Text-to-speech pronunciation annotation method and device, equipment and storage medium

Abstract

The application discloses a text-to-speech pronunciation annotation method and device, equipment and storage medium, and relates to the technical field of speech synthesis, which comprises the following steps: receiving a text corpus to be processed, performing text preprocessing on the text corpus to be processed, and obtaining a structured input text; identifying pronunciation risk segments in the structured input text, recording context information and risk categories of the pronunciation risk segments, and obtaining a risk segment identification result; sending the risk segment identification result and the context information to a large model to enable the large model to generate a candidate reading result; verifying the candidate reading result according to a preset pronunciation rule and a domain dictionary to obtain a verified pronunciation annotation; writing an annotation result that passes the review in the verified pronunciation annotation into the domain dictionary, and outputting a structured pronunciation annotation corpus, thereby improving the accuracy of text-to-speech pronunciation annotation and realizing continuous sedimentation and reuse of domain pronunciation knowledge.

Description

Technical Field

[0001] This application relates to the field of speech synthesis technology, and in particular to methods, apparatus, devices and storage media for text-to-speech pronunciation annotation. Background Technology

[0002] With the widespread application of intelligent voice technology in fields such as finance, customer service and public services, enterprises have put forward higher requirements for the pronunciation accuracy of text-to-speech corpora.

[0003] Existing text-to-speech systems mainly rely on fixed rules, static dictionaries, or manual annotation for pronunciation processing. Fixed rules are difficult to cover the pronunciation of number strings and English fragments in complex contexts, static dictionaries cannot cover newly added proprietary terms within enterprises in a timely manner, and manual error correction results have not formed reusable domain pronunciation knowledge, resulting in inaccurate pronunciation annotations in text-to-speech corpora and the recurrence of similar pronunciation errors.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and storage medium for text-to-speech pronunciation annotation, aiming to solve the technical problem of how to improve the accuracy of pronunciation annotation of text-to-speech corpora and realize the reuse of domain pronunciation knowledge.

[0006] To achieve the above objectives, this application proposes a text-to-speech pronunciation annotation method, which includes: Receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text; Identify pronunciation risk segments in the structured input text, and record the context information and risk category of the pronunciation risk segments to obtain the risk segment identification results; The risk segment identification result and the context information are sent to the large model so that the large model can generate candidate reading results; The candidate pronunciation results are verified according to the preset pronunciation rules and the domain dictionary to obtain the verified pronunciation annotations. The verified phonetic annotations that have passed the review are written into the domain dictionary, and the structured phonetic annotation corpus is output.

[0007] In one embodiment, the step of identifying pronunciation risk segments in the structured input text, recording the context information and risk category of the pronunciation risk segments, and obtaining the risk segment identification result includes: The structured input text is subjected to character type recognition to obtain the character type recognition result; Based on the preset pronunciation risk triggering rules and the character type recognition results, the pronunciation risk segments in the structured input text are determined, and the risk segment location information is obtained; Based on the location information of the risk segment, the context information and risk category corresponding to the pronunciation risk segment are extracted to obtain the risk segment identification result.

[0008] In one embodiment, the step of receiving the text corpus to be processed and performing text preprocessing on the text corpus to obtain structured input text includes: Receive the text corpus to be processed and assign a unique identifier to the text corpus to be processed to obtain identifiable corpus; The labeled corpus is processed into sentence segmentation and word segmentation to obtain word segmentation results; The word segmentation results are subjected to character type recognition and punctuation standardization to obtain standardized text; A context window is constructed based on the normalized text to obtain the structured input text.

[0009] In one embodiment, the step of sending the risk segment identification result and the context information to a large model so that the large model generates candidate reading results includes: The risk segment identification result and the context information are input into the large model, so that the large model can perform contextual understanding on the pronunciation risk segment based on the context information to obtain the contextual understanding result; Based on the context understanding results, candidate reading content is output based on the large model; Based on the candidate reading content, the reading type and judgment criteria are output according to the large model; The candidate reading content, the reading type, and the judgment criteria are combined to obtain the candidate reading result.

[0010] In one embodiment, the step of verifying the candidate pronunciation results according to preset pronunciation rules and a domain dictionary to obtain the verified pronunciation annotation includes: The candidate pronunciation results are filtered according to preset pronunciation rules to obtain the rule filtering results; The rule filtering results are matched with the domain dictionary to obtain the dictionary matching results; When the dictionary matching result matches the candidate reading result, the recommendation confidence of the candidate reading result is increased to obtain the first verification result; When the dictionary matching result conflicts with the candidate reading result, the candidate reading result is marked as pending verification, and a second verification result is obtained. Based on the first verification result and the second verification result, the verified pronunciation annotation is obtained.

[0011] In one embodiment, the step of marking the candidate reading result as pending verification and obtaining a second verification result when the dictionary matching result conflicts with the candidate reading result includes: When the dictionary matching result conflicts with the candidate reading result, the candidate reading result is marked as pending review, and the segment to be reviewed is obtained. The segment to be reviewed is sent to the review terminal so that the review terminal can display the original text, candidate readings, recommendation reasons and domain dictionary hit status of the segment to be reviewed, and receive the review confirmation result returned by the review terminal; Based on the audit confirmation results, generate the post-audit read method corresponding to the segment to be reviewed; Based on the post-audit reading method and the first verification result, the second verification result is obtained.

[0012] In one embodiment, the step of writing the approved annotation results from the verified phonetic annotations into a domain dictionary and outputting structured phonetic annotation corpus includes: The verified pronunciation annotations that have passed the review are converted into dictionary candidates. The dictionary candidates are reviewed to obtain the reviewed dictionary entries; The reviewed dictionary entries are written into the domain dictionary to obtain the updated domain dictionary; Based on the updated domain dictionary and the verified phonetic annotations, a structured phonetic annotation corpus is output.

[0013] Furthermore, to achieve the above objectives, this application also proposes a text-to-speech pronunciation annotation device based on self-learning of a large model and a domain dictionary, the device comprising: The corpus access module is used to receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text. The pronunciation risk identification module is used to identify pronunciation risk segments in the structured input text, record the context information and risk category of the pronunciation risk segments, and obtain the risk segment identification result; The candidate reading generation module is used to send the risk segment identification result and the context information to the large model so that the large model can generate candidate reading results. The rule and dictionary verification module is used to verify the candidate pronunciation results according to the preset pronunciation rules and domain dictionary to obtain the verified pronunciation annotation; The dictionary self-learning module is used to write the approved annotation results from the verified phonetic annotations into the domain dictionary and output structured phonetic annotation corpus.

[0014] In addition, to achieve the above objectives, this application also proposes a text-to-speech pronunciation annotation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the text-to-speech pronunciation annotation method as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the text-to-speech pronunciation annotation method described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the text-to-speech pronunciation annotation method described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: By generating candidate pronunciations through a large model combined with context, and using preset pronunciation rules and a domain dictionary for closed-loop verification, the system can process the pronunciation annotation of Chinese polyphonic characters, numerical texts, and English fragments within the same framework, reducing the complexity of maintaining multiple sets of rules in parallel. By automatically converting approved annotation results into domain dictionary entries, enterprise-specific pronunciations can be gradually accumulated and reused, and subsequent similar corpora will prioritize calling domain dictionary results, reducing the workload of repetitive manual review. At the same time, by leveraging the version management and conflict detection mechanism of the domain dictionary, traceable management of pronunciation knowledge can be achieved, thereby improving corpus processing efficiency and pronunciation consistency while ensuring the accuracy of pronunciation annotation. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the first embodiment of the text-to-speech pronunciation annotation method of this application; Figure 2 This is a flowchart of the TTS pronunciation annotation and domain dictionary self-learning process provided in Embodiment 1 of the text-to-speech pronunciation annotation method of this application; Figure 3 This is a flowchart illustrating Embodiment 2 of the text-to-speech pronunciation annotation method of this application; Figure 4 This is a schematic diagram of the module structure of the text-to-speech pronunciation annotation device according to an embodiment of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the text-to-speech pronunciation annotation method in the embodiments of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or text-to-speech pronunciation annotation device capable of the above functions. The following description uses a text-to-speech pronunciation annotation device as an example to illustrate this embodiment and the subsequent embodiments.

[0025] Based on this, embodiments of this application provide a text-to-speech pronunciation annotation method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the text-to-speech pronunciation annotation device method of this application.

[0026] In this embodiment, the text-to-speech pronunciation annotation device method includes steps S10~S50: Step S10: Receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text; like Figure 2 As shown, Figure 2The flowchart for TTS pronunciation annotation and domain dictionary self-learning process is as follows: First, enterprise text corpus is received, followed by text preprocessing and context construction. Then, risky speech segments are identified, and candidate pronunciations are generated through a large model. Next, rule verification and domain dictionary matching are performed. Then, it is determined whether manual review is required. If so, manual review and correction are performed; otherwise, recommended annotation results are generated. Both manual review and correction and the generated recommended annotation results are incorporated into the final structured pronunciation annotation. The approved results are entered into the dictionary candidate pool, and then the domain dictionary is added to the database, version management, and rollback are performed. Finally, the annotated corpus is output for TTS training, quality inspection, or subsequent preprocessing. Among these, the generation of recommended annotation results and the addition of the domain dictionary to the database, version management, and rollback are all fed back to the rule verification and domain dictionary matching stages through the reuse of the dictionary.

[0027] In a preferred embodiment, the present invention can be deployed as an offline annotation tool within an enterprise. This tool is not directly limited to a TTS synthesis engine, but rather outputs structured phonetic annotation corpora, domain dictionary entries, and phonetic risk analysis results for use by subsequent TTS training, quality control, or preprocessing systems.

[0028] It should be noted that the text corpus to be processed refers to enterprise text content that requires phonetic annotation. In this embodiment, it includes customer service scripts, SMS notifications, product descriptions, address information, invoice texts, and product introductions. This corpus needs to undergo accurate phonetic annotation before being converted into speech output to ensure the accuracy of the speech synthesis results.

[0029] It should be noted that text preprocessing refers to the process of segmenting raw text into sentences, words, recognizing character types, standardizing punctuation, and constructing context windows. Through text preprocessing, unstructured raw text can be transformed into structured data that can be used for subsequent pronunciation risk identification and contextual reasoning.

[0030] It should be noted that structured input text refers to standardized data that has been preprocessed and includes corpus identifiers, word segmentation results, character types, punctuation formats, and context window information. This data provides a complete input foundation for subsequent recognition of pronunciation risk segments.

[0031] Understandably, receiving the text corpus to be processed refers to acquiring the text content from external input. Preprocessing the text corpus to be processed involves performing sentence segmentation, word segmentation, character recognition, and context construction on the acquired text content to obtain structured input text.

[0032] Furthermore, the beneficial effect of the above steps is that by assigning a unique identifier to the corpus and constructing a context window, subsequent processing can accurately trace the original corpus and make pronunciation judgments in combination with the context.

[0033] In one feasible implementation, step S10 may include steps A11 to A14: Step A11: Receive the text corpus to be processed and assign a unique identifier to the text corpus to be processed to obtain the identifiable corpus; It should be noted that the unique identifier is a non-repeating number assigned to each piece of corpus data, used to distinguish and trace different corpora in subsequent processing. This identifier is used throughout the entire corpus processing workflow to ensure that each processing stage can correspond to the original corpus data.

[0034] It should be noted that labeled corpora refer to text corpora to be processed that have been assigned unique identifiers. These unique identifiers allow for accurate location and association of processing results for each corpus in subsequent processing.

[0035] As is understandable, receiving the text corpus to be processed refers to acquiring text content from external input. Assigning a unique identifier to the text corpus to be processed means assigning a unique number to each piece of acquired text content, resulting in identifiable corpus.

[0036] Furthermore, the beneficial effect of the above steps is that they enable full-process traceability of the corpus through unique identifiers, which facilitates subsequent review and dictionary source recording.

[0037] Step A12: Perform sentence segmentation and word segmentation on the labeled corpus to obtain the word segmentation results; It should be noted that sentence segmentation refers to the process of dividing text into multiple sentence units according to punctuation marks. Sentence segmentation can break down long texts into independent semantic units, making it easier to identify pronunciation risks for individual sentences later.

[0038] It should be noted that word segmentation refers to the process of dividing a sentence into the smallest lexical units with independent semantic meaning. The results of word segmentation can clearly define the boundaries of each word, providing a basis for identifying polyphonic characters and proper nouns.

[0039] It should be noted that the word segmentation result refers to the word sequence obtained after sentence segmentation and word segmentation. This sequence retains the word order and identification information of the original text and can be directly used for subsequent character type analysis.

[0040] Understandably, processing marked corpora into sentences and words involves dividing the marked corpora into sentences according to punctuation and then into words to obtain the word segmentation results.

[0041] Furthermore, the beneficial effect of the above steps is that by segmenting sentences and words, the original text is broken down into processable lexical units, providing a foundation for subsequent character type recognition and pronunciation risk localization.

[0042] Step A13: Perform character type recognition and punctuation standardization on the word segmentation results to obtain standardized text; It should be noted that character type recognition refers to the process of determining whether each character in a text belongs to a category such as Chinese character, number, English letter, or punctuation mark. Character type recognition can preliminarily determine the types of potentially problematic segments in the text.

[0043] It should be noted that punctuation standardization refers to the process of converting irregular punctuation marks in text into standard punctuation marks. A unified punctuation format helps reduce sentence segmentation errors and pronunciation ambiguities caused by punctuation differences.

[0044] It should be noted that standardized text refers to standard text that has undergone character type recognition and punctuation standardization. This text has a unified character classification and punctuation format, and can be directly used to construct context windows.

[0045] Understandably, character type identification and punctuation standardization of word segmentation results refers to identifying the category of each character in the word segmentation results and standardizing the punctuation mark format to obtain standardized text.

[0046] Furthermore, the beneficial effect of the above steps is that they can initially screen out the types of pronunciation risk segments by character type identification and eliminate processing errors caused by punctuation differences by punctuation standardization.

[0047] Step A14: Construct a context window based on the normalized text to obtain the structured input text.

[0048] It should be noted that the context window refers to the pre-defined range of text surrounding the target text segment, used to provide contextual information. The construction of the context window enables subsequent processing to determine the pronunciation of polyphonic characters and number strings based on the context.

[0049] Understandably, constructing a context window from normalized text refers to extracting the surrounding text of each segment from the normalized text as context, resulting in structured input text. This structured input text contains the original content, character types, and contextual information, and can be directly used for identifying pronunciation risk segments.

[0050] Furthermore, the beneficial effect of the above steps is that by constructing a context window, contextual support is provided for subsequent large-scale model reasoning and pronunciation judgment, thereby improving the accuracy of pronunciation annotation.

[0051] In its implementation, the system receives enterprise text corpora such as customer service scripts, SMS notifications, product descriptions, address information, billing texts, and product introductions. The system assigns a unique identifier to each corpus entry and performs processing such as sentence segmentation, word segmentation, character type recognition, punctuation standardization, and context window construction to form structured input that can be used for subsequent pronunciation risk identification and contextual reasoning.

[0052] Step S20: Identify pronunciation risk segments in the structured input text, and record the context information and risk category of the pronunciation risk segments to obtain the risk segment identification results; It should be noted that "pronunciation risk segments" refer to text segments that may cause pronunciation errors during text-to-speech conversion. In this embodiment, these include Chinese characters with multiple pronunciations, phone numbers, amounts, dates, addresses, product numbers, order numbers, English words, English abbreviations, website addresses, email addresses, and brand / model names, etc.

[0053] It should be noted that contextual information refers to the surrounding text content of the pronunciation risk segment. This information is consistent with the contextual information used by the large model when generating candidate pronunciations, and is used to help determine the correct pronunciation of the risk segment in a specific context.

[0054] It should be noted that risk category refers to the type of pronunciation risk segment, such as Chinese characters with multiple pronunciations, phone numbers, amounts, dates, address numbers, product numbers, order numbers, English words, English abbreviations, website addresses, email addresses, or brand models. Different risk categories correspond to different pronunciation handling strategies.

[0055] It should be noted that the risk segment identification result refers to comprehensive information including the location, contextual information, and risk category of the pronunciation risk segment. This result serves as the basic input for the subsequent generation of candidate pronunciations by the large model.

[0056] Understandably, identifying pronunciation risk segments in structured input text refers to locating segments within the structured input text that may contain pronunciation errors. Recording the contextual information and risk category of the pronunciation risk segments involves extracting the text surrounding the risk segment and determining its type to obtain the risk segment identification result.

[0057] Furthermore, the beneficial effect of the above steps is that by identifying pronunciation risk segments and recording their contextual information and risk categories, it provides a precise location and classification basis for subsequent large-scale model generation of candidate pronunciations.

[0058] In its implementation, the system identifies text fragments that may cause TTS pronunciation errors, including Chinese characters with multiple pronunciations, phone numbers, amounts, dates, addresses, product numbers, order numbers, English words, English abbreviations, website addresses, email addresses, and brand / model numbers. For each risky fragment, the system records its location in the original text, risk category, context, triggering rules, and initial confidence level.

[0059] In one feasible implementation, step S20 may include steps A21 to A23: Step A21: Perform character type recognition on the structured input text to obtain the character type recognition results; It should be noted that the character type recognition result refers to the character category information obtained after character type recognition. This result can be directly used to match preset pronunciation risk triggering rules, thereby locating text segments that may contain pronunciation errors.

[0060] Understandably, character type recognition of structured input text refers to determining the category of each character in the structured input text and obtaining the character type recognition result.

[0061] Furthermore, the beneficial effect of the above steps is that character type recognition can quickly filter out character combinations that may contain pronunciation risks, thereby improving the efficiency of locating pronunciation risk segments.

[0062] Step A22: Based on the preset pronunciation risk triggering rules and character type recognition results, determine the pronunciation risk segments in the structured input text and obtain the risk segment location information; It should be noted that the preset pronunciation risk trigger rules refer to the judgment conditions used to identify text segments that may cause pronunciation errors. These rules are set according to character type and combination pattern, and can identify risk types such as Chinese polyphonic characters, number strings, English fragments, and proper nouns.

[0063] It should be noted that the risk segment location information refers to the start and end positions of the pronunciation risk segment in the original text. This information is used to accurately locate and extract the risk segment and its context from the original text.

[0064] Understandably, determining the pronunciation risk segment in the structured input text based on the preset pronunciation risk triggering rules and character type recognition results means filtering out text segments that may have pronunciation errors according to the judgment conditions and obtaining the location information of the risk segment.

[0065] Furthermore, the beneficial effect of the above steps is that they automatically locate pronunciation risk segments through preset rules, reducing the workload of manual inspection.

[0066] Step A23: Based on the location information of the risk segment, extract the context information and risk category corresponding to the pronunciation risk segment to obtain the risk segment recognition result.

[0067] Understandably, extracting the context information and risk category corresponding to the pronunciation risk segment based on the risk segment location information means extracting the text surrounding the risk segment based on the location information and determining its type to obtain the risk segment recognition result.

[0068] Furthermore, the beneficial effect of the above steps is that by accurately locating and extracting the context and type of risk fragments, complete input information is provided for subsequent large-scale model inference.

[0069] Step S30: Send the risk segment identification results and context information to the large model so that the large model can generate candidate reading results; It should be noted that the large-scale model refers to a large-scale language model with contextual understanding capabilities. This model can infer the pronunciation of polyphonic characters, the pronunciation of number sequences, and the pronunciation of English passages based on the context of the input text, and generate candidate pronunciation content.

[0070] It should be noted that the candidate pronunciation result refers to a comprehensive result that includes the candidate pronunciation content, pronunciation type, and judgment criteria. This result is a complete pronunciation suggestion generated by the large model based on contextual understanding.

[0071] Understandably, sending the risk segment identification results and contextual information to the large model means sending the location, category, and surrounding text of the risk segment to the large model so that the large model can generate pronunciation suggestions based on the contextual information and obtain candidate pronunciation results.

[0072] Furthermore, the beneficial effect of the above steps is that they utilize the contextual understanding capabilities of large models to generate candidate readings, thus compensating for the shortcomings of fixed rules in covering complex contexts.

[0073] In its implementation, the system inputs risk segments and their context into a large model, which then outputs candidate readings, recommended readings, reading types, and judgment criteria. For example, in "Repayment Rules," "repay" should be read as huan3; "10086" should be read digit by digit in the context of a customer service call; and "A12" in "A12 Product" can be read as a sequence of letters and numbers.

[0074] In one feasible implementation, step S30 may include steps A31 to A34: Step A31: Input the risk segment identification results and context information into the large model so that the large model can perform contextual understanding of the pronunciation risk segment based on the context information and obtain the contextual understanding result; It should be noted that contextual understanding refers to the large model inferring the most likely pronunciation of a segment in the current context based on the surrounding text. This process relies on the large model's deep learning capabilities in semantics and context.

[0075] It should be noted that the contextual understanding result refers to the preliminary judgment on the pronunciation risk of a segment obtained by the large model after contextual reasoning. This result provides a basis for subsequently outputting candidate pronunciations.

[0076] Understandably, inputting the risk segment identification results and contextual information into the large model means sending the location, category, and surrounding text of the risk segment to the large model so that the large model can perform contextual understanding of the pronunciation risk segment based on the contextual information and obtain the contextual understanding result.

[0077] Further, the beneficial effect of the above step is that the accuracy of pronunciation judgment of polyphonic characters and digit strings in complex contexts is improved through the in-depth understanding of the context by the large model.

[0078] Step A32: Output candidate pronunciation content based on the large model according to the context understanding result; It should be noted that the candidate pronunciation content refers to the pronunciation representation generated by the large model for the pronunciation risk segment. In this embodiment, for example, for the character "hai" in the repayment rule, the large model outputs a candidate pronunciation that is pronounced as "huan".

[0079] It can be understood that outputting the candidate pronunciation content based on the large model according to the context understanding result means that the large model generates the pronunciation representation of the pronunciation risk segment according to the context reasoning result to obtain the candidate pronunciation content.

[0080] Further, the beneficial effect of the above step is that the candidate pronunciations are generated by the large model to provide pronunciation references for subsequent rule verification and manual review.

[0081] Step A33: Output a pronunciation type and a judgment basis based on the large model according to the candidate pronunciation content; It should be noted that the pronunciation type refers to the category to which the candidate pronunciation belongs, that is, reading digit by digit, reading by numerical value, reading by letter, or reading continuously by word, etc. The pronunciation type is used to distinguish the pronunciation modes of the same text segment in different scenarios.

[0082] It should be noted that the judgment basis refers to the context reasoning reason on which the large model generates the candidate pronunciation. This basis is used to assist reviewers in understanding the generation logic of the candidate pronunciation.

[0083] It can be understood that outputting the pronunciation type and the judgment basis based on the large model according to the candidate pronunciation content means that the large model outputs the category and the reasoning reason of the pronunciation while generating the pronunciation representation, so as to obtain the pronunciation type and the judgment basis.

[0084] Further, the beneficial effect of the above step is that by outputting the pronunciation type and the judgment basis, clear classification and reasoning support are provided for subsequent manual review and rule verification.

[0085] Step A34: Combine the candidate pronunciation content, the pronunciation type and the judgment basis to obtain a candidate pronunciation result.

[0086] It can be understood that combining the candidate pronunciation content, the pronunciation type and the judgment basis refers to integrating the pronunciation representation, category and reasoning reason generated by the large model into a unified format to obtain the candidate pronunciation result. This result contains complete pronunciation suggestions and supporting information, and can be directly used for subsequent rule verification.

[0087] Furthermore, the beneficial effect of the above steps is that they integrate the scattered reading information into a standardized format, which facilitates unified processing by subsequent modules.

[0088] Step S40: Verify the candidate pronunciation results according to the preset pronunciation rules and domain dictionary to obtain the verified pronunciation annotation; It should be noted that preset pronunciation rules refer to pronunciation constraints set for specific types of text fragments. For example, mobile phone numbers are read digit by digit, monetary amounts are read numerically and by currency, and address numbers are read according to the numerical rules in the address context.

[0089] It should be noted that the domain dictionary refers to a database that stores enterprise-specific readings. This dictionary contains information such as text fragments, standard readings, reading types, applicable contexts, source corpus, reviewers, review dates, and version numbers, and is used to verify candidate readings generated by the large model.

[0090] It should be noted that verification refers to the process of filtering obviously erroneous results through preset pronunciation rules and matching them with enterprise-specific pronunciations in the domain dictionary to verify the accuracy of candidate pronunciations. When the domain dictionary result matches the candidate pronunciation, the recommendation confidence is increased; when there is a conflict, it is marked as pending review.

[0091] It should be noted that phonetic annotation refers to the verified phonetic information, including the text segment, pronunciation type, standard pronunciation, and applicable context. This verified phonetic annotation can be directly used for speech synthesis or enter the review process.

[0092] Understandably, verifying candidate pronunciation results based on preset pronunciation rules and domain dictionaries means using pronunciation constraints to filter out erroneous results and match enterprise-specific pronunciations to obtain verified pronunciation annotations.

[0093] Furthermore, the beneficial effect of the above steps is that, through rule filtering and domain dictionary matching, the accuracy of pronunciation annotation is further ensured on the basis of candidate pronunciations in the large model, and the historical review results are reused by utilizing the domain dictionary.

[0094] In its implementation, the system filters obviously erroneous results using preset rules. For example, mobile phone numbers are read digit by digit or group by digit, monetary amounts are read numerically and by currency, and address numbers are read according to the numerical rules specific to the address context. Simultaneously, the system matches enterprise-specific pronunciations from the domain dictionary. When the domain dictionary results match the candidate results from the large model, the recommendation confidence is improved; when the two conflict, the relevant segments are marked as requiring manual review.

[0095] Step S50: Write the verified phonetic annotation results into the domain dictionary and output the structured phonetic annotation corpus.

[0096] It should be noted that the approved annotation results refer to pronunciation annotation information that has been manually or automatically verified as correct. This result contains accurate pronunciation and complete contextual information, and can be directly converted into a domain dictionary entry.

[0097] It should be noted that structured phonetic annotation corpus refers to standardized data that includes phonetic annotation results, a phonetic risk list, domain dictionary update records, and review logs. This corpus can be used for speech synthesis training, quality control, or preprocessing.

[0098] Understandably, writing the verified phonetic annotations into the domain dictionary means storing the confirmed pronunciation information in the domain dictionary. Outputting structured phonetic annotation corpus means organizing the final phonetic annotation results and related records into a standardized format for output.

[0099] Furthermore, the beneficial effect of the above steps is that by writing the approved annotation results into the domain dictionary, the enterprise's proprietary pronunciation can be continuously accumulated and reused, while outputting standardized corpus to support subsequent speech synthesis applications.

[0100] In one feasible implementation, step S50 may include steps A41-A44: Step A41: Convert the verified pronunciation annotations that have passed the review into dictionary candidate options; It should be noted that dictionary candidates refer to dictionary entries to be added to the database, which are generated from the approved annotation results. These candidates include the text fragment, standard pronunciation, pronunciation type, applicable context, source corpus, reviewer, review date, and version number.

[0101] Understandably, converting the verified pronunciation annotations into dictionary candidates means organizing the confirmed pronunciation information into a dictionary entry format to be added to the database, thus obtaining dictionary candidates.

[0102] Furthermore, the beneficial effect of the above steps is that by transforming the annotation results into standardized dictionary candidates, preparation is made for the standardized inclusion of domain dictionaries into the database.

[0103] Step A42: Review the dictionary candidates to obtain the reviewed dictionary entries; It should be noted that the reviewed dictionary entries refer to pronunciation knowledge entries that have undergone further review and confirmation before being included in the domain dictionary. These entries possess complete source traceability information, ensuring the accuracy and reliability of the domain dictionary.

[0104] Understandably, reviewing dictionary candidates means confirming the converted entries to be added to the database, thus obtaining the reviewed dictionary entries.

[0105] Furthermore, the beneficial effect of the above steps is that by reviewing the entries again, the accuracy of the dictionary entries is ensured, and errors in pronunciation are avoided from contaminating the domain dictionary.

[0106] Step A43: Write the approved dictionary entries into the domain dictionary to obtain the updated domain dictionary; It should be noted that the updated domain dictionary refers to the domain dictionary after the new, reviewed entries have been added. This dictionary can be used for rule validation of subsequent similar corpora, enabling the reuse of pronunciation knowledge.

[0107] Understandably, writing the reviewed dictionary entries into the domain dictionary means storing the confirmed pronunciation knowledge entries into the domain dictionary to obtain an updated domain dictionary.

[0108] Furthermore, the beneficial effect of the above steps is that by continuously adding new dictionary entries, the domain dictionary gradually covers the proprietary pronunciations newly added within the enterprise.

[0109] Step A44: Output structured phonetic annotation corpus based on the updated domain dictionary and the verified phonetic annotations.

[0110] Understandably, outputting structured phonetic annotation corpus based on the updated domain dictionary and verified phonetic annotations means organizing the final phonetic annotation results, domain dictionary update records, and related logs into a standardized output format.

[0111] Furthermore, the beneficial effect of the above steps is that by outputting standardized corpus containing dictionary update records, a traceable data foundation is provided for subsequent speech synthesis training and quality inspection.

[0112] In practice, reviewers examine the original text, risky segments, candidate pronunciations, reasons for recommendation, domain dictionary matching, and historical similar cases to confirm, modify, or reject recommended pronunciations. After confirmation, a structured pronunciation annotation result is generated, which includes at least the segment text, pronunciation type, standard pronunciation, applicable context, review status, reviewer, and review time.

[0113] The system converts manually approved annotations into dictionary candidates. Each dictionary entry includes at least the text fragment, standard pronunciation, pronunciation type, applicable context, source corpus, reviewer, review date, and version number. After review, the entries are added to the domain dictionary. When the same or similar corpus appears again, the system prioritizes using the domain dictionary results, thereby reducing repetitive manual review work and improving pronunciation consistency.

[0114] The system outputs structured annotated corpus, a pronunciation risk list, domain dictionary update records, and necessary review logs. These outputs can be used for TTS training, TTS preprocessing rule generation, pronunciation dictionary maintenance, speech quality inspection, and enterprise corpus governance, and are not limited to specific TTS vendors or specific speech synthesis engines.

[0115] This solution can also be implemented as a system. As shown in Table 1 below, this system includes at least the following modules: Table 1

[0116] This embodiment provides a text-to-speech pronunciation annotation method. By receiving the text corpus to be processed and performing sentence segmentation, word segmentation, character type recognition, punctuation standardization, and context window construction, it provides a structured input foundation containing unique identifiers and contextual information for subsequent pronunciation risk identification, thereby ensuring the traceability of the corpus throughout the entire process. By identifying pronunciation risk segments in the structured input text and recording their contextual information and risk categories, it can accurately locate potential pronunciation errors such as Chinese polyphonic characters, phone numbers, amounts, English fragments, and company-specific terms. By sending the risk segment identification results and contextual information to a large model, the large model's contextual understanding capabilities are used to generate candidate pronunciations, pronunciation types, and judgment criteria, effectively compensating for the inability of fixed rules to cover complex contexts and improving the accuracy of pronunciation judgment for polyphonic characters and number strings in specific contexts. The candidate pronunciation results are verified by pre-set pronunciation rules and a domain dictionary. When the domain dictionary matches, historical review results are reused first to increase recommendation confidence. In conflict scenarios, a pending review status is marked, thus balancing processing efficiency and pronunciation accuracy. By reviewing approved annotations and then adding them to the domain dictionary after approval by dictionary candidates, the system achieves continuous accumulation and version management of enterprise-specific pronunciations. When similar corpora reappear, the domain dictionary results are prioritized, reducing repetitive manual review workload and improving pronunciation consistency. The final output includes structured pronunciation annotation corpora, a pronunciation risk list, domain dictionary update records, and review logs, providing standardized data support for speech synthesis training, corpus quality inspection, and preprocessing.

[0117] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S40 includes steps S401 to S405: Step S401: Filter the candidate pronunciation results according to the preset pronunciation rules to obtain the rule filtering results; It should be noted that preset pronunciation rules refer to pronunciation constraints set for specific types of text fragments. In this embodiment, for example, mobile phone numbers are read digit by digit first, monetary amounts are read numerically and by currency first, and address numbers can be read according to the numeric rules in the address context. These rules are preset based on character type and common pronunciation patterns to filter out candidate pronunciations that clearly do not conform to basic pronunciation standards.

[0118] It should be noted that the rule filtering result refers to the candidate pronunciation results retained after being filtered by preset pronunciation rules. This result eliminates pronunciation suggestions that are clearly inconsistent with the basic pronunciation constraints, and retains candidate pronunciations that conform to the general pronunciation rules for subsequent matching with the domain dictionary.

[0119] Understandably, filtering candidate pronunciation results based on preset pronunciation rules means using pronunciation constraints to screen out obviously incorrect pronunciation suggestions and obtain the rule-filtered results.

[0120] Furthermore, the beneficial effect of the above steps is that they can quickly filter out unreasonable candidate pronunciations by setting up pronunciation rules, thereby reducing the workload of subsequent matching and review.

[0121] Step S402: Match the rule filtering results with the domain dictionary to obtain the dictionary matching results; It should be noted that the dictionary matching result refers to the comparison conclusion obtained by comparing the rule filtering result with the enterprise-specific pronunciation in the domain dictionary. This result is used to determine whether the candidate pronunciation is consistent with the enterprise pronunciation confirmed in the historical review, thereby determining whether the historical review result can be directly reused.

[0122] Understandably, matching the rule-filtered results with the domain dictionary means comparing the rule-filtered reading suggestions with the enterprise's proprietary reading database to obtain the dictionary matching results.

[0123] Furthermore, the beneficial effect of the above steps is that by reusing historically verified enterprise-specific pronunciations through matching domain dictionaries, the consistency and processing efficiency of pronunciation annotation are improved.

[0124] Step S403: When the dictionary matching result is consistent with the candidate reading result, increase the recommendation confidence of the candidate reading result to obtain the first verification result; It should be noted that the recommendation confidence score is a numerical marker that characterizes the reliability of the candidate reading result. When the dictionary matching result matches the candidate reading result, it indicates that the reading has been confirmed by historical review, and therefore its recommendation confidence score is increased to indicate that the reading has high reliability.

[0125] It should be noted that the first verification result refers to the verification conclusion formed when the dictionary matching result matches the candidate pronunciation result. This result includes candidate pronunciations after increasing the recommendation confidence, which can be directly used as the basis for pronunciation annotation.

[0126] Understandably, when the dictionary matching result is consistent with the candidate reading result, increasing the recommendation confidence of the candidate reading result means that when the enterprise-specific reading is the same as the candidate reading, the reliability label of the reading is increased to obtain the first verification result.

[0127] Furthermore, the beneficial effect of the above steps is to increase confidence when the domain dictionary is used to match the target word, to prioritize the use of verified enterprise-specific reading methods, and to reduce unnecessary duplicate verification.

[0128] Step S404: When the dictionary matching result conflicts with the candidate reading result, mark the candidate reading result as pending verification and obtain the second verification result; It should be noted that the "pending review" status refers to a state where a candidate reading result conflicts with the enterprise-specific reading in the domain dictionary, marking the candidate reading as requiring further manual confirmation. This status is used to distinguish between readings that can be used directly and those that require review and confirmation, preventing unconfirmed readings from directly entering the output stage.

[0129] It should be noted that the second verification result refers to the verification conclusion formed when the dictionary matching result conflicts with the candidate pronunciation result. This result includes candidate pronunciations marked as pending review, which need to be manually reviewed to determine the final pronunciation in order to ensure the accuracy of the pronunciation annotation.

[0130] Understandably, when the dictionary matching result conflicts with the candidate reading result, marking the candidate reading result as pending verification means that when the enterprise-specific reading is inconsistent with the candidate reading, the reading is marked as needing further confirmation to obtain a second verification result.

[0131] Furthermore, the beneficial effect of the above steps is that by marking conflicting readings as pending verification, unconfirmed readings are prevented from directly entering the output stage, thus ensuring the accuracy of pronunciation annotation.

[0132] In one feasible implementation, step S404 may include steps A51 to A54: Step A51: When the dictionary matching result conflicts with the candidate reading result, mark the candidate reading result as pending review and obtain the segment to be reviewed; Understandably, when dictionary matching results conflict with candidate reading results, marking candidate reading results as pending review means marking the conflicting readings as requiring confirmation, resulting in a segment awaiting review. This segment includes the original text, candidate readings, and conflict information, serving as the object of manual review.

[0133] Furthermore, the beneficial effect of the above steps is that by clearly identifying the segments to be reviewed, reviewers can quickly locate the text locations where there are pronunciation conflicts.

[0134] Step A52: Send the segment to be reviewed to the review terminal so that the review terminal can display the original text, candidate readings, recommendation reasons and domain dictionary hit status of the segment to be reviewed, and receive the review confirmation result returned by the review terminal; It should be noted that the review terminal refers to the device or interface used by reviewers to view and confirm the phonetic transcription results. This terminal displays the original text of the passage to be reviewed, candidate pronunciations, reasons for recommendation, and the accuracy of the domain dictionary, providing reviewers with complete criteria for judgment.

[0135] It should be noted that the review confirmation result refers to the result returned by the reviewer after reviewing the passage to be reviewed, confirming, modifying, or rejecting the candidate readings. This result is used to determine the final reading of the passage to be reviewed and serves as the basis for generating the post-review reading.

[0136] Understandably, sending the segment to be reviewed to the review terminal means pushing the segment with conflicting readings to the reviewers for viewing, so that the review terminal can display the original text, candidate readings, recommendation reasons, and domain dictionary hit status of the segment to be reviewed, and receive the review confirmation result returned by the review terminal.

[0137] Furthermore, the beneficial effect of the above steps is that by displaying the complete context and reasons for recommendation through the review terminal, reviewers can make accurate judgments about the reading comprehension.

[0138] Step A53: Based on the review and confirmation results, generate the post-review reading method corresponding to the segment to be reviewed; It should be noted that the post-review pronunciation refers to the final pronunciation generated based on the review and confirmation results. This pronunciation may be a candidate pronunciation confirmed by the reviewers, or it may be a new pronunciation revised by the reviewers, serving as the final phonetic transcription for this segment.

[0139] Understandably, generating the post-review reading method corresponding to the segment to be reviewed, based on the review confirmation results, means determining the final pronunciation of the segment based on the confirmation or modification results returned by the reviewer, and thus obtaining the post-review reading method.

[0140] Furthermore, the beneficial effect of the above steps is that by manually reviewing and confirming or correcting candidate pronunciations, the problem of pronunciation accuracy in scenarios where the output of large models is unstable and the domain dictionary is not covered can be solved.

[0141] Step A54: Based on the post-audit reading method and the first verification result, obtain the second verification result.

[0142] Understandably, the second verification result, based on the manually verified reading and the first verification result, refers to merging the manually verified reading with the first verification result generated when the dictionary hits the word, resulting in a complete second verification result. This second verification result includes the final manually verified reading and the confidence level information at the time of dictionary hit.

[0143] Furthermore, the beneficial effect of the above steps is that they incorporate the results of manual review into the verification system, so that the second verification result includes the final pronunciation confirmed by humans, thus ensuring the reliability of the output pronunciation annotation.

[0144] Step S405: Based on the first verification result and the second verification result, obtain the verified pronunciation annotation.

[0145] Understandably, based on the first and second verification results, the verified pronunciation annotation refers to integrating the pronunciation results with increased confidence from dictionary hits with the pronunciation results confirmed by manual review, resulting in the final verified pronunciation annotation. This verified pronunciation annotation includes the final pronunciation and applicable context for all pronunciation risk segments.

[0146] Furthermore, the beneficial effect of the above steps is that by integrating the results of rule filtering, dictionary matching, and manual review, accurate and reliable final pronunciation annotations are formed, providing a high-quality corpus foundation for subsequent speech synthesis.

[0147] This embodiment provides a text-to-speech pronunciation annotation method. By filtering candidate pronunciation results through preset pronunciation rules, it can eliminate pronunciation suggestions that clearly do not conform to basic pronunciation constraints, reducing the workload of subsequent matching and review. By matching the rule-filtered results with a domain dictionary, historically confirmed results can be directly reused when enterprise-specific pronunciations match candidate pronunciations, improving the recommendation confidence of the corresponding pronunciations and thus balancing processing efficiency and pronunciation consistency. When there is a conflict between the dictionary matching results and candidate pronunciation results, a pending review status is marked, and the segment to be reviewed is sent to the review terminal to display the original text, candidate pronunciations, recommendation reasons, and domain dictionary hit status. This allows reviewers to make accurate judgments based on complete contextual information and generate a post-reviewed pronunciation based on the review confirmation results, effectively compensating for the instability of large model output and the lack of accuracy in scenarios not covered by the domain dictionary. Finally, the first and second verification results are integrated to form the verified pronunciation annotation, ensuring that the output includes both directly reusable historically confirmed pronunciations and the final pronunciations reviewed by humans, providing a highly reliable corpus foundation for subsequent speech synthesis.

[0148] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the text-to-speech pronunciation annotation method of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0149] This application also provides a text-to-speech pronunciation annotation device, please refer to... Figure 4 The text-to-speech pronunciation annotation device includes: The corpus access module 10 is used to receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text. The pronunciation risk identification module 20 is used to identify pronunciation risk segments in structured input text, record the context information and risk category of the pronunciation risk segments, and obtain the risk segment identification results; The candidate reading generation module 30 is used to send the risk segment identification results and context information to the large model so that the large model can generate candidate reading results. The rule and dictionary verification module 40 is used to verify the candidate pronunciation results according to the preset pronunciation rules and domain dictionary, and obtain the verified pronunciation annotation; The dictionary self-learning module 50 is used to write the verified pronunciation annotation results into the domain dictionary and output structured pronunciation annotation corpus.

[0150] The text-to-speech pronunciation annotation device provided in this application, employing the text-to-speech pronunciation annotation method in the above embodiments, can solve the technical problem of how to improve the accuracy of pronunciation annotation of text-to-speech corpora and realize the reuse of domain pronunciation knowledge. Compared with the prior art, the beneficial effects of the text-to-speech pronunciation annotation device provided in this application are the same as the beneficial effects of the text-to-speech pronunciation annotation method provided in the above embodiments, and other technical features in the text-to-speech pronunciation annotation device are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0151] In one embodiment, the pronunciation risk identification module 20 is also used to perform character type identification on the structured input text to obtain the character type identification result; Based on the preset pronunciation risk triggering rules and character type recognition results, the pronunciation risk segments in the structured input text are identified, and the location information of the risk segments is obtained; Based on the location information of the risk segment, the context information and risk category corresponding to the pronunciation risk segment are extracted to obtain the risk segment identification result.

[0152] In one embodiment, the corpus access module 10 is further configured to receive the text corpus to be processed and assign a unique identifier to the text corpus to be processed to obtain identifiable corpus; The labeled corpus is processed into sentence segmentation and word segmentation to obtain the word segmentation results; The word segmentation results are processed by character type recognition and punctuation standardization to obtain standardized text; A context window is constructed based on the normalized text to obtain the structured input text.

[0153] In one embodiment, the candidate reading generation module 30 is further configured to input the risk segment identification result and context information into the large model, so that the large model can perform contextual understanding of the pronunciation risk segment based on the context information and obtain the contextual understanding result; Based on the context understanding results, candidate reading content is output based on the large model; Based on the candidate reading content, the reading type and judgment criteria are output based on the large model. The candidate reading content, reading type, and judgment criteria are combined to obtain the candidate reading results.

[0154] In one embodiment, the rule and dictionary verification module 40 is further configured to filter the candidate pronunciation results according to preset pronunciation rules to obtain rule filtering results; The rule filtering results are matched with the domain dictionary to obtain the dictionary matching results; When the dictionary matching result matches the candidate reading result, the recommendation confidence of the candidate reading result is increased to obtain the first verification result; When the dictionary matching result conflicts with the candidate reading result, the candidate reading result is marked as pending verification, and a second verification result is obtained. Based on the first and second verification results, the verified pronunciation annotations are obtained.

[0155] In one embodiment, the rule and dictionary verification module 40 is further used to mark the candidate reading result as pending verification when the dictionary matching result conflicts with the candidate reading result, thereby obtaining the segment to be verified; Send the segment to be reviewed to the review terminal so that the review terminal can display the original text, candidate readings, recommendation reasons and domain dictionary hit status of the segment to be reviewed, and receive the review confirmation result returned by the review terminal; Based on the review and confirmation results, generate the post-review reading method corresponding to the segment to be reviewed; Based on the post-audit reading method and the first verification result, the second verification result is obtained.

[0156] In one embodiment, the dictionary self-learning module 50 is also used to convert the verified pronunciation annotations that have passed the review into dictionary candidate options; The dictionary candidates are reviewed to obtain the reviewed dictionary entries; The reviewed dictionary entries are written into the domain dictionary to obtain the updated domain dictionary; Based on the updated domain dictionary and the validated phonetic annotations, output structured phonetic annotation corpus.

[0157] This application provides a text-to-speech pronunciation annotation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the text-to-speech pronunciation annotation method in the above embodiment 1.

[0158] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing the text-to-speech pronunciation annotation device of the embodiments of this application. The text-to-speech pronunciation annotation device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The text-to-speech pronunciation annotation device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0159] like Figure 5As shown, the text-to-speech pronunciation annotation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the text-to-speech pronunciation annotation device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the text-to-speech transcription device to communicate wirelessly or wiredly with other devices to exchange data. Although text-to-speech transcription devices with various systems are shown in the figures, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems can be implemented alternatively.

[0160] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0161] The text-to-speech pronunciation annotation device provided in this application, employing the text-to-speech pronunciation annotation method in the above embodiments, can solve the technical problem of how to improve the accuracy of pronunciation annotation of text-to-speech corpora and realize the reuse of domain pronunciation knowledge. Compared with the prior art, the beneficial effects of the text-to-speech pronunciation annotation device provided in this application are the same as the beneficial effects of the text-to-speech pronunciation annotation device method provided in the above embodiments, and other technical features in this text-to-speech pronunciation annotation device are the same as the features disclosed in the previous embodiment method, and will not be repeated here.

[0162] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0163] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0164] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the text-to-speech pronunciation annotation method in the above embodiments.

[0165] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0166] The aforementioned computer-readable storage medium may be included in a text-to-speech pronunciation annotation device; or it may exist independently and not be assembled into a text-to-speech pronunciation annotation device.

[0167] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the text-to-speech pronunciation annotation device, they cause the following to occur: receive a text corpus to be processed; perform text preprocessing on the text corpus to be processed to obtain structured input text; identify pronunciation risk segments in the structured input text and record the context information and risk category of the pronunciation risk segments to obtain risk segment identification results; send the risk segment identification results and context information to a large model so that the large model generates candidate pronunciation results; verify the candidate pronunciation results according to preset pronunciation rules and a domain dictionary to obtain verified pronunciation annotations; write the verified annotation results that have passed the review into the domain dictionary and output structured pronunciation annotation corpus.

[0168] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0169] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0170] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0171] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described text-to-speech pronunciation annotation method. This solves the technical problem of how to improve the accuracy of pronunciation annotation of text-to-speech corpora and realize the reuse of domain pronunciation knowledge. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the text-to-speech pronunciation annotation method provided in the above embodiments, and will not be repeated here.

[0172] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the text-to-speech pronunciation annotation method described above.

[0173] The computer program product provided in this application can solve the technical problem of how to improve the accuracy of pronunciation annotation of text-to-speech corpora and realize the reuse of domain pronunciation knowledge. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the text-to-speech pronunciation annotation method provided in the above embodiments, and will not be repeated here.

[0174] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for text-to-speech pronunciation annotation, characterized in that, The text-to-speech pronunciation annotation method includes: Receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text; Identify pronunciation risk segments in the structured input text, and record the context information and risk category of the pronunciation risk segments to obtain the risk segment identification results; The risk segment identification result and the context information are sent to the large model so that the large model can generate candidate reading results; The candidate pronunciation results are verified according to the preset pronunciation rules and the domain dictionary to obtain the verified pronunciation annotations. The verified phonetic annotations that have passed the review are written into the domain dictionary, and the structured phonetic annotation corpus is output.

2. The method as described in claim 1, characterized in that, The step of identifying pronunciation risk segments in the structured input text, recording the context information and risk category of the pronunciation risk segments, and obtaining the risk segment identification result includes: The structured input text is subjected to character type recognition to obtain the character type recognition result; Based on the preset pronunciation risk triggering rules and the character type recognition results, the pronunciation risk segments in the structured input text are determined, and the risk segment location information is obtained; Based on the location information of the risk segment, the context information and risk category corresponding to the pronunciation risk segment are extracted to obtain the risk segment identification result.

3. The method as described in claim 1, characterized in that, The steps of receiving the text corpus to be processed and performing text preprocessing on the text corpus to obtain structured input text include: Receive the text corpus to be processed and assign a unique identifier to the text corpus to be processed to obtain identifiable corpus; The labeled corpus is processed into sentence segmentation and word segmentation to obtain word segmentation results; The word segmentation results are subjected to character type recognition and punctuation standardization to obtain standardized text; A context window is constructed based on the normalized text to obtain the structured input text.

4. The method as described in claim 1, characterized in that, The step of sending the risk segment identification result and the context information to the large model so that the large model generates candidate reading results includes: The risk segment identification result and the context information are input into the large model, so that the large model can perform contextual understanding on the pronunciation risk segment based on the context information to obtain the contextual understanding result; Based on the context understanding results, candidate reading content is output based on the large model; Based on the candidate reading content, the reading type and judgment criteria are output according to the large model; The candidate reading content, the reading type, and the judgment criteria are combined to obtain the candidate reading result.

5. The method as described in claim 1, characterized in that, The step of verifying the candidate pronunciation results according to preset pronunciation rules and a domain dictionary to obtain the verified pronunciation annotation includes: The candidate pronunciation results are filtered according to preset pronunciation rules to obtain the rule filtering results; The rule filtering results are matched with the domain dictionary to obtain the dictionary matching results; When the dictionary matching result matches the candidate reading result, the recommendation confidence of the candidate reading result is increased to obtain the first verification result; When the dictionary matching result conflicts with the candidate reading result, the candidate reading result is marked as pending verification, and a second verification result is obtained. Based on the first verification result and the second verification result, the verified pronunciation annotation is obtained.

6. The method as described in claim 5, characterized in that, The step of marking the candidate reading result as pending verification when the dictionary matching result conflicts with the candidate reading result, and obtaining the second verification result, includes: When the dictionary matching result conflicts with the candidate reading result, the candidate reading result is marked as pending review, and the segment to be reviewed is obtained. The segment to be reviewed is sent to the review terminal so that the review terminal can display the original text, candidate readings, recommendation reasons and domain dictionary hit status of the segment to be reviewed, and receive the review confirmation result returned by the review terminal; Based on the audit confirmation results, generate the post-audit read method corresponding to the segment to be reviewed; Based on the post-audit reading method and the first verification result, the second verification result is obtained.

7. The method as described in claim 1, characterized in that, The step of writing the approved annotation results from the verified phonetic annotations into the domain dictionary and outputting the structured phonetic annotation corpus includes: The verified pronunciation annotations that have passed the review are converted into dictionary candidates. The dictionary candidates are reviewed to obtain the reviewed dictionary entries; The reviewed dictionary entries are written into the domain dictionary to obtain the updated domain dictionary; Based on the updated domain dictionary and the verified phonetic annotations, a structured phonetic annotation corpus is output.

8. A text-to-speech pronunciation annotation device based on self-learning of a large model and a domain dictionary, characterized in that, The device includes: The corpus access module is used to receive the text corpus to be processed, perform text preprocessing on the text corpus to be processed, and obtain structured input text. The pronunciation risk identification module is used to identify pronunciation risk segments in the structured input text, record the context information and risk category of the pronunciation risk segments, and obtain the risk segment identification result; The candidate reading generation module is used to send the risk segment identification result and the context information to the large model so that the large model can generate candidate reading results. The rule and dictionary verification module is used to verify the candidate pronunciation results according to the preset pronunciation rules and domain dictionary to obtain the verified pronunciation annotation; The dictionary self-learning module is used to write the approved annotation results from the verified phonetic annotations into the domain dictionary and output structured phonetic annotation corpus.

9. A text-to-speech pronunciation annotation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the text-to-speech pronunciation annotation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the text-to-speech pronunciation annotation method as described in any one of claims 1 to 7.

Images