A method for predicting the prosodic structure of tibetan language based on syntactic information

Abstract

The application discloses a Tibetan prosody structure prediction method based on syntax information, relates to the technical field of speech synthesis, and is applied to the prosody structure prediction field in Tibetan speech synthesis. The Tibetan prosody structure prediction method based on syntax information comprises the following steps: S1, Tibetan text analysis and virtual word connection quantization are completed, syntax probability, hierarchical focus and pause difference are extracted, and a Tibetan prosody prediction database is stored and constructed after preprocessing; S2, boundary discrimination is performed based on virtual word connection relationship strength, influence range hit marking and word sequence distance; S3, entropy feature analysis is performed through syntax role probability data and boundary determination data; S4, three-layer prosody boundary type sequences are generated according to boundary assignment, and over-dense pauses are inhibited; and S5, rhythm intensity evaluation is performed through pause duration difference and syntax information entropy data. The problems of pause confusion and increased understanding burden caused by the misplacement of prosody and semantic hierarchy in Tibetan long speech synthesis are solved.

Description

Technical Field

[0001] This invention relates to the field of speech synthesis technology, specifically to a method for predicting Tibetan prosodic structure based on grammatical information. Background Technology

[0002] In the field of Tibetan prosodic structure prediction, accurately identifying and adjusting pause positions and stress placements is crucial for improving the quality of text reading. Existing technologies largely focus on prosodic prediction based on grammatical information and phonetic data, typically employing simple pause rules or temporal models based on short-term observations.

[0003] For example, invention patent CN106294310B discloses a Tibetan tone prediction method and system, comprising: receiving Tibetan text to be processed; performing word segmentation on the Tibetan text to be processed to obtain word units; determining the part of speech of the word units based on the contextual information of the word units in the Tibetan text to be processed; predicting the prosodic boundaries of the Tibetan text to be processed, and adjusting the word unit boundaries at the prosodic boundaries based on the part of speech of the word units at the prosodic boundaries; and predicting the tone of the syllable units of the Tibetan text to be processed after adjusting the word unit boundaries based on the part of speech of each word unit to obtain the tone information of the Tibetan text to be processed. Using this invention, the tone sandhi problem of multi-tone pattern words at different prosodic boundaries can be solved, effectively improving the application effect of Tibetan speech systems.

[0004] However, existing methods for prosodic structure research mainly focus on the application of static rules, with a lack of modeling of the dynamic relationship between grammatical information and prosody. There is still no effective technical implementation on how to extract the grammatical functions of function words, clause hierarchy, and pause differences based on grammatical information, and combine these features for prosodic structure optimization.

[0005] Therefore, in order to address the above problems, there is an urgent need for a Tibetan prosodic structure prediction method based on grammatical information. Summary of the Invention

[0006] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a Tibetan prosodic structure prediction method based on grammatical information, which solves the problems of pause confusion and increased comprehension burden caused by the misalignment of prosodic grouping and semantic hierarchy in Tibetan long sentence speech synthesis.

[0007] Technical solution

[0008] To achieve the above objectives, this invention provides the following technical solution: a Tibetan prosodic structure prediction method based on grammatical information, comprising: S1, completing Tibetan text parsing and quantifying function word continuations, extracting grammatical probabilities, hierarchical focus, and pause differences, and storing the preprocessed data to construct a Tibetan prosodic prediction database; S2, performing boundary discrimination based on the strength of function word continuation relationships, the hit mark of the influence range, and word order distance, thereby achieving candidate boundary screening, continuation segment merging, pause label configuration, and grammatical function annotation update; S3, performing entropy feature analysis using grammatical role probability data and boundary judgment data, and performing function word anchor point marking, clause skeleton adjustment, and pause compression operations based on the entropy feature analysis results; S4, reading the grammatical annotations and hierarchical information of candidate boundaries, generating a three-layer prosodic boundary type sequence based on boundary modeling, and suppressing excessively dense pauses; S5, evaluating rhythm intensity using pause duration differences and grammatical information entropy data, and performing phrase hierarchical subdivision and pause mark adjustment based on the rhythm evaluation results.

[0009] Further, to complete Tibetan text parsing and quantification of function word continuations, and extract grammatical probabilities, hierarchical focus, and pause differences, the specific measures are as follows: Collect Tibetan text data: Receive the text to be processed through the input interface, perform word segmentation to obtain word sequences and part-of-speech sequences, and output dependency arc sets, headword index sets, and modification range sets based on syntactic dependency analysis; identify all function words in the word sequence, classify them into three categories based on grammatical features: case particles, general function words, and class function words, and mark and number each function word; collect function word continuation rules: Extract legally connectable function word combinations, synonym substitution forms, and affix and suffix attachment information from the function word continuation constraint table formed based on corpus statistics and grammatical rules, combine the word sequence to generate a function word continuation relationship strength matrix and a word order distance matrix to achieve function word quantification, encode the matrix synchronously with the text sequence, and construct a function word continuation relationship strength matrix and a word order distance matrix based on boundary assignment rules and candidate boundary setting conditions. The system employs an indicator function matrix to determine whether each pair of function words falls within the same candidate boundary influence range, thus obtaining the boundary influence range indicator value. It also collects syntactic probability data: performing grammatical role discrimination on the function word sequence, outputting the probability distribution of various grammatical roles that each function word may assume, and recording this as a grammatical role probability sequence. Furthermore, it extracts clause hierarchical structure and semantic focus information: based on dependency relations, modification scope, and clause division, it determines the clause hierarchical index, clause start and end labels, and central word chains, extracting the boundaries of topic components and modifying components and registering them as a list of semantic focus anchor points. Finally, it collects prosodic phrase and pause-related data: in Tibetan speech-to-text corpora with phrase segmentation and pause annotations, it records the boundary start and end time nodes of each prosodic phrase based on the phrase segmentation annotations, obtains the time nodes of each actual pause period by aligning the pause annotations with the time axis, and calculates the duration difference between the prosodic phrase and the pause period to obtain the pause duration difference.

[0010] Furthermore, the specific measures for preprocessing and storing the Tibetan prosody prediction database are as follows: function words with pause duration differences less than the pause threshold are labeled as short pauses, and function words with pause duration differences greater than the pause threshold are labeled as long pauses; the collected numerical data of function word continuation strength, word order distance, and grammatical role probability are normalized to unify the range of all dimensions; based on the normalized feature data, the structured fields of function word category, continuation strength, grammatical role probability, dependency index, distance, hierarchical label, and focus identifier are concatenated in text order to generate a homo-frequency feature vector, and function word sequence number, data batch identifier, and main control field range are appended, and the feature vector and function word grammatical labels are stored synchronously to construct the Tibetan prosody prediction database.

[0011] Furthermore, the specific measures for boundary discrimination based on the strength of function word continuation relationships, the hit marker of the influence range, and the word order distance are as follows: obtain the continuation relationship strength, the boundary influence range indicator value, and the word order distance corresponding to position i and all function word positions j; for each pair of positions i and j, multiply the continuation relationship strength by the boundary influence range indicator value to obtain the boundary contribution term; add the word order distance to a small constant and then square it to obtain the distance penalty term; divide the boundary contribution term by the distance penalty term to obtain the boundary contribution value corresponding to position i and position j; perform a summation operation on the boundary contribution values ​​of all positions j to finally obtain the boundary candidate value at position i.

[0012] Furthermore, the specific measures for implementing candidate boundary screening, concatenation segment merging, pause label configuration, and grammatical function annotation updates are as follows: By comparing the candidate boundary value and the boundary threshold in real time, when the candidate boundary value is less than the boundary threshold, the system sequentially checks whether there are any legally connectable combinations registered in the function word continuation constraint table between adjacent function words. For the detected continuous function word concatenation segments, the system performs prosodic phrase boundary merging, cancels the temporary pause labels between all function words in the segment, and reverts the grammatical function annotation triggered by the function word concatenation segment. When the candidate boundary value is greater than or equal to the boundary threshold, the system determines the current position as a valid prosodic boundary candidate, retains the position in the candidate boundary list and marks it as a prosodic boundary, and simultaneously performs legality verification on the function word combinations on the left and right sides of the position according to the function word continuation constraint table. The system assigns short pause labels to function word concatenation segments that conform to the constraint rules and simultaneously annotates the grammatical functions.

[0013] Furthermore, the specific measures for entropy feature analysis using grammatical role probability data and boundary determination data are as follows: obtain the probability value of the k-th function word in each grammatical role and the corresponding boundary candidate value; multiply the probability of each grammatical role by its own logarithm to obtain the information content term of a single grammatical role; sum the information content terms of all grammatical roles and take the negative to obtain the basic grammatical entropy value of the k-th function word; multiply the boundary adjustment factor by the corresponding boundary candidate value to obtain the boundary correction term and add one to obtain the entropy correction term; multiply the basic grammatical entropy value by the entropy correction term to finally obtain the grammatical information entropy value of the k-th function word.

[0014] Furthermore, based on the entropy feature analysis results, the specific measures for function word anchor point marking, clause skeleton adjustment, and pause compression are as follows: By comparing the grammatical information entropy value and the entropy threshold in real time, when the grammatical information entropy value is less than the entropy threshold, the corresponding function word is marked as a stable grammatical function anchor point, a fixed case relation marker, a clause connection marker, and a mood marker, and will no longer be downgraded or merged in the subsequent boundary optimization stage; at the same time, it expands along the dependency arc to both sides, classifying related components into the same clause skeleton, and registering them synchronously in the clause-level index and the focus anchor point list, so that the subsequent prosodic phrase boundary is prioritized to be close to the target position for segmentation; when the grammatical information entropy value is greater than or equal to the entropy threshold, the corresponding function word is marked as a unit with undetermined grammatical function, triggering the speech... The grammatical role mitigation strategy involves checking whether the preceding and following combinations in the function word continuation constraint table meet the legal order requirements. For inconsistent grammatical function annotations, a fallback to the three categories of case particles, general function words, and class function words is performed. Within the function word continuation constraint table, the legal connection combination of function word pairs before and after the position is reviewed. After updating the grammatical role probability distribution for the hit combination, the grammatical information entropy value is recalculated. If the recalculated grammatical information entropy value is still greater than or equal to the entropy threshold, the position is prohibited from participating in the selection of prosodic focus anchors. Instead, the function word with the nearest grammatical information entropy value to the left and right of the position that is less than the entropy threshold is used as the focus anchor. The pause duration label corresponding to the position is marked as a short pause level to avoid binding the pause and stress point to the position of grammatical function fluctuation, which would lead to misalignment of long sentence levels.

[0015] Furthermore, the specific measures for reading the grammatical annotations and hierarchical information of candidate boundaries, generating a three-layer prosodic boundary type sequence based on boundary typing, and suppressing excessively dense pauses are as follows: Within the candidate boundary list, for each candidate position, the grammatical function annotations of function words, clause-level indexes, semantic focus anchors, and grammatical information entropy values ​​are read, and a three-layer boundary type sequence is output according to the boundary typing rules. The boundary typing rules include: when the grammatical function annotation of a function word at a candidate position is a case particle, the candidate position is determined as a prosodic boundary, and the case particle and the syntactic component adjacent to the left of the candidate position are merged into the same prosodic word; when the grammatical function annotation of a function word at a candidate position is a tense particle or a nominalized suffix, the candidate position is determined as a prosodic boundary and the prosodic word closure is completed; when the grammatical function annotation of a function word at a candidate position is a modal particle, ... Candidate positions are identified as intonation phrase boundaries and written into sentence-end intonation control information. When a candidate position matches a sentence-end function word marker, the candidate position is identified as an intonation phrase boundary and written into sentence-crossing control information. When a candidate position's function word is grammatically labeled as a connective function word and matches a compound sentence association marker, the candidate position is identified as a prosodic phrase boundary. The clause start and end boundaries are aligned to the candidate position based on the clause-level index to ensure that the pause point is consistent with the modification range boundary. After the boundary type is determined, candidate positions with grammatical information entropy values ​​less than the entropy threshold are kept as boundary types and do not participate in downgrading and merging. Candidate positions with grammatical information entropy values ​​greater than or equal to the entropy threshold are suppressed when the candidate boundary density exceeds the limit, resulting in a three-layer boundary type sequence of prosodic word boundaries, prosodic phrase boundaries, and intonation phrase boundaries.

[0016] Furthermore, the specific measures for evaluating rhythmic intensity using pause duration differences and grammatical information entropy data are as follows: Obtain the duration difference between the m-th prosodic phrase and each pause period, the time nodes of each pause period, the start and end time nodes of the m-th prosodic phrase boundary, and the grammatical information entropy value; for each pause period, divide the duration difference by one and add the product of the adjustment weighting factor and the absolute value of the time node difference to obtain a single pause contribution; perform an integral operation on all pause contributions within a given time range to obtain the basic rhythmic intensity term; multiply the information weighting factor by the grammatical information entropy value and add one to obtain the rhythm correction factor; multiply the basic rhythmic intensity term by the rhythm correction factor to finally obtain the hierarchical rhythmic intensity value of the m-th prosodic phrase.

[0017] Furthermore, based on the rhythm assessment results, the specific measures for phrase hierarchical subdivision and pause label adjustment are as follows: By comparing the hierarchical rhythm intensity value and intensity threshold in real time, when the hierarchical rhythm intensity value is less than the intensity threshold, a conservative subdivision strategy is implemented. Within the determined prosodic phrase boundaries, the boundary type remains unchanged, and no new segmentation points are added. Only the pause duration around the same prosodic phrase boundary is converged, and the pause label is uniformly compressed to the short pause level and the continuous occurrence of adjacent short pauses is suppressed to avoid introducing redundant levels and causing rhythm fragmentation in segments with indistinct pause duration differences. When the hierarchical rhythm intensity value is greater than or equal to the intensity threshold, an enhanced subdivision strategy is implemented: the hierarchical rhythm intensity value is used as the subdivision priority basis. Candidate positions that are also prosodic phrase boundaries are sorted according to the hierarchical rhythm intensity value, and the positions with higher ranking are selected as prosodic large phrase boundaries. The similar boundaries with lower ranking are downgraded to prosodic small phrase boundaries, and the difference in pause duration drives the hierarchical separation of short pauses and long pauses. When the same position simultaneously hits the prosodic large phrase and prosodic small phrase conditions, the subdivision is carried out according to the priority table. The side with the rhythm intensity value closer to the intensity threshold determines the final level and locks the boundary type to avoid repeated switching. For excessively long continuous modification chains at the end of a sentence, the rhythm intensity value of the layered rhythm is calculated around the boundary of the sentence-end intonation phrase, and the segmentation point with the larger rhythm intensity value is selected to move forward the prosodic large phrase boundary, widening the distance between the sentence-end intonation phrase boundary and the adjacent prosodic large phrase boundary to tighten the end rhythm. When performing dynamic programming segmentation on the prosodic phrase candidate boundary sequence, the rhythm intensity value is used as a cost term in the path scoring. When the clause boundary coverage, the central word chain integrity, and the prosodic phrase length distribution constraint are all satisfied, the segmentation point with the larger rhythm intensity value is preferentially retained. The output is the code of prosodic word boundary, prosodic small phrase boundary, prosodic large phrase boundary, intonation phrase boundary and semantic focus anchor information. At the speech synthesis end, the prosodic word boundary is mapped to the pause symbol insertion. The prosodic small phrase and prosodic large phrase boundaries are combined with the semantic focus anchor to control the fundamental frequency curve and duration scaling. The intonation phrase boundary is used for sentence-end tone constraints and cross-sentence connection to achieve pause level subdivision and semantic level alignment.

[0018] Beneficial effects

[0019] The present invention has the following beneficial effects: (1) This invention effectively screens candidate positions of prosodic boundaries by using a boundary determination method based on the strength of function word continuation relationship and word order distance, thereby achieving accurate identification of prosodic boundaries and effectively solving the problem of inaccurate prosodic boundary determination in the prior art.

[0020] (2) This invention, through the analysis and adjustment of grammatical information entropy value, accurately marks the anchor point of function words, and performs clause skeleton adjustment and pause compression, thereby achieving stable anchoring of grammatical functions and optimization of long sentence levels, effectively solving the problems of unstable grammatical functions and disordered syntactic structure in the prior art.

[0021] (3) The present invention, through a strategy based on hierarchical rhythm intensity assessment, achieves effective subdivision and rhythm alignment of pause levels by dynamically adjusting phrase level subdivision and pause markers, thereby effectively solving the problems of unclear pause levels and unnatural rhythm in the prior art.

[0022] (4) This invention optimizes the generation of three-layer prosodic boundary types by applying boundary type rules and combining grammatical annotation and hierarchical information, thereby realizing the reasonable allocation and rhythm coordination of different prosodic levels, and effectively solving the problems of inaccurate boundary type division and unclear hierarchy in the prior art.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] Figure 1 This is a flowchart of a Tibetan prosodic structure prediction method based on grammatical information according to the present invention; Figure 2 This is a schematic diagram of the Tibetan prosodic structure prediction model of the present invention; Figure 3 This is an example diagram illustrating the generation of Tibetan prosodic hierarchical structure based on case particles in this invention; Figure 4 This is an example diagram illustrating the generation of Tibetan prosodic hierarchical structure based on function words in this invention; Figure 5 This is a waveform diagram of Tibetan example sentences from the present invention; Figure 6 This is a boundary analysis diagram of the Tibetan prosodic structure prediction driven by the grammatical information of this invention. Figure 7 This is a flowchart of the three-layer prosodic boundary shaping and suppression process of the present invention; Figure 8 This is an example diagram of the prosodic structure generation of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-8 This invention provides a technical solution: a method for predicting the prosodic structure of Tibetan based on grammatical information, such as... Figure 1This is a flowchart of the Tibetan prosodic structure prediction method based on grammatical information in this embodiment. It includes: S1, completing Tibetan text parsing and quantifying function word continuations, extracting grammatical probabilities, hierarchical focus, and pause differences, and storing the preprocessed data to construct a Tibetan prosodic prediction database; S2, performing boundary discrimination based on the strength of function word continuation relationships, the hit mark of the influence range, and word order distance, realizing candidate boundary screening, continuation segment merging, pause label configuration, and grammatical function annotation update; S3, performing entropy feature analysis through grammatical role probability data and boundary judgment data, and performing function word anchor point marking, clause skeleton adjustment, and pause compression operations based on the entropy feature analysis results; S4, reading the grammatical annotations and hierarchical information of candidate boundaries, generating a three-layer prosodic boundary type sequence based on boundary modeling, and suppressing overly dense pauses; S5, evaluating rhythm intensity through pause duration differences and grammatical information entropy data, and performing phrase hierarchical subdivision and pause mark adjustment based on the rhythm evaluation results.

[0027] Specifically, the specific measures for completing Tibetan text parsing and quantifying function word continuations, and extracting grammatical probabilities, hierarchical focus, and pause differences are as follows: The overall processing flow is as follows: Figure 2 As shown, Figure 2This is a schematic diagram of the Tibetan prosodic structure prediction model in this embodiment. Starting with the input Tibetan text, the model sequentially goes through function word identification and continuation constraint analysis, prosodic boundary discrimination, grammatical entropy feature constraint, prosodic boundary shaping, rhythmic layering refinement, and text output. This is a structured unfolding of the steps. Various arrows, connector labels, and intermediate node names are used to indicate the transmission relationship between function word continuation constraints, clause-level indexing, semantic focus anchors, and pause duration differences. These can all be considered as abstract expressions of the data flow path in the steps, without limiting specific implementation methods or parameter values. Tibetan text data is collected: the data to be processed is received through the input interface. The text is processed, and word segmentation is performed to obtain word sequences and part-of-speech sequences. Combined with syntactic dependency analysis, a set of dependency arcs, a set of headword indices, and a set of modification ranges are output. All function words in the word sequences are identified and classified into three categories based on grammatical features: case particles, general function words, and class-specific function words. The position of each function word is marked and numbered. Function word continuation rules are collected: legally connectable function word combinations, synonym substitution forms, and affix and suffix attachment information are extracted from a function word continuation constraint table based on corpus statistics and grammatical rules. These are then combined with the word sequences to generate a function word continuation relationship strength matrix and a word order distance matrix to quantify function words. The continuation relationship strength matrix represents the strength of the continuation relationship between function word i and function word j, reflecting the tightness of their grammatical association. Each element value in the matrix represents the grammatical association between function words; a larger value indicates a stronger association. The word order distance matrix represents the word order distance between function word i and function word j, reflecting their relative positions in the syntactic structure; a smaller value indicates that they are closer in the sentence. These matrices are encoded synchronously with the text sequence, and an indicator function matrix is ​​constructed based on boundary assignment rules and candidate boundary setting conditions to determine whether each pair of function words is within the same candidate boundary influence range, thus obtaining the boundary influence range indicator value. The element value in the indicator function matrix is ​​1 or 0, indicating whether function word i and function word j belong to the same candidate boundary influence range. When function word i and function word j are within the same candidate boundary influence range, the corresponding element in the indicator function matrix is ​​1; otherwise, it is 0. Syntactic probability data is collected: syntactic roles are determined for the function word sequence, outputting the probability distribution of various syntactic roles that each function word may assume, and recording it as a syntactic role probability sequence. Clause hierarchical structure and semantic focus information are extracted: based on dependency relations, modification scope, and clause division, the clause hierarchical index, clause start and end labels, and central word chains are determined. The boundaries of topic components and modifying components are extracted and registered as a list of semantic focus anchor points, such as... Figure 3 This is an example diagram of the Tibetan prosodic hierarchy generated based on case particles in this embodiment, using Tibetan sentences. For example, in the Chinese meaning, "The father sends his child to learn Tibetan culture," the case particle is located at the boundary of the prosodic word, and the boundaries of the prosodic phrase and intonation phrase both fall at the end of the sentence. Figure 4 This is an example diagram of the Tibetan prosodic hierarchy generated based on function words in this embodiment, using Tibetan sentences. For example, the Chinese meaning is the competition time: 8:00 AM to 4:00 PM. The boundary type is defined as an intonation phrase boundary, using the final function word as the dividing line. Data related to prosodic phrases and pauses is collected: In Tibetan speech-to-text corpora with phrase segmentation and pause annotations, the start and end time nodes of each prosodic phrase boundary are recorded according to the phrase segmentation annotations. The time nodes of each actual pause period are obtained by aligning the pause annotations with the timeline. The difference in duration between the prosodic phrase and the pause period is calculated to obtain the pause duration difference. Example sentence The Chinese meaning is to pass the time by singing happily, specifically... Figure 5 The above is a waveform diagram of a Tibetan example sentence in this embodiment.

[0028] In this implementation plan, through comprehensive collection and analysis of Tibetan text data, information such as the continuation relationships of function words, word order distances, and grammatical roles was quantified. This information was then used to construct a function word continuation relationship strength matrix, a word order distance matrix, and an indicator function matrix, further determining the boundary influence range between function words. This process not only accurately models the relationships between function words but also provides precise data support for subsequent prosodic boundary determination and grammatical function annotation. By combining grammatical probability data and pause duration differences, a more accurate basic data framework can be effectively established for Tibetan prosodic structure prediction, thus providing a reliable basis for subsequent operations such as boundary typing, pause marker adjustment, and phrase level subdivision. This process ultimately significantly improves the accuracy and robustness of Tibetan prosodic structure prediction.

[0029] Specifically, the preprocessing and storage of the Tibetan prosody prediction database involves the following measures: function words with pause duration differences less than a pause threshold are labeled as short pauses, while those with pause duration differences greater than the threshold are labeled as long pauses. This effectively distinguishes between short and long pauses, ensuring the reasonable classification of pause labels and providing a clear hierarchical structure for pause duration in subsequent prosodic analysis. The collected numerical data on function word continuation strength, word order distance, and grammatical role probability are normalized using maximum and minimum values ​​to unify all dimensions. During normalization, the maximum and minimum value method is used, and normalization can be performed by field or sentence dimension, selecting the appropriate dimension based on specific needs. Based on the normalized feature data, structured fields such as function word category, continuation strength, grammatical role probability, dependency index, distance, hierarchical label, and focus identifier are concatenated in text order to generate a homo-frequency feature vector. This vector not only contains basic grammatical information about function words but also integrates the location and relationship of each function word within the syntactic structure, enabling a more accurate reflection of the prosodic features of a sentence. Simultaneously, additional information such as function word sequence numbers, data batch identifiers, and key field ranges are appended to the feature vector to ensure efficiency and traceability during data storage and processing. Finally, these feature vectors are stored synchronously with the grammatical labels of the function words, and a Tibetan prosodic prediction database is constructed for use in subsequent analysis and model training.

[0030] In this implementation scheme, the pause duration of function words is labeled according to their differences. Function words with pause durations less than a pause threshold are labeled as short pauses, and those with pause durations greater than the threshold are labeled as long pauses, ensuring reasonable classification and hierarchy of pauses. Subsequently, the collected numerical data, such as the strength of function word continuation relationships, word order distance, and grammatical role probabilities, are normalized to their maximum and minimum values, unifying the data's dimensional range, and normalized according to field, sentence, or batch dimensions. Next, based on the normalized feature data, various grammatical information of function words are concatenated in text order to generate a unified frequency feature vector. Additional information such as function word sequence number and data batch identifier are added to the feature vector for subsequent storage and processing. Finally, all feature vectors and the grammatical labels of function words are stored synchronously, constructing a Tibetan prosody prediction database, providing a structured and unified data foundation for subsequent analysis and model training.

[0031] Specifically, the boundary discrimination based on the strength of function word continuation relationships, the hit marker of the influence range, and the word order distance is as follows: First, the strength of the continuation relationship, the boundary influence range indicator value, and the word order distance corresponding to position i and all function word positions j are obtained. Then, the strength of the continuation relationship and the boundary influence range indicator value for each pair of positions i and j are multiplied to obtain the boundary contribution term. This calculation reflects the tightness of the continuation relationship between function words and its influence on boundary judgment. Next, the word order distance is added to a small constant and squared to obtain a distance penalty term. This operation aims to ensure that the influence of word order distance on the boundary contribution term is reasonably suppressed, avoiding excessive influence of function words that are too far apart on boundary judgment. Then, the boundary contribution term is divided by the distance penalty term to obtain the boundary contribution value corresponding to position i and position j. This value is used to measure the potential contribution of this position to the prosodic boundary. Finally, the boundary contribution values ​​of all positions j are summed to obtain the candidate boundary value at position i. This candidate value is an important basis for determining whether this position is a valid prosodic boundary. By comprehensively considering the strength of continuation relationships, distance penalties, and the scope of boundary influence, this invention can more accurately evaluate candidate values ​​for prosodic boundaries, thereby improving the accuracy and robustness of Tibetan prosodic structure prediction.

[0032] The specific calculation method for the boundary candidate values ​​is as follows: ; In the formula, Indicates position The boundary candidate value at a given location is used to evaluate whether that location is a potential prosodic boundary candidate; Indicating function words and function words The strength of the sequential relationship between function words reflects the degree of grammatical connection between them. Represents the indicator function matrix, used to determine function words. and Whether it is within the influence range of the candidate boundary; Indicating function words and function words Word order distance in a sentence reflects the tightness of their grammatical structures; It is a tiny constant to avoid division by zero errors.

[0033] In this embodiment, the boundary candidate value of the function word position one is 0.83, its connection strength with positions one, two, and three is 0.69, 0.64, and 0.30, respectively, its boundary indicators are 1, 1, and 0, its word order distances are 1, 2, and 3, and its contribution items are 0.676, 0.158, and 0.000, respectively; the boundary candidate value of the function word position two is 0.71, its connection strength with positions one, two, and three is 0.53, 0.60, and 0.35, respectively, its boundary indicators are 1, 1, and 1, its word order distances are 1, 2, and 3, and its contribution items are 0.519, 0.148, and 0.039, respectively; the boundary candidate value of the function word position three is 0.45, its connection strength with positions one, two, and three is 0.30, 0.82, and 0.86, respectively, its boundary indicators are 1, 1, and 1, its word order distances are 3, 2, and 2, and its contribution items are 0.033, 0.203, and 0.213, respectively.

[0034]

[0035] like Figure 6 This is a boundary analysis diagram of Tibetan prosodic structure prediction driven by grammatical information provided in the embodiments of this application. See Table 1 and... Figure 6 It can be seen that the boundary candidate values ​​at different function word positions can effectively reflect their prosodic boundary potential in a sentence. When the connection strength matrix at function word position one is [0.69, 0.64, 0.30] and the word order distance is [1, 2, 3], the calculated boundary candidate value is 0.83, which is greater than the boundary threshold, indicating that this position is a strong prosodic boundary candidate point. Although the connection strength at function word position three is high [0.30, 0.82, 0.86], the word order distance is relatively large [3, 2, 2] and some connections are not activated, so the calculated boundary candidate value is only 0.45, which is less than the boundary threshold, indicating that the grammatical function at this position is relatively certain and it is not suitable as a major prosodic boundary. It can be seen that the size of the boundary candidate value depends not only on the connection strength between function words, but also on the synergistic regulation of word order distance and boundary indication. Function words with high connection strength, close word order distance and active boundary indication are more likely to make significant contributions to the prosodic boundary. It is evident that the determination of prosodic boundaries depends on the combined effect of multiple factors. By calculating candidate boundary values, potential prosodic segmentation points in sentences can be dynamically identified, improving the accuracy of prosodic grouping in Tibetan long sentence speech synthesis. This provides a basis for optimizing the Tibetan prosodic structure prediction system and allows for dynamic adjustment of prosodic boundary determination based on grammatical information, providing reliable support for speech synthesis.

[0036] In this implementation scheme, by comprehensively considering the strength of function word continuation relationships, the boundary influence range indicator value, and word order distance, this step can accurately assess the boundary contribution of each function word position to other positions and calculate the candidate boundary value for each position. This process ensures the accuracy of boundary determination and effectively suppresses the influence of distant function words on the boundary contribution by weighted calculation of continuation relationship strength and boundary influence range indicator value, and by penalizing word order distance. Ultimately, the obtained candidate boundary values ​​provide a reliable basis for determining effective prosodic boundaries, thereby improving the accuracy of Tibetan prosodic structure prediction.

[0037] Specifically, the measures for candidate boundary screening, contiguous segment merging, pause label configuration, and grammatical function annotation updates are as follows: By comparing the candidate boundary value and the boundary threshold in real time, when the candidate boundary value is less than the boundary threshold, it is checked whether there are any legally connectable combinations registered in the function word continuation constraint table between adjacent function words. For the detected continuous function word contiguous segments, prosodic phrase boundary merging is performed, temporary pause labels between all function words in the segment are canceled, and the grammatical function annotations triggered by the function word contiguous segments are rolled back. During the rollback, the grammatical function level is adjusted and duplicate annotations are cleaned up only within the scope of function words that have not been marked as stable grammatical function anchors in the grammatical information entropy analysis. The function word case relation markers, clause connection markers, and mood markers that have been registered as stable grammatical function anchors are kept unchanged, and the rollback is recorded. The retained grammatical function annotations are written into the clause skeleton record to ensure that the stable anchor points continue to be effective in subsequent steps. When the boundary candidate value is greater than or equal to the boundary threshold, the current position is determined as a valid prosodic boundary candidate. The position is retained in the candidate boundary list and marked as a prosodic boundary. At the same time, the legality of the function word combination on the left and right sides of the position is checked according to the function word continuation constraint table. Short pause labels are assigned to function word continuation segments that meet the constraint rules and grammatical functions are annotated simultaneously. During the annotation process, the stable anchor points of grammatical functions that have been locked by the grammatical information entropy analysis are skipped. Only pause level adjustment and grammatical function supplement records are added at non-anchor point positions. This allows the candidate boundary screening, continuation segment merging, and pause label configuration to complete the boundary discrimination and grammatical function annotation update without destroying the stable anchor points.

[0038] In this implementation scheme, by comparing boundary candidate values ​​with boundary thresholds, and with the support of the function word continuation constraint table, on the one hand, function word continuation segments that do not meet the boundary conditions are merged into prosodic phrase boundaries and redundant pause markers are removed. Under the premise of not affecting the stable anchor points obtained by the grammatical information entropy determination, the grammatical function annotations at non-anchor point positions are orderly backtracked and organized, so that redundant boundaries and overly dense pauses are converged into a more compact prosodic structure. On the other hand, candidate positions that meet the boundary conditions are stably retained as prosodic boundaries and short pause markers and grammatical function information are added, guiding the subsequent long sentence layering, focus positioning, and stress placement to complete the mapping on reliable boundaries. This improves the overall consistency of Tibetan sentence prosodic grouping and grammatical hierarchy, thereby alleviating the problems of messy long sentence pauses and excessive comprehension burden.

[0039] Specifically, the entropy feature analysis using grammatical role probability data and boundary judgment data involves the following steps: After obtaining the probability value of the k-th function word in each grammatical role and the corresponding boundary candidate value, the probability of each grammatical role is multiplied by its own logarithm to obtain the information content term for that grammatical role. A scoring function with a weighted sum of boundary recall and precision is constructed on a development corpus with standard prosodic boundary annotations. A grid search and fitting analysis is performed on the coupling curve between the boundary candidate value and the grammatical information entropy to obtain a boundary adjustment factor, ranging from 0.1 to 2.0. The information content terms of all grammatical roles are summed and their negatives are taken to obtain the basic grammatical entropy value of the k-th function word, reflecting the uncertainty of its grammatical function. Subsequently, the boundary adjustment factor is multiplied by the corresponding boundary candidate value to obtain a boundary correction term, and an entropy correction term is obtained by incrementing by one, which is used to adjust the calculation of the grammatical information entropy value. Finally, the basic grammatical entropy value is multiplied by the entropy correction term to obtain the grammatical information entropy value of the k-th function word. This process can comprehensively consider the grammatical role probability, boundary candidate value and its correction effect, accurately quantify the grammatical uncertainty of function words, and provide a more accurate basis for subsequent prosodic structure prediction.

[0040] The specific method for calculating the syntactic information entropy value is as follows: ; In the formula, Indicates the first The grammatical information entropy of a function word measures the uncertainty of its grammatical function. A function word with a lower entropy value indicates that its grammatical function is more certain. Indicates the first The function word may carry the first Each grammatical role probability reflects the type of role that a function word plays in syntax; Indicates the first The boundary candidate values ​​at each position affect the strength of the function word's role in the syntactic hierarchy, and function words with larger boundary candidate values ​​are given priority. It is a boundary adjustment factor, used to adjust the degree of influence of boundary candidate values ​​on information entropy; Indicates the number of grammatical roles that a function word may play.

[0041] In this implementation scheme, by calculating the first... The probability of a function word in each grammatical role and the corresponding boundary candidate values ​​are determined, and combined with boundary adjustment factors and entropy correction terms, to finally obtain the grammatical information entropy value of the function word. This process effectively quantifies the grammatical function uncertainty of function words, providing a precise basis for subsequent prosodic boundary determination and grammatical function labeling, thereby helping to improve the accuracy and reliability of Tibetan prosodic structure prediction.

[0042] Specifically, the measures for function word anchor point marking, clause skeleton adjustment, and pause compression based on entropy feature analysis results are as follows: By comparing the grammatical information entropy value with the set entropy threshold in real time, this step can effectively determine the grammatical functional stability of function words. When the grammatical information entropy value is less than the entropy threshold, the function word is marked as a grammatically stable anchor point, thereby fixing its case relation marker, clause connection marker, and mood marker, ensuring that such function words will not be downgraded or merged in subsequent boundary optimization stages. In addition, according to dependency relationship expansion, related components are grouped into the same clause skeleton and synchronously registered in the clause-level index and focus anchor list to ensure accurate and tight prosodic phrase boundary segmentation in subsequent stages. Conversely, when the grammatical information entropy value is greater than or equal to the entropy threshold, the function word will be marked as a unit with undetermined grammatical function, triggering a grammatical role mitigation strategy. By checking whether the preceding and following combinations in the function word continuation constraint table conform to the legal order, inconsistent markings are rolled back, and the grammatical role probability distribution is updated by reviewing legal connection combinations. If the entropy value still does not reach a stable state after recalculation during this process, the function word will be prohibited from participating in the selection of prosodic focus anchors. Instead, a function word with a lower entropy value will be used, and its pause duration will be marked as a short pause level. This will avoid the misalignment of long sentence levels due to fluctuations in grammatical function and ensure that pauses and stress points in speech synthesis are more natural and accurate.

[0043] In this implementation, by dynamically adjusting the grammatical function markers of function words, for function words with stable grammatical functions, their grammatical roles are fixed to ensure that subsequent boundary optimization does not require unnecessary adjustments. Furthermore, by expanding dependency relations, related components are grouped into the same clause skeleton, improving the accuracy of subsequent prosodic phrase boundary segmentation. For unstable function words, a grammatical role mitigation strategy is triggered through backtracking and recalculating entropy values. This ensures that only function words that conform to legal combinations can serve as prosodic focus anchors, effectively preventing grammatical function fluctuations from causing sentence-level misalignment, ensuring natural connections between pauses and stress points, and optimizing the structural hierarchy of long sentences and the speech synthesis effect.

[0044] Specifically, the grammatical annotations and hierarchical information of candidate boundaries are read, and a three-layer prosodic boundary type sequence is generated based on boundary typing to suppress excessively dense pauses. The specific measures are as follows: Within the candidate boundary list, for each candidate position, the grammatical function annotations of function words, clause-level indexes, semantic focus anchors, and grammatical information entropy values ​​are read. A boundary typing mapping relationship is first constructed, combining function word category labels, clause-level index values, semantic focus anchor labels, and grammatical information entropy values ​​as input features. A mapping table is used to provide the correspondence and priority between the input features and the three target labels: prosodic word boundaries, prosodic phrase boundaries, and intonation phrase boundaries. Based on this, a three-layer boundary type sequence is output according to the boundary typing rules, such as... Figure 7 This is a flowchart of the three-layer prosodic boundary assignment and suppression process in this embodiment. The boundary assignment rules include: when the grammatical function of a candidate position is marked as a case particle, the candidate position is determined as a prosodic boundary, and the case particle and the syntactic component adjacent to the left of the candidate position are merged into the same prosodic word; when the grammatical function of a candidate position is marked as a tense particle or a nominalized suffix, the candidate position is determined as a prosodic boundary and the prosodic word is closed; when the grammatical function of a candidate position is marked as a mood particle, the candidate position is determined as an intonation phrase boundary and sentence-end intonation control information is written to ensure that the rhythm of the sentence termination position is consistent with the mood marker; when the candidate position hits the sentence-end function word marker, the candidate position is determined as an intonation phrase boundary and cross-sentence connection control information is written to constrain the prosodic continuity relationship and intonation transition mode of the preceding and following sentences; when the grammatical function of a candidate position is marked as a connective particle and the sentence-end intonation phrase is closed ... When marking related components in complex sentences, candidate positions are identified as prosodic phrase boundaries. The start and end boundaries of clauses are aligned to candidate positions based on clause-level indexes to ensure that pauses and modification range boundaries are consistent. In the mapping table, these connecting components and clause-level information are mapped to prosodic phrase boundary labels. After determining the boundary type, the role of syntactic entropy in the mapping relationship is utilized. Candidate positions with syntactic entropy values ​​less than the entropy threshold retain their boundary type and do not participate in downgrading and merging. Candidate positions with syntactic entropy values ​​greater than or equal to the entropy threshold and located in the candidate boundary density excess range are suppressed. Boundaries aligned with semantic focus anchors and clause skeletons are preferentially retained according to priority rules, resulting in a three-layer boundary type sequence: prosodic word boundaries, prosodic phrase boundaries, and intonation phrase boundaries. This provides a clearly layered prosodic structure input for subsequent rhythm evaluation and speech synthesis stages.

[0045] In this implementation scheme, by uniformly mapping function word categories, clause-level indexes, semantic focus anchors, and grammatical information entropy to three types of labels—prosodic word boundaries, prosodic phrase boundaries, and intonation phrase boundaries—the pause positions are aligned with clause start and end boundaries, modification scope boundaries, and intonation termination positions under the triggering of case particles, tense particles, nominalized suffixes, mood-related function words, sentence-final function words, and conjunction-related function words. At the same time, the scheme maintains boundary stability for candidate positions with grammatical information entropy close to the lower limit and suppresses and downgrades dense candidate positions with high grammatical information entropy. This enables long sentences to form a boundary sequence with clear layers, prominent focus, and moderate pause density at the prosodic level, providing a structurally reliable prosodic grouping foundation for subsequent rhythm evaluation and speech synthesis stages.

[0046] Specifically, the measures for evaluating rhythm intensity using pause duration differences and grammatical information entropy data are as follows: Obtain the duration difference between the m-th prosodic phrase and each pause segment, the time nodes of each pause segment, the start and end time nodes of the m-th prosodic phrase boundary, and the grammatical information entropy value. Divide the continuous time axis into several discrete time slices with a fixed time granularity and assign corresponding pause segment records and grammatical information entropy records to each time slice. For each pause segment, divide the duration difference by one and add the product of the adjustment weight factor and the absolute value of the time node difference to obtain a single pause contribution item. The adjustment weight factor is set based on corpora with pause annotations through offline statistical analysis, with a value range between 0 and 1, used to control the increase of time intervals. The decay rate of pause contributions is determined; all pause contributions are accumulated in discrete time slices within a given time range, and the integral operation is discretized into a summation calculation to obtain the basic rhythm intensity term; the rhythm correction factor is obtained by multiplying the information weighting factor with the grammatical information entropy value and adding one. The information weighting factor is set by comparing the consistency between different grammatical levels and the auditory rhythm annotation during the training phase, and its value ranges from 0 to 1. It is used to adjust the influence weight of grammatical information entropy in rhythm evaluation; the basic rhythm intensity term is multiplied with the rhythm correction factor to finally obtain the hierarchical rhythm intensity value of the m-th prosodic phrase, where m represents the number of the prosodic phrase in the current sentence, which starts from 1 and increases and does not exceed the total number of prosodic phrases in the sentence.

[0047] The specific calculation method for the layered rhythm intensity value is as follows: ; In the formula, Indicates the first The hierarchical rhythm intensity value of a rhythmic phrase reflects whether the rhythmic phrase needs to be further subdivided into rhythmic sub-phrases or rhythmic major phrases. Indicates the first The first rhythmic phrase and the first The difference in the duration of each pause reflects the duration of the pause in the rhythmic phrase; This indicates an adjustment weighting factor used to balance the impact of pause duration on prosodic segmentation; This indicates the time point for each actual pause period. It indicates the start and end time points of each prosodic phrase, reflecting the specific moment of the pause; Entropy represents grammatical information and is used to weight the alignment of pause duration with grammatical levels. Function words with lower entropy values ​​have higher certainty at the grammatical level. This represents the information weighting factor, used to adjust the degree of influence of syntactic information entropy on the duration of subdivision pauses.

[0048] In this implementation scheme, a hierarchical rhythm intensity value is obtained by integrating the differences in pause duration, temporal position, and grammatical information entropy within the discrete time granularity. This provides a quantitative basis for distinguishing between short and long pause levels, controlling the boundary positions of prosodic large phrases and prosodic small phrases, tightening the rhythm at the end of sentences, and maintaining consistency between stress and semantic focus. This stabilizes the pause hierarchy structure, alleviates rhythm fragmentation, and improves the clarity of semantic hierarchy perception during the synthesis of long sentences in Tibetan.

[0049] Specifically, the measures for phrase level subdivision and pause label adjustment based on rhythm assessment results are as follows: By comparing the level rhythm intensity value and intensity threshold in real time, when the level rhythm intensity value is less than the intensity threshold, a conservative subdivision strategy is implemented. Within the determined prosodic phrase boundaries, the boundary type remains unchanged, and no new segmentation points are added. Only the pause duration around the boundary of the same prosodic phrase is converged, and the pause labels are uniformly compressed to the short pause level and the continuous occurrence of adjacent short pauses is suppressed to avoid introducing redundant levels and causing rhythm fragmentation in segments with indistinct pause duration differences. When the level rhythm intensity value reaches or exceeds the intensity threshold, an enhanced subdivision strategy is implemented. The level rhythm intensity value is used as the subdivision priority, and subdivision is performed on segments of the same rhyme group. Candidate positions for rhythmic phrase boundaries are sorted by hierarchical rhythmic intensity values, with the highest-ranked positions prioritized as prosodic large phrase boundaries. Lower-ranked similar boundaries are reclassified as prosodic small phrase boundaries. Pause duration differences drive hierarchical separation between short and long pauses. When both prosodic large and small phrase conditions are met at the same position, the final level and boundary type are determined by the side with the hierarchical rhythmic intensity value closer to the intensity threshold, according to a priority table. This priority table is statistically derived from Tibetan speech-to-text corpus with phrase segmentation and pause annotations. For scenarios where multiple boundary conditions are met at the same point, the frequency of prosodic large phrases, prosodic small phrases, intonation phrases, and prosodic words selected as the final level is statistically analyzed, combined with clause level. The matching relationships of indexes, semantic focus anchors, grammatical function annotations of function words, and differences in pause duration are used to generate a priority matrix with hit condition combinations as rows and boundary levels as columns. Matrix cells record priority numbers, ensuring a unique level decision for each hit condition and avoiding repeated switching. For excessively long continuous modification chains at the end of sentences, the hierarchical rhythm intensity value is calculated around the boundary of the sentence-end intonation phrase, and the segmentation point with the larger hierarchical rhythm intensity value is selected to move forward the prosodic large phrase boundary, increasing the distance between the sentence-end intonation phrase boundary and the adjacent prosodic large phrase boundary to tighten the end rhythm. When performing dynamic programming segmentation on the prosodic phrase candidate boundary sequence, the hierarchical rhythm intensity value is used as a cost term in path scoring. This cost term is determined by the clause boundary. The path cost is composed of penalties for missing clauses, broken head chains, prosodic phrase length deviations, and short pauses. These penalties are weighted and synthesized into a path cost. When clause boundary coverage, head chain integrity, and prosodic phrase length distribution constraints are simultaneously satisfied, segmentation points with higher hierarchical rhythm intensity values ​​are prioritized. The output includes prosodic word boundaries, prosodic short phrase boundaries, prosodic long phrase boundaries, and intonation phrase boundaries, along with semantic focus anchor information. At the speech synthesis end, prosodic word boundaries are mapped to pause symbols for insertion. The prosodic short phrase and prosodic long phrase boundaries are combined with semantic focus anchors to control the fundamental frequency curve and duration scaling. Intonation phrase boundaries are used for sentence-end intonation constraints and cross-sentence connections, achieving detailed pause hierarchy and semantic level alignment. Figure 8 This is an example diagram of the prosodic structure generation in this embodiment.

[0050] In this implementation, segmentation decisions are driven by comparing the hierarchical rhythm intensity value with the intensity threshold. This ensures that prosodic phrases maintain stable segmentation in segments with insignificant pause duration differences and suppresses consecutive short pauses. In segments with significant pause differences, hierarchical selection is completed according to the sorting and priority matrix. Similar candidate boundaries are decomposed into prosodic small phrase boundaries and prosodic large phrase boundaries, forming a hierarchical separation of short and long pauses. At the same time, the sentence-end modification chain constraint moves the prosodic large phrase boundary forward to tighten the end rhythm. In the dynamic programming segmentation, a path cost term consisting of clause boundary coverage, head word chain integrity, length distribution deviation, and continuous suppression of short pauses is introduced to stabilize the whole sentence segmentation path. The final output is boundary encoding and semantic focus anchor information that can be directly mapped to the speech synthesis end. This ensures that pause symbol insertion, fundamental frequency curve control, and duration scaling have a consistent hierarchical basis, thereby alleviating the increased understanding burden caused by chaotic pause positions in long sentences, rhythm fragmentation, sentence-end dragging, and semantic hierarchical misalignment.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting Tibetan prosodic structure based on grammatical information, characterized in that, Includes the following steps: S1 completes Tibetan text parsing and function word continuation quantification, extracts grammatical probability, hierarchical focus and pause differences, and stores them after preprocessing to construct a Tibetan prosody prediction database; S2 performs boundary discrimination based on the strength of function word continuation relationship, the hit mark of the influence range and the word order distance, and realizes candidate boundary filtering, concatenation segment merging, pause label configuration and grammatical function annotation update; The specific measures for boundary discrimination based on the strength of function word continuation relationships, the hit marker of the scope of influence, and the word order distance are as follows: Obtain the continuation relationship strength, boundary influence range indicator value, and word order distance corresponding to position i and all function word positions j; for each pair of positions i and j, multiply the continuation relationship strength by the boundary influence range indicator value to obtain the boundary contribution term; add the word order distance to a small constant and then square it to obtain the distance penalty term; divide the boundary contribution term by the distance penalty term to obtain the boundary contribution value corresponding to position i and position j. The boundary contribution values ​​at all positions j are summed to obtain the candidate boundary value at position i. S3 performs entropy feature analysis using grammatical role probability data and boundary judgment data, and performs function word anchor point marking, clause skeleton adjustment and pause compression operations based on the entropy feature analysis results. The specific measures for entropy feature analysis using grammatical role probability data and boundary determination data are as follows: Obtain the probability value of the k-th function word in each grammatical role and the corresponding boundary candidate value; multiply the probability of each grammatical role by its own logarithm to obtain the information content of a single grammatical role; The basic syntactic entropy value of the k-th function word is obtained by summing the information values ​​of all syntactic roles and taking their negative values; the boundary adjustment factor is multiplied by the corresponding boundary candidate value to obtain the boundary correction term, and one is added to obtain the entropy correction term; the basic syntactic entropy value and the entropy correction term are multiplied to finally obtain the syntactic information entropy value of the k-th function word. S4: Read the grammatical annotations and hierarchical information of the candidate boundaries, generate a three-layer prosodic boundary type sequence based on the boundary type and suppress excessively dense pauses; The specific measures for reading the syntactic annotations and hierarchical information of candidate boundaries, generating a three-layer prosodic boundary type sequence based on boundary typing, and suppressing overly dense pauses are as follows: Within the candidate boundary list, for each candidate position, the grammatical function annotation, clause-level index, semantic focus anchor, and syntactic information entropy value of the function words are read. A three-layer boundary type sequence is output according to the boundary assignment rules. These rules include: when the grammatical function annotation of a candidate position's function word is a case particle, the candidate position is determined as a prosodic word boundary, and the case particle is merged with the adjacent syntactic component to the left of the candidate position into the same prosodic word; when the grammatical function annotation of a candidate position's function word is a tense particle or a nominalized suffix, the candidate position is determined as a prosodic word boundary and prosodic word closure is completed; when the grammatical function annotation of a candidate position's function word is a mood particle, the candidate position is determined as an intonation phrase boundary and sentence-end intonation control information is written. When a candidate position is selected and a sentence-ending function word is marked, the candidate position is determined as an intonation phrase boundary and written into the cross-sentence cohesion control information. When a candidate position is marked as a connective function word and a compound sentence related component is marked, the candidate position is determined as a prosodic phrase boundary. The starting and ending boundaries of the subordinate clause are aligned to the candidate position according to the subordinate clause level index to ensure that the pause point is consistent with the modification range boundary. After the boundary type is determined, the boundary type of candidate positions with grammatical information entropy values ​​less than the entropy threshold is kept and does not participate in the downgrading and merging. When the candidate boundary density exceeds the limit, the boundary type of candidate positions with grammatical information entropy values ​​greater than or equal to the entropy threshold is suppressed, resulting in a three-level boundary type sequence of prosodic word boundary, prosodic phrase boundary, and intonation phrase boundary. S5 assesses rhythm intensity using pause duration differences and grammatical information entropy data, and performs phrase level subdivision and pause mark adjustment based on rhythm assessment results; The specific measures for evaluating rhythm intensity using pause duration differences and grammatical information entropy data are as follows: Obtain the duration difference between the m-th prosodic phrase and each pause, the time nodes of each pause, the start and end time nodes of the m-th prosodic phrase boundary, and the grammatical information entropy value; for each pause, divide the duration difference by one and add the product of the adjustment weighting factor and the absolute value of the time node difference to obtain a single pause contribution; perform an integral operation on all pause contributions within a given time range to obtain the basic rhythm intensity term; multiply the information weighting factor by the grammatical information entropy value and add one to obtain the rhythm correction factor; multiply the basic rhythm intensity term by the rhythm correction factor to finally obtain the hierarchical rhythm intensity value of the m-th prosodic phrase.

2. The Tibetan prosodic structure prediction method based on grammatical information according to claim 1, characterized in that: The specific measures for completing Tibetan text parsing and quantifying function word continuation, and extracting grammatical probabilities, hierarchical focus, and pause differences are as follows: Collect Tibetan text data: Receive the text to be processed through the input interface, perform word segmentation to obtain word sequences and part-of-speech sequences, and output a set of dependency arcs, a set of headword indexes, and a set of modification ranges based on syntactic dependency analysis; identify all function words in the word sequence, classify them into three categories based on grammatical features: case particles, general function words, and class-like function words, and mark and number each function word's position; collect function word continuation rules: extract legally connectable function word combinations, synonym substitution forms, and affix and suffix attachment information from the function word continuation constraint table formed based on corpus statistics and grammatical rules, generate a function word continuation relationship strength matrix and a word order distance matrix based on the word sequence to quantify function words, encode the matrix synchronously with the text sequence, and construct an indicator function matrix based on boundary assignment rules and candidate boundary setting conditions to determine whether each pair of function words is located within the same candidate boundary influence range to obtain the boundary influence range indicator value; collect syntactic probability data: perform grammatical role discrimination on the function word sequence, output the probability distribution of each function word's possible grammatical roles, and record it as a grammatical role probability sequence; Extracting clause hierarchical structure and semantic focus information: Based on dependency relations, modification scope, and clause division, determine the clause hierarchical index, clause start and end labels, and central word chains. Extract the boundaries of topic components and modifier components and register them as a list of semantic focus anchor points. Collecting prosodic phrase and pause-related data: In Tibetan speech-to-text corpus with phrase segmentation and pause annotations, record the boundary start and end time nodes of each prosodic phrase according to the phrase segmentation annotations. Obtain the time nodes of each actual pause period by aligning the pause annotations with the time axis. Calculate the pause duration difference between the prosodic phrase and the pause period.

3. The Tibetan prosodic structure prediction method based on grammatical information according to claim 1, characterized in that: The specific measures for constructing the Tibetan prosody prediction database after preprocessing are as follows: Function words with pause duration differences less than the pause threshold are labeled as short pauses, and function words with pause duration differences greater than the pause threshold are labeled as long pauses. The collected numerical data of function word continuation relationship strength, word order distance, and grammatical role probability are normalized to unify the range of all units. Based on the normalized feature data, the structured fields of function word category, continuation strength, grammatical role probability, dependency index, distance, hierarchical label, and focus identifier are concatenated in the text order to generate a homo-frequency feature vector. The function word sequence number, data batch identifier, and main control field range are then added. The feature vector and function word grammatical labels are stored synchronously to construct a Tibetan prosody prediction database.

4. The Tibetan prosodic structure prediction method based on grammatical information according to claim 1, characterized in that: The specific measures for implementing candidate boundary filtering, concatenation segment merging, pause label configuration, and syntax function annotation updates are as follows: By comparing the boundary candidate value and the boundary threshold in real time, when the boundary candidate value is less than the boundary threshold, the system checks whether there are any legally connectable combinations registered in the function word continuation constraint table between adjacent function words. For the detected continuous function word continuation segments, the system performs prosodic phrase boundary merging, cancels the temporary pause labels between all function words in the segment, and rolls back the grammatical function annotations triggered by the function word continuation segments. When the boundary candidate value is greater than or equal to the boundary threshold, the current position is determined as a valid prosodic boundary candidate. The position is retained in the candidate boundary list and marked as a prosodic boundary. At the same time, the legality of the combination of function words on the left and right sides of the position is checked according to the function word continuation constraint table. Short pause labels are assigned to function word continuation segments that meet the constraint rules, and grammatical functions are annotated simultaneously.

5. The Tibetan prosodic structure prediction method based on grammatical information according to claim 1, characterized in that: The specific measures for performing function word anchor point marking, clause skeleton adjustment, and pause compression operations based on the entropy feature analysis results are as follows: By comparing the grammatical information entropy value and the entropy threshold in real time, when the grammatical information entropy value is less than the entropy threshold, the corresponding function word is marked as a stable anchor point of grammatical function, a fixed case relation marker, a clause connection marker, and a mood marker, and will no longer be downgraded or merged in the subsequent boundary optimization stage; at the same time, it expands to both sides along the dependency arc, classifies the related components into the same clause skeleton, and registers them synchronously in the clause level index and the focus anchor list, so that the subsequent prosodic phrase boundary is prioritized to be close to the target position to complete the segmentation; When the grammatical information entropy value is greater than or equal to the entropy threshold, the corresponding function word is marked as a grammatical function pending unit, triggering a grammatical role mitigation strategy: the function word continuation constraint table checks whether the preceding and following combinations meet the legal order requirements, and for inconsistent grammatical function annotations, a fallback to the three categories of case particles, general function words, and class function words is performed. The function word continuation constraint table is then used to verify the legal connection combination of the function word pairs before and after the position. After updating the grammatical role probability distribution for the hit combination, the grammatical information entropy value is recalculated. If the grammatical information entropy value is still greater than or equal to the entropy threshold after recalculation, the position is prohibited from participating in the prosodic focus anchor selection. Instead, the function word with the nearest grammatical information entropy value to the left and right of the position that is less than the entropy threshold is used as the focus anchor, and the pause duration label corresponding to the position is marked as a short pause level to avoid binding the pause and stress point to the grammatical function fluctuation position, which would lead to misalignment of long sentence levels.

6. The Tibetan prosodic structure prediction method based on grammatical information according to claim 1, characterized in that: The specific measures for implementing phrase hierarchy subdivision and pause marker adjustment based on rhythm assessment results are as follows: By comparing the layered rhythm intensity value and intensity threshold in real time, when the layered rhythm intensity value is less than the intensity threshold, a conservative subdivision strategy is executed. Within the determined prosodic phrase boundary, the boundary type remains unchanged and no new segmentation points are added. Only the pause duration around the boundary of the same prosodic phrase is converged. The pause labels are uniformly compressed to the short pause level and the continuous occurrence of adjacent short pauses is suppressed, so as to avoid introducing redundant levels in segments with indistinct pause durations and causing rhythm fragmentation. When the layered rhythm intensity value is greater than or equal to the intensity threshold, an enhanced segmentation strategy is implemented: the layered rhythm intensity value is used as the segmentation priority, candidate positions that are both prosodic phrase boundaries are sorted according to their layered rhythm intensity values, and the positions with higher rankings are prioritized as prosodic large phrase boundaries, while the lower-ranked similar boundaries are downgraded to prosodic small phrase boundaries. The difference in pause duration drives the separation of short and long pauses into different levels. When the same position simultaneously meets the conditions for both prosodic large and prosodic small phrases, the final level is determined and the boundary type is locked according to the priority table based on the side with the layered rhythm intensity value closer to the intensity threshold, avoiding repeated switching. In segments with excessively long continuous modification chains at the end of a sentence, the layered rhythm intensity value is calculated around the boundary of the sentence-end intonation phrase, and the segmentation point with the larger layered rhythm intensity value is selected to move the prosodic large phrase forward. To tighten the rhythm at the end of a sentence, the boundary of the intonation phrase at the end of the sentence is widened, increasing the distance between it and the boundary of the adjacent prosodic phrase. When performing dynamic programming segmentation on the candidate boundary sequence of prosodic phrases, the hierarchical rhythm intensity value is used as a cost term in the path scoring. When the constraints of clause boundary coverage, central word chain integrity, and prosodic phrase length distribution are simultaneously satisfied, the segmentation point with the larger hierarchical rhythm intensity value is prioritized. The output includes the encoding of prosodic word boundaries, prosodic small phrase boundaries, prosodic large phrase boundaries, intonation phrase boundaries, and semantic focus anchor information. At the speech synthesis end, the prosodic word boundaries are mapped to pause symbols for insertion. The prosodic small phrase and prosodic large phrase boundaries are combined with the semantic focus anchor to control the fundamental frequency curve and duration scaling. The intonation phrase boundaries are used for sentence-end tone constraints and cross-sentence connections, achieving pause level subdivision and semantic level alignment.

Images