Explainable sentiment recognition method based on large models

Abstract

The application provides an interpretable emotion recognition method based on a large model, comprising: extracting multiple basic acoustic features for emotion classification from original speech data; performing feature screening processing on the basic acoustic features to obtain interpretable acoustic features closely related to a specific emotion; constructing a data set based on the specific emotion, the corresponding interpretable acoustic features and an emotion category label; performing model training using data in the data set to obtain an interpretable emotion recognition model; and performing audio emotion recognition through the interpretable emotion recognition model to obtain a complete emotion analysis result. The method aims to solve the problem of non-transparent decision-making process in current emotion recognition, improve the understanding ability of the large model for audio emotion, and further improve the accuracy of emotion recognition.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an interpretable emotion recognition method based on a large model. Background Technology

[0002] Speech is the fastest and most natural form of human communication, and therefore, it is often used as a convenient bridge for human-computer interaction. This signal contains a wealth of valuable emotional information. Currently, speech emotion recognition (SER) has become an indispensable component of human-computer interaction and other high-end speech processing systems. Typically, SER systems identify whether a speaker is exhibiting a specific emotion by extracting and classifying key features from preprocessed speech signals. However, significant quantitative and qualitative differences exist between humans and machines in how they recognize and associate emotions in speech signals. These differences pose a significant challenge when integrating interdisciplinary knowledge, especially in fields such as speech emotion recognition, applied psychology, and human-computer interaction.

[0003] With the development of deep learning, many studies have adopted neural networks for emotion recognition. However, current deep learning methods suffer from poor interpretability, often failing to explain the specific reasons and basis for the model's classification decisions. This "black box" nature poses a challenge in applications requiring transparent decision-making, such as emotion recognition. In recent years, large-scale models have demonstrated some capability in emotion understanding. For example, they have improved the interpretability of decisions through natural language processing and emotion reasoning; by generating and analyzing emotion descriptions, these models not only enhance the accuracy of emotion recognition but also provide transparency to the model's decision-making process, facilitating more natural emotional interaction applications. While deep learning-based emotion recognition methods have achieved good results, their lack of transparency in the decision-making process makes it difficult to understand how the model arrives at a particular emotion judgment. This "black box" nature limits their application in scenarios requiring precise emotion tracking and interpretation, such as psychological dialogue, and affects therapists' judgment of clients' emotional states.

[0004] Existing sentiment recognition technologies have several shortcomings that limit their effectiveness in practical applications. Most current sentiment recognition models output sentiment categories in a black-box manner, lacking a transparent decision-making process and detailed reasoning explanations, making it difficult for users such as psychologists to understand the basis of the sentiment recognition results. Furthermore, existing technologies often fail to effectively guarantee the logical coherence of sentiment analysis, resulting in sentiment results lacking a clear reasoning chain, thus limiting the interpretability and applicability of sentiment recognition. Summary of the Invention

[0005] In view of this, the present invention provides an interpretable emotion recognition method based on a large model to solve the above problems.

[0006] This invention provides an interpretable emotion recognition method based on a large model, comprising: extracting multiple basic acoustic features for emotion classification from raw speech data; performing feature filtering on the basic acoustic features to obtain interpretable acoustic features closely related to a specific emotion; constructing a dataset based on the specific emotion, the corresponding interpretable acoustic features, and the emotion category label; training a model using the data in the dataset to obtain an interpretable emotion recognition model; and performing audio emotion recognition through the interpretable emotion recognition model to obtain complete emotion analysis results.

[0007] In another implementation of the present invention, the basic acoustic features include prosodic features, timbre features, and spectral features.

[0008] In another implementation of the present invention, the prosodic features include duration-related features, fundamental frequency-related features, and energy-related features; the duration-related features include speech rate and short-time average zero-crossing rate; the fundamental frequency-related features include fundamental frequency and its mean, range of variation, rate of variation, and root mean square error; the energy-related features include short-time average energy and short-time average amplitude.

[0009] In another implementation of the present invention, the sound quality characteristics include time base error, amplitude perturbation, and harmonic noise ratio.

[0010] In another implementation of the present invention, the spectral features are a reflection of the correlation between changes in vocal tract shape and vocalization motion, including linear prediction cepstral coefficients and frequency cepstral coefficients.

[0011] In another implementation of the present invention, the step of using data in the dataset to train a model and obtain an interpretable emotion recognition model includes: inputting interpretable acoustic features from the dataset into a pre-trained model to obtain continuous high-dimensional acoustic feature encoding; decomposing the continuous high-dimensional acoustic feature encoding into a series of discrete speech units; and using the speech units as input and emotion tags and interpretable emotion cues as output to train the model and obtain an interpretable emotion recognition model.

[0012] In another implementation of the present invention, the loss function of the interpretable emotion recognition model includes the cross-entropy loss of emotion classification and the similarity loss of interpretability cues.

[0013] In another aspect of the present invention, an interpretable emotion recognition device based on a large model is provided, comprising: a feature extraction module for extracting various basic acoustic features for emotion classification from raw speech data; a feature processing module for performing feature filtering on the basic acoustic features to obtain interpretable acoustic features closely related to a specific emotion; a data preparation module for constructing a dataset based on the specific emotion, the corresponding interpretable acoustic features, and the emotion category label; a model training module for training a model using the data in the dataset to obtain an interpretable emotion recognition model; and an emotion recognition module for performing audio emotion recognition through the interpretable emotion recognition model to obtain complete emotion analysis results.

[0014] In another aspect of the invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of an interpretable emotion recognition method based on a large model as described in any of the preceding claims.

[0015] In another aspect of the invention, a computer storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of an interpretable emotion recognition method based on a large model as described in any of the preceding claims.

[0016] In the large-model-based interpretable emotion recognition method of this invention, a multi-level feature extraction mechanism is introduced. Combined with basic acoustic features, the model can more deeply understand complex emotional information, improve the overall emotion recognition effect, and increase the accuracy of recognition. By guiding the large model to generate emotion understanding descriptions, emotion recognition is not limited to simple classification results, but also includes detailed reasoning processes, which makes it easier for psychological counselors to understand the emotional state of clients. In addition, the use of the thought chain method enhances the logical coherence of emotion analysis, and the output includes not only emotion labels, but also reasoning processes, thereby significantly improving the interpretability of the model. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention.

[0018] In the attached diagram:

[0019] Figure 1 This is a schematic diagram of the process of an interpretable emotion recognition method based on a large model, according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the framework of an interpretable emotion recognition system based on a large model, according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the feature processing module according to an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of an interpretable emotional cue acquisition process according to an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the model training module flow according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0025] Figure 1 This is a schematic diagram of an interpretable emotion recognition method based on a large model, provided by an embodiment of the present invention. Figure 1 As shown, this embodiment mainly includes the following steps:

[0026] S101. Extract various basic acoustic features for emotion classification from the raw speech data.

[0027] For example, audio signals carry rich emotional information, and by extracting specific acoustic features, it is possible to analyze the potential changes in the speaker's emotions. To improve the accuracy of LLM in understanding speech emotions, a variety of basic acoustic features that can be used for emotion classification were extracted from the raw speech data. These basic acoustic features include, but are not limited to, pitch and pause duration, which show significant changes under different emotional states, thus aiding in the identification and tracking of emotions.

[0028] S102. Perform feature filtering on the basic acoustic features to obtain explanatory acoustic features closely related to specific emotions.

[0029] For example, such as Figure 3 As shown, feature selection is performed on the extracted features to obtain sentiment cues. By filtering out the interpretable acoustic features most closely related to specific emotions, LLM ensures that it can focus on features with strong sentiment discriminative power and use them in the data preparation module to generate interpretable sentiment analysis results.

[0030] S103. Construct a dataset based on the specific emotion, the corresponding interpretive acoustic features, and the emotion category label.

[0031] For example, the data preparation module is used to obtain interpretable sentiment cues. By introducing sentiment-related interpretable features into the dataset, large models can output reasonable explanations in sentiment judgment. Each data sample pair includes audio, interpretable output (sentiment cues), and sentiment category labels. This dataset design not only provides rich training data for sentiment recognition tasks but also provides a framework for output interpretation for large models, enabling them to provide reasoning processes while outputting category labels. This helps applications such as psychological counseling to more clearly understand the specific state and changing trends of users' emotions.

[0032] S104. Use the data in the dataset to train the model and obtain an interpretable emotion recognition model.

[0033] For example, such as Figure 4 As shown, the obtained acoustic feature set (such as pitch and loudness) is added to the Prompt to ensure that the large model can better understand the emotional information behind the acoustic signal when analyzing audio sentiment, rather than simply providing a sentiment classification result. Furthermore, the designed Prompt incorporates the Chain of Thought (CoT) method to obtain sentiment cues, enabling the model to provide a detailed explanation of the sentiment judgment process, output the sentiment category and its reasoning process, and further manually correct the obtained sentiment cues, ultimately obtaining interpretable sentiment cues corresponding to the speech. This guides the LLM to include not only the sentiment category in its output but also the corresponding sentiment analysis description, i.e., interpretable sentiment cues. These interpretable acoustic features can effectively reflect the expression of emotions, making the model's decision-making process more transparent.

[0034] S105. Perform audio emotion recognition using the explained emotion recognition model to obtain complete emotion analysis results.

[0035] For example, for emotion recognition tasks, a fine-tuned large model can be used to replace the traditional emotion recognition model, achieving high-precision and interpretable output for audio emotion recognition. This fine-tuned large model combines speech features and interpretable feature descriptions, thus not only accurately determining emotion categories in emotion recognition tasks but also generating explanatory descriptions during output. This makes the emotion recognition process more transparent, moving beyond simple category judgment to outputting complete emotion analysis results through interpretable reasoning. This can provide more reliable and transparent emotion recognition services in application scenarios requiring a deep understanding of emotions, such as psychological counseling, emotion tracking, and human-computer interaction.

[0036] In the large-model-based interpretable emotion recognition method of this invention, a multi-level feature extraction mechanism is introduced. Combined with basic acoustic features, the model can more deeply understand complex emotional information, improve the overall emotion recognition effect, and increase the accuracy of recognition. By guiding the large model to generate emotion understanding descriptions, emotion recognition is not limited to simple classification results, but also includes detailed reasoning processes, which makes it easier for psychological counselors to understand the emotional state of clients. In addition, the use of the thought chain method enhances the logical coherence of emotion analysis, and the output includes not only emotion labels, but also reasoning processes, thereby significantly improving the interpretability of the model.

[0037] In another implementation of the present invention, the basic acoustic features include prosodic features, timbre features, and spectral features.

[0038] For example, feature extraction and selection are crucial in speech emotion recognition tasks. Standard features in speech emotion recognition can be divided into three main categories: prosodic features, phonological features, and spectral features. Emotion-related acoustic features include multiple dimensions such as speech rate, pitch, and pause duration. Existing large models often struggle to fully extract and understand these complex acoustic emotion features. To address this issue, this invention proposes a multi-level feature extraction method that combines a large model with interpretable acoustic emotion factors. This enables the model to understand the emotional information extracted from audio, improving the large model's ability to understand audio emotions and thus enhancing the accuracy of emotion recognition.

[0039] By combining fundamental acoustic features with thought chains, the model gains a deeper understanding of complex acoustic emotional information. Interpretability methods are used to enhance the large model's ability to understand acoustic emotional features. Through fine-tuning of the open-source large model, it is made to understand and process audio emotional information, improving the accuracy and interpretability of emotion recognition. This makes it suitable for applications such as psychological counseling and emotion recognition.

[0040] In another implementation of the present invention, the prosodic features include duration-related features, fundamental frequency-related features, and energy-related features; the duration-related features include speech rate, short-time average zero-crossing rate, etc.; the fundamental frequency-related features include fundamental frequency and its mean, range of variation, rate of variation, and root mean square error, etc.; the energy-related features include short-time average energy, short-time average amplitude, etc.

[0041] For example, prosodic features are a commonly used feature type in speech recognition, involving changes in rhythm, intonation, amplitude, and frequency in a sentence.

[0042] In another implementation of the present invention, the sound quality characteristics include jitter, shimmer, harmonic noise ratio (HNR), etc.

[0043] For example, sound quality features are closely related to emotional state and have proven to be very effective in speech emotion recognition.

[0044] In another implementation of the present invention, the spectral features are a reflection of the correlation between changes in vocal tract shape and vocalization motion, including linear predictive cepstral coefficients (LPCC), frequency cepstral coefficients (MFCC), etc.

[0045] For example, spectral features are often used to supplement acoustic feature analysis, as different emotions exhibit different spectral morphologies. Important spectral features commonly used in emotion recognition include: Mel frequency cepstral coefficients (MFCC), linear prediction coefficients (LPC), linear prediction cepstral coefficients (LPCC), gamma-ton frequency cepstral coefficients (GFCC), perceptual linear prediction (PLP), and formants.

[0046] In another implementation of the present invention, the step of using data in the dataset to train a model and obtain an interpretable emotion recognition model includes: inputting interpretable acoustic features from the dataset into a pre-trained model to obtain continuous high-dimensional acoustic feature encoding; decomposing the continuous high-dimensional acoustic feature encoding into a series of discrete speech units; and using the speech units as input and emotion tags and interpretable emotion cues as output to train the model and obtain an interpretable emotion recognition model.

[0047] For example, large open-source models, such as GPT4 and Claude, are used to optimize audio emotion recognition tasks. To improve the model's performance in emotion recognition while ensuring high interpretability of its output, interpretable emotional cues and voice tokens are introduced for fine-tuning during model training, enabling the large model to effectively understand and process audio data.

[0048] Specifically, such as Figure 5 As shown, in the data preprocessing stage, the acoustic features are first extracted and tokenized. The acoustic features are used to generate a set of high-dimensional embeddings through the aforementioned pre-trained model, and then further transformed into a set of speech tokens. These tokens not only reflect the acoustic information of the audio, but also have the temporal structure required for emotion recognition, providing rich feature inputs for the subsequent training process.

[0049] During model fine-tuning, interpretable sentiment cues and voice tokens are used as inputs, and sentiment tags and interpretable sentiment cues are used as outputs to train the LLM. This ensures that the LLM not only incorporates sentiment categories during inference but also generates reasonable sentiment analysis to enhance the transparency of its decisions. Ultimately, the fine-tuned large model can achieve high accuracy and interpretable output in audio sentiment recognition.

[0050] It should be understood that while traditional emotion recognition models can output emotion categories, they cannot explain their decision-making process. To address this, this invention leverages the powerful natural language processing capabilities of large models (such as GPT-4 or LLaMA3) to design appropriate prompts that guide the model to perform emotion analysis in a way that is understandable to humans, generating corresponding emotion understanding descriptions. Furthermore, it incorporates the Chain of Thought (CoT) method to provide a detailed explanation of the emotion judgment process, enhancing the model's logical coherence in emotion analysis. The model not only outputs emotion categories but also generates emotion categories and detailed reasoning processes, achieving transparency in the emotion recognition decision-making process, solving the "black box" problem of emotion recognition, improving the interpretability of emotion recognition, helping psychological counselors understand and track clients' emotional changes, and providing evidence-based emotion descriptions to assist in personalized diagnosis.

[0051] In another implementation of the present invention, the loss function of the interpretable emotion recognition model includes the cross-entropy loss of emotion classification and the similarity loss of interpretability cues.

[0052] For example, during training, the model's ability to understand and recognize emotions is enhanced by jointly optimizing the loss function, enabling the model to effectively balance emotion classification and interpretable output. The loss function includes cross-entropy loss for emotion classification and similarity loss for interpretability cues. The latter is used to evaluate the similarity between the generated description and the expected interpretation, thereby improving the interpretability of the emotion analysis results. Through this dual-loss optimization strategy, the model gradually learns how to generate interpretable emotion output based on voice tokens. The model generated by this large audio model training module can accept audio token input and output category labels and corresponding interpretive descriptions during emotion classification, providing higher accuracy and transparency for emotion recognition tasks and meeting the high requirements for emotion analysis in application scenarios such as psychological counseling.

[0053] The present invention aims to make the emotion recognition process more transparent and accurate, thereby enhancing its practical value in applications such as psychological counseling.

[0054] In another aspect of the invention, an interpretable emotion recognition device based on a large model is provided, such as... Figure 2 As shown, it includes:

[0055] Feature extraction module: Extracts various basic acoustic features from raw speech data for emotion classification.

[0056] Feature processing module: Performs feature filtering on the basic acoustic features to obtain interpretive acoustic features closely related to specific emotions.

[0057] Data preparation module: Construct a dataset based on the specific emotion, the corresponding interpretive acoustic features, and the emotion category label.

[0058] Model training module: Use the data in the dataset to train the model and obtain an interpretable emotion recognition model.

[0059] Emotion Recognition Module: Audio emotion recognition is performed using the interpretable emotion recognition model to obtain complete emotion analysis results.

[0060] In the large-model-based interpretable emotion recognition device of this invention, a multi-level feature extraction mechanism is introduced. Combined with basic acoustic features, the model can more deeply understand complex emotional information, improve the overall emotion recognition effect, and increase the accuracy of recognition. By guiding the large model to generate emotion understanding descriptions, emotion recognition is not limited to simple classification results, but also includes detailed reasoning processes, which makes it easier for psychological counselors to understand the emotional state of clients. In addition, the use of the thought chain method enhances the logical coherence of emotion analysis, and the output includes not only emotion labels, but also reasoning processes, thereby significantly improving the interpretability of the model.

[0061] The present invention also provides an electronic device, which may include: a processor, a memory, a communication bus, and a communication interface.

[0062] in:

[0063] The processor, memory, and communication interface communicate with each other via a communication bus.

[0064] A communication interface is used to communicate with other electronic devices or servers.

[0065] The processor is used to execute programs, specifically the steps of any of the large-model-based interpretable emotion recognition methods described in the above embodiments.

[0066] Specifically, the program may include program code, which includes computer operation instructions.

[0067] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0068] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0069] Specifically, the program can be used to cause the processor to execute steps to implement any of the large-model-based interpretable emotion recognition methods described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed by any of the large-model-based interpretable emotion recognition methods described above, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0070] The present invention also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of the embodiments of the present application.

[0071] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0072] Specific embodiments of the invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing can be advantageous.

[0073] It should be noted that all directional indicators (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0074] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0076] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0077] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An interpretable sentiment recognition method based on a large model, characterized in that, include: Extracting various fundamental acoustic features for emotion classification from raw speech data; The basic acoustic features are subjected to feature filtering to obtain interpretive acoustic features closely related to specific emotions; A dataset is constructed based on the specific emotion, the corresponding interpretive acoustic features, and the emotion category label; Using the data in the aforementioned dataset to train a model, an interpretable emotion recognition model is obtained, including: The interpretable acoustic features from the dataset are input into the pre-trained model to obtain continuous high-dimensional acoustic feature encodings; The continuous high-dimensional acoustic feature encoding is decomposed into a series of discrete speech units; The speech unit is used as input, and the emotion label and interpretable emotion cue are used as output to train the model and obtain an interpretable emotion recognition model. The interpretable emotion cue is the emotion analysis description corresponding to the emotion category output by the interpretable emotion recognition model. The loss function of the interpretable emotion recognition model includes cross-entropy loss for emotion classification and similarity loss for interpretability cues; By performing audio emotion recognition using the aforementioned interpretable emotion recognition model, a complete emotion analysis result is obtained.

2. The method according to claim 1, characterized in that, The basic acoustic features include prosodic features, timbre features, and spectral features.

3. The method according to claim 2, characterized in that, The prosodic features include duration-related features, fundamental frequency-related features, and energy-related features; The duration-related features include speech rate and short-term average zero-crossing rate; The fundamental frequency-related features include the fundamental frequency and its mean, range of variation, rate of variation, and standard deviation; The energy-related features include short-time average energy and short-time average amplitude.

4. The method according to claim 2, characterized in that, The sound quality characteristics include time base error, amplitude perturbation, and harmonic noise ratio.

5. The method according to claim 2, characterized in that, The spectral characteristics reflect the correlation between changes in vocal tract shape and vocalization motion, including linear prediction cepstral coefficients and frequency cepstral coefficients.

6. An interpretable emotion recognition device based on a large model, characterized in that, include: Feature extraction module: Extracts various basic acoustic features from raw speech data for emotion classification; Feature processing module: Performs feature filtering on the basic acoustic features to obtain interpretive acoustic features closely related to specific emotions; Data preparation module: Construct a dataset based on the specific emotion, the corresponding interpretive acoustic features, and the emotion category label; Model training module: Uses the data in the dataset to train the model and obtain an interpretable emotion recognition model, including: inputting the interpretable acoustic features in the dataset into the pre-trained model to obtain continuous high-dimensional acoustic feature encoding; The continuous high-dimensional acoustic feature encoding is decomposed into a series of discrete speech units; the speech units are used as input, and emotion tags and interpretable emotion cues are used as output for model training to obtain an interpretable emotion recognition model. The interpretable emotion cues are the emotion analysis descriptions corresponding to the emotion categories output by the interpretable emotion recognition model; the loss function of the interpretable emotion recognition model includes the cross-entropy loss of emotion classification and the similarity loss of interpretable cues. Emotion Recognition Module: Audio emotion recognition is performed using the interpretable emotion recognition model to obtain complete emotion analysis results.

7. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of a large model-based interpretable emotion recognition method as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the interpretable emotion recognition method based on a large model as described in any one of claims 1 to 5.

Images