A data output method and system based on timbre and emotion simulation

Abstract

The application discloses a data output method and system based on timbre and emotion simulation, relates to the technical field of data processing, and comprises the following steps: receiving role-playing data and timbre data, preprocessing the role-playing data and the timbre data, obtaining a role-playing data set and a timbre data set; inputting the role-playing data set into a pre-established llama3 model for training, outputting a role-playing model, inputting the timbre data set into a pre-established GPT-SoVITS model for training, and outputting a timbre model; obtaining user real-time conversation voice data, carrying out noise removal processing on the user real-time conversation voice data, converting phonemes and Chinese characters of the processed user real-time conversation voice data based on a CTC greedy search algorithm, and obtaining recognized text; inputting the recognized text into a pre-established word segmentation model, obtaining a parameter file by adopting a byte pair encoding method through a subword marking strategy, slicing the recognized text, inputting the sliced text into the llama3 model, combining the parameter file for prediction, and outputting predicted text; inputting the predicted text into the timbre model, obtaining voice output results, and returning the voice output results to the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a data output method and system based on timbre and emotion simulation. BACKGROUND

[0002] Existing emotional companionship and timbre simulation technologies are mostly limited to general dialogue or simple interaction, and cannot fully adapt to the needs in specific situations, thus having certain deficiencies in meeting diversified needs. SUMMARY

[0003] To solve the deficiencies mentioned in the background, the purpose of the present application is to provide a data output method and system based on timbre and emotion simulation.

[0004] In a first aspect, the purpose of the present application can be achieved by the following technical solution: a data output method based on timbre and emotion simulation, the method comprising the following steps:

[0005] Receiving role-playing data and timbre data, preprocessing the role-playing data and timbre data to obtain a role-playing data set and a timbre data set;

[0006] Inputting the role-playing data set into a pre-established llama3 model for training to obtain a role-playing model, and inputting the timbre data set into a pre-established GPT-SoVITS model for training to obtain a timbre model;

[0007] Obtaining user real-time dialogue voice data, performing noise removal processing on the user real-time dialogue voice data, converting phonemes and Chinese characters of the processed user real-time dialogue voice data based on a CTC greedy search algorithm to obtain recognized text;

[0008] Inputting the recognized text into a pre-established word segmentation model, obtaining a parameter file by adopting a byte pair encoding method through a subword marking strategy, slicing the recognized text, inputting the sliced text into the llama3 model, and combining the parameter file to perform prediction to output predicted text;

[0009] Inputting the predicted text into the timbre model to obtain a voice output result, and returning the voice output result to the user.

[0010] In combination with the first aspect, in some implementations of the first aspect, the method further comprises: the preprocessing process of the role-playing data is as follows:

[0011] Selecting a personality image, personality, and speaking manner to be simulated, and customizing AI personality and speaking according to a template;

[0012] According to the initialization setting of the user, the simulation is performed, the role identity is obtained, and the character and dialogue mode are used to perform dialogue with the customer.

[0013] In combination with the first aspect, in some implementations of the first aspect, the method further includes that the pre-established llama3 model is as follows:

[0014] y = sigma (w2 * ReLu (w1 * x + b1) + b2)

[0015] In the formula, sigma (z) is an activation function of an output layer, x is an input feature vector, ReLu (z) is a nonlinear activation function, w1 and w2 are weight matrices, and b1 and b2 are bias vectors.

[0016] Role data input preprocessing:

[0017] X' = Normalize (X) -> Tokenize (X') -> Embedding (x')

[0018] Normalize (X) : normalization processing, mapping the feature value to [0, 1], Tokenize (X') : cutting the text input into sub-word units; Embedding (X') : using an embedding layer to convert discrete sub-word units into continuous vector representation;

[0019] Training and role generation:

[0020] The model is trained by supervising the role characteristic label Y, and the minimum loss function is obtained:

[0021]

[0022] L is a cross-entropy loss function; y i is a real label; is a model prediction output;

[0023] The final output of the model is a customized vector for the role characteristic: R = LLM (X')

[0024] R is the role characteristic output vector.

[0025] In combination with the first aspect, in some implementations of the first aspect, the method further includes that the timbre data is preprocessed as follows:

[0026] The timbre data is used as material, and the material is first subjected to UVR5 human voice separation and dereverberation, and then subjected to data noise reduction, screening, slicing preprocessing operation by using a speech segmentation tool to obtain a data set, resampling processing is performed on the data set by using a pre-established GPT-SoVITS model to generate a feature file, and then training is performed.

[0027] According to the obtained loss rate curve image control progress, the model is prevented from overfitting.

[0028] In combination with the first aspect, in some implementations of the first aspect, the method further comprises that the pre-established GPT-SoVITS model is expressed as follows:

[0029]

[0030] x is an input timbre data feature vector; w1 and b1 are weights and biases of an encoder; g(z) is a nonlinear transformation layer; h(z) is a convolution layer and a self-attention mechanism; w2 and b2 are weights and biases of a decoder; z is a random noise vector, is a timbre file.

[0031] In combination with the first aspect, in some implementations of the first aspect, the method further comprises that when the user processes the real-time dialogue voice data, a certain noise removal processing is first performed, the real-time processed voice is imported into the trained model file, offline automatic speech recognition (ASR) is used, and an end-to-end voice processing technology is used to process in combination with the characteristics of a convolutional neural network (CNN) and a recurrent neural network (RNN).

[0032] In combination with the first aspect, in some implementations of the first aspect, the method further comprises that the algorithm process of the CTC greedy search algorithm is as follows:

[0033] selecting an output label y with the maximum probability from a prediction distribution of each time frame t :

[0034]

[0035] integrating the labels into a whole sequence, deleting repeated labels in the sequence, removing CTC blank placeholders in the sequence, and converting the sequence into text.

[0036] Secondly, in order to achieve the above object, the application discloses a data output system based on timbre and emotion simulation, comprising:

[0037] a data processing module, configured to receive role-playing data and timbre data, pre-process the role-playing data and the timbre data, and obtain a role-playing data set and a timbre data set;

[0038] a model training module, configured to input the role-playing data set into a pre-established llama3 model for training, output a role-playing model, input the timbre data set into a pre-established GPT-SoVITS model for training, and output a timbre model;

[0039] The text generation module is configured to obtain real-time dialogue voice data of a user, perform noise removal on the real-time dialogue voice data of the user, convert phonemes and Chinese characters of the processed real-time dialogue voice data of the user based on a CTC greedy search algorithm, and obtain recognized text;

[0040] The text processing module is configured to input the recognized text into a pre-established word segmentation model, obtain a parameter file by using a byte pair encoding method and a subword tagging strategy, slice the recognized text, input the sliced text into a llama3 model, and perform prediction on the sliced text in combination with the parameter file to output predicted text.

[0041] The voice output module is configured to input the predicted text into a timbre model to obtain voice output results and return the voice output results to the user.

[0042] In another aspect of the present application, in order to achieve the above-mentioned purpose, a terminal device is disclosed, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, a data output method based on timbre and emotion simulation as described above is adopted.

[0043] In still another aspect of the present application, in order to achieve the above-mentioned purpose, a computer readable storage medium is disclosed, which stores a computer program, and when the computer program is loaded and executed by a processor, a data output method based on timbre and emotion simulation as described above is adopted.

[0044] The present application has the following beneficial effects:

[0045] The present application realizes the process of sum data output of timbre and emotion simulation by innovative role simulation and timbre generation technology, thereby meeting the diversified needs of users. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings;

[0047] Figure 1 is a method flowchart of the present application;

[0048] Figure 2 is a timbre simulation and space-time companion model flowchart of the present application;

[0049] Figure 3is a role model pre-training process schematic diagram of the present application;

[0050] Figure 4 is a system structure schematic diagram of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0052] Embodiment one:

[0053] As shown in Figure 1 A data output method based on timbre and emotion simulation, the method comprising the following steps:

[0054] S101: receiving role-playing data and timbre data, and preprocessing the role-playing data and the timbre data to obtain a role-playing data set and a timbre data set;

[0055] The preprocessing process of the role-playing data is as follows:

[0056] Specifically, first, a personality setting template (fixed options - including common personality characteristics and image characteristics, and specific content can also be added) is provided for the customer, and the user selects the personality image, personality, and speaking manner to be simulated. The user first customizes the AI personality and speaking style according to the template. More information can be input to enrich the AI settings.

[0057] The model will first simulate according to the user's initial settings, first understand the identity to be played, and then dialogue with the customer according to the personality and dialogue manner. The user can add corresponding events at any time to improve the model's understanding of the person to be played and increase the returnable output content.

[0058] The present step is based on llama3 model, and the role is constructed step by step. The present step uses five attributes in psychology to preliminarily measure the character of the role, and scores the five attributes of emotionalization, agreeableness, creativity, extroversion and responsibility. The model role is constructed by updating the role attributes through the initial attributes and data training model, and through the model interview question evaluation method. The ability of the model role playing is improved by adjusting and improving the model prompt words. The model will first simulate according to the user's initialization setting, first understand the identity to be played, and then have a dialogue with the customer according to the character and dialogue mode. The user can increase the corresponding events at any time to improve the model's understanding of the person to be played, and increase the returnable output content.

[0059] The pre-established llama3 model is as follows:

[0060] 1. Model formula

[0061] y = σ (w2·ReLu (w1·x+b1) +b2)

[0062] σ(z): activation function of output layer. Sigmoid is used in the application to normalize the output value for classification.

[0063] x: input feature vector, representing preprocessed role-playing data (personality label, dialogue style) or timbre data (spectrum feature).

[0064] ReLU(z): nonlinear activation function, used to introduce nonlinear features. In the application, negative values are set to 0, and positive values are z, to avoid gradient disappearance and improve the fitting effect of complex emotional data.

[0065] w1, w2: weight matrix, used for linear transformation of features.

[0066] w1 is responsible for projecting high-dimensional input x (text file in the application) to the dimension of the hidden layer, and adjusting the capacity of the hidden layer by controlling its size;

[0067] w2 is used to map the features of the hidden layer to the target dimension of the output layer. In the application, the features are emotion labels and role language style vectors.

[0068] b1, b2: bias vector, used to improve the expression ability of the model and avoid excessive dependence on the origin. In the application, it better captures the small differences in emotional simulation, such as different emotional expressions of the same timbre.

[0069] 2. Role data input preprocessing

[0070] The input role-playing data X is a data vector containing multiple features, which is processed by the following steps:

[0071] X' = Normalize(X) → Tokenize(X') → Embedding(x')

[0072] Normalize(X): normalization processing, mapping feature values to [0, 1].

[0073] Tokenize(X'): cutting the text input into sub-word units to facilitate model processing of context.

[0074] Embedding(X'): using an embedding layer to convert discrete sub-word units into continuous vector representations.

[0075] 3. Training and Role Generation

[0076] The model is trained with the supervision of role characteristic labels Y, obtaining the minimum loss function:

[0077]

[0078] L: cross-entropy loss function; y i : true label; Model prediction output.

[0079] The final output of the model is a customized vector for role characteristics: R = LLM(X')

[0080] R: role characteristic output vector, used as input for subsequent dialogue generation.

[0081] 4. Personalized Customization Process

[0082] The model generates role characteristics based on the initial settings of user customization, dynamically adjusts the weight distribution based on input, and finally generates a role model that adapts to user needs. Specifically,

[0083] Application Case:

[0084] 1. Emotional Accompaniment File for Bereaved Families

[0085] For families who have lost loved ones, an emotional accompaniment file is constructed to provide long-term psychological support.

[0086] Case 1: After losing a loved one, users can upload past voice recordings, photos, and other materials to generate a "loved one replacement personality." Users can recall shared life episodes through voice interaction, gradually establishing positive emotional connections and reducing the risk of depression.

[0087] Case 2: Elderly people feel lonely due to the geographical separation of their children. By recording and customizing a file of parental care, the system regularly reminds users of important family dates and events related to shared memories.

[0088] S102: input the role-playing dataset into the pre-established llama3 model for training, output the role-playing model, input the timbre dataset into the pre-established GPT-SoVITS model for training, and output the timbre model;

[0089] The process of preprocessing the timbre data is as follows:

[0090] The timbre data is used as material. The material is first subjected to UVR5 human voice separation and dereverberation, and then subjected to data noise reduction, screening and slicing preprocessing operations using a speech segmentation tool to obtain a dataset. The pre-established GPT-SoVITS model is used to resample and process the dataset to generate a feature file, and then the model is trained.

[0091] The progress is controlled according to the obtained loss rate curve image to prevent model overfitting.

[0092] The expression of the pre-established GPT-SoVITS model is as follows:

[0093] 1. Timbre data preprocessing

[0094] The UVR5 separation model and the speech segmentation tool are used for combined processing to achieve noise reduction, segmentation and resampling.

[0095] 1. Model formula

[0096]

[0097] x: input timbre data feature vector, representing an audio segment after preprocessing (such as noise reduction, slicing, etc.).

[0098] w1, b1: weights and biases of the encoder (Encoder) for extracting timbre feature representation.

[0099] g(z): nonlinear transformation layer (LeakyReLU is used in the present application to avoid negative values and improve convergence performance) for mapping timbre features to a hidden space.

[0100] h(z): convolutional layer (RNN+BiLSTM) with self-attention mechanism. The convolutional layer is used to model the temporal relationship between audio segments and capture short-term audio characteristics. The self-attention mechanism is used to capture long-distance dependencies.

[0101] w2, b2: weights and biases of the decoder (Decoder) for converting hidden space features to feature representations of target audio segments.

[0102] z: random noise vector added in the present application to enhance the diversity of generated timbres.

[0103] Timbre file.

[0104] S103: Obtain user real-time dialogue voice data, perform noise removal processing on the user real-time dialogue voice data, convert phonemes and Chinese characters based on a CTC greedy search algorithm on the processed user real-time dialogue voice data, and obtain recognized text;

[0105] When the user real-time dialogue voice data is processed, a certain noise removal processing is performed first, the processed real-time voice is imported into a trained model file, offline automatic speech recognition (ASR) is used, and an end-to-end voice processing technology is used to process in combination with the characteristics of a convolutional neural network (CNN) and a recurrent neural network (RNN).

[0106] The algorithm process of the CTC greedy search algorithm is as follows:

[0107] 1. Select the output label yt with the maximum probability from the prediction distribution of each time frame.

[0108]

[0109] Integrate the labels into a whole sequence.

[0110] 2. Delete the repeated labels in the sequence and remove the CTC blank placeholders in the sequence.

[0111] 3. Convert the sequence into text.

[0112] S104: Input the recognized text into a pre-established word segmentation model, obtain a parameter file by using a byte pair encoding method through a subword tagging strategy, slice the recognized text, input the sliced text into a llama3 model, and perform prediction in combination with the parameter file to output a predicted text.

[0113] S105: Input the predicted text into a voice model to obtain a voice output result, and return the voice output result to the user.

[0114] The client is in the form of software, API, and hardware, and is a medium for the user to interact with the method.

[0115] Specifically, the following embodiments are used to further illustrate the method:

[0116] The method can be applied to:

[0117] 1. Emotional marketing in customer service

[0118] In the service industry, the service quality is improved through emotional simulation technology, and a highly personalized interactive experience is provided for customers. The method can be applied to virtual customer service to simulate brand image or familiar style for dialogue.

[0119] 2. Educational assistance for knowledge interaction and personalized companionship

[0120] The method can simulate familiar role images (such as teachers and study partners) for students and provide them with customized learning companionship services. The user inputs the student's subject needs and learning goals, and sets the student's learning style (visual / auditory, etc.) as role data input. The method generates emotional voice interaction content, including answering questions, simulating classroom teaching, or providing emotional encouragement. The model dynamically adjusts the teaching progress and feedback tone by recognizing the student's voice input questions or emotional state, forms a learning companion profile, and records the student's growth trajectory and learning state.

[0121] Embodiment Two: In the second aspect, as shown in Figure 4 To achieve the above purpose, the present application discloses a data output system based on timbre and emotion simulation, comprising:

[0122] The data processing module 11 is used for receiving role-playing data and timbre data, preprocessing the role-playing data and timbre data, obtaining a role-playing data set and a timbre data set;

[0123] The model training module 12 is used for inputting the role-playing data set into a pre-established llama3 model for training, outputting a role-playing model, inputting the timbre data set into a pre-established GPT-SoVITS model for training, and outputting a timbre model;

[0124] The text generation module 13 is used for acquiring user real-time dialogue voice data, de-noising the user real-time dialogue voice data, converting phonemes and Chinese characters of the processed user real-time dialogue voice data based on a CTC greedy search algorithm, and obtaining recognized text;

[0125] The text processing module 14 is used for inputting the recognized text into a pre-established word segmentation model, obtaining a parameter file by using a byte pair encoding method through a sub-word marking strategy, slicing the recognized text, inputting the sliced text into the llama3 model, combining the parameter file for prediction, and outputting predicted text;

[0126] The voice output module 15 is used for inputting the predicted text into the timbre model to obtain voice output results, and returning the voice output results to the user.

[0127] Based on the same inventive concept, the present application further provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the program comprises program instructions, and the processor is configured to execute the program instructions stored in the memory. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are configured to implement one or more instructions, and are specifically configured to load and execute one or more instructions in the computer storage medium to implement the above method.

[0128] It needs to be further explained that, based on the same inventive concept, the present application further provides a computer storage medium, which stores a computer program, and the computer program is executed by the processor to perform the above method. The storage medium can adopt any combination of one or more computer readable media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium may, for example, be but is not limited to an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples (non-exhaustive list) of the computer readable storage medium include: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0129] In the description of the present application, the description of the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0130] The foregoing presents and describes the basic principles, main features and advantages of the present disclosure. It should be understood by those skilled in the art that the present disclosure is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, various changes and improvements can be made to the present disclosure, and all these changes and improvements fall within the scope of the present disclosure.

Claims

1. A data output method based on timbre and emotion simulation, characterized in that, The method includes the following steps: Receive role-playing data and timbre data, preprocess the role-playing data and timbre data to obtain role-playing dataset and timbre dataset; The role-playing dataset is input into the pre-built llama3 model for training and fine-tuning, and the output is the role-playing model. The timbre dataset is input into the pre-built GPT-SoVITS model for training, and the output is the timbre model. The system acquires real-time user dialogue voice data, processes the data, and uses the CTC greedy search algorithm to convert phonemes into Chinese characters to obtain the recognized text. The identified text is input into a pre-built word segmentation model. Through a sub-word tagging strategy and a byte-pair encoding method, a parameter file is obtained. The identified text is then sliced, and the sliced ​​text is input into the llama3 model. The model is then combined with the parameter file for prediction, and the predicted text is output. The predicted text is input into the timbre model to obtain the speech output, which is then returned to the user.

2. The data output method based on timbre and emotion simulation according to claim 1, characterized in that, The preprocessing procedure for the role-playing data is as follows: Select the personality, character, and speaking style you want to simulate, and customize the AI's personality and speech based on the template; The simulation is performed based on the user's initial settings to obtain the role, and the user engages in conversation with the customer according to their personality and dialogue style.

3. The data output method based on timbre and emotion simulation according to claim 1, characterized in that, The pre-established llama3 model is as follows: y = σ(w2·PeLu(w1·x+b1)+b2) In the formula, σ(z) is the activation function of the output layer, x is the input feature vector; ReLU(z) is the nonlinear activation function; w1, w2 are weight matrices; b1, b2 are bias vectors; Character data input preprocessing: X ′ =Normalize(X)→Tokenize(X ′ )→Embedding(x ′ ) Normalize(X): Normalizes the text input, mapping the feature values ​​to [0,1]. Tokenize(X'): Divides the text input into sub-word units. Embedding(X'): Uses an embedding layer to convert discrete sub-word units into continuous vector representations. Training and character generation: The model is trained under supervision using the character attribute label Y, and obtains the loss function that minimizes the loss function: L is the cross-entropy loss function; y i This is a real label; The model predicts the output; The model's final output is a customized vector based on the character's characteristics: R = LLM(X') R is the character trait output vector.

4. The data output method based on timbre and emotion simulation according to claim 1, characterized in that, The process of preprocessing the timbre data: Using timbre data as material, the material is first subjected to UVR5 voice separation and dereverb, and then preprocessed by using speech segmentation tools for data noise reduction, filtering, and slicing to obtain a dataset. The dataset is then resampled using a pre-built GPT-SoVITS model to generate feature files, and then trained. The progress is controlled based on the obtained loss rate curve image to prevent model overfitting.

5. The data output method based on timbre and emotion simulation according to claim 4, characterized in that, The pre-established GPT-SoVITS model expression is as follows: x is the input timbre data feature vector; w1 and b1 are the encoder weights and biases; g(z) is the nonlinear transform layer; h(z) is the convolutional layer with self-attention mechanism; w2 and b2 are the decoder weights and biases; z is a random noise vector. This is a sound file.

6. The data output method based on timbre and emotion simulation according to claim 1, characterized in that, The process of processing user real-time dialogue voice data: First, noise reduction is performed. The real-time processed speech is imported into the trained model file, and offline automatic speech recognition (ASR) is used. End-to-end speech processing technology is used, combining the characteristics of convolutional neural networks (CNN) and recurrent neural networks (RNN) for processing.

7. The data output method based on timbre and emotion simulation according to claim 1, characterized in that, The algorithm process of the CTC greedy search algorithm is as follows: Select the output label y with the highest probability from the prediction distribution of each time frame. t : The tags are integrated into a single sequence, duplicate tags are removed from the sequence, CTC whitespace placeholders are removed from the sequence, and the sequence is converted into text.

8. A data output system based on timbre and emotion simulation, characterized in that, include: The data processing module is used to receive role-playing data and timbre data, preprocess the role-playing data and timbre data, and obtain role-playing dataset and timbre dataset. The model training module is used to input the role-playing dataset into the pre-built llama3 model for training and output the role-playing model. The timbre dataset is input into the pre-built GPT-SoVITS model for training and output the timbre model. The text generation module is used to acquire real-time user dialogue voice data, remove noise from the real-time user dialogue voice data, and convert the processed real-time user dialogue voice data into phonemes and Chinese characters based on the CTC greedy search algorithm to obtain the recognized text. The text processing module is used to input the recognized text into a pre-established word segmentation model, obtain a parameter file by using a sub-word tagging strategy and byte pair encoding, slice the recognized text, input the sliced ​​text into the llama3 model, perform prediction based on the parameter file, and output the predicted text. The speech output module is used to input the predicted text into the timbre model, obtain the speech output result, and return the speech output result to the user.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs a data output method based on timbre and emotion simulation as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs a data output method based on timbre and emotion simulation as described in any one of claims 1 to 7.

Images