Style speech synthesis method and system based on large model, and storage medium
By fine-tuning parameter and style interpolation of style embedding features of pre-trained large models, the problem of insufficient diversity of speech synthesis styles in the prior art is solved, and more flexible and diverse speech style control is achieved.
Patent Information
- Application Number
- CN202411754362.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-02
- Publication Date
- 2025-02-28
AI Technical Summary
In the existing pronunciation synthesis technology, the pronunciation model can only synthesize a single style of pronunciation, resulting in poor diversity in pronunciation styles.
The style speech synthesis method based on the big model is adopted, and the pre-trained big model is fine-tuned and the speaker characteristics of the style reference speech are extracted to obtain the style embedded features. Then, style interpolate the style embedding features according to the style speech synthesis requirements, adjust the style level, and finally, the model to be synthesized after inputting the text to be synthesized and the style expected features are fine-tuned for speech synthesis.
The flexibility of adjusting the voice style according to user needs is achieved, the diversity of voice synthesis styles is improved, and the target style speech with the desired style or emotions can be generated.
Smart Images
Figure CN119541448A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of speech synthesis technology, and in particular to a style speech synthesis method, system and storage medium based on a large model. Background Art
[0002] Speech synthesis is a technology that converts text information into understandable and fluent spoken output. The speech synthesis process is to first convert text information into linguistic features or phonemes, and then convert the linguistic features or phonemes into audio waveforms.
[0003] In the existing speech synthesis process, the speech model can only synthesize speech of a single style, resulting in poor diversity in speech synthesis styles. Summary of the invention
[0004] The purpose of the embodiments of the present invention is to provide a speech synthesis method, system, terminal and storage medium to solve the problem of poor diversity of speech synthesis styles in the prior art.
[0005] The embodiment of the present invention is implemented as follows: a style speech synthesis method based on a large model, the method comprising:
[0006] Fine-tune the parameters of the pre-trained large model according to the style reference speech, and extract the speaker features of the style reference speech to obtain the style embedding features;
[0007] Acquire a style speech synthesis requirement, and perform style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain a style expected feature, wherein the style interpolation is used to adjust the style degree of the style embedding feature;
[0008] The text to be synthesized in the style speech synthesis requirement and the pre-trained large model after fine-tuning the style expected feature input parameters are used for speech synthesis to obtain the target style speech.
[0009] Preferably, style interpolation is performed on the style embedding feature according to the style speech synthesis requirement to obtain the style expected feature, including:
[0010] Classifying the style embedding features to obtain a feature classification set, and obtaining an average feature of the style embedding features in the feature classification set to obtain a neutral style feature;
[0011] Acquire a style type identifier and a style strength in the style speech synthesis requirement, and determine a target neutral feature in the neutral style feature according to the style type identifier;
[0012] The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
[0013] Preferably, the formula used to adjust the style of the target neutral feature according to the style strength includes:
[0014] s=α*s n +(1.0-α)*s h
[0015] Where α is the style strength, S n is the target neutral feature, S h is the style embedding feature with the maximum strength in the feature classification set corresponding to the target neutral feature, and S is the style expected feature.
[0016] Preferably, fine-tuning parameters of the pre-trained large model is performed according to the style reference speech, including:
[0017] Acquire a style text of the style reference speech, and extract text features of the style text;
[0018] Attention processing is performed on the style reference speech and the text features to obtain global features, and parameters of the pre-trained large model are fine-tuned according to the global features.
[0019] Preferably, performing attention processing on the style reference speech and the text features to obtain global features includes:
[0020] Performing convolution processing on the style reference speech to obtain a first convolution feature, and performing vector conversion on the first convolution feature to obtain a style vector;
[0021] Performing attention processing on the style vector as a key matrix and a value matrix and the text feature as a query matrix to obtain an attention feature;
[0022] Perform multi-convolution processing on the attention feature to obtain a second convolution feature and a third convolution feature, and perform weight processing on the second convolution feature and the third convolution feature to obtain the global feature.
[0023] Preferably, before fine-tuning the parameters of the pre-trained large model according to the style reference speech, the method further includes:
[0024] Acquire a speech-text pair, and perform feature extraction on the speech-text pair to obtain speech-text features;
[0025] Performing speaker embedding processing on the speech-text feature to obtain speaker embedding features, and inputting the speech-text pair, the speech-text feature and the speaker embedding features into a large model for speech synthesis to obtain synthesized speech;
[0026] The model loss is determined according to the synthesized speech, and the parameters of the large model are updated according to the model loss until the large model converges to obtain the pre-trained large model.
[0027] Preferably, after performing speaker embedding processing on the speech text feature to obtain the speaker embedding feature, the method further includes:
[0028] The speaker embedding features are normalized, and the normalized speaker embedding features are subjected to dimensionality reduction.
[0029] Another object of an embodiment of the present invention is to provide a style speech synthesis system based on a large model, the system comprising:
[0030] A model fine-tuning module is used to fine-tune the parameters of the pre-trained large model according to the style reference speech, and extract the speaker features of the style reference speech to obtain the style embedding features;
[0031] A style interpolation module, used to obtain a style speech synthesis requirement, and perform style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain a style expected feature, wherein the style interpolation is used to adjust the style degree of the style embedding feature;
[0032] The speech synthesis module is used to perform speech synthesis on the text to be synthesized in the speech synthesis requirement of the style and the pre-trained large model after fine-tuning the input parameters of the expected feature of the style to obtain the target style speech.
[0033] Preferably, the style interpolation module is also used for:
[0034] Classifying the style embedding features to obtain a feature classification set, and obtaining an average feature of the style embedding features in the feature classification set to obtain a neutral style feature;
[0035] Acquire a style type identifier and a style strength in the style speech synthesis requirement, and determine a target neutral feature in the neutral style feature according to the style type identifier;
[0036] The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
[0037] The embodiments of the present invention perform style interpolation on style embedding features according to style speech synthesis requirements, and can effectively adjust the style strength of style embedding features according to user requirements, thereby improving the flexibility of speech style control and improving the style diversity of target style speech generated based on the text to be synthesized and the expected style features. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 is a flow chart of a style speech synthesis method based on a large model provided by the first embodiment of the present invention;
[0039] Figure 2 is a schematic diagram of the structure of a style speech synthesis system based on a large model provided by a second embodiment of the present invention;
[0040] Figure 3 It is a schematic diagram of the structure of a terminal device provided in the third embodiment of the present invention. DETAILED DESCRIPTION
[0041] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0042] In order to illustrate the technical solution of the present invention, a specific embodiment is provided below for illustration.
[0043] Embodiment 1
[0044] See also Figure 1 , is a flow chart of a large-model-based style speech synthesis method provided in a first embodiment of the present invention. The large-model-based style speech synthesis method can be applied to any device or system. The large-model-based style speech synthesis method includes the following steps:
[0045] Step S10, fine-tuning the parameters of the pre-trained large model according to the style reference speech, and extracting the speaker features of the style reference speech to obtain the style embedding features;
[0046] Among them, the big model can be set according to needs. For example, the big model is set to a GPT model. In this step, a speech library of different styles or emotions is prepared as needed, and the pre-trained big model is fine-tuned using the speech library to obtain a GPT fine-tuned synthesis model.
[0047] Preferably, before fine-tuning the parameters of the pre-trained large model according to the style reference speech, the method further includes:
[0048] Acquire a speech-text pair, and perform feature extraction on the speech-text pair to obtain speech-text features; wherein the speech-text pair includes sample speech and sample text, the sample text pair can be set according to requirements, and the speech-text features can be selected according to requirements, for example, the speech-text features can adopt Mel frequency cepstral coefficients, linear prediction coefficients, linear prediction cepstral coefficients, perceptual linear prediction coefficients, Fbank based on filter bank and other features;
[0049] Perform speaker embedding processing on the speech-text feature to obtain speaker embedding features, and input the speech-text pair, the speech-text feature and the speaker embedding features into a large model for speech synthesis to obtain synthesized speech; wherein, by performing speaker embedding processing on the speech-text feature, the representation of the speaker's voice features can be effectively extracted;
[0050] The model loss is determined according to the synthesized speech, and the parameters of the large model are updated according to the model loss until the large model converges to obtain the pre-trained large model; wherein, an ultra-large-scale sound library (more than 100,000 hours of effective speech) is used, and the sample speech is discretized using a trained encoder codec to obtain a speech token, and the trained voiceprint model is used to extract speaker embedding features, and the speech-text pairs, speech-text features and speaker embedding features are used as inputs and the speech tokens are used as outputs to train the large model until the large model converges.
[0051] Optionally, fine-tune the parameters of the pre-trained large model based on the style reference speech, including:
[0052] Acquire the style text of the style reference speech, and extract text features of the style text; wherein the identifier of the style reference speech is matched with a text query table to obtain the style text, and the style text is convoluted to obtain the text features;
[0053] Attention processing is performed on the style reference speech and the text features to obtain global features, and parameters of the pre-trained large model are fine-tuned according to the global features; wherein, by performing attention processing on the style reference speech and the text features, the style reference speech and the text features can be effectively fused to obtain global features.
[0054] Furthermore, attention processing is performed on the style reference speech and the text features to obtain global features, including:
[0055] The style reference speech is convolved to obtain a first convolution feature, and the first convolution feature is vector-converted to obtain a style vector; wherein the convolution kernel in the convolution process of the style reference speech can be set according to demand, and the first convolution feature is vector-converted according to a preset mapping relationship to obtain a style vector;
[0056] The style vector is used as a key matrix and a value matrix, and the text feature is used as a query matrix for attention processing to obtain an attention feature; multi-convolution processing is performed on the attention feature to obtain a second convolution feature and a third convolution feature, and weight processing is performed on the second convolution feature and the third convolution feature to obtain the global feature.
[0057] Preferably, after performing speaker embedding processing on the speech text features to obtain speaker embedding features, the method further includes: normalizing the speaker embedding features and performing dimensionality reduction processing on the normalized speaker embedding features; wherein, by normalizing the speaker embedding features, the dimensional range of the speaker embedding features can be effectively unified, which facilitates subsequent processing of the speaker embedding features, and by performing dimensionality reduction processing on the normalized speaker embedding features, post-processing operations on the speaker embedding features are further facilitated.
[0058] Step S20, obtaining a style speech synthesis requirement, and performing style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain a style expected feature;
[0059] Among them, style interpolation is used to adjust the style degree of style embedding features. The style speech synthesis requirements include the text to be synthesized, the style type identifier and the style strength. The text to be synthesized is the speech content that the user needs to synthesize. The style type identifier is used to indicate the speech style type required by the user. The style strength is used to indicate the style emotion strength required by the user.
[0060] Optionally, style interpolation is performed on the style embedding feature according to the style speech synthesis requirement to obtain the style expected feature, including:
[0061] The style embedding features are classified to obtain a feature classification set, and the average features of the style embedding features in the feature classification set are obtained to obtain a neutral style feature; wherein a pre-trained classifier may be used to classify the style embedding features to obtain a feature classification set, and for each feature classification set, the average features of the style embedding features in the set are respectively calculated to obtain a neutral style feature;
[0062] The style type identifier and style strength in the style speech synthesis requirement are obtained, and the target neutral feature in the neutral style feature is determined according to the style type identifier; wherein the style type identifier is matched with a set query table to obtain a target classification set, the set query table stores the correspondence between different style type identifiers and corresponding target classification sets, and the neutral style feature in the target classification set is determined as the target neutral feature;
[0063] The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
[0064] Furthermore, the formula used to adjust the style of the target neutral feature according to the style strength includes:
[0065] s=α*s n +(1.0-α)*s h
[0066] Where α is the style strength, S n is the target neutral feature, S h is the style embedding feature with the maximum strength in the feature classification set corresponding to the target neutral feature, and S is the style expectation feature. α can control the strength of the style emotion. For example, when α is set to 1.0, it is a completely neutral emotion. As α decreases, the happy emotion becomes stronger and stronger. When α is a negative value, the happy emotion is stronger than the happy emotion in the speech library. It is even possible to generate styles or emotions that are difficult to describe qualitatively in practice by interpolating different emotions.
[0067] Step S30, performing speech synthesis on the text to be synthesized in the style speech synthesis requirement and the pre-trained large model after fine-tuning the style expected feature input parameters to obtain the target style speech;
[0068] Among them, the pre-trained large model is fine-tuned with the expected style features and the input parameters of the text to be synthesized to obtain the target style speech, which has the expected synthesis style or emotion, thereby improving the diversity of speech synthesis styles.
[0069] In this embodiment, by performing style interpolation on the style embedding feature according to the style speech synthesis requirements, the style strength of the style embedding feature can be effectively adjusted according to user needs, thereby improving the flexibility of speech style control and improving the style diversity of the target style speech generated based on the text to be synthesized and the expected style features.
[0070] Embodiment 2
[0071] See also Figure 2 , is a schematic diagram of the structure of a large model-based style speech synthesis system 100 provided in a second embodiment of the present invention, including:
[0072] The model fine-tuning module 10 is used to fine-tune the parameters of the pre-trained large model according to the style reference speech, and extract the speaker characteristics of the style reference speech to obtain the style embedding features. The large model can be set according to the needs. For example, the large model is set as a GPT model. In this module, a speech library of different styles or emotions is prepared as needed, and the pre-trained large model is fine-tuned using the speech library to obtain a GPT fine-tuned synthesis model.
[0073] Optionally, the model fine-tuning module 10 is further used to: obtain a style text of the style reference speech, and extract text features of the style text;
[0074] Attention processing is performed on the style reference speech and the text features to obtain global features, and parameters of the pre-trained large model are fine-tuned according to the global features.
[0075] Furthermore, the model fine-tuning module 10 is also used to: perform convolution processing on the style reference speech to obtain a first convolution feature, and perform vector conversion on the first convolution feature to obtain a style vector;
[0076] Performing attention processing on the style vector as a key matrix and a value matrix and the text feature as a query matrix to obtain an attention feature;
[0077] Perform multi-convolution processing on the attention feature to obtain a second convolution feature and a third convolution feature, and perform weight processing on the second convolution feature and the third convolution feature to obtain the global feature.
[0078] Furthermore, the model fine-tuning module 10 is also used to: obtain a speech-text pair, and perform feature extraction on the speech-text pair to obtain speech-text features;
[0079] Performing speaker embedding processing on the speech-text feature to obtain speaker embedding features, and inputting the speech-text pair, the speech-text feature and the speaker embedding features into a large model for speech synthesis to obtain synthesized speech;
[0080] The model loss is determined according to the synthesized speech, and the parameters of the large model are updated according to the model loss until the large model converges to obtain the pre-trained large model.
[0081] Preferably, the model fine-tuning module 10 is further used to: normalize the speaker embedding features, and perform dimensionality reduction on the normalized speaker embedding features.
[0082] The style interpolation module 11 is used to obtain the style speech synthesis requirements, and perform style interpolation on the style embedded features according to the style speech synthesis requirements to obtain the expected style features; wherein, the style interpolation is used to adjust the style degree of the style embedded features, and the style speech synthesis requirements include the text to be synthesized, the style type identifier and the style strength, the text to be synthesized is the speech content that the user needs to synthesize, the style type identifier is used to indicate the speech style type required by the user, and the style strength is used to indicate the style emotional strength required by the user.
[0083] Optionally, the style interpolation module 11 is further used to: classify the style embedding features to obtain a feature classification set, and obtain an average feature of the style embedding features in the feature classification set to obtain a neutral style feature;
[0084] Acquire a style type identifier and a style strength in the style speech synthesis requirement, and determine a target neutral feature in the neutral style feature according to the style type identifier;
[0085] The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
[0086] Furthermore, the formula used to adjust the style of the target neutral feature according to the style strength includes:
[0087] s=α*s n +(1.0-α)*s h
[0088] Where α is the style strength, S n is the target neutral feature, S h is the style embedding feature with the maximum strength in the feature classification set corresponding to the target neutral feature, and S is the style expected feature.
[0089] The speech synthesis module 12 is used to perform speech synthesis on the text to be synthesized in the style speech synthesis requirement and the pre-trained large model after fine-tuning the input parameters of the style expected feature to obtain the target style speech. The target style speech is obtained by fine-tuning the pre-trained large model after the expected style expected feature and the input parameters of the text to be synthesized. The target style speech has the expected synthesis style or emotion, thereby improving the diversity of speech synthesis styles.
[0090] In this embodiment, by performing style interpolation on the style embedding feature according to the style speech synthesis requirements, the style strength of the style embedding feature can be effectively adjusted according to user needs, thereby improving the flexibility of speech style control and improving the style diversity of the target style speech generated based on the text to be synthesized and the expected style features.
[0091] Embodiment 3
[0092] Figure 3 2 is a block diagram of a terminal device 2 provided in the third embodiment of the present application. Figure 3 As shown, the terminal device 2 of this embodiment includes: a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program of a style speech synthesis method based on a large model. When the processor 20 executes the computer program 22, the steps in each embodiment of the style speech synthesis method based on a large model are implemented.
[0093] Exemplarily, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, a processor 20 and a memory 21.
[0094] The processor 20 may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0095] The memory 21 may be an internal storage unit of the terminal device 2, such as a hard disk or memory of the terminal device 2. The memory 21 may also be an external storage device of the terminal device 2, such as a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal device 2. Further, the memory 21 may also include both an internal storage unit of the terminal device 2 and an external storage device. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 may also be used to temporarily store data that has been output or is to be output.
[0096] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional unit.
[0097] If the integrated module is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium can be non-volatile or volatile. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in computer-readable storage media can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable storage media do not include electrical carrier signals and telecommunication signals.
[0098] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
Claims
1. A style speech synthesis method based on a large model, characterized in that: The method comprises: Fine-tune the parameters of the pre-trained large model according to the style reference speech, and extract the speaker features of the style reference speech to obtain the style embedding features; Acquire a style speech synthesis requirement, and perform style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain a style expected feature, wherein the style interpolation is used to adjust the style degree of the style embedding feature; The text to be synthesized in the style speech synthesis requirement and the pre-trained large model after fine-tuning the style expected feature input parameters are used for speech synthesis to obtain the target style speech.
2. The method for synthesizing style speech based on a large model as claimed in claim 1, characterized in that: Performing style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain the style expected feature, including: Classifying the style embedding features to obtain a feature classification set, and obtaining an average feature of the style embedding features in the feature classification set to obtain a neutral style feature; Acquire a style type identifier and a style strength in the style speech synthesis requirement, and determine a target neutral feature in the neutral style feature according to the style type identifier; The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
3. The method for synthesizing style speech based on a large model as claimed in claim 2, characterized in that: The formula used to adjust the style of the target neutral feature according to the style strength includes: s=α*s n +(1.0-α)*s h Where α is the style strength, S n is the target neutral feature, S h is the style embedding feature with the maximum strength in the feature classification set corresponding to the target neutral feature, and S is the style expected feature.
4. The method for synthesizing style speech based on a large model as claimed in claim 1, characterized in that: Fine-tune the parameters of the pre-trained large model based on the style reference speech, including: Acquire a style text of the style reference speech, and extract text features of the style text; Attention processing is performed on the style reference speech and the text features to obtain global features, and parameters of the pre-trained large model are fine-tuned according to the global features.
5. The method for synthesizing style speech based on a large model as claimed in claim 4, characterized in that: Performing attention processing on the style reference speech and the text features to obtain global features, including: Performing convolution processing on the style reference speech to obtain a first convolution feature, and performing vector conversion on the first convolution feature to obtain a style vector; Performing attention processing on the style vector as a key matrix and a value matrix and the text feature as a query matrix to obtain an attention feature; Perform multi-convolution processing on the attention feature to obtain a second convolution feature and a third convolution feature, and perform weight processing on the second convolution feature and the third convolution feature to obtain the global feature.
6. The method for synthesizing style speech based on a large model as claimed in claim 1, characterized in that: Before fine-tuning the parameters of the pre-trained large model according to the style reference speech, it also includes: Acquire a speech-text pair, and perform feature extraction on the speech-text pair to obtain speech-text features; Performing speaker embedding processing on the speech-text feature to obtain speaker embedding features, and inputting the speech-text pair, the speech-text feature and the speaker embedding features into a large model for speech synthesis to obtain synthesized speech; The model loss is determined according to the synthesized speech, and the parameters of the large model are updated according to the model loss until the large model converges to obtain the pre-trained large model.
7. The method for synthesizing style speech based on a large model as claimed in claim 6, characterized in that: After performing speaker embedding processing on the speech text feature to obtain the speaker embedding feature, the method further includes: The speaker embedding features are normalized, and the normalized speaker embedding features are subjected to dimensionality reduction.
8. A style speech synthesis system based on a large model, characterized in that: The system comprises: A model fine-tuning module is used to fine-tune the parameters of the pre-trained large model according to the style reference speech, and extract the speaker features of the style reference speech to obtain the style embedding features; A style interpolation module, used to obtain a style speech synthesis requirement, and perform style interpolation on the style embedding feature according to the style speech synthesis requirement to obtain a style expected feature, wherein the style interpolation is used to adjust the style degree of the style embedding feature; The speech synthesis module is used to perform speech synthesis on the text to be synthesized in the speech synthesis requirement of the style and the pre-trained large model after fine-tuning the input parameters of the expected feature of the style to obtain the target style speech.
9. The style speech synthesis system based on a large model as claimed in claim 8, characterized in that: The style interpolation module is also used to: Classifying the style embedding features to obtain a feature classification set, and obtaining an average feature of the style embedding features in the feature classification set to obtain a neutral style feature; Acquire a style type identifier and a style strength in the style speech synthesis requirement, and determine a target neutral feature in the neutral style feature according to the style type identifier; The target neutral feature is style-adjusted according to the style strength to obtain the style desired feature.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.