Accent simulation method and device, computer equipment and storage medium

By obtaining the fuzzy sound data set and the fusion feature vector set, embedding the mask weight value, and training the accent simulation model, the problem that language big models are difficult to imitate individual accents is solved, and a more accurate speech imitation effect is achieved.

CN119091853APending Publication Date: 2024-12-06SHENZHEN ANJINUO INTELLIGENT ROBOT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411230144.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2024-12-06

Smart Images

  • Figure CN119091853A_ABST
    Figure CN119091853A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an accent simulation method and device, computer equipment and a storage medium. The method comprises the steps that a standard text set is provided for an imitated person, an original voice set provided by the imitated person is acquired, and a fuzzy voice data set is acquired based on the standard text set and the original voice set; obtaining a standard voice set of the standard text set, and obtaining a fusion feature vector set based on the standard text set and the standard voice set; acquiring a mask weight value corresponding to each fusion feature vector in the fusion feature vector set based on the fuzzy sound data set, and embedding the mask weight values into the corresponding fusion feature vectors to generate training data; constructing an accent simulation model, inputting the training data into the accent simulation model for training, and obtaining a trained accent simulation model; and obtaining a target text, obtaining a target vector according to the target text, and inputting the target vector into the trained accent simulation model to obtain a target voice. According to the invention, accurate simulation of accent can be realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an accent simulation method, device, computer equipment and storage medium. Background Art

[0002] In recent years, with the advancement of science and technology, the field of large language models has made significant progress. Due to their powerful learning and information processing capabilities, these large language models have been widely used in the field of human-computer interaction. They can accurately handle ambiguous sound interference in input and realize speech-to-text functions.

[0003] Everyone's speech habits and accents are unique, and large language models often find it difficult to capture and accurately imitate these subtle differences. This results in the output of large language models that imitate the speaking habits of the person being imitated being inaccurate, and there is still room for improvement. Summary of the invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide an accent simulation method, apparatus, computer device and storage medium to solve the problem in the prior art that the speech output by the large language model that imitates the speaking habits of the person being imitated is not accurate enough.

[0005] The present invention discloses an accent simulation method, comprising:

[0006] Providing a standard text set to the person being imitated, obtaining an original voice set provided by the person being imitated based on the standard text set, and obtaining an ambiguous sound data set based on the standard text set and the original voice set;

[0007] Acquire a standard speech set corresponding to the standard text set, and acquire a fusion feature vector set based on the standard text set and the standard speech set;

[0008] Based on the fuzzy sound data set, a mask weight value corresponding to each fused feature vector in the fused feature vector set is obtained, and the mask weight value is embedded in the corresponding fused feature vector to generate training data;

[0009] Constructing an accent simulation model, inputting the training data into the accent simulation model for training, and obtaining a trained accent simulation model;

[0010] A target text is obtained, a target vector is obtained according to the target text, and the target vector is input into the trained accent simulation model to obtain a target speech.

[0011] Optionally, the step of acquiring the ambiguous sound data set based on the standard text set and the original speech set includes:

[0012] For each standard text in the standard text set, obtaining the corresponding original speech, performing speech recognition on the original speech, and obtaining the original text;

[0013] Comparing the standard text with the original text to obtain inconsistent fuzzy text and corresponding fuzzy speech;

[0014] Obtaining similarity levels corresponding to fuzzy sound types of the standard text and the fuzzy text;

[0015] For each of the fuzzy texts, all fuzzy sound types and similarity levels are counted, and based on all the fuzzy sound types and the similarity levels, a fuzzy sound degree value of the fuzzy text is obtained;

[0016] The more types of fuzzy sounds there are, the lower the fuzzy sound degree value is.

[0017] Optionally, the step of inputting the training data into the accent simulation model for training comprises:

[0018] Inputting the training data into the accent simulation model, comparing the corresponding original text with the output data of the accent simulation model, and obtaining a loss function value;

[0019] The model parameters of the accent simulation model are adjusted based on the loss function value until the loss function value meets the preset requirements.

[0020] Optionally, the step of acquiring a mask weight value corresponding to each fused feature vector in the fused feature vector set based on the fuzzy sound data set includes:

[0021] Obtain the fuzzy position of the fuzzy text corresponding to the fused feature vector;

[0022] Obtaining a mask weight value corresponding to the standard text based on the fuzzy sound degree value of the corresponding fuzzy text;

[0023] The mask weight values ​​are mapped to the blur locations.

[0024] Optionally, the step of acquiring a target vector according to the target text includes:

[0025] The target standard speech corresponding to the target text is obtained, and the target text and the target standard speech are fused into the target vector.

[0026] Optionally, the step of inputting the target vector into the trained accent simulation model to obtain the target speech comprises:

[0027] Obtain the ambiguous text of each word in the target text, generate an ambiguous speech text, and use the speech of the ambiguous speech text as the target speech.

[0028] Optionally, the step of using the speech of the ambiguous speech text as the target speech includes:

[0029] The fuzzy speech text is input into the timbre simulation model to obtain the target speech.

[0030] Optionally, the step of acquiring a fused feature vector set based on the standard text set and the original speech set includes:

[0031] Obtaining text features of each of the standard texts and speech features of the standard speech corresponding to each of the standard texts through a bidirectional LSTM;

[0032] The text feature and the speech feature are fused using an attention mechanism to obtain a fused feature vector.

[0033] The present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0034] The present invention also discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method described above.

[0035] Compared with the prior art, the beneficial effect of the accent simulation method provided by the embodiment of the present invention is that: the present invention obtains a fuzzy sound data set based on the standard text set and the original speech set provided by the person being imitated, obtains a fused feature vector set based on the standard text set and the corresponding standard speech set, obtains the mask weight value corresponding to each fused feature vector in the fused feature vector set based on the fuzzy speech data set, and embeds the mask weight value into the fused feature vector, which can guide the model to pay more attention to the features of accented pronunciation, thereby improving the ability to accurately imitate the accent. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The scheme of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments, in which:

[0037] Figure 1 is a flow chart of an embodiment of the accent simulation method provided by the present invention;

[0038] Figure 2 It is a schematic diagram of the flow of the accent simulation model for training provided by the present invention;

[0039] Figure 3This is a schematic diagram of the process of obtaining fuzzy speech text provided by the present invention

[0040] Figure 4 It is a structural schematic diagram of an embodiment of a computer device provided by the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. Now, in conjunction with the accompanying drawings, the preferred embodiments of the present invention are described in detail.

[0042] See also Figure 1 , Figure 1 1 is a flow chart of an embodiment of the accent simulation method provided by the present invention. The accent simulation method provided by the present invention comprises the following steps:

[0043] S101: providing a standard text set to the person being imitated, obtaining an original voice set provided by the person being imitated based on the standard text set, and obtaining a fuzzy sound data set based on the standard text set and the original voice set.

[0044] In a specific implementation scenario, a standard text set is provided to the person being imitated, the standard text set includes multiple standard texts, and the standard text includes at least one of a single word, a phrase, and a sentence. The person being imitated reads the standard texts in sequence, so that the original speech corresponding to each standard text can be collected, and the original speech corresponding to all the standard texts is used as the original speech set.

[0045] For the same standard text, the person being imitated can read it aloud multiple times, and collect multiple original voices. The multiple original voices can be screened and one of them can be selected as the original voice. The multiple original voices can also be averaged and integrated into one original voice.

[0046] Obtain a fuzzy sound data set based on a standard text set and an original voice set, obtain the standard text in the standard text set, and the original voice corresponding to the standard text in the original voice set, perform a matching operation on the standard text and the original voice, find the mismatching part between the two, and obtain the fuzzy sound data based on the mismatching part. Collect all the fuzzy sound data to obtain a fuzzy sound data set.

[0047] The original speech can be subjected to speech recognition to obtain the original text corresponding to the original speech. The original text is compared with the standard text. If the two are the same, there is no fuzzy sound data. If they are different, the parts where the two texts are inconsistent are obtained, and the fuzzy sound data is obtained based on the inconsistent parts. The fuzzy sound data includes the content of the inconsistent parts. For example, if the standard text is "one, two, three, four" and the original text is "one, two, mountain, four", the inconsistent parts are "three" and "mountain", that is, the third single character in this sentence has a fuzzy sound, while the other single characters do not have a fuzzy sound.

[0048] In other implementation scenarios, the standard text can also be subjected to speech conversion to obtain the corresponding standard speech. The standard speech is compared with the original speech. If the two are the same, there is no fuzzy sound data. If they are different, the parts where the two texts are inconsistent are obtained, and the fuzzy sound data is obtained based on the inconsistent parts. The fuzzy sound data includes the content of the inconsistent parts. For example, if the standard speech is "yi er san si" and the original speech is "yi er shan si", the inconsistent parts are "san" and "shan". Through such comparison, it can also be found that the third single character in this sentence has a fuzzy sound, while the other single characters do not have a fuzzy sound.

[0049] Generate fuzzy sound data according to the fuzzy position where the fuzzy sound is located. For example, record the fuzzy position or generate a corresponding vector. In one embodiment, the standard text is compared with the corresponding original text, and the fuzzy sound data is obtained based on the inconsistent fuzzy text and the corresponding fuzzy speech. Specifically, obtain the standard speech corresponding to the standard text and the fuzzy speech corresponding to the fuzzy text, and obtain the fuzzy sound type of the fuzzy text and the standard text based on the difference between the standard speech and the fuzzy speech. Each fuzzy sound type has a corresponding similarity level. For example, the fuzzy sound types can include flat and rolled tongue sounds, tones, and front and back nasal sounds. Different similarity levels are assigned to different fuzzy sound types. For example, the pronunciation is the same but the tones are different ("le" and "yue") can be set as fuzzy sound level 1, and the flat and rolled tongue and front and back nasal sounds are fuzzy ("zhi" and "zi") are set as fuzzy sound level 2.

[0050] It can be understood that the same fuzzy text can correspond to multiple fuzzy sound types. For example, for "zhang", there can be two fuzzy sound types at the same time: flat and rolled tongue and front and back nasal sounds are fuzzy: "zan". The lower the degree value of the fuzzy sound when there are more fuzzy sound types. If the fuzzy text and its corresponding standard text are similar in multiple fuzzy sound types (flat and rolled tongue sounds, tones, front and back nasal sounds, etc.), it means that these two texts are more similar in pronunciation characteristics.

[0051] Statistically analyze all the types and levels of fuzzy sounds in the fuzzy text compared to the standard text, and obtain the fuzzy sound degree value of the standard text based on all the types of fuzzy sounds and similar levels. An increase in the similar level indicates a higher similarity between the fuzzy text and the standard text, so the degree of fuzzy sound will be relatively lower. In other words, if the fuzzy text and the standard text are similar in multiple features, then the degree of fuzzy sound will be lower.

[0052] Obtain the fuzzy sound degree value of each single character in the standard text. If a single character has no corresponding fuzzy text, its fuzzy sound degree value is 0. If a single character has a corresponding fuzzy sound cost, obtain its fuzzy sound degree value. Obtain the fuzzy sound degree value sequence according to the arrangement order of the single characters in the standard text, and use the fuzzy sound degree value sequence as the fuzzy sound data of the standard text.

[0053] For example, if the standard text is "one two three four" and the original text is "one two mountain four", the inconsistent parts are "three" and "mountain". The difference between "san" and "shan" lies in the flat and curled tongues, so the fuzzy sound degree value can be set to 2, and the fuzzy sound degree values of the other single characters are 0. Therefore, the corresponding fuzzy data obtained is 0020.

[0054] S102: Obtain the standard speech set corresponding to the standard text set, and obtain the fused feature vector set based on the standard text set and the standard speech set.

[0055] In a specific implementation scenario, obtain the standard text in the standard text set, and obtain the corresponding standard speech for each standard text. Input the standard text and the corresponding standard speech into the neural network for extraction to obtain the fused feature vector. Take all the fused feature vectors as the fused feature vector set.

[0056] For example, the standard text and the standard audio can be encoded into vector formats and input into a bidirectional long short-term memory network to obtain the text vector of the standard text and the speech vector of the standard speech respectively. The bidirectional long short-term memory network (BiLSTM) is widely used in tasks such as speech recognition and natural language processing. It can learn complex feature representations in the input sequence and helps to extract more meaningful features. Fuse these two feature vectors based on the attention mechanism to obtain the fused feature vector, which contains the information of the standard text and the corresponding standard audio.

[0057] S103: Based on the fuzzy sound data set, obtain the mask weight value corresponding to each fused feature vector in the fused feature vector set, and embed the mask weight value into the corresponding fused feature vector to generate the training data.

[0058] In a specific implementation scenario, for each fuzzy sound data in the fuzzy sound dataset, obtain the corresponding fusion feature vector in the fusion feature vector set, and assign a matching mask weight value to the corresponding fusion feature vector based on the fuzzy information of the fuzzy sound data. In this implementation scenario, the mask weight value can be obtained based on the fuzziness value of the fuzzy sound data. The higher the fuzziness value of the fuzzy sound data, the greater the mask weight value. The specific mapping relationship between the two values can be adjusted according to the actual application scenario, which can be a linear relationship or a non-linear relationship.

[0059] In an implementation scenario, obtain the fuzzy position of the fuzzy text corresponding to the fusion feature vector. For example, in the above-mentioned embodiment, the standard text is "one two three four", the original text is "one two mountain four", then the fuzzy text is "mountain", the obtained fuzzy sound data is 0020, the fuzziness value corresponding to the fuzzy text is 2, and the corresponding mask weight value is 4. The remaining single characters have no fuzzy text, that is, the fuzziness value is 0, and the corresponding mask weight value is 1. Map the mask weight value to the corresponding fuzzy position, that is, the position corresponding to the third single character. The values in this embodiment are only examples, and no specific limitations are imposed on the specific values and mathematical operation relationships of the fuzziness value and the mask weight value. Embed the mask weight value into the fusion feature vector. For example, it can be achieved by multiplying the unit corresponding to the fuzzy position in the fusion feature vector by the mask weight value. In this way, when training the accent simulation model, the accent simulation model will learn according to the input fusion feature vector and the corresponding mask weight value to optimize the performance of the model when processing fuzzy text.

[0060] S104: Construct an accent simulation model, input the training data into the accent simulation model for training, and obtain the trained accent simulation model.

[0061] In a specific implementation scenario, construct an accent simulation model, input the training data into the accent simulation model for training, and use a fully connected layer and an activation function in the output layer of the accent simulation model to output the character with the highest probability for each speech position. The fully connected layer can map the input features to the output space, and the activation function introduces non-linearity to help the model learn complex patterns. The model learns through training how to predict the most likely character at each speech position based on the input feature vector. The output of the model is the probability distribution of the characters at each position, and the character with the highest probability is selected as the prediction result.

[0062] In an accent simulation model, the role of the embedding layer is to convert discrete data (such as words) into dense vector representations so that the neural network can better process this data. The embedding layer usually contains two parts: the weighted sum of the word embeddings of the words at the corresponding positions in the original text and the word embeddings of the masks. The purpose of the weighted sum is to obtain a more comprehensive representation, where the weights of different parts can control the importance of different word embeddings. In this way, the model can better combine the words and mask information in the original text to obtain a richer representation.

[0063] In this embodiment, after inputting the training data into the accent simulation model, at least one fuzzy text corresponding to the currently input at least one training data is used as the output target, the output of the current accent simulation model is compared with the output target to obtain the loss function value, and the parameters of the accent simulation model are adjusted based on the loss function value.

[0064] Verify the accent simulation model with adjusted parameters, input at least one training data, and continue to compare the output of the accent simulation model with the corresponding at least one fuzzy text to obtain the loss function value. Determine whether the loss function value meets the preset requirements. If not, continue to repeat the above steps to adjust the parameters of the accent simulation model based on the loss function value until the loss function value meets the preset requirements. At this time, the training of the accent simulation model is completed. Please refer to Figure 2 , Figure 2 which is a schematic flowchart of the process for training the accent simulation model provided by the present invention.

[0065] S105: Obtain the target text, obtain the target vector according to the target text, and input the target vector into the trained accent simulation model to obtain the target speech.

[0066] In a specific implementation scenario, obtain the target text, obtain the target vector according to the target text, and input the target vector into the trained accent simulation model. According to the steps described above, the most likely fuzzy characters corresponding to each single character in the target text will be obtained. When there are no fuzzy characters for some single characters, the characters themselves are used as the fuzzy characters. Combine all the fuzzy characters into a fuzzy speech text, and use the speech of the fuzzy speech text as the target speech. For example, if the target text is "three hundred and thirty-three", the obtained fuzzy speech text is "shan bai shan shishan", and play the speech of "shan bai shan shishan" as the target speech. Please refer to Figure 3 , Figure 3 which is a schematic flowchart of the process for obtaining the fuzzy speech text provided by the present invention.

[0067] In other embodiments, the target text to be converted into speech is input into the selected speech synthesis engine through speech synthesis technology, and the target standard speech corresponding to the target text is generated by the speech synthesis engine. The target text and the target standard speech are merged into a target vector, specifically, a text representation vector corresponding to the target text and a speech representation vector corresponding to the standard speech are obtained, and the text representation vector and the speech representation vector are connected or stacked together in a certain way to form an input feature vector, and the input feature vector is used as the target vector.

[0068] In other implementation scenarios, after obtaining the fuzzy speech text, the fuzzy speech text is input into a pre-trained timbre simulation model to obtain the target speech. This ensures that the output target speech can retain the user's accent while having the user's timbre characteristics. Specifically, processing accent simulation and timbre simulation separately helps to focus on different tasks. Accent simulation focuses on the pronunciation and intonation of speech, while timbre simulation focuses on the texture and characteristics of the sound. Separate processing allows each model to focus more on its specific task. By processing the audio at different levels, a more realistic and natural audio effect can be produced.

[0069] If a model is used to directly output audio, the model may ignore the important features of the accent because it focuses on the timbre. The model may then have learning bias and fail to fully learn multiple aspects of speech during training, resulting in the generated audio being similar in timbre but with significant differences in other aspects such as accent. It may also reduce the generalization ability of the model, causing it to perform poorly when dealing with different speech features or different speakers.

[0070] From the above description, it can be seen that in this embodiment, a fuzzy sound data set is obtained based on the standard text set and the original speech set provided by the imitated person, a fused feature vector set is obtained based on the standard text set and the corresponding standard speech set, and a mask weight value corresponding to each fused feature vector in the fused feature vector set is obtained based on the fuzzy speech data set. The mask weight value is embedded in the fused feature vector, which can guide the model to pay more attention to the features of accented pronunciation, thereby improving the ability to accurately imitate the accent.

[0071] Figure 4 FIG. 1 shows a schematic diagram of the internal structure of a computer device in one embodiment. Figure 3As shown, the computer device includes a processor, a memory and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor may implement the accent simulation method. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor may implement the accent simulation method. Those skilled in the art will understand that Figure 4 The structure shown in the figure is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the electronic device to which the scheme of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0072] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the above steps.

[0073] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor executes the above steps.

[0074] The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROMElectrically Erasable Programmable Read-Only Memory), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory may be a disk memory or a tape memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), direct memory bus random access memory (DRRAM). The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0075] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0076] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0077] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0078] Those skilled in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical disks, and other media that can store program codes.

[0079] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0080] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0081] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0082] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0083] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. An accent simulation method, characterized in that: include: Providing a standard text set to the person being imitated, obtaining an original voice set provided by the person being imitated based on the standard text set, and obtaining an ambiguous sound data set based on the standard text set and the original voice set; Acquire a standard speech set corresponding to the standard text set, and acquire a fusion feature vector set based on the standard text set and the standard speech set; Based on the fuzzy sound data set, a mask weight value corresponding to each fused feature vector in the fused feature vector set is obtained, and the mask weight value is embedded in the corresponding fused feature vector to generate training data; Constructing an accent simulation model, inputting the training data into the accent simulation model for training, and obtaining a trained accent simulation model; A target text is obtained, a target vector is obtained according to the target text, and the target vector is input into the trained accent simulation model to obtain a target speech.

2. The accent simulation method according to claim 1, characterized in that: The step of obtaining the ambiguous sound data set based on the standard text set and the original speech set comprises: For each standard text in the standard text set, obtaining the corresponding original speech, performing speech recognition on the original speech, and obtaining the original text; Comparing the standard text with the original text to obtain inconsistent fuzzy text and corresponding fuzzy speech; Obtaining similarity levels corresponding to fuzzy sound types of the standard text and the fuzzy text; For each of the fuzzy texts, all fuzzy sound types and similarity levels are counted, and based on all the fuzzy sound types and the similarity levels, a fuzzy sound degree value of the fuzzy text is obtained; The more types of fuzzy sounds there are, the lower the fuzzy sound degree value is.

3. The accent simulation method according to claim 2, characterized in that: The step of inputting the training data into the accent simulation model for training comprises: Inputting the training data into the accent simulation model, comparing the corresponding original text with the output data of the accent simulation model, and obtaining a loss function value; The model parameters of the accent simulation model are adjusted based on the loss function value until the loss function value meets the preset requirements.

4. The accent simulation method according to claim 2, characterized in that: The step of obtaining the mask weight value corresponding to each fused feature vector in the fused feature vector set based on the fuzzy sound data set includes: Obtain the fuzzy position of the fuzzy text corresponding to the fused feature vector; Obtaining a mask weight value corresponding to the standard text based on the fuzzy sound degree value of the corresponding fuzzy text; The mask weight values ​​are mapped to the blur locations.

5. The accent simulation method according to claim 1, characterized in that: The step of obtaining a target vector according to the target text comprises: The target standard speech corresponding to the target text is obtained, and the target text and the target standard speech are fused into the target vector.

6. The accent simulation method according to claim 5, characterized in that: The step of inputting the target vector into the trained accent simulation model to obtain the target speech comprises: Obtain the ambiguous text of each word in the target text, generate an ambiguous speech text, and use the speech of the ambiguous speech text as the target speech.

7. The accent simulation method according to claim 6, characterized in that: The step of using the speech of the ambiguous speech text as the target speech comprises: The fuzzy speech text is input into the timbre simulation model to obtain the target speech.

8. The accent simulation method according to claim 2, characterized in that: The step of obtaining a fused feature vector set based on the standard text set and the original speech set comprises: Obtaining text features of each of the standard texts and speech features of the standard speech corresponding to each of the standard texts through a bidirectional LSTM; The text feature and the speech feature are fused using an attention mechanism to obtain a fused feature vector.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.