solve the problem in the prior art that background noise interferes with the quality of the voice signal, resulting in reduced accuracy in emotion feature extraction, and more directly affecting the overall performance of emotion classification.

[0004] In a first aspect, an embodiment of the present application provides a method for training an emotion recognition model, the method comprising:

[0005] Performing noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal, wherein the speech sample is a sample with an emotion label;

[0006] Inputting the enhanced speech signal into a deep learning model to obtain speech coding features;

[0007] Decoupling the speech coding features to separate speech features and noise features;

[0008] Performing emotion classification on the speech features using the deep learning model and calculating the emotion classification loss;

[0009] Performing noise classification on the noise features and calculating auxiliary task loss;

[0010] Based on the emotion classification loss and the auxiliary task loss, the deep learning model is iteratively updated to obtain an emotion recognition model.

[0011] In some embodiments, after inputting the enhanced speech signal into the deep learning model to be trained to obtain speech coding features and before decoupling the speech coding features, the training method further includes:

[0012] Passing the speech coding features to the discriminator;

[0013] evaluating the speech coding features using the discriminator to calculate an adversarial loss;

[0014] The adversarial loss is passed to the deep learning model to optimize the deep learning model.

[0015] In some embodiments, the step of passing the adversarial loss to the deep learning model further includes:

[0016] The adversarial loss is passed to the feature extraction layer in the deep learning model using a gradient reversal mechanism.

[0017] In some embodiments, performing noise addition and enhancement processing on the acquired speech sample to obtain an enhanced speech signal includes:

[0018] Performing noise processing on the speech sample to generate a noisy speech signal;

[0019] Preprocessing the noisy speech signal to obtain a preprocessed noisy speech signal;

[0020] Calculating a frequency spectrum based on the preprocessed noisy speech signal;

[0021] generating a spectral mask based on the spectrum;

[0022] Adjusting the spectrum according to the spectrum mask to obtain an adjusted spectrum;

[0023] Spectrum reconstruction is performed according to the adjusted spectrum to obtain the enhanced speech signal.

[0024] In some embodiments, decoupling the speech coding features to separate speech features and noise features includes:

[0025] The speech coding features are decoupled based on an orthogonal constraint algorithm to separate the speech features and the noise features.

[0026] In some embodiments, using spectral feature information generated by the speech sample with the emotion label as a supervisory signal, and performing emotion classification on the speech feature using the deep learning model includes:

[0027] The supervisory signal is input into the deep learning model, so that the deep learning model performs emotion classification on the speech features according to the supervisory signal.

[0028] In a second aspect, an embodiment of the present application provides an emotion recognition method, comprising:

[0029] The speech signal is input into an emotion recognition model for processing to obtain an emotion recognition result corresponding to the speech signal; wherein the emotion recognition model is obtained by using the above-mentioned emotion recognition model training method.

[0030] In a third aspect, an embodiment of the present application provides an emotion recognition model training system, comprising:

[0031] A sample processing module is used to perform noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal, wherein the speech sample is a sample with an emotion label;

[0032] A feature coding module, configured to input the enhanced speech signal into a deep learning model to obtain speech coding features;

[0033] A feature decoupling module, configured to decouple the speech coding features to separate speech features and noise features;

[0034] An emotion classification module, configured to perform emotion classification on the speech features using the deep learning model and calculate the emotion classification loss;

[0035] A noise classification module, configured to perform noise classification on the noise features and calculate the auxiliary task loss;

[0036] A model optimization module is used to iteratively update the deep learning model based on the emotion classification loss and the auxiliary task loss to obtain an emotion recognition model.

[0037] In a fourth aspect, an embodiment of the present application further provides a computer device, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the emotion recognition model training method or emotion recognition method.

[0038] In a fifth aspect, an embodiment of the present application further provides a readable storage medium storing a computer program, which, when executed, implements the emotion recognition model training method or emotion recognition method.

[0039] The embodiments of the present application have the following beneficial effects:

[0040] The embodiment of the present application provides a training method for an emotion recognition model, comprising: performing noise addition and enhancement processing on an acquired speech sample to generate an enhanced speech signal, wherein the speech sample is a sample with an emotion label; inputting the enhanced speech signal into a deep learning model to obtain speech coding features; decoupling the speech coding features to separate speech features and noise features; performing emotion classification on the speech features through the deep learning model and calculating the emotion classification loss; performing noise classification on the noise features and calculating the auxiliary task loss; iteratively updating the deep learning model based on the emotion classification loss and the auxiliary task loss to obtain an emotion recognition model. The noisy speech signal is processed by enhancement technology to improve the quality of the speech signal; and the robustness and accuracy of the model are significantly improved by training the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 A first flow chart of a method for training an emotion recognition model according to an embodiment of the present application is shown;

[0043] Figure 2 A second flow chart of the method for training an emotion recognition model according to an embodiment of the present application is shown;

[0044] Figure 3 A third flow chart of the method for training an emotion recognition model according to an embodiment of the present application is shown;

[0045] Figure 4 A fourth flow chart of the emotion recognition method according to an embodiment of the present application is shown;

[0046] Figure 5 A structural diagram of the emotion recognition model training system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0048] The components of the embodiments of the present application generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but rather merely represents selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0049] Hereinafter, the terms "including", "having" and their cognates used in various embodiments of the present application are intended only to indicate specific features, numbers, steps, operations, elements, components or combinations of the aforementioned items, and should not be understood as excluding the existence of one or more other features, numbers, steps, operations, elements, components or combinations of the aforementioned items or adding the possibility of one or more features, numbers, steps, operations, elements, components or combinations of the aforementioned items. In addition, the terms "first", "second", "third" and the like are only used to distinguish descriptions and should not be understood as indicating or implying relative importance.

[0050] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by those skilled in the art to which the various embodiments of the present application belong. The terms (such as those defined in generally used dictionaries) will be interpreted as having the same meaning as in the context of the relevant technical field and will not be interpreted as having an idealized meaning or an overly formal meaning unless clearly defined in the various embodiments of the present application.

[0051] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0052] Considering the problem that background noise in the prior art interferes with the quality of speech signals, resulting in reduced accuracy in emotion feature extraction and more directly affecting the overall performance of emotion classification, the present application provides a training method, system, and emotion recognition method for an emotion recognition model. The method improves the quality of noisy speech signals by processing them using enhancement technology. The method also significantly improves the robustness and accuracy of the model through model training.

[0053] The following describes the training method of the emotion recognition model in conjunction with some examples in one implementation.

[0054] Figure 1 A flow chart of a method for training an emotion recognition model according to an embodiment of the present application is shown. Exemplarily, the training method includes S101-S106:

[0055] S101 , performing noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal, where the speech sample is a sample with an emotion label.

[0056] It is understandable that the voice sample should be a clean voice sample without noise, which can be obtained using a voice acquisition device such as a microphone. Each voice segment is equipped with at least one clear emotion label, such as anger, sadness, calmness, etc.

[0057] The samples with emotion labels are subjected to noise processing to generate noisy speech signals. It is understandable that the noise data of the noise processing includes but is not limited to additive noise, convolution noise or impulse noise, etc. The speech samples can be subjected to noise processing using a generative adversarial network. In one embodiment, the generative adversarial network consists of two parts: a generator and a discriminator. The generator is used to generate samples that are as close as possible to the real data distribution, while the discriminator is used to distinguish between the generated samples and the real samples. The two compete with each other during the training process and eventually reach a state of equilibrium. Applying the generative adversarial network to noise generation can generate noise signals with dynamic characteristics, thereby achieving more flexible noise addition operations.

[0058] First, various types of noise data, such as white noise, pink noise, and ambient noise, can be collected and used as real noise samples. At the same time, random noise vectors are sampled from a standard distribution and used as input to the generator. Using an alternating optimization strategy, the generator parameters are first fixed to update the discriminator, and then the discriminator parameters are fixed to update the generator. After multiple rounds of iteration, the generator is able to generate high-quality noise signals. Furthermore, when generating noise, control parameters such as noise type weights and intensity factors can be introduced and set according to the actual application. The generator is used to generate a noise signal with the same length as the speech signal. The generated noise signal is superimposed on the speech sample to form a noisy speech signal.

[0059] Spectral enhancement processing is performed on the noisy speech signal to obtain an enhanced speech signal. In one embodiment, the spectral enhancement processing includes time-frequency domain noise reduction, and a spectral mask can be generated by a convolutional recurrent network or a deep complex convolutional network to amplify the speech component and suppress the noise component.

[0060] For example, in one embodiment, Figure 2 As shown in FIG, performing spectrum enhancement processing on a noisy speech signal includes the following sub-steps:

[0061] S201 , preprocessing the noisy speech signal to obtain a preprocessed noisy speech signal.

[0062] In one embodiment, preprocessing may include framing and windowing. Framing divides a continuous speech signal into short-duration signal frames so that each frame can be analyzed separately. The framed signal may have discontinuities at frame boundaries, which can introduce high-frequency noise in the frequency domain. To mitigate this effect, windowing applies a window function to each frame to reduce discontinuities at frame boundaries.

[0063] After framing and windowing, the noisy speech signal is converted into a series of short-term stationary signal frames, and the discontinuity at the frame boundaries is effectively suppressed.

[0064] S202: Calculate a frequency spectrum according to the preprocessed noisy speech signal.

[0065] The spectrum includes the amplitude spectrum and the phase spectrum. A discrete Fourier transform is performed on each frame of the signal to calculate the amplitude spectrum and the phase spectrum. Through calculation, the noisy speech signal can be converted into a time-frequency domain representation, providing a basis for subsequent processing.

[0066] S203: Generate a spectrum mask according to the spectrum.

[0067] Spectral masks are used to amplify speech components and suppress noise components. They can be generated using deep learning models such as convolutional recurrent networks or deep complex convolutional networks. Furthermore, spectral masks can be trained through self-supervised learning or contrastive learning, enhancing the model's adaptability to complex noise scenarios.

[0068] S204: Adjust the spectrum according to the spectrum mask to obtain an adjusted spectrum.

[0069] Exemplarily, spectrum enhancement technology is used to perform random masking on the spectrum graph. In one embodiment, spectrum masking rules are set based on actual application conditions. For example, the masking ratio may be set to mask a maximum of 30% of the bandwidth; the masking may be a continuous frequency band; the masking position and range may be randomly generated, etc. Based on the spectrum masking rules, the continuous frequency band is randomly masked in the frequency dimension to obtain an adjusted spectrum.

[0070] S205 , performing spectrum reconstruction according to the adjusted spectrum to obtain an enhanced speech signal.

[0071] The adjusted spectrum is subjected to inverse discrete Fourier transform to convert the frequency domain signal into a time domain signal, and the complete enhanced speech signal is restored by combining overlap-addition.

[0072] By performing spectral enhancement on noisy speech signals, the quality of the speech signal is improved, bringing it closer to clean speech. This process also preserves the semantic information required for tasks like emotion recognition and generates robust features more suitable for emotion recognition tasks. This provides better input for subsequent emotion recognition tasks and effectively reduces noise interference, improving the overall performance of the emotion recognition model, particularly its robustness in complex noisy environments.

[0073] S102: Input the enhanced speech signal into a deep learning model to obtain speech coding features.

[0074] A deep learning model is constructed using a convolutional neural network architecture, and the enhanced speech signal is input into the model. The model extracts high-dimensional, robust emotion-related features from the input speech signal to form speech coding features.

[0075] Furthermore, the discriminator can be used to generate adversarial losses to optimize the deep learning model. For example, in one embodiment, Figure 3 As shown, using the discriminator to generate adversarial loss, the deep learning model is optimized including S301-S303:

[0076] S301: passing the speech coding features to the discriminator.

[0077] In one embodiment, the speech coding features include noisy speech coding features and clean speech coding features.

[0078] Furthermore, gradient-free features can be created and passed to the discriminator to prevent gradients from flowing back to the deep learning model, ensuring that the update of the discriminator does not directly affect the learning process of the feature encoder.

[0079] S302: Evaluate the speech coding features using a discriminator to calculate the adversarial loss.

[0080] The discriminator distinguishes whether the feature comes from clean speech or noisy speech, and uses the cross-entropy function to calculate the adversarial loss to make the noisy speech deceive the discriminator as much as possible.

[0081] S303: passing the adversarial loss to the deep learning model to optimize the deep learning model.

[0082] An adversarial loss is used to update the discriminator's parameters, maximizing the loss to better distinguish between clean and noisy speech. The adversarial loss is then backpropagated back to the model to update its parameters, making the noise characteristics in the generated speech features more obscure and difficult for the discriminator to identify. This process effectively enables the model to learn more robust feature representations during training, thereby improving its performance in complex noisy environments.

[0083] The adversarial loss is passed to the feature extraction layer of the deep learning model using the gradient reversal mechanism. Only the gradient from the discriminator is allowed to propagate back to the model, while preventing the gradient from affecting the discriminator itself.

[0084] By performing adversarial training on the model, we improve its robustness and generalization capabilities, ensuring that the model can extract high-dimensional, robust emotion-related features from the input speech signal. These features need to minimize noise interference while retaining emotion-related semantic information.

[0085] S103: Decouple the speech coding features to separate speech features and noise features.

[0086] In one embodiment, the speech coding features are decoupled based on an orthogonal constraint algorithm to separate speech features and noise features. The orthogonal constraint algorithm achieves decoupling by minimizing the correlation between the speech features and the noise features.

[0087] S104: emotion classification is performed on the speech features through a deep learning model, and the emotion classification loss is calculated.

[0088] The spectral feature information generated by emotion-labeled speech samples can be used as a supervisory signal and fed into a deep learning model, which then uses the supervisory signal to classify the speech features into emotions. In one embodiment, the model predicts the emotion category of the input speech features by learning from the features of the spectrogram. Each speech feature is matched with at least one explicit emotion label, and the emotion classification loss is calculated using a cross-entropy function. The supervisory signal serves as a reference to optimize the quality of the speech features and ensure that the model learns emotion-related features that are invariant to noise.

[0089] S105, performing noise classification on the noise features and calculating the auxiliary task loss.

[0090] The noise features are classified and the auxiliary task loss is calculated based on the classification results using the cross entropy function.

[0091] S106: Based on the emotion classification loss and the auxiliary task loss, the deep learning model is iteratively updated to obtain an emotion recognition model.

[0092] A dual loss function is constructed based on emotion classification loss and auxiliary task loss. According to the dual loss function, the parameters of the model are updated to minimize the emotion classification loss to improve the accuracy of emotion recognition, and maximize the auxiliary loss to encourage the model to generate feature representations that are insensitive to noise.

[0093] The model is trained through a dual loss function consisting of the auxiliary task loss of noise classification and the emotion classification loss, which not only improves the accuracy of emotion recognition but also further enhances the model's robustness to noise in complex noise environments.

[0094] Figure 4 This is a flow chart of an emotion recognition method according to an embodiment of the present application. Exemplarily, the emotion recognition method includes:

[0095] S401: Acquire a voice signal.

[0096] For example, in one environment, a voice collection device with a sound pickup function can be used to obtain a user's voice signal. It is understood that the voice signal generally includes the user's actual voice signal and environmental noise.

[0097] S402: Input the speech signal into the emotion recognition model for processing to obtain an emotion recognition result corresponding to the speech signal.

[0098] Specifically, the emotion recognition result is that each speech segment in the speech signal is equipped with at least one clear emotion label, such as anger, sadness, calmness, etc. The emotion recognition model is trained through the training method of the above embodiment, and can ignore the influence of background noise such as wind, keyboard sound, engine sound, etc., and focus on the key information related to rain emotion in the speech signal, avoiding the interference of noise on the emotion recognition results, and realizing accurate speech emotion recognition in a noisy environment.

[0099] It can be understood that the emotion recognition method of this embodiment can be applied to children's education sentiment analysis scenarios, and can adapt to different speaking objects and different home environment noises; the emotion recognition method of this embodiment can also be used for intelligent emotion recognition in the whole house, and the emotion recognition method of this embodiment can be applied to any scenario that requires emotion recognition of voice signals.

[0100] Among them, the training method of the emotion recognition model can be found in the description of the above embodiment, which will not be repeated here. The optional options in the above embodiment are also applicable to this embodiment, so they will not be repeated here.

[0101] like Figure 5 As shown, based on the method of the above embodiment, this embodiment provides an emotion recognition model training system. Exemplarily, the emotion recognition model training system 100 includes:

[0102] The sample processing module 110 is used to perform noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal. The speech sample is a sample with an emotion label;

[0103] A feature coding module 120 is configured to input the enhanced speech signal into a deep learning model to obtain speech coding features;

[0104] A feature decoupling module 130 is used to decouple speech coding features to separate speech features and noise features;

[0105] The emotion classification module 140 is used to classify the speech features into emotions using a deep learning model and calculate the emotion classification loss;

[0106] A noise classification module 150 is used to classify noise features and calculate auxiliary task losses;

[0107] The model optimization module 160 is used to iteratively update the deep learning model based on the emotion classification loss and the auxiliary task loss to obtain an emotion recognition model.

[0108] It can be understood that the system of this embodiment corresponds to the control method of the above embodiment, and the optional items in the above embodiment are also applicable to this embodiment, so they will not be described again here.

[0109] The present application also provides a computer device. Exemplarily, the computer device includes a processor and a memory, wherein the memory stores a computer program, and the processor runs the computer program, thereby enabling the device to execute the above-mentioned emotion recognition model training method or the above-mentioned emotion recognition method.

[0110] Among them, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including a central processing unit (CPU), a graphics processing unit (GPU) and a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or at least one of other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application.

[0111] The memory may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), and electrically erasable programmable read-only memory (EEPROM). The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving an execution instruction.

[0112] The present application also provides a computer-readable storage medium for storing a computer program used in the terminal device. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and structure diagrams in the accompanying drawings show the possible architectures, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in an alternative implementation, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the structure diagram and / or flowchart, and the combination of boxes in the structure diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0114] In addition, the functional modules or units in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0115] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a smart phone, personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0116] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A training method for an emotion recognition model, characterized in that: The method comprises: Performing noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal, wherein the speech sample is a sample with an emotion label; Inputting the enhanced speech signal into a deep learning model to obtain speech coding features; Decoupling the speech coding features to separate speech features and noise features; Performing emotion classification on the speech features using the deep learning model and calculating the emotion classification loss; Performing noise classification on the noise features and calculating auxiliary task loss; Based on the emotion classification loss and the auxiliary task loss, the deep learning model is iteratively updated to obtain an emotion recognition model.

2. The method for training an emotion recognition model according to claim 1, wherein: After inputting the enhanced speech signal into the deep learning model to be trained to obtain speech coding features and before decoupling the speech coding features, the training method further includes: Passing the speech coding features to the discriminator; evaluating the speech coding features using the discriminator to calculate an adversarial loss; The adversarial loss is passed to the deep learning model to optimize the deep learning model.

3. The method for training an emotion recognition model according to claim 2, wherein: The step of transferring the adversarial loss to the deep learning model further includes: The adversarial loss is passed to the feature extraction layer in the deep learning model using a gradient reversal mechanism.

4. The method for training an emotion recognition model according to claim 1, wherein: The step of adding noise and enhancing the acquired speech sample to obtain an enhanced speech signal includes: Performing noise processing on the speech sample to generate a noisy speech signal; Preprocessing the noisy speech signal to obtain a preprocessed noisy speech signal; Calculating a frequency spectrum based on the preprocessed noisy speech signal; generating a spectrum mask based on the spectrum; Adjusting the spectrum according to the spectrum mask to obtain an adjusted spectrum; Spectrum reconstruction is performed according to the adjusted spectrum to obtain the enhanced speech signal.

5. The method for training an emotion recognition model according to claim 1, wherein: Decoupling the speech coding features to separate speech features and noise features includes: The speech coding features are decoupled based on an orthogonal constraint algorithm to separate the speech features and the noise features.

6. The method for training an emotion recognition model according to claim 1, wherein: Using the spectral feature information generated by the speech sample with the emotion label as a supervisory signal, and performing emotion classification on the speech feature by using the deep learning model includes: The supervisory signal is input into the deep learning model, so that the deep learning model performs emotion classification on the speech features according to the supervisory signal.

7. An emotion recognition method, characterized in that: include: Acquire voice signals; The speech signal is input into an emotion recognition model for processing to obtain an emotion recognition result corresponding to the speech signal; wherein the emotion recognition model is obtained by using the training method of the emotion recognition model according to any one of claims 1 to 6.

8. An emotion recognition model training system, characterized in that: include: A sample processing module is used to perform noise addition and enhancement processing on the acquired speech sample to generate an enhanced speech signal, wherein the speech sample is a sample with an emotion label; A feature coding module, configured to input the enhanced speech signal into a deep learning model to obtain speech coding features; A feature decoupling module, configured to decouple the speech coding features to separate speech features and noise features; An emotion classification module, configured to perform emotion classification on the speech features using the deep learning model and calculate the emotion classification loss; A noise classification module, configured to perform noise classification on the noise features and calculate the auxiliary task loss; A model optimization module is used to iteratively update the deep learning model based on the emotion classification loss and the auxiliary task loss to obtain an emotion recognition model.

9. A computer device, characterized in that: The computer device comprises: a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed, the method according to any one of claims 1 to 7 is implemented.

Images