Tibetan written language character-voice conversion method based on voice structure characteristics

Abstract

The invention, which belongs to the technical field of data processing, discloses a Tibetan language written language character-voice conversion method based on voice structure characteristics, comprising the following steps: S1, obtaining a Tibetan language text; s2, performing text segmentation processing on the Tibetan text by taking syllables as units, and extracting font structure elements of each syllable based on the font structure of the Tibetan to obtain a syllable font structure element set; s3, performing matching processing on initial consonant items and final consonant items of each syllable based on the syllable font structure element set to obtain a voice structure feature vector of each syllable; and S4, obtaining an initial consonant conversion coding table and a vowel conversion coding table, and mapping the voice structure feature vector of each syllable into a corresponding syllable phoneme sequence, thereby obtaining the Tibetan written language audio. According to the method, the problems that the basic characteristics of the Tibetan speech cannot be accurately described in a Tibetan language Latin transcription mode in the existing Tibetan speech synthesis technology and the like, and the calculation amount is too large and the character and speech conversion speed is low due to multilayer combination of Tibetan language phoneme sequences are solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for converting Tibetan written language into phonetic characters based on phonological structural features. Background Technology

[0002] Tibetan written text phono-translation refers to converting semantically descriptive original text into phonetic text form, that is, transcribing Tibetan syllables into corresponding phoneme sequences. This process accurately depicts the initial and final consonant formation rules of modern Tibetan, revealing the basic phonological structure of written Tibetan. Current Tibetan written text phono-translation methods obtain Tibetan syllable sequences by identifying the structure of Tibetan characters, then further determine the semantic granularity between syllables based on the part-of-speech differentiation and gradient distance between syllables. The semantic granularity and syllable recognition loss are used to determine the word contribution of each syllable. By combining the syllable sequences at multiple scales, syllable segments are obtained, and segment priorities are determined using granular entropy and segment semantic loss, ultimately achieving gradient translation of Tibetan text.

[0003] For example, Chinese invention patent CN118898259B discloses a pre-trained model-based Chinese-Tibetan translation device and Tibetan translation method, which includes: obtaining a Tibetan syllable sequence by recognizing the structure of Tibetan characters; determining the semantic granularity between Tibetan syllables by the part-of-speech discrimination of Tibetan syllables and the gradient distance between each Tibetan syllable; determining the word contribution of each Tibetan syllable in the Tibetan translation process by the semantic granularity and the loss of syllable recognition; performing multi-scale combination of Tibetan syllables in the Tibetan syllable sequence to obtain syllable segments; determining the granularity entropy of Tibetan syllables in each syllable segment; determining the segment priority of each syllable segment in the Tibetan translation process by the granularity entropy and the semantic loss of the syllable segment; and performing gradient translation of the Tibetan text based on the contribution of all words and the priority of all segments.

[0004] For example, Chinese invention patent CN118761389B discloses a Tibetan machine translation system and an automatic Tibetan text segmentation method, which includes: acquiring Tibetan text information; extracting sentence representations from all Tibetan sentences in the Tibetan text information to obtain multiple Tibetan sentence representation sequences; further determining the sentence granularity of each Tibetan sentence, thereby dividing all Tibetan sentences into topics to obtain multiple semantic topic clusters; extracting part-of-speech vectors of different Tibetan keywords from the Tibetan text information to determine the inter-word transfer features of each Tibetan keyword in each Tibetan sentence, and further determining the inter-word transfer degree of different Tibetan keywords; and then determining the paragraph boundary features of the Tibetan text information, thereby adaptively segmenting the Tibetan text information based on the paragraph boundary features.

[0005] The above-mentioned technology has at least the following technical problems: Existing technologies mainly achieve word-to-speech conversion and text processing through syllable or sentence structure analysis, semantic granularity division, and multi-scale combination. However, on the one hand, due to the inaccurate word-to-speech mapping, it is impossible to fully characterize the basic features of Tibetan speech, resulting in pronunciation errors, training difficulties, and word-to-speech conversion errors in the case of sparse data. On the other hand, Tibetan phoneme sequences contain a large number of complex structures caused by multi-layered combinations of initials and finals, which drastically increases the computational load and memory consumption when processing phoneme sequences, resulting in slow word-to-speech conversion speed. Summary of the Invention

[0006] To address the problem that existing technologies cannot accurately depict the basic phonetic features of Tibetan through Tibetan Latin transliteration, this invention provides a method for converting written Tibetan characters based on phonetic structure features. The method includes the following steps: S1, obtaining a set of written Tibetan text to be processed and standardizing it to obtain Tibetan text; S2, segmenting the Tibetan text into syllables and extracting the glyph structure elements of each syllable based on the glyph structure of Tibetan characters to obtain a syllable glyph structure element set; S3, matching the initial consonant and final vowel items of each syllable based on the syllable glyph structure element set to obtain the phonetic structure feature vector of each syllable; S4, obtaining an initial consonant conversion encoding table and a final vowel conversion encoding table, thereby mapping the phonetic structure feature vector of each syllable to the corresponding syllable phoneme sequence, and synthesizing all syllable phoneme sequences to obtain a phonetic symbol sequence; S5, performing consistency verification and homophone merging based on the phonetic symbol sequence to obtain a phoneme text set, thus obtaining the audio of written Tibetan.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. The Tibetan written language pronunciation conversion method based on speech structure features provided by this invention sequentially performs standardization processing, syllable segmentation processing, and extraction of syllable glyph structure elements based on Tibetan glyph structure to identify the initial consonant and final vowel items of each syllable. Furthermore, it obtains speech structure feature vectors by matching the initial consonant and final vowel items, thereby revealing the composition rules of initial consonants and final vowels in modern Tibetan written language. This achieves high accuracy and consistency in pronunciation conversion and effectively solves the problem that the existing technology cannot accurately depict the basic features of Tibetan speech through Tibetan Latin transliteration.

[0008] 2. This invention obtains the initial consonant conversion encoding table and the final vowel conversion encoding table, and maps the speech structure feature vector to the corresponding syllable phoneme sequence based on the conversion relationship, thereby realizing the standardized generation of speech symbol sequences; then, through consistency verification and homophone merging processing, a phoneme text set is formed, which effectively solves the problem of inaccurate word pronunciation conversion caused by data sparsity and unstable phoneme sequences in the prior art.

[0009] 3. This invention extracts and optimizes the multi-layered combination structure of initials and finals in Tibetan phoneme sequences by performing hierarchical feature extraction and optimization, thereby effectively reducing the pressure on computing resources caused by sequence dependency span. This achieves a significant reduction in computation and memory usage while maintaining high-precision phonetic conversion, effectively solving the problems of synthesis delay accumulation and response jitter caused by excessively long sequences in the prior art.

[0010] 4. This invention obtains text processing parameters based on phoneme text set analysis and calculates phoneme density coefficients and cross-structural correlation spans to obtain conversion feasibility prediction labels. Then, it adjusts the attention window length, frame shift, and batch processing bucket width of the phoneme conversion processing unit to achieve initial adjustment of phoneme conversion processing. This effectively solves the problem of excessive computation and slow phoneme conversion speed caused by multi-layer combination of Tibetan phoneme sequences in the prior art. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of a Tibetan written language phonetic conversion method based on phonological structure features provided in this application embodiment; Figure 2 A schematic diagram of Tibetan phonetic structure for the Tibetan written language phonetic conversion method based on phonetic structure features provided in this application embodiment; Figure 3 A distribution map of Tibetan written consonant initials for the Tibetan written language phonetic conversion method based on phonological structure features provided in this application embodiment; Figure 4 A distribution diagram of the Tibetan written vowel system for the Tibetan written language phonetic conversion method based on phonological structure features provided in this application embodiment; Figure 5 The Tibetan syllable structure writing rules and syllable conversion structure diagram of the Tibetan written language phonology conversion method based on phonological structure features provided in the embodiments of this application; Figure 6 A schematic diagram of the basic composition structure of Tibetan syllables in the Tibetan written language pronunciation conversion method based on speech structure features provided in the embodiments of this application. Detailed Implementation

[0013] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0014] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0015] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0016] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, Figure 1 The flowchart of the Tibetan written language phonetic conversion method based on phonetic structure features provided in this application embodiment includes the following steps: S1, obtaining a Tibetan written language text set to be processed and performing standardization processing to obtain Tibetan text; S2, segmenting the Tibetan text into syllables and extracting the glyph structure elements of each syllable based on the glyph structure of Tibetan, obtaining a syllable glyph structure element set; S3, using the first vowel of the Tibetan syllable as the boundary, matching the initial consonant and final vowel items of each syllable based on the syllable glyph structure element set to obtain the phonetic structure feature vector of each syllable, wherein the matching processing is to decompose the glyph structure within each syllable in a phonetic sense, converting it from "written structure" into "phonetic structure feature vector"; S4, obtaining The initial consonant conversion encoding table and the final vowel conversion encoding table are used to map the speech structure feature vector of each syllable to the corresponding syllable phoneme sequence. The phoneme sequences of all syllables are then synthesized to obtain the speech symbol sequence. Specifically, a unified phoneme pronunciation symbol is used to obtain the initial consonant pronunciation symbol and the final vowel pronunciation symbol. The speech structure feature vector of each syllable (note: the speech structure feature here refers to the Tibetan initial consonant set and final vowel set) is then mapped to the corresponding initial consonant phoneme sequence and final vowel phoneme sequence. The phoneme sequences of all syllables are then synthesized to obtain the speech symbol sequence of the Tibetan text. The speech symbol sequence is a string used to represent the pronunciation form of the Tibetan text. S5. Based on the speech symbol sequence, consistency verification and homophone merging are performed to obtain the phoneme text set, thereby obtaining the Tibetan written audio.

[0018] In this embodiment, the standardization process specifically involves converting the input text into a unified Unicode encoding format (e.g., Tibetan segments such as U+0F40–0FBC) to ensure that characters can be correctly recognized during subsequent glyph extraction. This means converting non-Tibetan characters (i.e., special symbols) such as Arabic numerals and English letters into their corresponding Tibetan characters; removing erroneously entered numbers, English letters, and miscellaneous symbols from the text to avoid abnormal boundaries in syllable recognition, thereby obtaining pure Tibetan text; and decomposing and standardizing the complex encodings generated by traditional compound characters, prefix characters, and superscript characters to ensure that prefix characters, base characters, and other elements can be identified one by one during glyph extraction.

[0019] Tibetan texts are segmented into syllable units. Specifically, according to Tibetan orthography, continuous text is divided into standard syllable units. The segmentation is based on the fact that the structure of a Tibetan syllable is fixed as “prefix + superfix + base + subfix + vowel + suffix + further suffix”. It should be noted that not all Tibetan syllables are composed of these elements; some elements may be missing, i.e., it varies from character to character.

[0020] Traditional Tibetan grammar, including the *Orthography* and *Character Grammar*, comprehensively reflects the phonetic features of the Tibetan language. They profoundly reveal the core elements of universal linguistics: the concepts of "initial consonant" and "final vowel," which constitute the basic framework of phonetics. In the structure of Tibetan syllables, the vowel, as the core of the syllable, is divided into two main groups: the base syllable before the vowel, the prefix syllable, the superscript syllable, and the subscript syllable. These four components collectively reflect the Tibetan initial consonant system. The vowels and their subsequent suffixes, and so on, collectively reflect the Tibetan vowel system. ).

[0021] like Figure 2 As shown, Figure 2 A schematic diagram of Tibetan phonetic structure for the Tibetan written language phonetic conversion method based on phonetic structure features provided in this application embodiment. Figure 2 It visually presents the blueprint of modern Tibetan phonetics. Using syllables as an example... For example, add the word "" before it. Adding a character above "、base character" "and underlined words" (“ The consonant initial (composed of the variant form of "") is then added to form the word "". "and then add words" "Combined to form" This vividly showcases the unique charm of consonant endings, among which " The '☐' symbol serves as a vowel marker, embellishing the vowels. This phonetic structure almost completely preserves the characteristics of the ancient Tibetan phonological system from around the 9th century. In the representative work "Classical Tibetan Language," which provides a relatively comprehensive study of Tibetan language from the perspective of modern linguistics, the compositional structure of Tibetan phonemes is divided into three parts: onset (the group of consonants before the core vowel, consisting of up to four consonants), nucleus (the monophthong), and coda (the consonant following the core vowel, usually one or two).

[0022] In Tibetan syllable structure, the base syllable, superscript syllable, subscript syllable, and prescript syllable combine to express consonant initials. Based on this characteristic, Tibetan consonant initials can be divided into 29 simple consonants (…). ) and 202 consonant clusters ( Based on this, consonant clusters are further subdivided into 124 diphthongs, 72 triphthongs, and 6 tetraphthongs.

[0023] Furthermore, the glyph structure of Tibetan includes initials and finals. Initials include prefixes, superscripts, bases, subscripts, and sub-subscripts, while finals include vowels, suffixes, sub-suffixes, consonants, diphthongs, second consonants, third consonants, and third vowels.

[0024] Furthermore, the syllable glyph structure element set is obtained. The specific method is as follows: the Tibetan text is segmented into syllables, and syllables are identified based on the syllable boundary rules of Tibetan orthography. Syllable identification specifically includes identifying whether each syllable contains prefixes, superscripts, subscripts, vowels, suffixes, and further superscripts; the corresponding glyph structure elements are extracted from each syllable to obtain the glyph structure elements of each syllable; the glyph structure elements of each syllable are arranged according to the glyph hierarchy of the syllables to obtain the syllable glyph structure element set.

[0025] In this embodiment, the corresponding glyph structure elements are extracted from each syllable to obtain the glyph structure elements of each syllable. The specific steps are as follows: locate the base character of the syllable as the core of the current syllable, scan and identify the preceding and following characters from the left side of the base character; identify the following characters from below the base character; identify the following characters from the right side of the base character in writing order: vowel marks, following characters, and further following characters; then sort the seven types of characters identified in a fixed hierarchical order to generate the "syllable glyph structure element set" as the parsing input.

[0026] It should be noted that the structure of Tibetan script is based on a single consonant letter, the "base letter," upon which all other characters are built. Syllables can be formed by horizontal combinations from left to right or vertical stacking from top to bottom, creating a two-dimensional structure. The specific hierarchical order of the syllable shapes is as follows: Figure 6 As shown, Figure 6This is a schematic diagram of the basic composition structure of Tibetan syllables in the Tibetan written language pronunciation conversion method based on phonological structure features provided in the embodiments of this application. In the composition structure of Tibetan syllables, the prefix, base, suffix, and subsequent suffix are spelled in a horizontal combination form, while the supernumerary, base, sub-suffix, and vowel are spelled in a vertical superposition form. The vertical arrangement form (the combination of superimposed characters) in a syllable is called a "character".

[0027] Furthermore, the speech structure feature vectors of each syllable are obtained. The specific method is as follows: locate the initial consonant item of the syllable based on the syllable glyph structure element set; locate the final vowel item of the syllable based on the syllable glyph structure element set; arrange the initial consonant item and the final vowel item according to the syllable structure combination order to form the initial-final structure of the syllable; perform combination calculation on the internal encoding of the initial consonant item and the final vowel item to obtain the speech structure feature vector of each syllable. The speech structure feature vector of each syllable is used to characterize the initial-final combination relationship of the syllable.

[0028] In this embodiment, the internal codes of the initial consonant and final vowel items are combined and calculated. The internal code combination calculation method is as follows: the internal codes of the initial consonant items are concatenated in order; the internal codes of the final vowel items are concatenated in order; and they are combined into a single vector in the manner of "initial consonant code sequence + final vowel code sequence".

[0029] Furthermore, the initial consonant conversion encoding table and the final vowel conversion encoding table are obtained as follows: The set of consonant initials and the set of finals in Tibetan written language are obtained. The consonant initial set includes single consonant initials and consonant compound initials; the consonant compound initials include diphthongs, triphthongs, and tetraphthongs; the final set includes vowel finals and consonant codas; the vowel finals include monophthongs, diphthongs, and triphthongs; the consonant codas include single consonant codas and consonant codas; the consonant codas include structures formed by combinations of suffixes and further suffixes. The initial consonant set and the final vowel set are internally encoded, and internal encoding numbers are generated based on the positional relationship of the initials and finals in the syllable structure. Character mapping is then performed based on the internal encoding numbers and preset mapping rules (such as the mapping rules in Table 1), thereby obtaining the initial consonant conversion encoding table and the final vowel conversion encoding table.

[0030] In this embodiment, a single consonant initial is a syllable consisting of only one consonant letter, such as " Originally, it covered 30 consonant letters. But as mentioned earlier, " "At the phonetic level, it mainly functions as a vowel component and is an indispensable part of the pronunciation of the other 29 consonant letters. In this chapter, it is classified as a final rather than an initial. Therefore, strictly speaking, there are 29 single consonant initials in written Tibetan."

[0031] Two-component consonant initials include four different structures: the first is the "additional syllable + base syllable" structure ( ),like" "etc., a total of 33 consonant clusters; secondly, the "base character + additional character" structure ( ),like" There are a total of 42 consonant clusters, including "etc." Among them, as stated in *Tibetan Language Reading: A Beginner's Guide*: , meaning to add a word below. "It can combine with all consonants, but only 15 characters are counted for now. Therefore, it can be seen that it combines with underscores..." "The initial consonant formed by combining consonants has a certain degree of uncertainty; this chapter only uses the generally accepted..." "These 15 combinations, but like..." Some initials, such as "", are rarely seen in modern texts and are trending towards disappearance; thirdly, the "prefix + base character" structure, such as " "etc.", totaling 48. Fourth, the newly added consonant initial " ", composed of the consonant " "and" It is composed of "combinations".

[0032] The three-component consonant clusters also include four structures: one is the "upper consonant + base consonant + lower consonant" structure ( ),like" There are 15 in total, including "etc."; the second type is the "preceding character + superscript character + base character" structure, such as " There are 20 in total, including "etc."; the third is the structure of "preceding character + base character + following character", such as " "etc., a total of 34; fourth, the special structure of "base character + additional character + additional character" ( ),Right now There are a total of 3.

[0033] The four-component consonant clusters have only one structure: "prefix + superposition + base + subposition," and the consonant combination in this structure is specific, consisting only of the prefix "..." Adding a character above "or" “、Additional characters” "or" ", respectively: There are a total of 6.

[0034] The above analysis shows that the analysis and statistical scope of the Tibetan initial consonant system are related to the so-called basic characters in modern Tibetan linguistics and other related studies. There is some correlation, but slight differences exist in the statistical results. In summary, there are 231 basic characters in written Tibetan, corresponding to Tibetan consonant initials. To comprehensively cover all possible phenomena, this paper did not perform special processing such as frequency statistics on initials that appear very rarely in the text (referred to as "stop words" in some research fields). The distribution of the initial consonant system in modern written Tibetan is as follows: Figure 3 As shown, Figure 3 The distribution map of Tibetan written consonant initials for the Tibetan written language phonetic conversion method based on phonological structure features provided in the embodiments of this application.

[0035] As stated in "The Fundamentals of Tibetan Phonetics": ", meaning that when vowels overlap, they form a compound vowel ( When consonants come together, they form a consonant cluster ( ). In the Tibetan phonological system, vowels and consonant codas intertwine to form a complex and sophisticated vowel system. Vowels are categorized into two main types: monophthongs and diphthongs. Consonant codas are manifested through the flexible use of syllable elements such as adding syllables after consonants, and further divided into simple consonant codas and complex consonant codas based on differences in syllable structure. A detailed analysis follows: Vowel phonology: The syllable structure in modern Tibetan texts is not a simple two-dimensional combination, but can be summarized into seven glyph structures, primarily stemming from the diversity of vowel finals. Based on these seven glyph structures, Tibetan syllables contain at least one, and at most three, vowels, thus giving modern written Tibetan vowels a variety of classifications: monophthongs, diphthongs, and triphthongs. The latter two are collectively referred to as diphthongs. Vowel phonology is divided into monophthongs and diphthongs, among which monophthongs (… ), that is, a syllable contains only one basic vowel. It is composed of the dominant " "4 vowel symbols and 1 hidden vowel" "The structure showcases the fundamental beauty of the Tibetan vowel system. Among them, the compound vowels are as exquisitely described in *The Fundamentals of Tibetan Phonetics*:" When two or more vowels coexist harmoniously in the same syllable, a diphthong is created. Based on the number of vowels, diphthongs can be further subdivided into diphthongs and triphthongs. It is worth noting that Tibetan syllables do not contain more than three vowels.

[0036] Diphthongs A diphthong is a compound vowel formed by two vowels within a single syllable. Based on structural differences, diphthongs can be divided into two categories: firstly, according to the Tibetan grammar text *Thirty Verses*... The record in the document states: This reveals the Tibetan genitive particle ( ) "、Terminator ( ) “、Decorative Collection of Words ( ) Separable words ( ) , and the word-forming suffix "" indicating "tiny". These five function words can be combined with the preceding syllable to form a simplified form (referred to as "agglutinative words" or "contracted words" in other related studies) without affecting their meaning. Among them, " "is another form of vowel expression, while" This is another manifestation of a consonant cluster final. Specifically, when the preceding syllable (i.e., an "open syllable") has no following syllable or the following syllable is " When “”, any one of these five function words can be combined with it to form a compound vowel containing two vowels (excluding “”). "and" There are 15 such compound vowel rhymes, with the vowel " "function words" Taking the combination of an open syllable preceding a vowel containing any one of the five basic vowels as an example, its structure is as follows: In these merged syllable structures, the word "" appears. The vowel combination of "" creates a harmonious rhyme of compound vowels. The other four function words " The merger structure of "" can be deduced in the same way. Furthermore, " "and" "Besides usually indicating the end of a word and insignificant, they are also used to spell foreign sounds, for example:" (secretary), (Liaoning) (Jiangsu) (Ministry of Education), etc., demonstrate the Tibetan vowel system's inclusiveness and integration of foreign words. Secondly, in the process of transliterating words from other languages ​​into Tibetan, the component "..." is often added after the preceding open syllable. This gives rise to a new syllable form, for example, "Shanghai" is transliterated as "…". "Guizhou" is transliterated as " "Anhui" is transliterated as " These syllables also contain diphthongs, forming another example of a diphthong. There should be a total of 5 such diphthongs, and their structure is similar to the first type, namely: However, this syllable structure does not conform to the traditional spelling rules of Tibetan, and can be considered to reflect the flexibility and innovation of the Tibetan phonetic system in transliterating foreign words.

[0037] Triphthongs ( ) are compound vowels that further develop on the basis of diphthongs. They are usually formed by adding the affix component " " that indicates "tiny" to form a diphthong, and then merging with the genitive particle " ". For example, in syllables such as " ", three vowels appear simultaneously, jointly weaving a magnificent chapter of compound vowels. There should be a total of 5 such compound vowel rhymes, and their spelling structures are respectively: .

[0038] In addition, there is another form of compound vowel of triphthong that is formed by adding the affix component " " to form a diphthong and then merging with the terminal word " ". For example, " ", etc., which usually appears at the end of a sentence to indicate termination. There should also be a total of 5 such compound vowel rhymes, and their structures are quite similar to the above triphthongs, respectively: . The emergence of these compound vowel rhymes further enriches the expressiveness and appeal of the Tibetan vowel system.

[0039] Based on the above analysis, it can be seen that the number of vowels in modern Tibetan syllables is flexible, with a minimum of one and a maximum of three. This diversity fully demonstrates the complexity and harmony of the Tibetan vowel system. Based on the above theoretical analysis and combined with the true features of modern Tibetan texts, there should be 35 vowel finals in the Tibetan language in this article.

[0040] Consonant finals. In the Tibetan vowel system, in addition to the situation where the vowel alone serves as the final when there are no postfixes and re-postfixes and other consonant letters in a syllable, there are also finals composed of vowels and consonant finals, which are called "consonant finals". According to the differences in syllable structures, the representation forms of consonant finals can be divided into two categories: one is the finals formed by the combination of vowels and single-consonant finals, and the other is the finals formed by the combination of vowels and complex-consonant finals. Among them, single-consonant finals (abbreviated as "single-tail finals"), that is, only adding one "postfix" as the consonant final after the vowel. According to the record of the character nature of the postfix: " ", its meaning is that all postfixes can be attached after all "basic characters". However, from the writing form of Tibetan syllables, this statement does not fully reveal the specific attachment position of the postfix, nor can it completely depict the spelling process of Tibetan syllables. If understood according to the above description, it seems that the postfix is directly added after the basic character, and the vowel is located after the postfix. But in fact, in the spelling process of Tibetan syllables, the postfix is placed after the vowel final as the final, so it is called "consonant final". Thus, the syllable structure formed by single-consonant finals should be: "consonant initial + vowel + consonant final".

[0041] Based on the number of vowels in a syllable and their combination with consonant codas, finals formed by single consonant codas can be further divided into two categories: first, finals formed by the combination of monophthongs and single consonant codas. Theoretically, based on the five monophthong finals, combined with ten suffixes, there should be 50 single-coda finals. However, according to traditional Tibetan grammar records: ", means that in addition to syllables consisting of a prefix and a base word, a suffix needs to be added after them. "In addition, if the syllable structure contains superscript or subscript syllables, or vowel symbols (excluding implicit vowels)," If any one or more of the constituent elements are included, then the postfix "" will not be added. "With vowels" "and" For example, the finals formed by combining it with 10 additional characters are: The remaining monophthongs combined with the 10 consonant finals can be deduced similarly. Based on the above analysis, it can be seen that, from the perspective of the rules governing the formation of Tibetan vowels, since modern Tibetan generally does not have "…", "These are the four writing forms. Therefore, to be precise, 5 monophthongs combined with 10 consonant codas form 46 vowels with single consonant codas. Secondly, there are vowels in the form of a combination of diphthongs and single consonant codas. As before, open syllables are followed by word-forming elements indicating "tiny"..." "It can form a compound vowel, which can be followed by a consonant letter." "and" "Then it forms another syllable structure. From the perspective of spelling structure, " "and" "Here, consonant endings are expressed. From a grammatical perspective, they do not function as suffixes in syllables, but rather as grammatically meaningful function words (contract words), namely, ergative particles (..." ) "and positional particles ( ) The agglutinative form of "". For example: in the sentence " (The calf lets the lamb graze on the green grass) In the phrase "(the minister who loves birds)," the consonant ending " "and" "" respectively represent the doer (subject) and the object (object) of the action. Therefore, the combination of diphthongs and single consonant finals forms 10 finals with single consonant finals, which are: ".

[0042] A consonant-final consonant, or simply "compound consonant," refers to the phenomenon of adding two consonant letters after a vowel. In syllable structure, these two consonant letters typically represent a suffix and a further suffix, with the latter expressing a consonant-final consonant. Tibetan consonant-final consonants can be divided into two main types: one where the two consonant letters following a vowel function as a suffix and a further suffix, respectively. The rules for adding suffixes are as described above (except for "..."). While there are no strict rules for adding words after a given word, there are strict rules for adding words after a given word; they cannot be added arbitrarily after any given word. According to traditional grammar: , which means to add the word "" after it. "Add on" "After three additional characters, and then another additional character." Then add it to " "After four additional characters. Therefore, add another character." "respectively with " "Combined to form 3 consonant endings" (Its usage has weakened and is on the verge of extinction), and then add the word " "respectively with " "Combined to form 4 compound rhyme endings" Although the addition of a word after the vowel can be ignored when analyzing the modern Tibetan vowel system. The weakening of the compound final rhyme formed by "" still plays an irreplaceable role in grammatical analysis (such as distinguishing the "three-times-one-style" of verbs and the rules for adding function words). Therefore, to comprehensively cover linguistic phenomena, this chapter will temporarily add the word "" after it. The compound consonant endings formed by these 7 consonant endings are included in the statistical analysis. Based on the above analysis, each of the 7 consonant endings can combine with 5 monophthongs to form 35 different vowels (because "... The weakening of compound finals (in practice, there are usually 20 compound finals). Another type has been mentioned in the analysis of compound vowel formation, such as the decorative words " "and separable words" "Syllable structures that combine with the preceding syllable can form compound rhymes, which can be further divided into two types: one is the function word " "and" Each syllable consists of two consonant letters, placed in the positions of the following and subsequent syllables respectively in spelling. However, these two consonant letters do not represent the following and subsequent syllables of the syllable. For example, the vowel structures formed by combining it with the five monophthongs are: Therefore, there are 10 compound finals of this type. Secondly, there are open syllables followed by word-forming elements indicating "tiny" or "small". "Syllables that can form compound vowels, and these compound vowel syllables then combine with function words." "and" "Combined into a final consisting of a diphthong and a consonant cluster. For example: " There should be a total of 10 vowels with this type of consonant cluster ending, namely "etc." ".

[0043] It is worth noting that the second type of single consonant finals (partial syllables formed by the combination of a diphthong and a consonant final) and the second type of compound consonant finals (" "and" "with additional word-forming elements" Structures such as "" followed by a syllable (forming a partial syllable by merging the preceding syllable) are relatively rare in actual texts. For example, adding a word after "" Similar to compound consonant endings formed by combining three single consonant endings, although described in traditional grammar, their use has gradually weakened and is on the verge of disappearing with the development of the language. However, these structures not only fully conform to the spelling rules described in Tibetan grammar, but also truly reflect the appearance of modern Tibetan texts and the evolution of the language.

[0044] The aforementioned consonant codas (especially consonant compound codas) are omitted in spoken Tibetan, but to maintain the integrity and standardization of syllable structure, they are still strictly written according to orthography in written records. Therefore, this study focuses on the practical needs of text-to-speech conversion and conducts a relatively systematic analysis of the compositional structure of the vowel system in modern written Tibetan. Based on the above analysis, this paper concludes that the vowel system of modern written Tibetan consists of 35 vowel codas and 111 consonant codas, totaling 146 vowel units, the specific distribution of which is as follows: Figure 4 As shown, Figure 4 A distribution diagram of the Tibetan written vowel system for the Tibetan written language phonetic conversion method based on phonological structure features provided in this application embodiment.

[0045] The initials and finals of Sanskrit Tibetan script: The composition of modern Tibetan texts is not limited to standard modern Tibetan syllables; it also often incorporates elements of Sanskrit Tibetan script, some of which are used very frequently, primarily for spelling foreign words. Therefore, from the perspective of Tibetan information processing technology development, written texts jointly represented by Tibetan syllables and Sanskrit Tibetan script become a key research object in Tibetan text analysis tasks and an indispensable research focus in the field of Tibetan natural language processing. Ignoring this aspect will prevent text analysis results from fully and accurately reflecting the characteristics of modern Tibetan written texts.

[0046] The spelling structure of Tibetan Sanskrit does not include concepts such as prefixes, superscripts, subscripts, and suffixes. However, just as all human languages ​​exhibit certain commonalities in their phonetic structure, Tibetan and Sanskrit are closely linked in their relationship and definition of vowels and consonants. Specifically, the grammar of both languages ​​defines the relationship between consonants and vowels as: This means that vowels can express clear pronunciation through their unique independent articulation function, while consonants, if separated from their vowel accompaniment, cannot produce clear sounds independently. Furthermore, although Sanskrit Tibetan script currently lacks a clearly defined spelling rule, Tibetan texts not only use Tibetan characters but also often employ the spelling of Tibetan syllables. Further analysis of the representation of Sanskrit Tibetan script and the constituent elements of Sanskrit itself reveals that Sanskrit is composed of 34 basic consonant letters and 13 basic vowels. In Sanskrit Tibetan script… In the syllable structure of Tibetan characters, apart from their unique 11 consonant letters and 8 vowels, the rest are all contained within the 34 consonant letters and vowel system of Tibetan script. This analysis fully demonstrates that the analytical method for Tibetan phonological structure is equally applicable to the phonological structure of Sanskrit Tibetan script. Therefore, in order to more comprehensively and deeply reveal the diversity of modern Tibetan written texts, this paper, based on the Tibetan phonological system consisting of 231 initials and 146 finals analyzed above, further introduces the 11 consonant initials of Sanskrit Tibetan script (… ) and 8 vowel finals ( This study aims to further enrich and improve the representation of the Tibetan phonetic system, in order to achieve a more comprehensive and in-depth understanding and processing of modern Tibetan written texts. Based on the above analysis, the resulting consonant initials and finals are cleverly combined, and under the strict constraints of Tibetan spelling rules, all Tibetan syllables and potential syllable organization structures can be derived.

[0047] Furthermore, the phonetic symbol sequence is obtained. Specifically, based on the initial consonant structure encoding in the initial consonant conversion encoding table and the preset mapping rules (as shown in Table 1), and the final vowel structure encoding in the final vowel table and the preset mapping rules, the phonetic structure feature vectors of each syllable are mapped and converted to obtain the corresponding phonetic symbol strings for each syllable. Simultaneously, a Tibetan initial consonant-final vowel-phonetic symbol mapping table is generated. This table records the correspondence between the initial consonant structure encoding and the final vowel structure encoding in the written structure and the phonetic symbols, serving as the retrieval basis and mapping index for generating the phoneme sequence. Based on the Tibetan initial consonant-final vowel-phonetic symbol mapping table, the phonetic symbols are combined in the order of initial consonant symbols and final vowel symbols to generate the corresponding syllable phoneme sequence, thus obtaining the phoneme sequence for each syllable. The phoneme sequences of each syllable are synthesized to obtain the phonetic symbol sequence. The phonetic symbol sequence is used to represent the syllable-by-syllable phoneme expression of written Tibetan, and is a combination of all initial consonant phonemes and final vowel phonemes in a syllable.

[0048] In this embodiment, the Tibetan phonetic transcription task differs from that of languages ​​such as Chinese and English. For example, Chinese can rely on its corresponding pinyin system (including initials, finals, and tones) to provide direct phonemic basis for phonetic transcription and effectively address the challenges of polyphonic characters; while English can directly use the words themselves, whether as phoneme components or character sequences, to effectively represent the text. However, Tibetan phoneme annotation faces three core challenges: Firstly, while Tibetan is indeed a phonetic script similar to English, and phonetic scripts share certain commonalities in their spelling structure—for example, the horizontal combination in Tibetan spelling is not significantly different from other phonetic scripts—Tibetan's unique vertical stacking form is distinctly different. Tibetan uses consonant letters as its foundation, with other letters stacked vertically or horizontally around it, forming a unique two-dimensional syllable structure. From an information processing perspective, this complex spelling structure undoubtedly categorizes Tibetan as a "complex script," making it impossible to directly represent phonemes or characters using a one-dimensional alphabet, as is the case with English or Chinese Pinyin. More importantly, the components of a Tibetan syllable do not possess the characteristics of computer-analyzable phonemes. Therefore, the primary challenge in Tibetan phoneme annotation is the complexity of structural dimensionality reduction—that is, the conversion from two-dimensional Tibetan syllables to a one-dimensional structure. This is one of the key considerations in developing the two national standards, "Information Technology Tibetan Coded Character Set Standard Extension A" and "Information Technology Tibetan Coded Character Set Standard Extension B" (hereinafter referred to as Tibetan character extension set).

[0049] Secondly, Tibetan, as a uniquely structured agglutinative language, possesses a complex syllable structure, and existing analytical methods struggle to accurately depict the special characteristics of this agglutinative structure. The claim that modern Tibetan syllables consist of only six consonants and one vowel, forming a mere two-dimensional spelling structure, is questionable. Without a systematic description of these linguistic phenomena, the unique morphological structure of modern Tibetan texts and its syllable spelling cannot be fully reflected, thus affecting the accurate description of Tibetan phonetic structure and ultimately hindering the rationality and accuracy of word-to-sound conversion.

[0050] Third, although Tibetan constructs all its syllables using 34 consonant letters and vowels, these consonant letters and vowels are not the smallest units of Tibetan pronunciation, but rather the smallest constituent elements of the Tibetan script expressed in symbolic form. Phonetically, as a phonemic alphabet, the basic unit of Tibetan pronunciation is the initial and final phonemes that make up the syllable. This means that the Tibetan syllable itself is not the smallest unit of pronunciation. Therefore, in the study of Tibetan speech synthesis, neglecting the phonemic features with initials and finals as the smallest units of pronunciation will have a certain impact on the accurate representation of segmental features and the accuracy of prosodic modeling.

[0051] To address the aforementioned challenges, while transliteration methods such as Latin characters can be used to attempt to solve the Tibetan phonetic conversion task, especially with the potential to establish a unified standard for Tibetan Latin transliteration, this approach does possess good data universality and portability, and can effectively represent Tibetan syllables in a one-dimensional form. However, it cannot accurately characterize and depict the phonemic pronunciation rules with initials and finals as the smallest unit of pronunciation; that is, it is difficult to analyze the unique segmental features of Tibetan, merely serving as a substitute for Tibetan characters using Latin characters.

[0052] Furthermore, a phoneme text set is obtained. The specific method is as follows: the phoneme sequence is subjected to consistency verification, which includes checking whether the initial consonant combination, final vowel combination, and syllable initial-final combination in the phoneme sequence conform to the syllable formation rules of Tibetan orthography; the phoneme sequence is subjected to homophonic merging processing, where if there are several syllables that correspond to generate completely identical phoneme sequences, they are represented as unique phoneme text units; the phoneme text units after consistency verification and homophonic merging processing are then arranged according to the original text syllable order to obtain the phoneme text set.

[0053] In this embodiment, as Figure 5 As shown, Figure 5 This document presents the Tibetan syllable structure writing rules and syllable conversion structure diagram for the Tibetan written language phonological conversion method based on phonological structure features provided in this application embodiment. This diagram not only profoundly reflects the Tibetan phonological composition and syllable writing rules based on seven syllable structures, but also further illustrates the practical application of these rules. In this diagram, besides the "base character" and "vowel" as essential basic elements constituting Tibetan syllables, the components in other positions vary depending on the character. Furthermore, it is worth noting that the elements used to represent initials (i.e., syllable components) in the diagram all use names widely used in academia, while the final components following the "vowel" are temporarily named "consonant," "diphthong," "second consonant," "third consonant," etc., according to the syllable writing order. These names are only for ease of understanding and do not have universal applicability; their application is limited to the content of this document and directly corresponds to the Tibetan written language phonological structure analyzed in Section 1. The numerical labels in the diagram clearly indicate the writing order of Tibetan syllables and the phoneme pronunciation rules in the phonological structure.

[0054] This method first uses the seven previously revealed Tibetan syllable glyph structures as a blueprint, and closely combines them with the phonetic change rules reflected in traditional Tibetan grammar to conduct an in-depth analysis of the phonetic structure of modern written Tibetan, systematically summarizing the specific constituent elements of the Tibetan phonological system. Secondly, given that Tibetan, like other languages, does not have a one-to-one correspondence between the initials and finals of each syllable and unique phonetic phenomena, this method uses the principle of "homophones" and "homophones (referring to the same phonetic symbol)" to perform phonetic conversion on the initials and finals that make up Tibetan syllables, as shown in Table 1. That is, the mapping relationship between the initials and finals of syllables in the text and the predefined phonetic symbols obtains the phonetic symbol sequence of the input text. The phonetic symbol sequence refers to a string that uniformly uses Latin letters to represent the pronunciation form of Tibetan text.

[0055]

[0056] Based on the above analysis of the phonetic structure features of written Tibetan, a Tibetan phonetic conversion model was finally realized using a permutation and combination generation algorithm via computer program. This model was then used to automatically transcribe experimental data for testing. This method addresses the problem of representing homophones using homophones, effectively alleviating the data sparsity problem common in low-resource and agglutinative languages, as well as the challenges posed by phoneme dispersion to speech synthesis model training. The model uses Latin letters to represent Tibetan phonemes, providing solid support for acoustic modeling in Tibetan speech synthesis research and ensuring the accuracy and reliability of subsequent model training results.

[0057] In summary, Tibetan text standardization and phonetic conversion form the cornerstone of Tibetan speech synthesis systems, playing a crucial role in promoting the development of Tibetan speech interaction systems. The structural features of written Tibetan speech revealed in this paper focus on solving the mapping problem from semantic syllables to specific phonemes in Tibetan speech synthesis, providing accurate textual representations for acoustic modeling in Tibetan speech synthesis tasks—serving as input data and basic modeling units for acoustic model training. This phonemic textual representation not only ensures that the acoustic model can effectively capture the key features of Tibetan speech but also greatly improves the coverage of Tibetan speech segment features, thereby enhancing the model's generalization ability and adaptability.

[0058] Furthermore, the conversion feasibility prediction label is obtained. The specific method is as follows: Based on phoneme text set analysis, the text processing parameters of the phoneme text set are obtained. From this analysis, the phoneme density coefficient and cross-structural association span of the phoneme text set are obtained. The text processing parameters include the number of phonemes, the number of syllables, and the number of consecutive occurrences of consonant clusters. The number of phonemes refers to the number of characters in the phoneme text set. A syllable is composed of multiple factors, representing a complete Tibetan character; therefore, the number of syllables represents the number of syllables composed of phonemes in the phoneme text set. A preset cross-structural association span threshold is obtained from the database and compared with the cross-structural association span. If the cross-structural association span is above the cross-structural association span threshold, the conversion feasibility prediction label is defined as the first abnormal label. If the cross-structural association span is less than the cross-structural association span threshold, the audio conversion demand index of the phoneme text set is obtained through analysis. The audio conversion demand index is used to characterize the computational resources required by the phoneme text set in the acoustic generation stage. The sound pressure amplitude values ​​of each sampling point in the audio frame at the current moment are obtained by the phoneme conversion processing unit, and the total energy of the audio frame at the current moment is analyzed. Based on the total energy of the audio frame at the current moment, mapping and matching are performed to obtain the upper limit of audio conversion performance. This upper limit is then compared with the audio conversion demand index of the phoneme text set to obtain the conversion feasibility prediction label. If the audio conversion demand index of the phoneme text set is greater than the upper limit of audio conversion performance, the conversion feasibility prediction label is the second abnormal label; otherwise, the conversion feasibility prediction label is predicted to be feasible.

[0059] In this embodiment, the phoneme density coefficient of the phoneme text set is obtained by dividing the number of phonemes in the phoneme text set by the number of syllables. The cross-structural association span is specifically the number of consecutive occurrences of consonant clusters within a preset Tibetan text template.

[0060] The specific method for obtaining the audio conversion requirement indicators of the phoneme text set is as follows: ; In the formula, XQ represents the audio conversion requirement index of the phoneme text set, α1 represents the phoneme density coefficient weight, D1 represents the phoneme density coefficient of the phoneme text set, α2 represents the cross-structure association span weight, and D2 represents the cross-structure association span of the phoneme text set.

[0061] The phoneme density coefficient weight and cross-structure association span weight can be obtained by querying historical data. The historical audio conversion demand index evaluation dataset is obtained, which includes each historical phoneme density coefficient and each historical cross-structure association span, as well as their corresponding weights. The historical phoneme density coefficient weight and historical cross-structure association span weight corresponding to the last moment in the historical audio conversion demand index evaluation dataset are selected and used as the phoneme density coefficient weight and cross-structure association span weight, respectively.

[0062] The sound pressure amplitude values ​​at each sampling point in the audio frame at the current moment are obtained from the phonetic conversion processing unit. The sound pressure amplitude value is the digitized amplitude quantization value of the instantaneous sound pressure at that sampling moment. The total energy of the audio frame at the current moment is then analyzed. The specific method for obtaining the total energy of the audio frame at the current moment is as follows: ; In the formula, Et represents the total energy of the audio frame at the current moment, and x t (n) represents the sound pressure amplitude value of the nth sampling point in the tth frame, where t represents the current time, n represents the sampling point number, and t represents the audio frame number.

[0063] The upper limit of audio conversion performance is obtained by mapping and matching based on the total energy of the audio frame at the current moment. The specific method is as follows: obtain the stored dataset of the total energy of the audio frames at the current moment in history, which includes the total energy of several audio frames at the current moment. Compare the total energy of the audio frames at the current moment with the total energy of the audio frames at the current moment in history. The total energy of the audio frame at the current moment that is closest to the total energy of the audio frame at the current moment in history is used as the total energy of the audio frame at the current moment in history. The upper limit of audio conversion performance corresponding to the total energy of the audio frame at the current moment in history is obtained as the upper limit of audio conversion performance.

[0064] By analyzing phoneme text sets to obtain text processing parameters and calculating phoneme density coefficients and cross-structural correlation spans, the feasibility of phoneme-to-word conversion can be accurately assessed, enabling a quantitative judgment of phoneme sequence complexity and audio conversion requirements. By comparing the cross-structural correlation span with a preset threshold and analyzing the total energy of the current audio frame and mapping it to obtain an upper limit for audio conversion performance, conversion feasibility prediction labels can be clearly defined, including first and second anomaly labels. This effectively avoids phoneme-to-word conversion errors caused by complex phoneme combinations or audio requirements exceeding processing capacity. This method can identify potential anomalies in advance, providing feasibility warnings before conversion, offering a scientific basis for subsequent processing steps, and improving the stability and accuracy of phoneme-to-word conversion.

[0065] Further, an initial adjustment is performed on the phonetic conversion processing unit. Specifically, based on conversion feasibility prediction label analysis, if the conversion feasibility prediction label is the first abnormal label, a difference analysis is conducted between the cross-structural association span and the cross-structural association span threshold to obtain the cross-structural association span difference value. Specifically, the cross-structural association span threshold is subtracted from the cross-structural association span to obtain the cross-structural association span difference value. This difference value is then divided by the cross-structural association span to obtain the cross-structural association span difference value. Based on this difference value, mapping and matching are performed to obtain an attention window length adjustment value. The attention window length of the phonetic conversion processing unit is then reduced by subtracting the current attention window length adjustment value to obtain the adjusted attention window length. If the conversion feasibility prediction label is the first abnormal label, then... If the label is identified as the second abnormal label, a difference analysis is performed between the audio conversion requirement index and the upper limit of audio conversion performance based on the phoneme text set. This yields the audio conversion performance difference value. Specifically, the audio conversion requirement index is subtracted from the upper limit of audio conversion performance to obtain the audio conversion requirement difference. The audio conversion requirement difference is then divided by the audio conversion requirement index to obtain the audio conversion performance difference value. Based on the audio conversion performance difference value, a mapping and matching process is performed to obtain the frame shift adjustment amount and the batch processing bucket width adjustment amount. This is then used to increase the frame shift and increase the batch processing bucket width of the phonetic conversion processing unit. Specifically, the frame shift adjustment amount is added to the current frame shift amount to obtain the increased frame shift amount, and the batch processing bucket width adjustment amount is added to the current batch processing bucket width to obtain the increased batch processing bucket width. This completes the initial adjustment of the phonetic conversion processing unit.

[0066] In this embodiment, the attention window length adjustment value is obtained by mapping and matching based on the cross-structure association span difference value. Specifically, the method is as follows: obtain each historical cross-structure association span difference value stored in history, match it with the cross-structure association span difference value, select the historical cross-structure association span difference value that is closest to the cross-structure association span difference value, and mark it as the historical reference cross-structure association span difference value. Thus, obtain the historical attention window length adjustment value corresponding to the historical reference cross-structure association span difference value, and use it as the attention window length adjustment value.

[0067] Based on the audio conversion performance difference degree value, mapping and matching are performed to obtain the frame shift adjustment amount and the batch processing bucket width adjustment amount. The specific method is as follows: Obtain each historical audio conversion performance difference degree value stored in history, compare it with the audio conversion performance difference degree value, and select the historical audio conversion performance difference degree value closest to the audio conversion performance difference degree value as the historical reference audio conversion performance difference degree value. Obtain the historical frame shift adjustment amount and the historical batch processing bucket width adjustment amount corresponding to the historical reference audio conversion performance difference degree value, and thus use the historical frame shift adjustment amount and the historical batch processing bucket width adjustment amount as the frame shift adjustment amount and the batch processing bucket width adjustment amount respectively.

[0068] By initially adjusting the phonetic conversion processing unit based on the conversion feasibility prediction label, adaptive parameter optimization for different abnormal situations can be achieved. For the first abnormal label, by analyzing the cross-structure association span difference degree and adjusting the attention window length, the problem of excessive dependence caused by long-span phoneme combinations can be reduced, the computational load can be reduced, and the processing efficiency can be improved; for the second abnormal label, by analyzing the audio conversion performance difference degree and adjusting the frame shift amount and the batch processing bucket width, the parallelism and throughput capacity of the processing unit can be increased, and the problems of delay and response jitter caused by high audio requirements can be alleviated.

[0069] Furthermore, the verification result is obtained. The specific method is as follows: Obtain the performance verification parameters within a preset time period. The performance verification parameters include the change rate of the hidden layer activation energy, the adjacent frame correlation degree, and the change rate of the hidden state gradient. Based on the analysis of the performance verification parameters, the determination results of the change rate of the hidden layer activation energy, the adjacent frame correlation degree determination result, and the hidden state gradient determination result are obtained respectively. If the determination results of the change rate of the hidden layer activation energy, the adjacent frame correlation degree determination result, and the hidden state gradient determination result are all qualified, the verification result is qualified; otherwise, the verification result is unqualified.

[0070] In this embodiment, the performance verification parameters include the change rate of the hidden layer activation energy, the adjacent frame correlation degree, and the change rate of the hidden state gradient. Among them, the change rate of the hidden layer activation energy is used to characterize whether the internal activation intensity of the phonetic conversion processing unit converges to a stable state within a preset time period after the initial adjustment of the phonetic conversion processing unit. It can be obtained by calculating the change rate of the hidden layer activation energy at each moment within the preset time period, that is, subtracting the hidden layer activation energy of the previous moment from the hidden layer activation energy of the next moment to obtain the hidden layer activation energy difference, and then dividing the hidden layer activation energy difference by the hidden layer activation energy of the next moment to obtain the change rate of the hidden layer activation energy. If the change rate of the hidden layer activation energy gradually decreases, the determination result of the change rate of the hidden layer activation energy is qualified; otherwise, it is unqualified.

[0071] The correlation between adjacent frames indicates whether the continuity of the acoustic state in the time dimension is being restored or tending to smooth out after the initial adjustment. This can be achieved by calculating the fluctuation amplitude between two adjacent moments within a preset time window, and by using cosine similarity to calculate the hidden state vectors of two adjacent frames. If the correlation between adjacent frames gradually increases over time within the preset time window, the correlation determination result is considered satisfactory; otherwise, the correlation determination result is considered unsatisfactory.

[0072] The hidden state gradient change rate indicates whether the number of jumps in the time dimension of the phonology conversion processing unit is decreasing after the initial adjustment, that is, whether the internal state is converging from "disorder" to "stability". If the gradient gradually decreases, it indicates stable convergence, and the hidden state gradient determination result is qualified; otherwise, it is unqualified. The specific calculation method of the hidden state gradient change rate is as follows: ; ; In the formula, Gt represents the internal change amplitude at time t, t represents the time number, t=1,2,...,T,T represents the total time, ht represents the hidden state vector of the phonetic conversion processing unit at time t, ht-1 represents the hidden state vector of the phonetic conversion processing unit at time t-1, Gt+k represents the internal change amplitude after time t+k, Gt+k-Gt represents the total amplitude of the internal change amplitude within the k-frame time span, and ΔGt represents the relative increase or decrease of the hidden state change amplitude within the preset k-frame time window.

[0073] A second adjustment is performed to obtain the Tibetan written audio. The specific method is as follows: the conversion feasibility prediction label is traced back to the initial adjustment of the phonetic conversion processing unit. If the conversion feasibility prediction label is the first abnormal label, the sparse trigger interval reduction processing is performed. The sparse trigger interval reduction processing is performed by subtracting the sparse trigger interval adjustment amount from the current sparse trigger interval to obtain the reduced sparse trigger interval. If the conversion feasibility prediction label is the second abnormal label, the attention head number reduction processing is performed. The attention head number reduction processing is performed by subtracting the attention head number adjustment amount from the current attention head number to obtain the reduced attention head number.

[0074] The hidden layer activation energy change rate determination result, adjacent frame correlation determination result, and hidden state gradient determination result are obtained respectively. If any determination result is unqualified, the corresponding numerical result is obtained as the numerical influence value of that determination result. If any determination result is qualified, the numerical influence value of the corresponding numerical result is defined as zero. The numerical results include the hidden layer activation energy change rate numerical result, adjacent frame correlation numerical result, and hidden state gradient change rate numerical result. Specifically, the hidden layer activation energy change rate numerical result is the hidden layer activation energy change rate at the first and last moments within the preset time period of the initial adjustment; the adjacent frame correlation numerical result is the adjacent frame correlation at the first and last moments within the preset time period of the initial adjustment; and the hidden state gradient change rate numerical result is the hidden state gradient change rate at the first and last moments within the preset time period of the initial adjustment.

[0075] A preset verification reference set is obtained, which includes reference values ​​for the rate of change of activation energy in the hidden layer, the correlation between adjacent frames, and the rate of change of gradient in the hidden state. The verification reference set is compared with the corresponding numerical results, and combined with the corresponding influence coefficient analysis, to obtain the initial adjustment deviation value.

[0076] The method for obtaining the initial adjustment deviation value is as follows: ; In the formula, PL represents the initial adjustment deviation value, ZY represents the numerical result of the hidden layer activation energy change rate, τZY represents the reference value of the hidden layer activation energy change rate, β1 represents the influence coefficient of the numerical result of the hidden layer activation energy change rate, ZL represents the numerical result of the correlation between adjacent frames, τZL represents the reference value of the correlation between adjacent frames, β2 represents the influence coefficient of the numerical result of the correlation between adjacent frames, ZT represents the numerical result of the hidden layer activation energy change rate, τZT represents the reference value of the hidden state gradient change rate, and β3 represents the influence coefficient of the numerical result of the hidden layer activation energy change rate.

[0077] It should be noted that the influence coefficients of the hidden layer activation energy change rate, the adjacent frame correlation, and the hidden layer activation energy change rate can be obtained through historical mapping. For example, the influence coefficient of the hidden layer activation energy change rate can be obtained by acquiring the stored historical hidden layer activation energy change rate results, comparing them with the current hidden layer activation energy change rate results, selecting the historical hidden layer activation energy change rate result that is closest to the current result as the historical reference hidden layer activation energy change rate result, and obtaining the historical hidden layer activation energy change rate result influence coefficient corresponding to the historical reference result as the hidden layer activation energy change rate influence coefficient. The methods for obtaining the adjacent frame correlation and the hidden layer activation energy change rate influence coefficients are the same as those for the hidden layer activation energy change rate influence coefficients; both can be obtained by mapping and matching the corresponding numerical results with historical data.

[0078] Mapping and matching are performed based on the initial adjustment deviation value to obtain the sparse trigger interval adjustment and the attention head number adjustment. The stored historical initial adjustment deviation values ​​are obtained and compared with the current initial adjustment deviation value. The historical initial adjustment deviation value that is closest to the current initial adjustment deviation value is selected as the historical control initial adjustment deviation value. The historical sparse trigger interval adjustment and historical attention head number adjustment corresponding to the historical control initial adjustment deviation value are obtained and used as the sparse trigger interval adjustment and attention head number adjustment.

[0079] By acquiring performance verification parameters within a preset time period, including the rate of change of hidden layer activation energy, the correlation between adjacent frames, and the rate of change of hidden state gradient, and making judgments based on these parameters, the operational status of the phonetic conversion processing unit can be effectively verified. Quantitative analysis of key performance indicators can determine whether the current conversion process meets the acceptable standards, ensuring that hidden layer activation, inter-frame continuity, and state gradient changes remain stable. This method achieves real-time performance monitoring and verification. When the verification result is unacceptable, potential anomalies can be detected in a timely manner, ensuring the reliability and continuity of the phonetic conversion process, thereby avoiding phonetic conversion errors or delays caused by processing unit malfunctions, and improving the stability and controllability of the system.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for converting the pronunciation of Tibetan written language characters based on phonetic structure features, characterized in that, Includes the following steps: S1. Obtain the Tibetan written text set to be processed and perform standardization processing to obtain the Tibetan text; S2. The Tibetan text is segmented into syllables, and the glyph structure elements of each syllable are extracted based on the glyph structure of Tibetan characters to obtain a set of syllable glyph structure elements. S3. Based on the syllable glyph structure element set, the initial consonant and final vowel items of each syllable are matched to obtain the speech structure feature vector of each syllable; S4. Obtain the initial consonant conversion encoding table and the final vowel conversion encoding table, thereby mapping the speech structure feature vector of each syllable to the corresponding syllable phoneme sequence, and synthesizing all syllable phoneme sequences to obtain the speech symbol sequence; S5. Based on the phonetic symbol sequence, perform consistency verification and homophone merging to obtain the phoneme text set, thereby obtaining the Tibetan written audio.

2. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The Tibetan language's written structure includes initials and finals. The initials include prefixes, superscripts, bases, subscripts, and sub-subscripts. The finals include vowels, suffixes, sub-suffixes, consonants, diphthongs, second consonants, third consonants, and third vowels.

3. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The specific method for obtaining the syllable glyph structure element set is as follows: Tibetan text is segmented into syllables. Syllables are identified based on the syllable boundary rules of Tibetan orthography. Specifically, the identification of syllables includes identifying whether each syllable contains Tibetan glyph structural elements such as prefixes, superscripts, subscripts, vowels, suffixes, and further suffixes, based on the vertical and horizontal writing order of the base character. Extract the corresponding glyph structural elements from each syllable to obtain the glyph structural elements of each syllable; Arrange the glyph structural elements of each syllable according to the glyph hierarchy of the syllables to obtain the set of glyph structural elements of the syllables.

4. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The specific method for obtaining the speech structure feature vector of each syllable is as follows: The initial consonant of a syllable is located based on the set of syllable glyph structure elements; Locating the vowel item of a syllable based on the set of syllable glyph structure elements; Arrange the initial consonant and final vowel items according to the syllable structure combination order to form the initial consonant and final vowel structure of the syllable; The internal codes of the initial consonant and final vowel items are combined and calculated to obtain the speech structure feature vector of each syllable. The speech structure feature vector of each syllable is used to characterize the speech structure feature vector of the initial consonant combination relationship of the syllable.

5. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The specific methods for obtaining the initial consonant conversion encoding table and the final vowel conversion encoding table are as follows: Obtain the set of consonant initials and the set of finals in written Tibetan, wherein the set of consonant initials includes simple consonant initials and compound consonant initials; The consonant clusters include diphthongs, triphthongs, and tetraphthongs; The set of finals includes vowel finals and consonant finals; The vowel finals include monophthongs, diphthongs, and triphthongs; The consonant codas include simple consonant codas and compound consonant codas; The consonant ending includes a structure formed by combining a suffix and a further suffix; The set of consonant initials and the set of finals are internally encoded, and internal code numbers are generated based on the positional relationship of the initials and finals in the syllable structure. Based on the internal encoding number and the preset mapping rules, character mapping processing is performed to obtain the initial consonant conversion encoding table and the final vowel conversion encoding table.

6. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The specific method for obtaining the speech symbol sequence is as follows: Based on the initial consonant structure encoding in the initial consonant conversion encoding table and the preset mapping rules, as well as the final vowel structure encoding in the final vowel table and the preset mapping rules, the speech structure feature vectors of each syllable are mapped and converted to obtain the corresponding speech symbol strings of each syllable, and a Tibetan initial consonant and final vowel mapping table and speech symbol mapping table are generated simultaneously. The Tibetan initials and finals mapping table is used to record the correspondence between the initial structure code and the final structure code in the written structure and the phonetic symbol, and is used as the retrieval basis and mapping index for generating phoneme sequences; Based on the mapping table of Tibetan initials and finals to phonetic symbols, the phonetic symbols are combined in the order of initial symbols and final symbols to generate the corresponding syllable phoneme sequence, thereby obtaining the phoneme sequence of each syllable; The phoneme sequences of each syllable are synthesized to obtain a sequence of phonetic symbols; The phonetic symbol sequence is used to represent the syllable-by-syllable phoneme expression of Tibetan written language, and is a combination of all initial consonant phonemes and final consonant phonemes in a syllable.

7. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 1, characterized in that: The method for obtaining the phoneme text set and thus the audio of Tibetan written language is as follows: A consistency check is performed on the phonetic symbol sequence, which includes checking whether the initial consonant combination, final vowel combination and syllable initial-final combination in the phonetic symbol sequence conform to the syllable formation rules of Tibetan orthography; Homophone merging is performed on the sequence of phonetic symbols. If there are several syllables that correspond to a completely identical phoneme sequence, they are represented as a unique phoneme text unit. Therefore, the phoneme text units after consistency verification and homophonic merging are arranged according to the original text syllable order to obtain the phoneme text set; Based on phoneme text set analysis, the phoneme density coefficient and cross-structural association span of the phoneme text set are obtained, and conversion feasibility prediction tags are obtained from this analysis. If the conversion feasibility prediction label indicates that the conversion is feasible, the Tibetan written language phonetic conversion will continue to be performed to obtain the Tibetan written language audio. Otherwise, the phonetic conversion processing unit will be initially adjusted, and a prediction verification will be performed after the adjustment to obtain the verification result. If the verification result is qualified, the Tibetan written language audio will be obtained; otherwise, a second adjustment will be performed to obtain the Tibetan written language audio.

8. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 7, characterized in that: The specific method for obtaining the conversion feasibility prediction label is as follows: Based on the analysis of the phoneme text set, the text processing parameters of the phoneme text set are obtained. From this analysis, the phoneme density coefficient and cross-structural association span of the phoneme text set are obtained. The text processing parameters include the number of phonemes, the number of syllables, and the number of consecutive occurrences of consonant clusters. Obtain the preset cross-structure association span threshold in the database and compare it with the cross-structure association span. If the cross-structure association span is above the cross-structure association span threshold, define the conversion feasibility prediction label as the first abnormal label. If the cross-structure association span is less than the cross-structure association span threshold, analyze and obtain the audio conversion demand index of the phoneme text set. The audio conversion demand index is used to characterize the computing resources required by the phoneme text set in the acoustic generation stage. Obtain the sound pressure amplitude values ​​of each sampling point in the audio frame at the current moment from the phonetic conversion processing unit, and analyze to obtain the total energy of the audio frame at the current moment; Based on the total energy of the audio frame at the current moment, a mapping and matching is performed to obtain the upper limit of audio conversion performance. This is then compared with the audio conversion demand index of the phoneme text set to obtain a conversion feasibility prediction label. If the audio conversion demand index of the phoneme text set is greater than the upper limit of audio conversion performance, the conversion feasibility prediction label is the second abnormal label; otherwise, the conversion feasibility prediction label is predicted to be feasible.

9. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 8, characterized in that: The initial adjustment of the phonetic conversion processing unit is specifically performed as follows: Based on the conversion feasibility prediction label analysis, if the conversion feasibility prediction label is the first abnormal label, then the degree of difference analysis is performed based on the cross-structure association span and the cross-structure association span threshold to obtain the degree of difference value of cross-structure association span, and the attention window length adjustment value is obtained by mapping and matching based on the degree of difference value of cross-structure association span, thereby reducing and adjusting the attention window length of the phonetic conversion processing unit. If the conversion feasibility prediction label is the second abnormal label, then the difference between the audio conversion demand index and the upper limit of audio conversion performance based on the phoneme text set is analyzed to obtain the audio conversion performance difference value. Based on the audio conversion performance difference value, the frame shift adjustment amount and the batch processing bucket width adjustment amount are obtained through mapping and matching. Then, the frame shift of the phonetic conversion processing unit is increased and the batch processing bucket width is increased to complete the initial adjustment of the phonetic conversion processing unit.

10. The Tibetan written language phonetic conversion method based on phonological structure features as described in claim 7, characterized in that: The specific method for obtaining the verification result is as follows: Obtain performance verification parameters within a preset time period, including the hidden layer activation energy change rate, adjacent frame correlation, and hidden state gradient change rate. Based on the performance verification parameter analysis, the results of the hidden layer activation energy change rate determination, the adjacent frame correlation determination, and the hidden state gradient determination were obtained respectively. If the hidden layer activation energy change rate determination result, the adjacent frame correlation determination result, and the hidden state gradient determination result are all qualified, then the verification result is qualified; otherwise, the verification result is unqualified.

Images