A text pronunciation optimization method and system for speech synthesis

Abstract

This invention discloses a text pronunciation optimization method and system for speech synthesis, belonging to the field of speech synthesis management technology. In this method, after emotional processing, long sentences are split into segments based on a triple rule system of semantic blocks, classical Chinese function words, and character length. The English portion of the text undergoes layered processing, distinguishing between pure uppercase letter combinations and regular English words, adding splitting and prosodic markers to pure uppercase letter combinations, and generating optimized text pronunciation information. This invention breaks through the limitations of existing single-rule adaptation in speech synthesis text processing, pioneering a multi-dimensional layered optimization framework that integrates technologies such as semantic parsing of numbers and operators, scene-based matching of polyphonic characters, emotional markers for literary texts, and semantic long sentence splitting. It designs multiple exclusive optimization rules for the speech synthesis and reading needs of literary and everyday spoken texts, solving the core pain point of existing technologies that emphasize generality but neglect specific scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of speech synthesis management, and specifically to a text pronunciation optimization method and system for speech synthesis. Background Art

[0002] Text-to-Speech (TTS) technology is widely used in fields such as audiobooks, intelligent voice assistants, and educational audio production. Its core process is "text preprocessing → pinyin conversion → prosody generation → speech synthesis". Among them, text preprocessing is the key link that determines the TTS reading quality, directly affecting pronunciation accuracy, semantic integrity, and emotional naturalness.

[0003] With the popularization of TTS technology, users' requirements for reading quality have been upgraded from "being able to read" to "reading accurately, with emotion, and in line with the scenario". Especially for the reading requirements of literary texts (including polyphonic characters, reduplicated words, punctuation emotions, long sentence semantics), educational texts (including numbers, arithmetic expressions), and daily spoken texts (including modal particles, weak suffixes), higher requirements are put forward for the refinement degree of text preprocessing.

[0004] Existing TTS text preprocessing technologies mainly focus on basic processing in general scenarios (such as converting simple numbers to Chinese numbers, punctuation pauses), and do not design exclusive rules for the characteristics of literary and educational texts, resulting in poor reading effects and easy semantic confusion for special texts; specifically manifested as: in digital processing, only integers are converted to Chinese numbers, and decimals are directly read as characters (such as "1.3" read as "one point three" but the decimal places are easily lost, or "10%" read as "ten %" instead of "ten percent"); there is no discrimination rule for operator processing, both the negative sign and the minus sign are uniformly replaced by "minus", and the division sign is directly read as "divide"; in English processing, either all English words are split into single letters, or they are completely left unprocessed, and there is no prosody marking for pure capital letter combinations (such as "AI"); in polyphonic character processing, a simple mapping of "single character - pinyin" is adopted (such as "hao" is only mapped to "hǎo"), without distinguishing the pronunciation according to the context, and the pinyin tones use original characters (such as "ā"), which cannot be recognized by some TTS engines (such as Piper); in punctuation processing, only fixed-duration pause marks are added according to the punctuation type (such as a 0.5s pause for a comma and a 1s pause for a period), without emotional intonation marks; in long sentence splitting, only split according to a fixed character length (such as 50 characters), without combining semantic, function word and other features, and it is easy to split in the middle of the semantics (such as "I like reading but don't like writing" split into "I like reading but / don't like writing"); in the processing of reduplicated words and modal particles, there are no exclusive rules, and they are only read as ordinary words (such as "manman" read as "slow - slow", without light reading or lengthening marks; the intonation of "ma", "ne", "ba" is the same). Therefore, there are significant defects in existing speech synthesis technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a text pronunciation optimization method and system for speech synthesis to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a text pronunciation optimization method for speech synthesis, comprising: S1. Perform operator semantic discrimination processing on the input text, and convert the minus sign and the minus sign into their corresponding Chinese semantics according to the context; and perform hierarchical conversion processing on the numbers in the input text, distinguishing between integers, decimals and numbers with units, and converting them into Chinese pronunciations according to the integer hierarchical conversion rules, the decimal digit-by-digit conversion rules and the unit adaptation rules respectively. S2. For the text processed by S1, perform contextualized polyphonic character matching, select the correct pronunciation for polyphonic characters based on the context; and convert the original characters with tones into a format compatible with the target speech synthesis engine and containing pinyin, numerical tones and prosodic markers; S3. Perform emotional processing on the text processed by S2, add markers containing pause duration and emotional tone to punctuation marks; add stressed, unstressed, or intonation markers to preset literary keywords; apply exclusive emotional reading rules based on different structural types to reduplicated words; add differentiated intonation markers to different tone words; add unstressed reading markers to unstressed suffixes. S4. For the emotionally processed text, perform long sentence splitting based on three rules: semantic blocks, classical Chinese function words, and character length; and perform layered processing on the English part of the text, distinguishing between pure uppercase letter combinations and regular English words, adding splitting and prosodic markers to pure uppercase letter combinations; and generate optimized text pronunciation information.

[0007] The text optimization process of this invention follows the logic of "semantic parsing (operators / numbers) → pronunciation calibration (polyphonic characters / pinyin) → emotional enhancement (punctuation / reduplicated words / keywords) → format adaptation (English / long sentences)". The specific steps are as follows: input text → operator conversion → number to Chinese → polyphonic character / pinyin adaptation → punctuation emotional processing → literary keyword repetition → reduplicated word emotional optimization → modal particle differentiation → neutral tone suffix optimization → long sentence semantic splitting → English layered processing → output optimized text.

[0008] Furthermore, the operator semantic differentiation processing in S1 specifically includes: When the "-" sign is preceded or followed by numbers or decimal points, the corresponding Chinese semantic conversion result is determined to be subtraction. When the "-" sign is preceded by an equal sign or is located at the beginning of the text and followed by a number, the corresponding Chinese semantic conversion result will be determined to be negative; The semantic differentiation process of the operator also includes: recording the Chinese semantic conversion result corresponding to the "÷" sign as "divided by".

[0009] In the present invention, the semantic differentiation process of the operator is carried out because in the prior art, for the "-" sign, it is often converted to "minus", and it is impossible to effectively distinguish according to the context scenario, resulting in a relatively high probability of abnormal conversion results; for the "÷" sign, in the prior art, it is often translated as "divide", while in the mathematical sense, the actual meaning corresponding to this sign is "divided by". Although there is only a one-character difference, the meanings represented are completely opposite, and this conversion result is likely to cause serious misguidance to users.

[0010] Further, the integer grading conversion rule in S1 is used to segment and convert integers according to the preset levels of mega - billion - ten - thousand, and eliminate the redundancy of consecutive zeros; the decimal digit - by - digit conversion rule is to independently convert each digit of the decimal part; the unit adaptation rule at least includes converting "℃" to "degrees Celsius" and converting "㎡" to "square meters".

[0011] Further, the scenario - based polyphonic character matching in S2 includes: screening out the pronunciations corresponding to each polyphonic character in the corresponding text, as well as the pronunciations corresponding to each polyphonic character when in the preset format combinations or vocabulary combinations in the database. Each preset format combination or each vocabulary combination in the database is a context scenario for the corresponding polyphonic character; When converting the original characters with tones into a format compatible with the target speech synthesis engine and including pinyin, digital tones, and prosody marks, each original character with a tone corresponds to a preset format combination result including pinyin, digital tones, and prosody marks, and the format combination results corresponding to the same original character with different tones are different.

[0012] The scenario - based polyphonic character matching in the present invention at least includes the differentiation of the character "了": when "了" is at the end of a sentence and connected with the modal particles "吗,呢,吧,啊", it is converted to "le"; when "了" appears in words such as "了结,了解", it is converted to "liǎo"; the adaptation rule for the original characters with tones includes: replacing "ü" with "v", and adding a stress mark "°" and a lengthening mark "~" to the converted pinyin.

[0013] Further, the specific implementation steps for the emotional processing of the text processed by S2 in S3 include: When performing emotional processing on punctuation marks, at least for commas, full stops, exclamation marks, question marks, dashes, book titles, and quotation marks, different pause durations and intonation trajectories composed of rising and falling tones are set respectively; Exclusive emotional reading rules for reduplicated words, including: for AA - type reduplicated words, use light reading and lengthening; for ABB - type reduplicated words, use heavy reading for the first character and light reading for the last character; for ABAB - type reduplicated words, use rhythm splitting. When adding intonation marks for modal particles, different modal particles respectively correspond to different preset intonation mark results.

[0014] In this invention, "ma" at the end of a sentence is marked with a rising tone, "ne" is marked with a stronger rising tone, "ba" is marked with a falling tone, and "ai" is marked with a deep falling tone.

[0015] Furthermore, the optimization of the weak - sound suffix includes adding light - reading marks to the suffixes of preset words (such as: "tou", "zi", "men", "me").

[0016] Furthermore, the S4 performs long - sentence splitting based on triple rules of semantic chunks, classical Chinese function words, and character length, specifically including: The first rule: According to preset semantic - turning or progressive keywords, split the text after emotional processing into different semantic chunks; The second rule: Further split the results of the semantic chunks split by the first rule according to preset classical Chinese function words to obtain new different semantic chunks; The third rule: Determine the length of the semantic chunks in the results of the semantic chunks split by the second rule. If the length of a semantic chunk is greater than the preset value, then according to different preset punctuation marks in the corresponding semantic chunk, split the corresponding semantic chunk to reduce the length of the semantic chunk; in this invention, when encountering a comma, semicolon, or colon, force - split the text content in the semantic chunk. If the length of the semantic chunk is less than or equal to the preset value, do not split the corresponding semantic chunk according to different preset punctuation marks in the corresponding semantic chunk.

[0017] Furthermore, the specific implementation steps for the S4 to perform hierarchical processing on the English part in the text include: for pure uppercase - letter combinations composed of two or more uppercase letters, split the corresponding pure uppercase - letter combination into multiple single letters, and add prosody marks with heavy - reading and lengthening states to each letter; for regular English words with the first letter capitalized or all lowercase, keep the integrity of the corresponding English words.

[0018] A text pronunciation optimization system for speech synthesis, including: An operator conversion module, used to perform semantic - distinction processing on minus signs, subtraction signs, and division signs in the input text; A number - to - Chinese module, used to perform hierarchical conversion processing on integers, decimals, and numbers with units; A polyphonic - character and pinyin adaptation module, used to perform scenario - based polyphonic - character matching and convert original characters with tones into a format compatible with the speech synthesis engine and including pinyin, digital tones, and prosody marks; The punctuation emotion processing module is used to add pause and emotional tone marks to various categories of punctuation marks; The literary keyword emphasis module is used to add emphasis, de-emphasis, or intonation markers to preset literary keywords; The reduplicated word emotion optimization module is used to apply exclusive emotion-based reading rules based on the structural type of reduplicated words; The modal particle differentiation module is used to add differentiated intonation markers for different modal particles; The neutral tone suffix optimization module is used to add a neutral tone mark to neutral tone suffixes; The long sentence semantic segmentation module is used to segment long sentences based on a three-tiered rule system of semantic blocks, classical Chinese function words, and character length. The English layered processing module is used to distinguish and differentiate between pure uppercase letter combinations and regular English words; The synthesized speech output module is used to generate optimized text pronunciation information corresponding to the input text based on the processing results of the operator conversion module, number to Chinese conversion module, polyphonic character and pinyin adaptation module, punctuation emotion processing module, literary keyword repetition module, reduplicated word emotion optimization module, modal particle differentiation module, neutral tone suffix optimization module, long sentence semantic segmentation module and English layered processing module.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention breaks through the limitation of the existing speech synthesis text processing of "single rule adaptation" and creates the first "multi-dimensional hierarchical optimization framework". It integrates technologies such as semantic parsing of numbers and operators, scene matching of polyphonic characters, emotional marking of literary texts, and semantic long sentence splitting. It designs multiple exclusive optimization rules for speech synthesis reading needs of literary and everyday spoken texts, and solves the core pain point of existing technologies that emphasize generality and neglect specific scenarios. (2) This invention proposes for the first time an operator semantic differentiation algorithm (scenario-based matching rules for minus and division signs) to solve the problem of logical errors in the reading of educational texts; and constructs adaptation conversion rules for pinyin tones and speech synthesis engines (such as ü→v, characters with tones converted to "pinyin + number + prosodic marker") and exclusive emotional reading rules for reduplicated words (AA / ABB / ABAB), so that the reading of literary texts is emotional and layered, and everyday spoken language is more natural; at the same time, by constructing a three-fold long sentence splitting rule of semantic blocks + function words + length, the semantic splitting results are avoided from being fragmented, and the comprehensibility of long sentence reading is improved. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a text pronunciation optimization method for speech synthesis according to the present invention. Figure 2 This is a schematic diagram of the scene-based matching rules for polyphonic characters in a text pronunciation optimization method for speech synthesis according to the present invention; Figure 3 This is a schematic diagram of the long sentence splitting process using the three-rule approach in a text pronunciation optimization method for speech synthesis according to the present invention. Detailed Implementation

[0021] 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.

[0022] Please see Figure 1 This embodiment provides a text pronunciation optimization method for speech synthesis, including: S1. Perform operator semantic discrimination processing on the input text, and convert the minus sign and the minus sign into their corresponding Chinese semantics according to the context; and perform hierarchical conversion processing on the numbers in the input text, distinguishing between integers, decimals and numbers with units, and converting them into Chinese pronunciations according to the integer hierarchical conversion rules, the decimal digit-by-digit conversion rules and the unit adaptation rules respectively. The operator semantic differentiation processing in S1 specifically includes: When the "-" sign is preceded or followed by numbers or decimal points, the corresponding Chinese semantic conversion result is determined to be subtraction. When the "-" sign is preceded by an equal sign or is located at the beginning of the text and followed by a number, the corresponding Chinese semantic conversion result will be determined to be negative; Specifically, match "number / decimal point +- + number" (e.g., "4.6-32.5") and replace it with "number minus number"; Matches "equal to +-+ number" (e.g., "equal to -37.1") or "- number" at the beginning (e.g., "-5"), and replaces them with "equal to negative number" or "negative number"; The operator semantic differentiation processing also includes: recording the Chinese semantic conversion result corresponding to the "÷" sign as division by.

[0023] The integer tiered conversion rule in S1 is used to segment and convert integers according to a preset range of trillions to hundreds of millions to tens of thousands, and to eliminate the redundancy of consecutive zeros; the decimal digit-by-digit conversion rule is to convert each digit of the decimal part independently; the unit adaptation rule includes at least converting "℃" to degrees Celsius and "㎡" to square meters.

[0024] In this embodiment, it supports converting integers within the range of 0 to 99999999999999 into Chinese, and performs hierarchical processing according to "trillion - billion - ten thousand" to avoid redundancy of consecutive zeros (e.g., "1001" is converted to "one thousand and one" instead of "one thousand zero zero one"); Regarding decimal conversion: In this embodiment, the integer part and the decimal part are split. The integer part is converted according to the integer rule, and the decimal part is converted digit by digit (e.g., "1.3" is converted to "one point three", "2.56" is converted to "two point five six"), and the complete structure of the decimal is retained; Regarding the unit adaptation rule: For units such as ℃ (converted to "degrees Celsius"), X% (converted to "X percent"), ㎡ (converted to "square meters"), etc., first convert the number and then splice the unit. For example, "25℃" is converted to "twenty - five degrees Celsius", and "50%" is converted to "fifty percent".

[0025] S2: For the text processed by S1, perform context - aware homophone matching to select the correct pronunciation for homophones based on the context; and convert the original characters with tones into a format compatible with the target speech synthesis engine and containing pinyin, digital tones, and prosody marks The context - aware homophone matching in S2 includes: screening out the pronunciations corresponding to each homophone in the corresponding text, as well as the pronunciations corresponding to each homophone when in the preset format combinations or vocabulary combinations in the database. Each preset format combination or each vocabulary combination in the database is a context scenario for the corresponding homophone; as Figure 2 shown, the context - aware homophone matching in this embodiment at least includes the discrimination of the character "le": when "le" is at the end of a sentence and connected with modal particles "ma, ne, ba, a", it is converted to "le"; when "le" appears in words such as "了结 (resolve), 了解 (understand)", it is converted to "liǎo"; The adaptation rule for the original characters with tones includes: replacing "ü" with "v", and adding a stress mark "°" and a lengthening mark "~" to the converted pinyin.

[0026] When converting the original characters with tones into a format compatible with the target speech synthesis engine and containing pinyin, digital tones, and prosody marks, each original character with a tone corresponds to a preset format combination result containing pinyin, digital tones, and prosody marks, and the format combination results corresponding to the same original character with different tones are different. For example: the format combination corresponding to "ā" is "°a1~", and the format combination corresponding to "ǖe" is "°ve1~" (ü is replaced by v), adding stress (°) and lengthening (~) marks to enhance the prosody.

[0027] S3. Perform sentiment processing on the text processed in S2, adding markers including pause duration and emotional intonation for punctuation marks; adding stress, unstressed, or intonation markers for preset literary keywords; applying exclusive sentiment-based reading rules for reduplicated words based on different structural types; adding differentiated intonation markers for different modal particles; adding unstressed markers for weak-ending suffixes. The specific implementation steps for performing sentiment processing on the text processed in S2 in S3 include: When performing sentiment processing on punctuation marks, at least set different pause durations and intonation trajectories composed of rising and falling tones for commas, full stops, exclamation marks, question marks, dashes, book titles, and quotation marks respectively. Specifically, add markers including pause duration and emotional intonation for punctuation marks: The marker corresponding to a comma is ", · ↘" (short pause + falling tone); The marker corresponding to a full stop is ". · · · ↘↘" (long pause + deep falling tone); The marker corresponding to an exclamation mark is "! · · · ↗↗" (long pause + deep rising tone); The marker corresponding to a question mark is "? · · · ↘↗" (long pause + falling then rising tone); The marker corresponding to a dash is "—— · · · · · · · ·" (extremely long pause); The marker correspondence for book titles is as follows: "《" corresponds to the marker "《 · ↗" (light pause + rising tone); "》" corresponds to the marker "》 · ↘" (light pause + falling tone).

[0028] When adding stress, unstressed, or intonation markers for preset literary keywords, add stress (°), unstressed (_), or intonation markers for more than 50 literary keywords classified by scenarios (visual, auditory, tactile, emotional, action, etc.): Visual category (e.g., "嫩绿" is marked as "°嫩绿°↗", "朦胧" is marked as "°朦胧°↘"); Emotional category (e.g., "欣喜" is marked as "°欣喜°↗↗", "惆怅" is marked as "°惆怅°↘↘"); Function words are unstressed (e.g., "的 / 地 / 得" is marked as "的 / 地 / 得").

[0029] The exclusive sentiment-based reading rules for reduplicated words include: for AA-style reduplicated words, use unstressed and elongated processing; for ABB-style reduplicated words, use stressed first character and unstressed last character processing; for ABAB-style reduplicated words, use rhythm splitting processing; Specifically, the reading rule for AA-style (e.g., "慢慢, 轻轻") is "\1\1~" (unstressed + elongated); The reading rule for ABB-style (e.g., "怯生生, 扑棱棱") is "°\1\1\1°_" (stressed + unstressed tail); The reading rule for the ABAB pattern (such as "you zai you zai, hu shan hu shan") is "°\1\2°・\1\2_" (rhythm splitting).

[0030] When adding intonation marks for modal particles, different modal particles correspond to different preset intonation mark results respectively.

[0031] In this embodiment, "ma" at the end of a sentence is marked with a rising tone, "ne" is marked with a stronger rising tone, "ba" is marked with a falling tone, and "ai" is marked with a deep falling tone.

[0032] The optimization of the light sound suffix includes adding a light reading mark to the suffix of preset characters (such as: "tou", "zi", "men", "me").

[0033] Specifically, for example: "zhi tou / shi tou" is marked as "\1_tou_"; "hai zi / zhuo zi" is marked as "\1_zi_"; "wo men / ni men" is marked as "\1_men_"; "shen me" is marked as "shen_me".

[0034] S4. For the text after emotional processing, perform long sentence splitting based on the triple rules of semantic blocks, classical Chinese function words, and character length; and perform hierarchical processing on the English part in the text, distinguish pure capital letter combinations from regular English words, and add splitting and rhythm marks to pure capital letter combinations; generate optimized text pronunciation information.

[0035] The S4 performs long sentence splitting based on the triple rules of semantic blocks, classical Chinese function words, and character length, specifically including: The first rule: According to preset semantic transition or progressive keywords (more than 20 words such as "but", "yet", "suddenly", "originally", etc.), split the text after emotional processing into different semantic blocks; The second rule: Further split the results of the semantic blocks split by the first rule according to preset classical Chinese function words (zhi, hu, zhe, ye, etc.) to obtain new different semantic blocks; The third rule: Determine the length of the semantic blocks in the results of the semantic blocks split by the second rule. If the length of the semantic block is greater than the preset value, then according to different preset punctuation marks in the corresponding semantic block, split the corresponding semantic block to reduce the length of the semantic block; in this embodiment, when encountering a comma, semicolon or colon, forcefully split the text content in the semantic block. If the length of the semantic block is less than or equal to the preset value (in this embodiment, this preset value is taken as 30), then do not split the corresponding semantic block according to different preset punctuation marks in the corresponding semantic block, as Figure 3 shown.[[ID=3�]]

[0036] The specific execution steps of S4 for performing layered processing on the English part of the text include: for pure uppercase letter combinations consisting of two or more uppercase letters, the corresponding pure uppercase letter combinations are split into multiple individual letters, and rhythmic marks with stressed and elongated states are added to each letter; for regular English words with the first letter capitalized or all lowercase, the integrity of the corresponding English words is maintained.

[0037] This embodiment also provides a text pronunciation optimization system for speech synthesis, including: The operator conversion module is used to perform semantic differentiation processing on the minus sign, division sign and other symbols in the input text; The number-to-Chinese module is used to perform hierarchical conversion processing on integers, decimals, and numbers with units; The polyphonic character and pinyin adaptation module is used to perform scenario-based polyphonic character matching and convert native characters with tones into a format that is compatible with the speech synthesis engine and includes pinyin, numerical tones and prosodic markers; The punctuation emotion processing module is used to add pause and emotional tone marks to various categories of punctuation marks; The literary keyword emphasis module is used to add emphasis, de-emphasis, or intonation markers to preset literary keywords; The reduplicated word emotion optimization module is used to apply exclusive emotion-based reading rules based on the structural type of reduplicated words; The modal particle differentiation module is used to add differentiated intonation markers for different modal particles; The neutral tone suffix optimization module is used to add a neutral tone mark to neutral tone suffixes; The long sentence semantic segmentation module is used to segment long sentences based on a three-tiered rule system of semantic blocks, classical Chinese function words, and character length. The English layered processing module is used to distinguish and differentiate between pure uppercase letter combinations and regular English words; The synthesized speech output module is used to generate optimized text pronunciation information corresponding to the input text based on the processing results of the operator conversion module, number to Chinese conversion module, polyphonic character and pinyin adaptation module, punctuation emotion processing module, literary keyword repetition module, reduplicated word emotion optimization module, modal particle differentiation module, neutral tone suffix optimization module, long sentence semantic segmentation module and English layered processing module.

[0038] 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.

[0039] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A text pronunciation optimization method for speech synthesis, characterized in that, include: S1. Perform operator semantic discrimination processing on the input text, and convert the minus sign and the minus sign into their corresponding Chinese semantics according to the context; and perform hierarchical conversion processing on the numbers in the input text, distinguishing between integers, decimals and numbers with units, and converting them into Chinese pronunciations according to the integer hierarchical conversion rules, the decimal digit-by-digit conversion rules and the unit adaptation rules respectively. S2. For the text processed by S1, perform contextualized polyphonic character matching, select the correct pronunciation for polyphonic characters based on the context; and convert the original characters with tones into a format compatible with the target speech synthesis engine and containing pinyin, numerical tones and prosodic markers; S3. Perform emotional processing on the text processed by S2, add markers containing pause duration and emotional tone to punctuation marks; add stressed, unstressed, or intonation markers to preset literary keywords; apply exclusive emotional reading rules based on different structural types to reduplicated words; add differentiated intonation markers to different tone words; add unstressed reading markers to unstressed suffixes. S4. For the emotionally processed text, perform long sentence splitting based on three rules: semantic blocks, classical Chinese function words, and character length; and perform layered processing on the English part of the text, distinguishing between pure uppercase letter combinations and regular English words, adding splitting and prosodic markers to pure uppercase letter combinations; and generate optimized text pronunciation information.

2. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The operator semantic differentiation processing in S1 specifically includes: When the "-" sign is preceded or followed by numbers or decimal points, the corresponding Chinese semantic conversion result is determined to be subtraction. When the "-" sign is preceded by an equal sign or is located at the beginning of the text and followed by a number, the corresponding Chinese semantic conversion result will be determined to be negative; The operator semantic differentiation processing also includes: recording the Chinese semantic conversion result corresponding to the "÷" sign as division by.

3. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The integer tiered conversion rule in S1 is used to segment and convert integers according to a preset range of trillions to hundreds of millions to tens of thousands, and to eliminate the redundancy of consecutive zeros; the decimal digit-by-digit conversion rule is to convert each digit of the decimal part independently; the unit adaptation rule includes at least converting "℃" to degrees Celsius and "㎡" to square meters.

4. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The scenario-based polyphonic character matching in S2 includes: filtering out the pronunciation of each polyphonic character in the corresponding text, as well as the pronunciation of each polyphonic character in the database when they are combined in a preset format or in a word combination. Each preset format or word combination of the corresponding polyphonic character in the database constitutes a contextual scenario for the corresponding polyphonic character. When converting native characters with tones into a format compatible with the target speech synthesis engine and containing pinyin, numeric tones, and prosodic markers, each native character with tones corresponds to a preset format combination result containing pinyin, numeric tones, and prosodic markers, and the format combination result containing pinyin, numeric tones, and prosodic markers corresponding to the same native character with different tones is different.

5. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The specific steps for performing emotional processing on the text processed by S2 in S3 include: When applying emotional processing to punctuation marks, at least commas, periods, exclamation marks, question marks, dashes, book titles, and quotation marks should be given different pause durations and intonation trajectories consisting of rising and falling tones. The specific emotional reading rules for reduplicated words include: for AA-type reduplicated words, a light reading with a lengthened tone is used; for ABB-type reduplicated words, the first character is stressed and the last character is read lightly is used; and for ABAB-type reduplicated words, rhythm splitting is used. When adding intonation markers for modal particles, different modal particles correspond to different preset intonation marker results.

6. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The optimization of the neutral tone suffix includes adding a neutral tone mark to the preset text suffix.

7. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The S4 process performs long sentence segmentation based on a triple rule system of semantic blocks, classical Chinese function words, and character length, specifically including: The first rule: Based on preset semantic transition or progressive keywords, the emotionally processed text is split into different semantic blocks; The second rule: Based on the preset classical Chinese function words, the semantic blocks split by the first rule are further split to obtain new and different semantic blocks; The third rule: Determine the length of the semantic blocks in the semantic block results split by the second rule. If the length of the semantic block is greater than the preset value, then split the corresponding semantic block according to the different preset punctuation marks in the corresponding semantic block to reduce the length of the semantic block; if the length of the semantic block is less than or equal to the preset value, then do not split the corresponding semantic block according to the different preset punctuation marks in the corresponding semantic block.

8. The text pronunciation optimization method for speech synthesis according to claim 1, characterized in that, The specific execution steps of S4 for performing layered processing on the English part of the text include: for pure uppercase letter combinations consisting of two or more uppercase letters, the corresponding pure uppercase letter combinations are split into multiple individual letters, and rhythmic marks with stressed and elongated states are added to each letter; for regular English words with the first letter capitalized or all lowercase, the integrity of the corresponding English words is maintained.

9. A text pronunciation optimization system for speech synthesis, characterized in that, include: The operator conversion module is used to perform semantic differentiation processing on the minus sign, division sign and other symbols in the input text; The number-to-Chinese module is used to perform hierarchical conversion processing on integers, decimals, and numbers with units; The polyphonic character and pinyin adaptation module is used to perform scenario-based polyphonic character matching and convert native characters with tones into a format that is compatible with the speech synthesis engine and includes pinyin, numerical tones and prosodic markers; The punctuation emotion processing module is used to add pause and emotional tone marks to various categories of punctuation marks; The literary keyword emphasis module is used to add emphasis, de-emphasis, or intonation markers to preset literary keywords; The reduplicated word emotion optimization module is used to apply exclusive emotion-based reading rules based on the structural type of reduplicated words; The modal particle differentiation module is used to add differentiated intonation markers for different modal particles; The neutral tone suffix optimization module is used to add a neutral tone mark to neutral tone suffixes; The long sentence semantic segmentation module is used to segment long sentences based on a three-tiered rule system of semantic blocks, classical Chinese function words, and character length. The English layered processing module is used to distinguish and differentiate between pure uppercase letter combinations and regular English words; The synthesized speech output module is used to generate optimized text pronunciation information corresponding to the input text based on the processing results of the operator conversion module, number to Chinese conversion module, polyphonic character and pinyin adaptation module, punctuation emotion processing module, literary keyword repetition module, reduplicated word emotion optimization module, modal particle differentiation module, neutral tone suffix optimization module, long sentence semantic segmentation module and English layered processing module.

Images