Speech synthesis sample set construction method and device and speech synthesis model training method and device

Abstract

The embodiment of the invention discloses a speech synthesis sample set construction method and device and a speech synthesis model training method and device. The construction method comprises the following steps: firstly, for any first voice segment in a plurality of voice segments, acquiring a plurality of first annotation values corresponding to a plurality of voice style characteristics, which are obtained by annotating the first voice segment; then, combining the plurality of first annotation values into a first voice style description text; then, taking a first voice recognition text corresponding to the first voice segment as a to-be-synthesized text, taking the to-be-synthesized text and the first voice style description text as sample features, taking the first voice segment as label voice, constructing a first voice synthesis sample, and classifying the first voice synthesis sample into a voice synthesis sample set; furthermore, a speech synthesis model can be trained based on the constructed speech synthesis sample set, so that the trained model supports the style of speech synthesis controlled by using a text describing speech style features, and the embedded representation of a speaker does not need to be used as input.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of speech intelligent generation technology, and in particular, to a method and device for constructing a speech synthesis sample set, a method and device for training a speech synthesis model, a computer-readable storage medium, and a computing device. Background Art

[0002] The Text To Speech (TTS) task aims to convert the input text into waveform speech. It has a wide range of applications in the field of human-computer interaction, such as smart speakers, smart outbound calls, voice assistants, voice navigation, audio books, radio broadcasts, interactive games, digital humans, and animation / game character voice synthesis.

[0003] In recent years, with the rapid development of Large Speech Model (LSM) and Generative Artificial Intelligence (GenAI), more and more speech synthesis systems have begun to draw on the structure of Large Generative Language Model (LGLM) to generate richer and more diverse speech.

[0004] Therefore, the embodiments of this specification propose a speech synthesis solution that can meet higher requirements in practical applications, for example, supporting the use of text descriptions to control the style of synthesized speech. Summary of the invention

[0005] The embodiments of this specification describe a method and device for constructing a speech synthesis sample set, and a method and device for training a speech synthesis model, which can realize the customized construction of a speech synthesis sample set, and, using the speech synthesis model trained based on the sample set, can realize the use of text describing speech style features to control the style of synthesized speech.

[0006] According to a first aspect, a method for constructing a speech synthesis sample set is provided. The method comprises: for any first speech segment among multiple speech segments, obtaining a plurality of first annotation values ​​corresponding to a plurality of speech style features obtained by annotating the first speech segment; combining the plurality of first annotation values ​​into a first speech style description text; using a first speech recognition text corresponding to the first speech segment as a text to be synthesized, and using the first speech style description text together as sample features, and using the first speech segment as a labeled speech, to construct a first speech synthesis sample, and to include it in a speech synthesis sample set.

[0007] In some embodiments, obtaining the multiple first annotation values ​​includes: inputting the first speech segment into multiple trained machine learning models respectively to obtain the output multiple first annotation values; the multiple machine learning models are used to correspondingly identify the multiple speech style features.

[0008] In some embodiments, obtaining the multiple first annotation values ​​includes: displaying an annotation interface for the first voice segment, the annotation interface including the multiple voice style features and multiple alternative values ​​for each feature; receiving the multiple first annotation values ​​selected by the staff based on the annotation interface.

[0009] In some embodiments, the multiple first annotation values ​​are combined into a first voice style description text, including: using the multiple first annotation values ​​to fill in a preset prompt word template to obtain a first prompt word; inputting the first prompt word into a large language model to obtain the output of the first voice style description text.

[0010] In some embodiments, combining the plurality of first annotation values ​​into a first voice style description text includes: filling the plurality of first annotation values ​​into a preset voice style description template to obtain the first voice style description text.

[0011] In some embodiments, the acquisition of the multiple voice segments includes: acquiring multiple original voices and performing the following preprocessing on them: extracting human voices from each of the original voices, and slicing and removing silence from the extracted human voices to obtain multiple voice segments; screening out voice segments that meet predetermined sound quality standards, belong to a predetermined language and have a single speaker from the multiple voice segments, and including them in the multiple voice segments.

[0012] In some embodiments, the plurality of speech style features include at least one of the following: speaking language, speaker gender, speaker age, speaker emotion, speaking rhythm, and ambient sound field. It should be understood that the speaker gender and speaker age may also be replaced by the speaker timbre.

[0013] According to a second aspect, a method for training a speech synthesis model is provided. The method comprises: obtaining a speech synthesis sample set constructed by the method described in the first aspect; for any speech synthesis sample in the speech synthesis sample set, inputting the text to be synthesized and the corresponding speech style description text into the speech synthesis model to obtain a first predicted speech segment; and training the speech synthesis model using the first predicted speech segment and the corresponding labeled speech.

[0014] According to the third aspect, a device for constructing a speech synthesis sample set is provided. The device includes: a labeling data acquisition module, configured to obtain a plurality of first labeling values ​​corresponding to a plurality of speech style features obtained by labeling any first speech segment among a plurality of speech segments; a description text generation module, configured to combine the plurality of first labeling values ​​into a first speech style description text; a sample construction module, configured to use the first speech recognition text corresponding to the first speech segment as the text to be synthesized, and the first speech style description text as the sample features, and use the first speech segment as the label speech, to construct a first speech synthesis sample, and include it in the speech synthesis sample set.

[0015] According to a fourth aspect, a training device for a speech synthesis model is provided. The device comprises: a sample acquisition module configured to acquire a speech synthesis sample set obtained by using the device structure described in the third aspect; a prediction module configured to input the text to be synthesized and the corresponding speech style description text into the speech synthesis model for any speech synthesis sample in the speech synthesis sample set, and obtain a first predicted speech segment; and a training module configured to train the speech synthesis model using the first predicted speech segment and the corresponding labeled speech.

[0016] According to a fifth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method provided in the first aspect or the second aspect.

[0017] According to a sixth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method provided in the first aspect or the second aspect is implemented.

[0018] In summary, by using the above method and device disclosed in the embodiments of this specification, a batch of data for model training can be constructed, wherein each sample contains the target speech, the text to be synthesized, and the description text describing the speech style characteristics. Furthermore, the speech synthesis model obtained based on the constructed data can synthesize natural and clear speech according to the input text sequence and the description of the style characteristics, and the content in the speech is consistent with the input content text, and the style characteristics of the speech correspond to the input style feature description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 A flowchart of the construction process of the speech synthesis sample set disclosed in the embodiments of this specification; Figure 2 The following is a flow chart of the method for constructing a speech synthesis sample set disclosed in the embodiments of this specification; Figure 3 The template for manually marking speech style features disclosed in the embodiments of this specification is shown; Figure 4 This is a flowchart of preprocessing original voice data disclosed in the embodiments of this specification; Figure 5 A schematic diagram of the process steps of the training method of the speech synthesis model disclosed in the embodiments of this specification; Figure 6 A schematic diagram illustrating the functional modules of the apparatus for constructing a speech synthesis sample set disclosed in an embodiment of this specification; Figure 7 A schematic diagram illustrating the functional modules of a training device for a speech synthesis model disclosed in an embodiment of this specification. DETAILED DESCRIPTION

[0021] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0022] As mentioned above, the task of speech synthesis aims to convert the input text into waveform speech. With the increasing application of speech synthesis, it is hoped that more diverse speech can be generated by controlling the speech style.

[0023] Currently, the industry has proposed the task of voice cloning, which aims to clone a person's voice and convert any text into a corresponding speech with a similar voice. In the voice cloning task, most speech synthesis models use a style embedding to characterize the target speaker's voice, overall voice style and other features.

[0024] There are usually two ways to obtain style embedding: 1) Use a fixed known speaker ID to obtain a style embedding from a lookup table. This embedding representation is bound to the speaker ID, that is, the same speaker can only have one embedding representation. 2) Extract the style embedding representation from the reference speech, usually using an additional voiceprint recognition model or a speaker feature encoder embedded in the speech synthesis system. That is, a reference speech has one embedding representation. If the generalization of the embedding representation extractor is not high, the final synthesized speech may have a low similarity.

[0025] Based on the above observations and analysis, the applicant proposed an improved speech synthesis scheme, the target task of which is to control the style features in the synthesized speech by using descriptive text, where the descriptive text can be in the form of Chinese natural language such as "a powerful middle-aged male voice is giving an impassioned speech", "a kind and gentle young female anchor is quickly introducing the product in the live broadcast room", etc. Using this descriptive text as the style control input no longer requires finding a real voice as a reference, and the input text has a higher degree of freedom. Therefore, compared with the previous method of using speaker ID or reference voice to obtain style embedding representation, it is more convenient for users to interact directly and can provide richer and more varied timbre choices.

[0026] Under the above target task setting, a corresponding dataset is required for training, where each sample contains the following three contents: target speech, text to be synthesized, and style feature description text. However, there is no dataset that meets the requirements and can be used directly in the industry. Therefore, the improvement plan designs a dataset construction process to construct a large-scale speech synthesis dataset suitable for this task, and uses this dataset to achieve the task of using style feature description text to control the speaker's timbre and speaking style in speech synthesis after fine-tuning the existing large-scale speech synthesis model.

[0027] The following first introduces the construction process of the speech synthesis dataset (or speech synthesis sample set), and then introduces the process of training the speech synthesis model using the constructed dataset.

[0028] Figure 2 The following is a flow chart of the method for constructing a speech synthesis sample set disclosed in the embodiments of this specification. It should be noted that the execution subject of the construction method can be any server, platform or device cluster with computing and processing capabilities. Figure 2 For a description, see Figure 1 The flow chart is shown. Figure 2 The construction method shown in FIG. 1 includes the following steps: Step S210, for any first speech segment among the multiple speech segments, obtain multiple first annotation values ​​corresponding to multiple speech style features obtained by annotating it; Step S220, combine the multiple first annotation values ​​into a first speech style description text; Step S230, use the first speech recognition text corresponding to the first speech segment as the text to be synthesized, and use it together with the first speech style description text as sample features, and use the first speech segment as the labeled speech to construct a first speech synthesis sample, and include it in the speech synthesis sample set.

[0029] The above steps are expanded as follows: First, in step S210, for any first voice segment among the plurality of voice segments, a plurality of first annotation values ​​corresponding to the plurality of voice style features obtained by annotating the first voice segment are obtained. It should be noted that the "first" in "first voice segment", "first annotation value", and the second, third and other similar terms in the text are all for distinguishing similar things and do not have other limiting functions such as sorting.

[0030] The above-mentioned multiple speech style features may be preset, for example, Figure 1 The n style features shown may include: speaker's timbre, speaker's emotion, speaking rhythm, ambient sound field, and the like.

[0031] The annotation of speech style can be achieved by using a trained machine learning model, or by manual labeling, or a combination of the two methods.

[0032] In one implementation, obtaining the multiple first annotation values ​​in this step may include: using the trained multiple machine learning models to process the first speech segment respectively to obtain the corresponding output multiple first annotation values.

[0033] The above-mentioned multiple machine learning models are used to identify the above-mentioned multiple voice style features. In some embodiments, the above-mentioned multiple machine learning models include: a speaker gender recognition model, a speaker age recognition model, and a speaker emotion recognition model. In one example, the speaker age recognition model can identify the age group of the speaker. In one example, the speaker emotion recognition model can identify the speaker's emotions, such as happiness, anger, sadness, and calmness. Exemplarily, the speaker emotion recognition model can adopt FunASR, SPEAKER_VGG_CCT.

[0034] It should be understood that the training process and use process of each machine learning model in this embodiment are similar. The main difference lies in whether labels are used for parameter adjustment. Therefore, the training process will not be described in detail.

[0035] In another embodiment, the acquisition of the multiple first annotation values ​​in this step may include: first displaying an annotation interface for the first voice segment, the annotation interface including the multiple voice style features and multiple candidate values ​​for each feature; and then receiving the multiple first annotation values ​​selected by the staff based on the annotation interface. Figure 3 , which includes manually annotated templates.

[0036] From the above, multiple first annotation values ​​of the first voice segment for multiple voice style features can be obtained.

[0037] Next, in step S220, the plurality of first annotation values ​​are combined into a first voice style description text.

[0038] In one implementation, a plurality of first annotation values ​​may be combined into a first voice style description text using a fixed sentence pattern, or in other words, a plurality of first annotation values ​​may be filled into a preset voice style description template to obtain the first voice style description text.

[0039] Exemplarily, the voice style description template includes: "A ${speaking style} of ${age} ${gender} is ${emotionally} ${behaving}", and the first voice style description text obtained by filling multiple first annotation values ​​into the template may be: A young male voice speaking passionately is excitedly commenting on a sports game.

[0040] In another embodiment, a large language model can be used to construct sentences. Specifically, some prompt word templates can be set, and then the preset prompt word templates can be filled with the multiple first annotation values ​​to obtain a first prompt word, and then the first prompt word is input into the large language model to obtain an output first voice style description text.

[0041] For example, the first prompt word may be: "Please combine some of these words into a coherent sentence. The content of the sentence is to describe the characteristics of a certain sound. The words are: passionate, youth, male voice, excitement, sports game commentary." From the above, a first voice style description text for the first voice segment can be obtained.

[0042] Afterwards, in step S230, the first speech recognition text corresponding to the first speech segment is used as the text to be synthesized, and together with the first speech style description text as sample features, and the first speech segment is used as the labeled speech to construct a first speech synthesis sample, which is included in the speech synthesis sample set.

[0043] It should be understood that the first speech recognition text can be obtained by processing the first speech segment using Automatic Speech Recognition (ASR). ASR is a technology that converts speech signals into readable text, allowing computers and devices to understand and process human speech input. In addition, for the implementation of this step, please refer to Figure 1 .

[0044] The above describes the construction of a corresponding speech synthesis sample based on any speech segment in a plurality of speech segments. It should be noted that the plurality of speech segments may be obtained by preprocessing the original speech data. For example, Figure 4 As shown, for each of the multiple original voices, the following preprocessing may be performed on it: a. Voice extraction: Perform voice extraction on each long speech, that is, use a voice extraction tool to extract the pure voice part from a piece of original speech. For example, tools such as Sleepter and HTdemucs can be used.

[0045] b. Slicing and removing silence: Slice the extracted pure human voice and remove the silence segment. A long speech will be cut into multiple short speech segments without long silence segments. For example, tools such as silcer and silero-vad can be used.

[0046] c. Sound quality screening: The sliced ​​segments are screened for sound quality, and some tools are used to evaluate speech clarity, signal-to-noise ratio and other indicators. Speech with higher sound quality is selected from all speech segments, and speech with poor sound quality is removed.

[0047] d. Single speaker screening: Use the speaker recognition model to identify each small segment of the speech obtained in the previous step, and extract the speaker vector of each small segment. If the similarity of multiple speaker vectors in the same speech is low, it is considered that there is more than one speaker in this speech. Use this method to screen out the speech with only one speaker from all the speech. Remove the speech with multiple speakers.

[0048] e. Language identification and screening: Use the language identification model to obtain the language of each speech obtained in the previous step, retain only the speech in Chinese and English, and remove the speech containing other languages.

[0049] After the above preprocessing, multiple voice clips with high availability can be obtained. It should be understood that the above is an exemplary introduction to the possible steps of the preprocessing, and in practice, any selection and combination in different orders can be performed as needed.

[0050] From the above, by adopting the dataset construction process designed in the improved solution, a batch of data that can be used for model training can be accumulated, or in other words, a speech synthesis dataset that can be used for model training can be obtained. Each training sample contains three contents: target speech (or labeled speech), text to be synthesized, and description text describing the speech style characteristics. Therefore, the speech synthesis dataset can be used to train the speech synthesis model.

[0051] Figure 5 The figure is a schematic diagram of the process steps of the training method of the speech synthesis model disclosed in the embodiment of this specification. The execution subject of the training method can be any server, platform or device cluster with computing and processing capabilities. Figure 5 As shown, the method comprises the following steps: Step S510, obtaining the adopted Figure 2The speech synthesis sample set is constructed by the method shown in the step S520, for any speech synthesis sample in the speech synthesis sample set, the text to be synthesized and the corresponding speech style description text are input into the speech synthesis model to obtain a first predicted speech segment; step S530, the speech synthesis model is trained using the first predicted speech segment and the corresponding labeled speech.

[0052] In a typical implementation of the above steps, the speech synthesis model can be a pre-trained large-scale speech model, such as Seed-TTS, CosyVoice, etc. Figure 5 The training method shown can be a fine-tuning method for the pre-trained model. It can be understood that the original input of the large-scale speech model (or the input during pre-training) includes the embedded representation of the speaker, while the embedded representation is not included during fine-tuning.

[0053] From the above, a trained speech synthesis model can be obtained, which supports generating synthesized speech that conforms to the style description based on the style description text input by the user without using the speaker ID or speaker embedding representation.

[0054] The above introduces the implementation steps of the improvement plan. The implementation mainly includes two stages: data construction and model training (such as model fine-tuning). The obtained speech synthesis model can synthesize natural and clear speech based on the input text sequence and the description of style features. The content of the speech is consistent with the input text, and the style features of the speech correspond to the input style feature description.

[0055] Corresponding to the above-mentioned data set construction method and model training method, the embodiments of this specification also disclose a construction device and a training device. Specifically, they are as follows: Figure 6 The functional module diagram of the device for constructing a speech synthesis sample set disclosed in the embodiment of this specification is shown, wherein the construction device 600 shown includes the following functional modules: The annotation data acquisition module 610 is configured to obtain a plurality of first annotation values ​​corresponding to a plurality of speech style features obtained by annotating any first speech segment among the plurality of speech segments. The description text generation module 620 is configured to combine the plurality of first annotation values ​​into a first speech style description text. The sample construction module 630 is configured to use the first speech recognition text corresponding to the first speech segment as the text to be synthesized, and the first speech style description text as sample features, and use the first speech segment as the labeled speech, to construct a first speech synthesis sample, and include it in the speech synthesis sample set.

[0056] In some embodiments, the annotation data acquisition module 610 is specifically configured as follows: the first speech segment is input into a plurality of trained machine learning models respectively to obtain the output of the plurality of first annotation values; the plurality of machine learning models are used to correspondingly identify the plurality of speech style features.

[0057] In some embodiments, the annotation data acquisition module 610 is specifically configured to: display an annotation interface for the first voice segment, the annotation interface including the multiple voice style features and multiple alternative values ​​for each feature; receive the multiple first annotation values ​​selected by the staff based on the annotation interface.

[0058] In some embodiments, the description text generation module 620 is specifically configured to: fill a preset prompt word template with the multiple first annotation values ​​to obtain a first prompt word; input the first prompt word into a large language model to obtain the output first voice style description text.

[0059] In some embodiments, the description text generation module 620 is specifically configured to: fill the plurality of first annotation values ​​into a preset voice style description template to obtain the first voice style description text.

[0060] In some embodiments, the construction device 600 also includes a voice segment acquisition module 640, which is configured to: perform the following preprocessing on multiple original voices: extract human voice from each of the original voices, and slice and remove silence on the extracted human voice to obtain multiple voice segments; filter out voice segments that meet predetermined sound quality standards, belong to a predetermined language and have a single speaker from the multiple voice segments, and include them in the multiple voice segments.

[0061] In some embodiments, the plurality of speech style features include at least one of the following: speaker's timbre, speaker's emotion, speaking rhythm, and ambient sound field.

[0062] Figure 7 The functional module diagram of the training device of the speech synthesis model disclosed in the embodiment of this specification is shown, wherein the training device 700 shown includes the following functional modules: The sample acquisition module 710 is configured to acquire the sample using Figure 6 The device shown constructs a speech synthesis sample set; a prediction module 720, configured to input the text to be synthesized and the corresponding speech style description text into the speech synthesis model for any speech synthesis sample in the speech synthesis sample set, to obtain a first predicted speech segment; a training module 730, configured to train the speech synthesis model using the first predicted speech segment and the corresponding labeled speech.

[0063] It should be noted that for the introduction of the above functional modules, reference can also be made to the relevant introduction of the process method in the above embodiments.

[0064] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute Figure 2 or Figure 5 The method described.

[0065] According to another embodiment of the present invention, there is also provided a computing device, comprising a memory and a processor, wherein the memory stores an executable code, and when the processor executes the executable code, Figure 2 or Figure 5 The method described.

[0066] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0067] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a speech synthesis sample set, comprising: For any first speech segment among the plurality of speech segments, obtaining a plurality of first annotation values ​​corresponding to a plurality of speech style features obtained by annotating the first speech segment; Combining the plurality of first annotation values ​​into a first voice style description text; The first speech recognition text corresponding to the first speech segment is used as the text to be synthesized, and is used together with the first speech style description text as sample features. The first speech segment is used as the labeled speech to construct a first speech synthesis sample, which is included in the speech synthesis sample set.

2. The method according to claim 1, wherein: The obtaining of the plurality of first annotation values ​​comprises: The first speech segment is input into multiple trained machine learning models respectively to obtain the multiple first annotation values ​​as output; the multiple machine learning models are used to correspondingly identify the multiple speech style features.

3. The method according to claim 1, wherein: The obtaining of the plurality of first annotation values ​​comprises: Displaying a marking interface for the first voice segment, the marking interface including the plurality of voice style features and a plurality of candidate values ​​for each of the features; The multiple first annotation values ​​selected by the reception staff based on the annotation interface.

4. The method according to claim 1, wherein: Combining the plurality of first annotation values ​​into a first voice style description text comprises: Filling a preset prompt word template with the multiple first annotation values ​​to obtain a first prompt word; The first prompt word is input into a large language model to obtain the first voice style description text as output.

5. The method according to claim 1, wherein: Combining the plurality of first annotation values ​​into a first voice style description text comprises: The plurality of first annotation values ​​are filled into a preset voice style description template to obtain the first voice style description text.

6. The method according to claim 1, wherein: The acquisition of the plurality of voice segments includes: A plurality of original speech pieces are obtained and preprocessed as follows: a human voice is extracted from each of the original speech pieces, and the extracted human voice is sliced ​​and muted to obtain a plurality of speech segments; and speech segments that meet a predetermined sound quality standard, belong to a predetermined language and have a single speaker are screened out from the plurality of speech segments and are included in the plurality of speech segments.

7. The method according to any one of claims 1 to 6, wherein: The multiple speech style features include at least one of the following: speaking language, speaker gender, speaker age, speaker emotion, speaking rhythm, and environmental sound field.

8. A method for training a speech synthesis model, comprising: Acquire a speech synthesis sample set constructed by the method according to claim 1; For any speech synthesis sample in the speech synthesis sample set, input the text to be synthesized and the corresponding speech style description text into the speech synthesis model to obtain a first predicted speech segment; The speech synthesis model is trained using the first predicted speech segment and the corresponding labeled speech.

9. A device for constructing a speech synthesis sample set, comprising: The annotation data acquisition module is configured to acquire, for any first speech segment among the plurality of speech segments, a plurality of first annotation values ​​corresponding to a plurality of speech style features obtained by annotating the first speech segment; A description text generation module, configured to combine the plurality of first annotation values ​​into a first voice style description text; The sample construction module is configured to use the first speech recognition text corresponding to the first speech segment as the text to be synthesized, and the first speech style description text as sample features, and use the first speech segment as the labeled speech to construct a first speech synthesis sample and include it in the speech synthesis sample set.

10. A training device for a speech synthesis model, comprising: A sample acquisition module, configured to acquire a speech synthesis sample set obtained by using the device structure as claimed in claim 9; A prediction module is configured to input the to-be-synthesized text and the corresponding speech style description text into a speech synthesis model for any speech synthesis sample in the speech synthesis sample set, so as to obtain a first predicted speech segment; The training module is configured to train the speech synthesis model using the first predicted speech segment and the corresponding labeled speech.

11. A computer-readable storage medium having a computer program stored thereon, wherein: When the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 8.

12. A computing device comprising a memory and a processor, wherein: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 8 is implemented.

Images