A text processing method and system for speech synthesis

Abstract

The application provides a text processing method and system for speech synthesis, which can meet the prosody requirements of synthesized speech in various business scenarios, and does not require a large amount of manual design and debugging of prosody by humans, thereby saving the cost of landing of a speech synthesis scheme. The method is applied to a text processing system for speech synthesis, and the text processing system for speech synthesis comprises a personification transformation module. The method comprises the following steps: determining a first prosody label according to a business keyword of a to-be-processed text obtained by the personification transformation module, wherein the business keyword is defined based on a scene to which speech corresponding to the to-be-processed text is applied; performing prosody label prediction processing on the to-be-processed text based on a neural network model by the personification transformation module to obtain a second prosody label; performing prosody labeling on the to-be-processed text based on the first prosody label and the second prosody label to obtain a target text; and synthesizing speech corresponding to the target text according to a phoneme sequence of the target text.

Description

Technical Field

[0001] This application relates to the field of speech synthesis technology, and in particular to a text processing method and system for speech synthesis. Background Technology

[0002] With the development of artificial intelligence, the application of AI-based text-to-speech (TTS) technology is becoming increasingly widespread. This involves preprocessing the text to be processed (referred to as front-end processing) before the text is synthesized by a speech synthesis engine. Preprocessing the text allows the speech synthesis engine to more accurately recognize and process it, converting it into natural and fluent speech for output. The naturalness of the speech (i.e., a human-like effect) is usually achieved through prosody. To achieve this effect, the text to be processed needs to be tagged with appropriate prosodic labels, allowing the speech synthesis engine to determine the prosody of the synthesized speech at the corresponding positions based on these labels.

[0003] Currently, in speech synthesis methods, the commonly used text processing approach is rule-based front-end processing. Rule-based front-end processing requires the formulation of a series of rules and algorithms based on linguistic knowledge to preprocess the text to be processed.

[0004] However, since different business scenarios have different prosodic requirements for synthesized speech, when a speech synthesis solution needs to be applicable to multiple business scenarios, rule-based front-end processing requires researchers to manually design and debug the prosodic of the text to be processed for each business scenario, and then manually annotate the prosodic of the text to be processed by manually labeling the prosodic. This text processing method makes the implementation of speech synthesis methods costly. Summary of the Invention

[0005] This application provides a text processing method and system for speech synthesis, which can meet the prosodic requirements of synthesized speech in various business scenarios, and eliminates the need for extensive manual design and debugging of prosody, thus saving the cost of implementing speech synthesis solutions.

[0006] Firstly, a text processing method for speech synthesis is provided, applied to a text processing system for speech synthesis, the text processing system for speech synthesis including a human-like transformation module, and the method including:

[0007] Based on the business keywords of the text to be processed obtained by the anthropomorphic transformation module, the first prosodic tag is determined. The business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied.

[0008] The anthropomorphic transformation module performs prosodic label prediction processing on the text to be processed based on a neural network model to obtain the second prosodic label;

[0009] The text to be processed is prosodicly labeled based on the first prosodic label and the second prosodic label to obtain the target text;

[0010] Based on the phoneme sequence of the target text, synthesize the speech corresponding to the target text.

[0011] In one feasible design, the first prosodic tag is determined based on the business keywords of the text to be processed obtained by the anthropomorphic transformation module, including:

[0012] The business keywords of the text to be processed are obtained through the anthropomorphic transformation module;

[0013] The anthropomorphic transformation module obtains business scenario tags based on business keywords;

[0014] Through the anthropomorphic transformation module, a first business dictionary suitable for the target scenario of business scenario tag identification is selected from multiple business dictionaries. The first business dictionary is used to associate at least one business keyword with at least one first prosody tag.

[0015] The anthropomorphic transformation module determines the first prosody tag that matches the business keywords from the first business dictionary.

[0016] In one feasible design, the speech corresponding to the target text is synthesized based on the phoneme sequence of the target text, including:

[0017] Based on the first neural network model, the target text is processed by pinyinization and polyphonic character disambiguation to determine the target pinyin corresponding to characters with different pinyin in the scene and generate a pinyin sequence.

[0018] Perform phoneme decomposition on the pinyin sequence to obtain the phoneme sequence;

[0019] Input a phoneme sequence into the speech synthesis engine to synthesize the speech corresponding to the target text.

[0020] In a feasible design, the target text undergoes pinyinization and polyphonic character disambiguation, including:

[0021] Get the pinyin corresponding to each character in the target text;

[0022] Encode each character to obtain the context vector representation of each character;

[0023] Calculate the probability of the pinyin corresponding to each character based on the context vector representation of each character;

[0024] Based on the probability of the pinyin corresponding to each character, determine the target pinyin with the highest probability from the pinyin corresponding to each character.

[0025] In one feasible design, business scenario tags determined based on business keywords are obtained, including:

[0026] Obtain a neural network classifier trained with business keywords and business scenario labels;

[0027] Input the text to be processed into the neural network classifier and obtain the business scenario label corresponding to the text.

[0028] In one feasible design, business scenario tags determined based on business keywords are obtained, including:

[0029] Determine the first association relationship, which is used to associate at least one business keyword and at least one business scenario tag;

[0030] Based on the first association relationship and the business keywords, determine the business scenario tags corresponding to the business keywords.

[0031] In one feasible design, the pinyin sequence is split into phonemes to obtain a phoneme sequence, including:

[0032] Based on the business scenario tags, determine whether to perform pinyin correction on the pinyin sequence;

[0033] If pinyin correction is performed, the pinyin sequence corresponding to the target text is updated, and the updated pinyin sequence is processed by phoneme decomposition to obtain the phoneme sequence;

[0034] If no pinyin correction is performed, the pinyin sequence is split into phonemes to obtain the phoneme sequence.

[0035] In one feasible design, based on business scenario tags, it is determined whether to perform pinyin correction on the pinyin sequence, including:

[0036] If the business scenario label indicates that the text to be processed includes names or place names, it is determined that pinyin text will be corrected. Specifically, if pinyin correction is performed, the pinyin sequence corresponding to the target text will be updated, including:

[0037] If pinyin correction is performed, the pinyin corresponding to the name or place name in the target text is re-determined to update the pinyin sequence corresponding to the target text;

[0038] If the business scenario label indicates that the text to be processed does not include a name or place name, then it is determined not to perform pinyin correction on the pinyin text.

[0039] In one feasible design, the text to be processed is prosodicly labeled based on a first prosodic tag and a second prosodic tag to obtain the target text, including:

[0040] The anthropomorphic transformation module is used to obtain the text to be processed with the second prosody tag;

[0041] By using the anthropomorphic modification module, the first prosodic tag is added to the text to be processed that already has a second prosodic tag, and the target text is obtained.

[0042] If the same position in the target text corresponds to both a first prosodic tag and a second prosodic tag, the second prosodic tag corresponding to that position is deleted using the anthropomorphic modification module.

[0043] Secondly, a text processing system for speech synthesis is provided, comprising:

[0044] The anthropomorphic transformation module is used to determine the first prosodic tag based on the business keywords of the acquired text to be processed. The business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied.

[0045] The anthropomorphic transformation module is also used to perform prosodic label prediction processing on the text to be processed based on a neural network model to obtain a second prosodic label;

[0046] The anthropomorphic transformation module is also used to perform prosodic annotation on the text to be processed based on the first prosodic tag and the second prosodic tag to obtain the target text;

[0047] A speech synthesis engine is used to synthesize the speech corresponding to a target text based on the phoneme sequence of the target text.

[0048] The embodiments described above in this application enhance the human-like quality of synthesized speech by adding a humanization modification module to a text processing system used for speech synthesis. Specifically, firstly, based on the business keywords of the text to be processed obtained by the humanization modification module, a first prosodic tag is determined. Since the business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied, the first prosodic tag corresponding to the business keywords can fit the scenario in which the text to be processed is applied. That is, by adding a first prosodic tag to the text to be processed through the humanization modification module, the final synthesized speech can conform to the pausing habits of that scenario, thus improving the human-like quality of the speech. For example, when broadcasting an address, a real person usually adds a pause after words such as province, city, and county. Therefore, the first prosodic tag added to the text to be processed by the humanization modification module in this application embodiment can meet the prosodic requirements of synthesized speech in various business scenarios, and eliminates the need for extensive manual design and debugging of the prosody, saving the cost of implementing the speech synthesis solution.

[0049] Furthermore, considering that the text to be processed typically includes not only business keywords related to the business scenario but also other general text, this application uses a human-like modification module to perform prosodic tag prediction processing on the text to be processed based on a neural network model to obtain a second prosodic tag. This achieves prosodic annotation of the general text in the text to be processed, improving the accuracy of the prosodic tags added to the text. This ensures that the target text obtained based on the prosodic annotation of the first and second prosodic tags includes highly accurate prosodic tags, thereby making the speech synthesized from the target text more natural. Attached Figure Description

[0050] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of a text processing system for speech synthesis provided in an exemplary embodiment of this application;

[0052] Figure 2 This is a schematic diagram of another example of a text processing system for speech synthesis provided in an exemplary embodiment of this application;

[0053] Figure 3 This is a schematic flowchart of a text processing method for speech synthesis provided in an exemplary embodiment of this application;

[0054] Figure 4 This is a schematic diagram of an anthropomorphic modified module structure provided in an exemplary embodiment of this application. Detailed Implementation

[0055] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0056] Figure 1 This application provides an exemplary embodiment of a text processing system for speech synthesis, such as... Figure 1As shown, the text processing system for speech synthesis includes a humanization module. This humanization module, combined with the target scene, obtains prosodic tags for the text to be processed, enhancing its anthropomorphic effect, and outputs the target text annotated with these prosodic tags. This facilitates the output of humanized speech based on the phoneme sequence of the anthropomorphized target text.

[0057] For example, such as Figure 2 As shown, the text processing system for speech synthesis also includes a pinyin sequence generation module, a phoneme segmentation module, and a speech synthesis engine. The pinyin sequence generation module converts target text tagged with prosodic symbols into a pinyin sequence. The phoneme segmentation module splits the pinyin sequence into a phoneme sequence based on the initials and finals of the pinyin. The speech synthesis engine synthesizes speech based on the phoneme sequences.

[0058] based on Figure 1 and Figure 2 The text processing system shown is for speech synthesis, such as Figure 3 As shown, this application proposes a text processing method for speech synthesis, which is applied to the text processing system for speech synthesis. The method includes:

[0059] S110, determine the first prosody tag based on the business keywords of the text to be processed obtained by the anthropomorphic transformation module.

[0060] Among them, business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied.

[0061] Prosodic tags are used to indicate pauses. In synthesized speech, pauses are required at the prosodic tag positions to enhance the anthropomorphic effect of the speech. For example, prosodic tags #1, #2, #3, and #4 represent different pauses. #1 marks the boundary of a prosodic word, indicating a short pause; #2 marks the boundary of a prosodic phrase, indicating a prolonged or short pause; #3 indicates a more obvious pause with semantic completeness and a drop in intonation; and #4 marks the end of a sentence, a tag at the end of the sentence corresponding to each number. The first text after prosodic tagging is “XX Province #1 XX City #1 XX County #1”. It can be seen that there is a prosodic tag between the word segment “XX Province” and “XX City”, indicating that after reading the first segment, a pause of the duration corresponding to #1 should be made before reading the second segment.

[0062] In a feasible design, such as Figure 4 As shown, the anthropomorphic transformation module includes a business classification submodule and an anthropomorphic effect enhancement submodule. The business classification submodule includes a rule classifier or a neural network classifier, and the anthropomorphic effect enhancement submodule includes a pattern matcher and a prosodic label predictor. Based on this, the first prosodic label can be determined in the following way:

[0063] The business keywords of the text to be processed are obtained through a rule classifier.

[0064] The rule classifier is used to obtain business scenario tags determined based on business keywords;

[0065] The pattern matcher selects a first business dictionary from multiple business dictionaries that is suitable for identifying the target scenario using business scenario tags. The first business dictionary is used to associate at least one business keyword with at least one first prosody tag.

[0066] The pattern matcher determines the first prosody tag that matches the business keyword from the first business dictionary.

[0067] To improve the efficiency of matching the corresponding first prosody tag from the business dictionary based on business keywords, this application first classifies multiple business dictionaries according to the category of business scenario tags, that is, each business scenario tag corresponds to one business dictionary. Then, based on the business scenario tag, a first business dictionary suitable for that scenario is selected from multiple business dictionaries. This ensures that only the first prosody tag corresponding to the business keyword needs to be matched in the first business dictionary, avoiding the problem of reducing efficiency by matching from all business dictionaries.

[0068] In a feasible design, the methods for obtaining business keywords from the text to be processed through a rule classifier include:

[0069] A keyword list is obtained through a rule classifier. This keyword list includes multiple keywords statistically analyzed based on various business scenarios applicable to speech synthesis.

[0070] The rule classifier matches the text to be processed for speech synthesis with a keyword list to determine the corresponding business keywords from the text to be processed for speech synthesis.

[0071] In the example above, the keyword list enables the quick and accurate identification of business keywords from the text to be processed.

[0072] Examples of scenarios for voice applications of the text to be processed include address broadcasting, name broadcasting, weather broadcasting, etc.

[0073] For example, the keyword list includes keywords such as "province, city, county, may I ask, are you?". It can be seen that "province, city, county" is suitable for address announcements, for example, when the text to be processed is "XX province XX city XX county"; "may I ask, are you?" is suitable for name announcements, for example, when the text to be processed is "Hello, may I ask if you are Li XX?".

[0074] In a feasible design, business scenario tags can be obtained in the following ways:

[0075] Obtain a neural network classifier trained with business keywords and business scenario labels;

[0076] Input the text to be processed into the neural network classifier and obtain the business scenario label corresponding to the text.

[0077] For example, such as Figure 4 As shown, the neural network classifier is deployed in the business classification submodule.

[0078] The neural network classifier is a neural network model based on a classification algorithm, which can be trained using training samples. The training samples include multiple business keywords labeled with business scenario tags, allowing the classifier to learn the semantic features of the business keywords and business scenario tags, thereby calculating the matching business scenario tags based on the business keywords.

[0079] In the example above, the neural network classifier can accurately determine the business scenario label corresponding to the text to be processed semantically.

[0080] In a feasible design, business scenario tags can be obtained using a rule classifier in the following way:

[0081] Determine the first association relationship, which is used to associate at least one business keyword and at least one business scenario tag;

[0082] Based on the first association relationship and the business keywords, determine the business scenario tags corresponding to the business keywords.

[0083] In the example above, the corresponding business scenario tag can be quickly determined based on the business keywords through the first association relationship.

[0084] S120, through the anthropomorphic transformation module, performs prosodic label prediction processing on the text to be processed based on a neural network model to obtain the second prosodic label.

[0085] The second prosody tag corresponds to the text in the text to be processed, excluding business keywords.

[0086] For example, a prosodic label predictor is used to perform prosodic label prediction processing on the text to be processed based on a neural network model to obtain a second prosodic label.

[0087] Since the text to be processed includes not only business keywords related to the business scenario, but also other general text, the second prosodic tags corresponding to the general text can be added using the prosodic tag predictor.

[0088] Specifically, the prosodic labeling model is generated by fine-tuning the pre-trained model of the Bidirectional Encoder Representation from Transformers (BERT) in the Transformers layer using a prosodic label predictor, and then trained using general text with prosodic labels.

[0089] The prosodic label predictor is used to input the text to be processed into the trained prosodic labeling model to obtain the text to be processed with a second prosodic label.

[0090] In the example above, the prosodic annotation model is used to add a second prosodic label to the text to be processed, which realizes the prosodic annotation of the text other than business keywords in the text to be processed, and improves the accuracy of the prosodic labels added to the text to be processed.

[0091] S130, perform prosodic annotation on the text to be processed based on the first prosodic label and the second prosodic label to obtain the target text.

[0092] In a feasible design, based on Figure 4 The anthropomorphic modification module shown can obtain the target text in the following ways:

[0093] The text to be processed is obtained with a second prosodic label using a prosodic label predictor.

[0094] By adding a submodule with an anthropomorphic effect, the first prosodic tag is added to the text to be processed with the second prosodic tag, and the target text is obtained.

[0095] If the same position in the target text corresponds to both a first prosodic tag and a second prosodic tag, a submodule is added using an anthropomorphic effect, and the second prosodic tag corresponding to that position is deleted.

[0096] The addition of the first prosody tag can be achieved in the following ways:

[0097] Add a first prosodic tag based on the position of the business keyword in the text to be processed. That is, add the corresponding first prosodic tag after the business keyword or between two business keywords.

[0098] For example, the text to be processed is "Hello, are you Li XX?", where the business keywords include "please" and "are you?". After being annotated by the prosodic annotation model, the text to be processed is "Hello #3, are you Li XX #4?". The target text after adding the first prosodic tag is "Hello #3, please #1, are you Li XX #4?".

[0099] For example, the text to be processed is "He is in XX County, XX City, XX Province", where the business keywords include "province", "city", and "county". After being annotated by the prosodic annotation model, the text to be processed is "He is in #2XX Province, XX City, XX County". The target text after adding the first prosodic tag is "He is in #2XX Province, #1XX City, #1XX County, #1".

[0100] It's important to note that since the prosodic labeling model described above is trained on general text with prosodic tags, the second prosodic tag added to the text being processed is usually placed after the general text. However, it's possible that a second prosodic tag might exist after a business keyword. In this case, if the same position in the target text corresponds to both a first and a second prosodic tag, a submodule can be added using a personification effect to delete the corresponding second prosodic tag. This method can improve the accuracy of the prosodic tags in the target text.

[0101] S140, synthesize the speech corresponding to the target text based on the phoneme sequence of the target text.

[0102] In Chinese, polyphonic characters refer to characters that have different pronunciations in different contexts. Polyphonic character disambiguation is an important task in the field of natural language processing, and it is of great significance for applications such as speech recognition, pinyin input methods, and machine translation. Traditional polyphonic character disambiguation methods rely on manually written rules and dictionaries, which are difficult to adapt to large-scale corpora and complex linguistic environments. Current machine learning-based front-end processing automatically learns text features and patterns by training models to achieve text preprocessing, but it still struggles to solve the problem of polyphonic characters.

[0103] To address the aforementioned issues, in a feasible design where the text processing system used for speech synthesis also includes a pinyin sequence generation module, a phoneme decomposition module, and a speech synthesis engine, the synthesized speech corresponding to the target text can be achieved through the following methods:

[0104] The Pinyin sequence generation module, based on the first neural network model, performs Pinyinization and polyphonic character disambiguation on the target text to determine the target Pinyin corresponding to characters with different Pinyin in the scene and generate Pinyin sequences.

[0105] The phoneme sequence is obtained by performing phoneme segmentation on the pinyin sequence using the phoneme segmentation module.

[0106] Input a phoneme sequence into the speech synthesis engine to synthesize the speech corresponding to the target text.

[0107] For example, the pinyin sequence generation module can perform pinyinization and polyphonic character disambiguation on the target text in the following ways:

[0108] Obtain the pinyin corresponding to each character in the target text;

[0109] Encode each character to obtain the context vector representation of each character;

[0110] Calculate the probability of the pinyin corresponding to each character based on the context vector representation of each character;

[0111] Determine the target pinyin corresponding to the maximum probability from the pinyin corresponding to each character according to the probability of the pinyin corresponding to each character.

[0112] Exemplarily, the ways to encode each character include:

[0113] Use the bert model to encode the input sentence.

[0114] Exemplarily, the first neural network model is g2pW (from the paper: g2pW: A Conditional Weighted Softmax BERT for Chinese Multisyllabic Disambiguation).

[0115] In the above example, through the first neural network model, the target pinyin corresponding to each character in the target text can be recognized, and when there are multiple pinyins corresponding to a character, the correct target pinyin can be determined from the multiple pinyins according to the context.

[0116] Exemplarily, the phoneme splitting process of the pinyin sequence is implemented through the following method to obtain the phoneme sequence:

[0117] Based on the business scenario label, the pinyin sequence generation module determines whether to correct the pinyin of the pinyin sequence;

[0118] If pinyin correction is performed, the pinyin sequence generation module updates the pinyin sequence corresponding to the target text, and uses the phoneme splitting module to perform phoneme splitting on the updated pinyin sequence to obtain the phoneme sequence;

[0119] If pinyin correction is not performed, the phoneme splitting module is used to perform phoneme splitting on the pinyin sequence to obtain the phoneme sequence.

[0120] It should be understood that both the pinyin sequence and the phoneme sequence carry corresponding prosody labels. For example, taking "ni hao" as an example, the pinyin sequence [ni3hao3#1] is split into the phoneme sequence ["n", "i3", "h", "ao3", "#1"].

[0121] Exemplarily, the pinyin sequence generation module determines whether to correct the pinyin of the pinyin sequence through the following method:

[0122] If the business scenario label indicates that the text to be processed includes a name or a place name, determine to correct the pinyin of the pinyin text. Among them, if pinyin correction is performed, update the pinyin sequence corresponding to the target text, including:

[0123] If pinyin correction is performed, re-determine the pinyin corresponding to the name or place name in the target text to update the pinyin sequence corresponding to the target text;

[0124] If the business scenario label indicates that the text to be processed does not include a name or a place name, determine not to correct the pinyin of the pinyin text.

[0125] Exemplarily, a list including the pinyin of each surname and a list including the pinyin of each place name can be counted, and the pinyin corresponding to the name or place name in the pinyin text can be re-determined according to each list.

[0126] For example, for the text to be processed "Are you Zeng XX?", the pinyin of "Zeng" recognized by the neural network model is "ceng2" (2 indicates the rising tone), and after correction, the pinyin of "Zeng" is "zeng1" (1 indicates the high-level tone).

[0127] This application takes into account that for the text to be processed including a name or a place name, the pinyin of the "surname" or the pinyin of the "place name" may be a polyphonic character. Since the pronunciation of a name or a place name usually has nothing to do with the context, therefore, if the character in the surname of a name or a place name is a polyphonic character, the neural network model may not be able to predict the pronunciation of the "surname" or the "place name" based on the name. In the above embodiments, by correcting the pinyin of the text corresponding to the "surname" or the "place name", the deficiencies of the neural network model can be compensated, and the accuracy of the pinyin of the text can be improved.

[0128] In a feasible design, the following method is used to perform phoneme splitting processing on the pinyin sequence:

[0129] Obtain the pinyin of each character in the pinyin text;

[0130] According to the initial and final sounds of the pinyin of each character, split the pinyin of each character.

[0131] In the above example, the initial sound represents the beginning part of a syllable, and the final sound represents the remaining part of a syllable. The result after pinyin conversion can separate the initial and final sounds by simple rules or by looking up a table. For example, for the pinyin "zhong1", the initial sound is "zh", and the final sound is "ong1".

[0132] The embodiments described above in this application enhance the human-like quality of synthesized speech by adding a humanization modification module to a text processing system used for speech synthesis. Specifically, firstly, based on the business keywords of the text to be processed obtained by the humanization modification module, a first prosodic tag is determined. Since the business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied, the first prosodic tag corresponding to the business keywords can fit the scenario in which the text to be processed is applied. That is, by adding a first prosodic tag to the text to be processed through the humanization modification module, the final synthesized speech can conform to the pausing habits of that scenario, thus improving the human-like quality of the speech. For example, when broadcasting an address, a real person usually adds a pause after words such as province, city, and county. Therefore, the first prosodic tag added to the text to be processed by the humanization modification module in this application embodiment can meet the prosodic requirements of synthesized speech in various business scenarios, and eliminates the need for extensive manual design and debugging of the prosody, saving the cost of implementing the speech synthesis solution.

[0133] Furthermore, considering that the text to be processed typically includes not only business keywords related to the business scenario but also other general text, this application uses a human-like modification module to perform prosodic tag prediction processing on the text to be processed based on a neural network model to obtain a second prosodic tag. This achieves prosodic annotation of the general text in the text to be processed, improving the accuracy of the prosodic tags added to the text. This ensures that the target text obtained based on the prosodic annotation of the first and second prosodic tags includes highly accurate prosodic tags, thereby making the speech synthesized from the target text more natural.

[0134] In conjunction with the methods described above, this application also provides a text processing system for speech synthesis, such as... Figure 2 As shown, the system includes:

[0135] The anthropomorphic transformation module is used to determine the first prosodic tag based on the business keywords of the acquired text to be processed. The business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied.

[0136] The anthropomorphic transformation module is also used to perform prosodic label prediction processing on the text to be processed based on a neural network model to obtain a second prosodic label;

[0137] The anthropomorphic transformation module is also used to perform prosodic annotation on the text to be processed based on the first prosodic tag and the second prosodic tag to obtain the target text;

[0138] A speech synthesis engine is used to synthesize the speech corresponding to a target text based on the phoneme sequence of the target text.

[0139] In a feasible design, the anthropomorphic modification module determines the first prosodic label in the following ways:

[0140] Obtain the business keywords of the text to be processed;

[0141] Obtain business scenario tags determined based on business keywords;

[0142] Select a first business dictionary from multiple business dictionaries that is suitable for identifying the target scenario using business scenario tags. The first business dictionary is used to associate at least one business keyword with at least one first prosody tag.

[0143] Identify the first prosody tag that matches the business keyword from the first business dictionary.

[0144] In one feasible design, the device further includes a pinyin sequence generation module and a phoneme splitting module. The pinyin sequence generation module is used to perform pinyinization and polyphonic character disambiguation processing on the target text based on the first neural network model, so as to determine the target pinyin corresponding to the characters with different pinyin in the scene and generate a pinyin sequence.

[0145] The phoneme splitting module is used to perform phoneme splitting on the pinyin sequence to obtain the phoneme sequence;

[0146] A speech synthesis engine is used to synthesize the speech corresponding to a target text based on a phoneme sequence.

[0147] In a feasible design, the pinyin sequence generation module performs pinyinization and polyphonic character disambiguation on the target text in the following way:

[0148] Get the pinyin corresponding to each character in the target text;

[0149] Encode each character to obtain the context vector representation of each character;

[0150] Calculate the probability of the pinyin corresponding to each character based on the context vector representation of each character;

[0151] Based on the probability of the pinyin corresponding to each character, determine the target pinyin with the highest probability from the pinyin corresponding to each character.

[0152] In a feasible design, the anthropomorphic transformation module obtains business scenario tags determined based on business keywords in the following way:

[0153] Obtain a neural network classifier trained with business keywords and business scenario labels;

[0154] Input the text to be processed into the neural network classifier and obtain the business scenario label corresponding to the text.

[0155] In a feasible design, the anthropomorphic transformation module is implemented by obtaining business scenario tags determined based on business keywords in the following way:

[0156] Determine the first association relationship, which is used to associate at least one business keyword and at least one business scenario tag;

[0157] Based on the first association relationship and the business keywords, determine the business scenario tags corresponding to the business keywords.

[0158] In one feasible design, the pinyin sequence generation module is used to determine whether to perform pinyin correction on the pinyin sequence based on the business scenario tags;

[0159] If pinyin correction is performed, the pinyin sequence generation module is used to update the pinyin sequence corresponding to the target text, and the phoneme splitting processing module is used to perform phoneme splitting processing on the updated pinyin sequence to obtain the phoneme sequence.

[0160] If no pinyin correction is performed, the phoneme splitting module is used to split the pinyin sequence into phonemes and obtain the phoneme sequence.

[0161] In a feasible design, the pinyin sequence generation module is implemented by determining whether to correct the pinyin sequence based on the business scenario tags:

[0162] If the business scenario label indicates that the text to be processed includes names or place names, it is determined that pinyin text will be corrected. Specifically, if pinyin correction is performed, the pinyin sequence corresponding to the target text will be updated, including:

[0163] If pinyin correction is performed, the pinyin corresponding to the name or place name in the target text is re-determined to update the pinyin sequence corresponding to the target text;

[0164] If the business scenario label indicates that the text to be processed does not include a name or place name, then it is determined not to perform pinyin correction on the pinyin text.

[0165] In one feasible design, an anthropomorphic modification module is used to acquire the text to be processed with a second prosodic tag;

[0166] The anthropomorphic modification module is used to add the first prosodic tag to the text to be processed that has the second prosodic tag, and to obtain the target text.

[0167] If the same position in the target text corresponds to both a first prosodic tag and a second prosodic tag, the anthropomorphic modification module is used to delete the second prosodic tag corresponding to that position.

[0168] Other implementations and effects of this device can be found in the description of the text processing method embodiment for speech synthesis, and will not be repeated here.

[0169] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0170] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0171] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0172] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0173] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0174] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A text processing method for speech synthesis, characterized in that, A text processing system for speech synthesis, comprising a human-like characterization module, is applied to the text processing system for speech synthesis, the method comprising: Based on the business keywords of the text to be processed obtained by the anthropomorphic transformation module, a first prosodic tag is determined. The business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied. The determination of the first prosodic tag based on the business keywords of the text to be processed obtained by the anthropomorphic transformation module includes: The anthropomorphic transformation module is used to obtain the business keywords of the text to be processed; The anthropomorphic transformation module is used to obtain business scenario tags determined based on the business keywords; Through the anthropomorphic transformation module, a first business dictionary suitable for the target scenario identified by the business scenario tag is selected from multiple business dictionaries. The first business dictionary is used to associate at least one business keyword with at least one first prosody tag. The anthropomorphic transformation module determines a first prosodic tag that matches the business keyword from the first business dictionary. The anthropomorphic modification module performs prosodic tag prediction processing on the text to be processed based on a neural network model to obtain a second prosodic tag. The text to be processed is tactilely annotated based on the first prosodic tag and the second prosodic tag to obtain the target text; Based on the phoneme sequence of the target text, the corresponding speech is synthesized.

2. The method according to claim 1, characterized in that, The step of synthesizing the speech corresponding to the target text based on the phoneme sequence of the target text includes: Based on the first neural network model, the target text is processed by pinyinization and polyphonic character disambiguation to determine the target pinyin corresponding to characters with different pinyin in the scenario and generate a pinyin sequence. The phoneme sequence is split into phonemes to obtain a phoneme sequence. The phoneme sequence is input into the speech synthesis engine to synthesize the speech corresponding to the target text.

3. The method according to claim 2, characterized in that, The process of performing pinyinization and polyphonic character disambiguation on the target text includes: Obtain the pinyin corresponding to each character in the target text; Encode each of the aforementioned characters to obtain the context vector representation of each character; Calculate the probability of the pinyin corresponding to each character based on the context vector representation of each character; Based on the probability of the pinyin corresponding to each character, determine the target pinyin with the highest probability from the pinyin corresponding to each character.

4. The method according to claim 2 or 3, characterized in that, The step of obtaining the business scenario tags determined based on the business keywords includes: Obtain a neural network classifier trained with business keywords and business scenario labels; Input the text to be processed into the neural network classifier to obtain the business scenario label corresponding to the text to be processed.

5. The method according to claim 2 or 3, characterized in that, The step of obtaining the business scenario tags determined based on the business keywords includes: Determine a first association relationship, which is used to associate at least one business keyword and at least one business scenario tag; Based on the first association relationship and the business keyword, determine the business scenario tag corresponding to the business keyword.

6. The method according to claim 2 or 3, characterized in that, The step of performing phoneme decomposition on the pinyin sequence to obtain a phoneme sequence includes: Based on the business scenario tags, determine whether to perform pinyin correction on the pinyin sequence; If pinyin correction is performed, the pinyin sequence corresponding to the target text is updated, and the updated pinyin sequence is subjected to phoneme decomposition to obtain a phoneme sequence; If no pinyin correction is performed, the pinyin sequence is split into phonemes to obtain a phoneme sequence.

7. The method according to claim 6, characterized in that, The step of determining whether to perform pinyin correction on the pinyin sequence based on the business scenario tag includes: If the business scenario label indicates that the text to be processed includes a name or place name, it is determined that the pinyin sequence should be corrected. If pinyin correction is performed, the pinyin sequence corresponding to the target text is updated, including: If pinyin correction is performed, the pinyin corresponding to the name or place name in the target text is re-determined to update the pinyin sequence corresponding to the target text; If the business scenario label indicates that the text to be processed does not include a name or place name, it is determined that no pinyin correction will be performed on the pinyin sequence.

8. The method according to any one of claims 1-3, characterized in that, The step of performing prosodic annotation on the text to be processed based on the first prosodic tag and the second prosodic tag to obtain the target text includes: The anthropomorphic modification module is used to obtain the text to be processed, which carries the second prosody tag. The anthropomorphic modification module adds the first prosody tag to the text to be processed, which has the second prosody tag, to obtain the target text. If the same position in the target text corresponds to both the first prosodic tag and the second prosodic tag, the anthropomorphic modification module deletes the second prosodic tag corresponding to that position.

9. A text processing system for speech synthesis, characterized in that, include: The anthropomorphic modification module is used to determine a first prosodic label based on the business keywords of the acquired text to be processed. The business keywords are defined based on the scenario in which the speech corresponding to the text to be processed is applied. Determining the first prosodic label based on the business keywords of the text to be processed acquired by the anthropomorphic modification module includes: The anthropomorphic transformation module is used to obtain the business keywords of the text to be processed; The anthropomorphic transformation module is used to obtain business scenario tags determined based on the business keywords; Through the anthropomorphic transformation module, a first business dictionary suitable for the target scenario identified by the business scenario tag is selected from multiple business dictionaries. The first business dictionary is used to associate at least one business keyword with at least one first prosody tag. The anthropomorphic transformation module determines a first prosodic tag that matches the business keyword from the first business dictionary. The anthropomorphic transformation module is also used to perform prosodic tag prediction processing on the text to be processed based on a neural network model to obtain a second prosodic tag; The anthropomorphic modification module is also used to perform prosodic annotation on the text to be processed based on the first prosodic tag and the second prosodic tag to obtain the target text; A speech synthesis engine is used to synthesize the speech corresponding to the target text based on the phoneme sequence of the target text.

Images