A method of obtaining training data for a speech synthesis model

Abstract

The application discloses a method for obtaining training data of a speech synthesis model, comprising the following steps: screening part of corpus with high frequency syllables from a corpus of a target language; recording and sampling the screened corpus and marking the corpus to obtain an initial training set, and training the speech synthesis model through the initial training set; synthesizing speech of unmarked corpus in the corpus according to the speech synthesis model to obtain quality information of the speech synthesis; determining decision attributes of the unmarked corpus according to the quality information, taking the occurrence times of basic letters in the unmarked corpus as conditional attributes, constructing a decision table, and obtaining a dominant rough set. The application effectively solves the problems of few training data samples, high professional requirement, high cost and heavy workload in the recording and marking of training data samples in the target language synthesis task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method for obtaining training data of a speech synthesis model. BACKGROUND

[0002] The end-to-end speech synthesis model not only has a high requirement for the audio quality of the sample, but also needs the training data to cover the common pronunciation of the used language as much as possible from the phoneme level. Therefore, the requirement for the training data of the speech synthesis is much higher than that in other machine learning fields. How to select appropriate samples from the corpus as the training data of the speech synthesis is very important. For the end-to-end speech synthesis model, it is actually a kind of supervised learning, and the mel-frequency spectrum vector of the speech data is a kind of label of the training data. Due to the dialect difference or regional pronunciation characteristics of the target language, it is difficult to obtain speech synthesis training data with high audio quality and standard pronunciation. In order to make the speech synthesis training data meet the requirements, firstly, the samples most beneficial to the speech synthesis model need to be selected from the corpus as the training data scientifically, and secondly, professional persons need to be hired to use professional equipment to sample and label the selected training data. If the unmarked samples in the corpus are randomly selected as the training data, the selected training data may not be the optimal training data for the speech synthesis model. In this way, a large amount of time cost and a large amount of human cost are wasted. SUMMARY

[0003] The present application aims to provide a method for obtaining training data of a speech synthesis model, so as to quickly select the most valuable samples as the training data of the speech synthesis model.

[0004] The present application provides a method for obtaining training data of a speech synthesis model, comprising:

[0005] selecting part of the corpus with high-frequency syllables from the corpus of the target language;

[0006] sampling and labeling the selected corpus to obtain an initial training set, and training the speech synthesis model through the initial training set;

[0007] performing speech synthesis on the unmarked corpus in the corpus according to the speech synthesis model to obtain quality information of the speech synthesis;

[0008] determining the decision attribute of the unmarked corpus according to the quality information, taking the occurrence frequency of the basic letters in the unmarked corpus as the condition attribute, constructing a decision table, and obtaining an advantage rough set;

[0009] perform attribute reduction on the advantage rough set to obtain an intersection of attribute reduction, find necessary elements from the intersection, and screen new corpora from the unlabeled corpora by using the necessary elements to obtain new training data.

[0010] Further, the quality information is a MOS score, and determining the decision attribute of the unlabeled corpora according to the quality information comprises:

[0011] For any unlabeled corpus, when the corresponding MOS score is greater than a preset threshold, the decision attribute of the unlabeled corpus is determined to be a first type of attribute, and when the corresponding MOS score is less than or equal to the preset threshold, the decision attribute of the unlabeled corpus is determined to be a second type of attribute.

[0012] Further, using the number of occurrences of a basic letter in the unlabeled corpus as a condition attribute comprises:

[0013] For any unlabeled corpus, the corpus is converted into a text represented by syllable symbols of Wylie Romanization, and the number of occurrences of each syllable symbol of Wylie Romanization in the converted text is determined.

[0014] Each syllable symbol of Wylie Romanization and the number of occurrences of the syllable symbol in the text are used as condition attributes of the unlabeled corpus.

[0015] Further, performing attribute reduction on the advantage rough set to obtain an intersection of attribute reduction, and finding necessary elements from the intersection comprises:

[0016] Using heuristic attribute reduction, a reduced intersection core in which the decision attribute is of the second type of attribute and necessary elements in the reduced intersection core are found by using an element algorithm HARCC.

[0017] Further, screening new corpora from the unlabeled corpora by using the necessary elements to obtain new training data comprises:

[0018] Using the necessary elements to sort data in which the decision attribute is of the second type of attribute in the unlabeled corpora, and screening new corpora based on the sorting result, wherein the proportion of syllables corresponding to the necessary elements in the screened new corpora is higher than that in un-screened corpora.

[0019] Performing audio sampling and labeling on the new corpora to obtain new training data.

[0020] Further, the target language is any one of Tibetan, Naxi, Hani, Qiang, Miao, and Buyi.

[0021] Furthermore, the new training data is used to update and train the speech synthesis model, and the method further includes:

[0022] The process of repeatedly performing speech synthesis on the unlabeled corpus in the corpus according to the speech synthesis model to obtain the quality information of the speech synthesis, to selecting new corpus from the unlabeled corpus through the necessary elements for recording, sampling and labeling to obtain new training data, and updating the speech synthesis model according to the new training data, continues until the speech synthesis model meets the convergence condition.

[0023] The present invention has at least the following beneficial effects:

[0024] This invention proposes a method for selecting the optimal samples from a target language corpus as training data for a target language speech synthesis model, based on a limited amount of speech synthesis training data. This invention effectively solves the problems of limited training data samples, the high cost and workload associated with selecting highly specialized labeled training data samples in target language synthesis tasks. By selecting the most effective samples for the model, this invention effectively reduces the sample size. Compared to randomly selecting unlabeled samples from a target language corpus as training data, this method saves more time and manpower while maintaining the accuracy of the speech synthesis model.

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0026] Figure 1 A flowchart illustrating a method for acquiring training data for a speech synthesis model, provided by the present invention;

[0027] Figure 2 A schematic diagram illustrating the Wylie transliteration of Tibetan letters provided by this invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0029] This embodiment describes a method for obtaining training data for a speech synthesis model, including:

[0030] Select a portion of the corpus containing high-frequency syllables from the target language corpus;

[0031] In this embodiment, Tibetan is selected as the target language. This invention uses a small amount of Tibetan speech synthesis training data to select the optimal samples from a Tibetan corpus as training data for the Tibetan speech synthesis model. This invention improves the efficiency and accuracy of subsequent model training by identifying and extracting frequently occurring syllables (such as syllable combinations and common pronunciation fragments) from the Tibetan corpus, focusing on the most commonly used pronunciation patterns in the language.

[0032] The selected corpus is recorded, sampled, and labeled to obtain an initial training set, and the speech synthesis model is trained using the initial training set.

[0033] By recording actual speech from the selected high-frequency syllable corpus, corresponding audio samples are obtained, and each audio sample is labeled with corresponding text content (i.e., the correspondence between speech and text), forming labeled training data. The initial training set consists of these labeled audio and text data, which are used to initially train speech synthesis models (such as WaveNet, TTS models, etc.).

[0034] The speech synthesis model is used to synthesize speech from the unlabeled corpus to obtain the quality information of the speech synthesis.

[0035] Using the pre-trained speech synthesis model, speech synthesis is performed on unlabeled corpus data (i.e., unrecorded / unannotated text data) and the quality information of the speech synthesis is obtained.

[0036] The decision attributes of the unlabeled corpus are determined based on the quality information, and the occurrence frequency of basic letters in the unlabeled corpus is used as the condition attribute to construct a decision table and obtain the advantage rough set.

[0037] Based on quality information, decision labels (such as "high quality," "low quality," etc.) are assigned to unlabeled corpora as a basis for subsequent classification. The frequency of basic letters in the unlabeled corpora is used as a feature attribute. For example, the frequency of consonants, vowels, or specific letter combinations in each corpus is counted. Then, the decision attributes and conditional attributes are combined into a table to form a decision system. Each unlabeled corpus is treated as a record, containing its letter frequency features and quality label. Rough set theory (a mathematical method for handling uncertainty and data classification) is then used to analyze the decision table, identifying which conditional attributes (letter frequencies) have a greater impact on the decision (quality label), thus obtaining the set of dominant attributes, i.e., the dominant rough set.

[0038] The rough set of advantages is reduced in terms of attributes to obtain the intersection of the reduced attributes. Necessary elements are found from the intersection to select new corpus from the unlabeled corpus for recording sampling and labeling, thereby obtaining new training data.

[0039] The advantage rough set is simplified by an algorithm (such as a genetic algorithm, a heuristic algorithm), redundant attributes are removed, and the most critical features for decision-making are retained. For example, it is found that the frequency of some letter combinations has a significant impact on quality, while other attributes can be ignored. If there are multiple reduction results, find the common key attributes in these results to obtain the intersection of attribute reduction. Then find the necessary elements from the intersection, that is, the features essential to classification decision-making, and use the necessary elements to further filter a higher-quality subset from the unlabeled corpus. For example, preferentially select the text in which the necessary elements appear frequently. Then record and text label the filtered new corpus to obtain new high-quality training data.

[0040] The method effectively solves the problems of few training data samples, high cost and large workload of professional recording and labeling training data samples in the Tibetan language synthesis task. The present application selects the most effective samples for the model as training data, which reduces the samples, saves time and labor costs compared to training the Tibetan language synthesis model by randomly selecting unlabeled samples from the Tibetan corpus as training data for the Tibetan language synthesis model, while ensuring the accuracy of the Tibetan speech synthesis model.

[0041] Further, the quality information is a MOS score, and determining the decision attribute of the unlabeled corpus according to the quality information comprises:

[0042] For any unlabeled corpus, when the corresponding MOS score is greater than a preset threshold, the decision attribute of the unlabeled corpus is determined to be a first type of attribute, and when the corresponding MOS score is less than or equal to the preset threshold, the decision attribute of the unlabeled corpus is determined to be a second type of attribute.

[0043] The decision attribute set D is obtained by performing speech synthesis on the unlabeled sample set U by the speech synthesis model and performing speech synthesis evaluation MOS scoring one by one, and the value of the decision attribute set D includes: when the speech synthesis evaluation MOS score is greater than or equal to 4, D i =1, and when the speech synthesis evaluation MOS score is less than 4, D i =2.

[0044] Further, the occurrence frequency of the basic letters in the unlabeled corpus is used as the condition attribute, which comprises:

[0045] For any of the unlabeled corpora, the corpus is converted into a text represented by syllable symbols of the Wylie Romanization, and the occurrence frequency of each syllable symbol of the Wylie Romanization in the converted text is determined.

[0046] Each syllable symbol of the Wylie Romanization appearing in the text and the number of times the syllable symbol appears are taken as the conditional attributes of the unlabeled corpus.

[0047] As Figure 2 shown, the Tibetan language is composed of 30 basic consonant letters and 4 vowels. In order to express conveniently, the 34 Tibetan basic letters are transcribed into Wylie Romanization, and the 34 Wylie Romanization are numbered as , and all the 34 letters form a non-empty finite conditional attribute set C:

[0048] ,

[0049] Among them, the value of the conditional attribute is the number of times the corresponding letter appears in each unlabeled sample .

[0050] Assume that the existing Tibetan speech synthesis model trained based on the labeled data, use this model to estimate each unlabeled data in the unlabeled sample set U to get the decision attribute set D, the value of the decision attribute is only two, when the speech synthesis evaluation MOS score is greater than or equal to 4, D i =1, when the speech synthesis evaluation MOS score is less than 4, D i =2. Thus, a non-empty finite attribute set A is obtained:

[0051] ,

[0052] From this, it can be deduced that:

[0053] ,

[0054] Among them, is the value range of the conditional attribute , and the information function is:

[0055] ,

[0056] For any conditional attribute belongs to the non-empty finite attribute set A and the unlabeled sample data x i belongs to the unlabeled data set U, it can be obtained:

[0057] ,

[0058] From the above formula, a four-tuple decision information system (decision table) is obtained:

[0059] , ​

[0060] Suppose the unlabeled data samples are 5000, then the decision information system (decision table) can be represented as follows:

[0061] Table 1 Decision information system (decision table) of 15000 unlabeled samples

[0062]

[0063] Further, the attribute reduction of the advantage rough set is performed to obtain an intersection of attribute reduction, and necessary elements are found from the intersection, including:

[0064] The heuristic attribute reduction is used to find a reduction intersection core of the decision table in which the decision attribute is the second type of attribute by using an element algorithm HARCC, and necessary elements in the reduction intersection core.

[0065] Further, the new corpus is screened from the unlabeled corpus by using the necessary elements to perform audio sampling and labeling to obtain new training data, including:

[0066] The data in which the decision attribute is the second type of attribute in the unlabeled corpus is sorted by using the necessary elements, and new corpus is screened based on the sorting result, wherein the syllable corresponding to the necessary element in the screened new corpus accounts for a higher proportion than that in the un-screened corpus;

[0067] The new corpus is subjected to audio sampling and labeling to obtain new training data.

[0068] The biggest advantage of the advantage rough set compared with the ordinary rough set is that it can solve the ordered decision information system, in the present application, the number of times that each syllable appears once in the single factor covering information decision system is cumulatively added in turn, thereby forming an ordered decision information system with different importance degrees, and the strong correlation of the syllable in the single phoneme covering decision information system can be found by using the advantage rough set.

[0069] Attribute reduction of rough set is the key to find the most valuable unlabeled samples, but there are more than one minimal set in a decision information system, a decision information system may have multiple reductions, then the intersection of multiple reductions must be found, that is, the core, denoted as core, so as to find the irreplaceable criteria in the core, that is, indispensable criteria. Generally, there are two big methods of attribute reduction of rough set, one is attribute reduction based on discernibility matrix, and the other is heuristic attribute reduction. The present application uses heuristic attribute reduction, and uses HARCC (heuristic attribute reduction with computing core) algorithm to calculate core elements,

[0070] According to the characteristics of the Tibetan phonetic alphabet and the characteristics of the speech synthesis model, in order to cover the Tibetan phonemes as much as possible, the Tibetan speech synthesis model M p The speech synthesis evaluation MOS score is lower than 4, that is, the sample of decision attribute D i =2 is paid special attention, and data is screened from it. That is, the indispensable syllable of decision information system D i =2 is found by HARCC algorithm, and the advantage rough set ordered characteristics are used to screen the unlabeled samples with a large proportion of indispensable syllables, and multiple most valuable training data are screened to obtain new training data N, including the following steps:

[0071] Input: decision system S (data set containing condition attribute and decision attribute), decision system .

[0072] Output: reduction A of system S, that is, the final key attribute set.

[0073] Step 1: initialize the reduction A of system S as an empty set, that is, set ;

[0074] Step 2: calculate the importance of each attribute, and calculate the importance score of each attribute a in classification (for example, the stronger the ability of attribute a to distinguish samples, the higher the score): ;

[0075] Step 3: select core elements (indispensable elements), if the score of an attribute a meets a certain condition (for example, the highest score or reaches a certain threshold), it will be added to the reduction A: if ,

[0076] The attribute a is selected into the reduction A;

[0077] Step 4: initialize the reduction A: ;

[0078] Step 5-7: Add attributes (heuristics selection) iteratively.

[0079] while loop condition: current reduction A has not met the termination condition (e.g., classification accuracy meets the requirement or attribute set size meets the limit), select the most suitable attribute a from the remaining attributes (according to heuristics rules, such as selecting the attribute with the highest score or the attribute with the highest dependency), add attribute a to reduction A.

[0080] Step 8: Attribute verification and deletion (reduction strategy), remove attribute a from reduction A, if the system classification ability does not decrease after deletion (e.g., by verifying the classification consistency of reduction A), it means that attribute a is not necessary and can be deleted. This step ensures that each attribute a in the final reduction A is necessary, i.e., the reduced set is minimized.

[0081] Step 9: Output the final reduction set A.

[0082] This algorithm uses the "add-reduce" strategy of kernel elements, through loop steps 2 and 3, all kernel elements can be obtained, and the importance of each condition attribute C a is calculated to verify the indispensability of each criterion. Through steps 5 to 7, attributes are added to the initialized as the starting point, in each while iteration loop, according to heuristics to select the most suitable criterion until the condition is met, step 8 ensures that the reduced set does not contain redundant attributes by deleting attributes and checking whether the classification ability has decreased.

[0083] Suppose there is a more complex decision table containing 4 condition attributes (A, B, C, D) and 1 decision attribute (MOS score classification).

[0084] Table 2 Decision table

[0085]

[0086] Calculate the importance of each attribute:

[0087] Calculate the importance of each attribute by some algorithm (such as HARCC). Suppose we find:

[0088] The importance of attribute A is 0.6,

[0089] The importance of attribute B is 0.8,

[0090] The importance of attribute C is 0.4,

[0091] The importance of attribute D is 0.2,

[0092] Select the most important attribute:

[0093] Select the most important attribute B to join the reduction set A.

[0094] Check if the classification ability is met:

[0095] Check if all sentences can still be correctly classified when only attribute B is retained. It is found that:

[0096] Sentence 1: B = 1, classified as 1 (high quality);

[0097] Sentence 2: B = 0, classified as 2 (low quality);

[0098] Sentence 3: B = 1, classified as 1 (high quality);

[0099] Sentence 4: B = 1, classified as 1 (high quality);

[0100] Sentence 5: B = 0, classified as 2 (low quality);

[0101] It is found that when only attribute B is retained, all sentences can still be correctly classified. By using only the currently selected attributes (here only attribute B), rejudge whether the MOS classification of each sentence is consistent with the original table. If all classifications are correct, it means that the attribute is sufficient to distinguish different categories, and other attributes can be further reduced.

[0102] Continue to select other important attributes:

[0103] Select the less important attribute A to join the reduction set A.

[0104] Check if all sentences can still be correctly classified when attributes A and B are retained. It is found that:

[0105] Sentence 1: A = 1, B = 1, classified as 1 (high quality);

[0106] Sentence 2: A = 1, B = 0, classified as 2 (low quality);

[0107] Sentence 3: A = 0, B = 1, classified as 1 (high quality);

[0108] Sentence 4: A = 1, B = 1, classified as 1 (high quality);

[0109] Sentence 5: A = 0, B = 0, classified as 2 (low quality);

[0110] When multiple attribute combinations are used, it is necessary to ensure that all possible value combinations can uniquely correspond to the correct classification. It is found that when attributes A and B are retained, all sentences can still be correctly classified.

[0111] Check if other attributes can be removed:

[0112] It is found that after removing attributes C and D, all sentences can still be correctly classified. Therefore, attributes C and D are not necessary.

[0113] Core: The core is the intersection of all reductions and contains irreplaceable attributes. For example, if there are multiple reduction schemes (such as {A, B} or {B, C}), the core is their intersection. In this example, the only feasible reduction is {A, B}, so the core is {A, B}.

[0114] Indispensable Criteria: Indispensable criteria are key attributes in the core that affect classification ability if removed. Although the core is {A, B}, further analysis finds that if A is removed alone (B is retained), correct classification is still possible. However, if B is removed (A is retained), some sentences cannot be distinguished (such as sentences 2 and 5 when B=0). Therefore, in this example, the indispensable criteria is {B}. The new training data is to find a corpus containing B with a syllable symbol ratio greater than the threshold.

[0115] Further, the target language is any one of Tibetan, Naxi, Hani, Qiang, Miao, and Buyi.

[0116] Further, the new training data is used to update the training of the speech synthesis model, and the method further comprises:

[0117] The steps of performing speech synthesis on the unlabeled corpus according to the speech synthesis model, obtaining quality information of the speech synthesis, filtering new corpus from the unlabeled corpus through the indispensable criteria, obtaining new training data by recording sampling and labeling, and updating the speech synthesis model according to the new training data are repeatedly performed until the speech synthesis model meets the convergence condition.

[0118] After obtaining the most beneficial training data, the training data is recorded, sampled and labeled, and then added to the initial training set for incremental training and updating of the speech synthesis model, thereby strengthening the model and making the model more robust.

[0119] The present application collects a large amount of modern Tibetan from websites and books through experiments, establishes a Tibetan corpus, and evaluates the method proposed by the present application based on this. A mature end-to-end speech synthesis model Tacotron-2 is used as the test model for this experiment. The same batch of experts performs MOS scoring on the model on the test set.

[0120] In this experiment, 3000 samples were first selected from the Tibetan corpus as initial training data using statistical methods. Recordings were then sampled and labeled to obtain the initial model M. p 3000 data points were selected from the Tibetan corpus as test data, and the initial model M was tested using two different methods. p The first method for training optimization is to use a random method to select unlabeled samples from the remaining Tibetan corpus as training data to train the initial model M. p The second method is to use the method provided in this invention for obtaining training data for a speech synthesis model to select unlabeled samples from the remaining Tibetan corpus as training data to train the initial model M. p Here, the most favorable training data N is set to 50, meaning that the first fifty data points are selected each time for recording and annotation, and then added to the initial training set. The experimental results are recorded every 1000 and 2000 training data points added, and the results are shown in Tables 3 and 4.

[0121] Table 3 Comparison of random selection and DRS-AL algorithms when 1000 additional training data points are added.

[0122]

[0123] Table 4 Comparison of random selection and DRS-AL algorithms when 2000 additional training data points are added.

[0124]

[0125] As shown in Tables 3 and 4, the experimental results demonstrate that, regardless of the scenario, using the method provided in this invention to select unlabeled data as training data for training the speech synthesis model is more efficient than randomly selecting unlabeled data. This method achieves excellent training results after the number of training data points increases to 2000. In contrast, the random selection method only achieves good results with a large sample size. The proposed method for obtaining training data for speech synthesis models reduces the human and time costs of recording, sampling, and labeling Tibetan speech synthesis, and significantly improves model efficiency.

[0126] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for acquiring training data for a speech synthesis model, characterized in that, include: Select a portion of the corpus containing high-frequency syllables from the target language corpus; The selected corpus is recorded, sampled, and labeled to obtain an initial training set, and the speech synthesis model is trained using the initial training set. The speech synthesis model is used to synthesize speech from the unlabeled corpus to obtain the quality information of the speech synthesis. The decision attributes of the unlabeled corpus are determined based on the quality information, and the occurrence frequency of basic letters in the unlabeled corpus is used as the condition attribute to construct a decision table and obtain the advantage rough set. The rough set of advantages is reduced in terms of attributes to obtain the intersection of the reduced attributes. Necessary elements are found from the intersection to select new corpus from the unlabeled corpus for recording sampling and labeling, thereby obtaining new training data.

2. The method for obtaining training data for a speech synthesis model according to claim 1, characterized in that, The quality information is the MOS score, and the step of determining the decision attributes of the unlabeled corpus based on the quality information includes: For any unlabeled corpus, if the corresponding MOS score is greater than a preset threshold, the decision attribute of the unlabeled corpus is determined to be a first type attribute; if the corresponding MOS score is less than or equal to the preset threshold, the decision attribute of the unlabeled corpus is determined to be a second type attribute.

3. The method for obtaining training data for a speech synthesis model according to claim 1, characterized in that, The conditional attribute, which uses the frequency of occurrence of basic letters in the unlabeled corpus, includes: For any of the unlabeled corpora, the corpus is converted into text represented by syllable symbols of Willy Romanization, and the number of times each Willy Romanization syllable symbol appears in the converted text is determined. The number of times each syllable symbol of the Willy Romanization appears in the text, and the number of times that syllable symbol appears, are used as conditional attributes of the unlabeled corpus.

4. The method for obtaining training data for a speech synthesis model according to claim 2, characterized in that, The step of performing attribute reduction on the dominant rough set to obtain the intersection of the attribute reductions, and finding necessary elements from the intersection, includes: Using heuristic attribute reduction, the HARCC algorithm is used to find the reduced intersection kernel in the decision table that makes the decision attribute a second type of attribute, as well as the necessary elements in the reduced intersection kernel.

5. The method for obtaining training data for a speech synthesis model according to claim 4, characterized in that, The process of selecting new corpus from the unlabeled corpus using the necessary elements, performing recording sampling and labeling, and obtaining new training data includes: The necessary elements are used to sort the data in the unlabeled corpus whose decision attribute is the second type attribute, and new corpus is selected based on the sorting results. The proportion of syllables corresponding to the necessary elements in the selected new corpus is higher than that in the unselected corpus. The new corpus is recorded, sampled, and labeled to obtain new training data.

6. A method for obtaining training data for a speech synthesis model according to any one of claims 1-5, characterized in that, The target language is any one of Tibetan, Naxi, Hani, Qiang, Miao, or Buyi.

7. A method for obtaining training data for a speech synthesis model according to any one of claims 1-5, characterized in that, The new training data is used to update the training of the speech synthesis model, and the method further includes: The process of repeatedly performing speech synthesis on the unlabeled corpus in the corpus according to the speech synthesis model to obtain the quality information of the speech synthesis, to selecting new corpus from the unlabeled corpus through the necessary elements for recording, sampling and labeling to obtain new training data, and updating the speech synthesis model according to the new training data, continues until the speech synthesis model meets the convergence condition.

Images