A recognition method for information-based synthesized speech

Abstract

The present invention discloses a method for recognizing information-based synthesized speech, which relates to the technical field of intelligent speech and is used to solve the problem of low accuracy in the recognition and judgment of information-based synthesized speech. The specific steps include: collecting sample speech and performing preprocessing to obtain processed speech; extracting audio features of the processed speech through a speech content analysis model, and converting the processed speech into text to generate processed speech text; extracting text content features of the processed speech text, and formulaically analyzing to generate an excitement coefficient; setting an excitement coefficient comparison threshold, comparing the excitement coefficient comparison threshold with the excitement coefficient, and generating an emotion identifier; analyzing the audio features to generate a sound-text difference value, performing threshold comparison on the sound-text difference value, and then generating a sound-text conformity identifier; integrating the emotion identifier and the sound-text conformity identifier to generate a recognition target, and marking the corresponding sample speech.

Description

Technical Field

[0001] The present invention relates to the field of intelligent speech technology, and in particular to a method for recognizing information-based synthesized speech. Background Art

[0002] Information-based synthetic speech is a technology that converts text or other forms of information into sound. It can convert text information such as books, documents, and emails into sound, providing auditory assistance for the visually impaired while providing accessibility. For those who have reading difficulties or cannot read, information-based synthetic speech can convert text content into speech, providing them with an easier way to understand and obtain information, including the elderly, the disabled, and those with learning difficulties.

[0003] Reference publication number CN113921012A proposes a synthetic speech recognition method, system, intelligent device and storage medium, which obtain target text data to be recognized based on the target speech to be recognized, obtain speech emotion data of each pronunciation phoneme in the target speech data to be recognized and text emotion data of each word in the target text data to be recognized, and judge whether the speech emotion data and the text emotion data match. If the speech emotion data and the text emotion data match, the target speech data to be recognized is determined to be non-synthetic speech data, and whether the speech is synthetic can be judged from the emotional aspect, which can improve the accuracy and reliability of the judgment.

[0004] Combining the above solutions with existing technologies, the following issues need to be addressed:

[0005] However, with the widespread use of intelligent AI, the emotions of information-based synthesized speech are becoming increasingly similar to those of real people. Emotional assessment alone cannot identify the characteristics of information-based synthesized speech, making the method of using tone to assess emotion too limited.

[0006] 2. In the existing technology, the recognition of information-based synthesized speech has become increasingly difficult with the advancement of technology, and the accuracy of the recognition and judgment of information-based synthesized speech has been relatively low.

[0007] In order to solve the above-mentioned problems, a recognition method for information-based synthesized speech is proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for recognizing information-based synthesized speech to address the shortcomings of the background technology.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] The method for recognizing information-based synthesized speech comprises the following steps:

[0011] Step S100: collecting sample speech and performing preprocessing to obtain processed speech;

[0012] Step S200: extracting audio features from the processed speech using a speech content analysis model, and converting the processed speech into text to generate a processed speech text;

[0013] Step S300, extracting text content features from the processed speech text, and formulaically analyzing and generating an excitement coefficient;

[0014] Step S400: setting an excitement coefficient comparison threshold, comparing the excitement coefficient comparison threshold with the excitement coefficient, and generating an emotion identifier;

[0015] Step S500: Analyze the audio features to generate a voice-text difference value, perform a threshold comparison on the voice-text difference value, and then generate a voice-text consistency indicator;

[0016] Step S600 : Integrate the emotion identifier and the audio-text consistency identifier to generate a recognition target, and mark the corresponding sample speech.

[0017] In a preferred embodiment, the text content features include emotion expression words and emotion symbols, and the text content analysis process is as follows:

[0018] The processed speech text is divided into n analysis intervals, where one sentence represents one analysis interval. The emotion expression words and symbols in the n analysis intervals are extracted and integrated for analysis. The emotion expression words are classified according to the degree of emotional excitement. Emotions such as happiness, excitement, and anger are set as first-level emotion words. Emotions such as sadness, fear, disgust, and anxiety are set as second-level emotion words. Calm emotions are set as third-level emotion words. First-level emotion words have greater emotional reflection than second-level emotion words, and so on.

[0019] Classify "," and "." as first-level emotional symbols, "..." and "~" as second-level emotional symbols, and "?" and "!" as third-level emotional symbols. First-level emotional symbols reflect less emotion than second-level emotional symbols, and so on.

[0020] The emotion expression words and emotion symbols in n analysis intervals are matched and analyzed. When a single analysis interval contains both first-level emotion words and third-level emotion symbols, the analysis interval is marked as a high emotion object. When a single analysis interval contains both second-level emotion words and second-level emotion symbols, first-level emotion words and second-level emotion symbols, or second-level emotion words and third-level emotion symbols, the analysis interval is marked as a moderate emotion object. When a single analysis interval contains both second-level emotion words and first-level emotion symbols, first-level emotion words and second-level emotion symbols, or first-level emotion words and first-level emotion symbols, the analysis interval is marked as a low emotion object. The matching analysis between other emotion expression words and emotion symbols does not mark the degree of emotion objects. High emotion objects have a greater degree of emotion expression than moderate emotion objects, that is, the emotions analyzed in the text are stronger.

[0021] In a preferred embodiment, the generation logic of the excitement coefficient is:

[0022] Assume that the number of analysis intervals marked as high emotion objects is a1, the number of analysis intervals marked as medium emotion objects is a2, and the number of analysis intervals marked as low emotion objects is a3. Given that the total number of analysis intervals for processing speech text is n, the excitement coefficient β is obtained by formula analysis:

[0023] β= (wherein a1, a2, and a3 are all greater than or equal to 0, and n is an integer greater than 1);

[0024] Through formula processing, it can be seen that when the proportion of the number of analysis intervals a1 containing highly emotional objects is greater, the sum of the proportion of the number of analysis intervals a2 containing moderate emotional objects and the number of analysis intervals a3 containing low emotional objects is smaller, the excitement coefficient β is larger. The larger the excitement coefficient β, the greater the degree of emotional expression in the processed speech text, and vice versa.

[0025] In a preferred embodiment, the emotion marker includes a shallow emotion marker, a balanced emotion marker, and a deep emotion marker, and the emotion marker generation step is:

[0026] Set the excitement coefficient comparison thresholds to qx1 and qx2, where qx1 < qx2, and both qx1 and qx2 are greater than 0, substitute the excitement coefficient β in the processed speech text into the excitement coefficient comparison thresholds qx1 and qx2 for comparison analysis. If the excitement coefficient β is less than the excitement coefficient comparison threshold qx1, generate a shallow emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx1 and less than the excitement coefficient comparison threshold qx2, generate a balanced emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx2, generate a deep emotion label for the sample speech corresponding to the processed speech text;

[0027] Among them, the processed speech text has deep emotional labels, which means that the degree of emotional expression in the processed speech text is high. The emotional labels in the processed speech text decrease in degree from deep emotional labels to shallow emotional labels. Compared with information-based synthesized speech, the degree of change of human voice is more obvious, that is, the deeper the emotional labels, the lower the degree of information-based synthesis, and so on.

[0028] In a preferred embodiment, the audio features include audio speech rate and pitch intensity, and the steps of analyzing the audio features are:

[0029] The sample speech is divided into m detection intervals, and each sentence in the speech is represented as a single interval. It can be seen that in the m detection intervals of the sample speech and the n analysis intervals of the processed speech text, the number of detection intervals m is the same as the number of analysis intervals n;

[0030] Extract the audio speed and pitch intensity of a single detection interval in the sample speech, where the pitch is divided into high-speed rhythm, medium-speed rhythm and low-speed rhythm, and the pitch intensity can be divided into high-intensity tone, medium-intensity tone and low-intensity tone;

[0031] The audio speaking speed and tone intensity in m detection intervals are matched and analyzed. When a single detection interval has both high-speed rhythm and high-intensity tone, the detection interval is marked as a high-amplitude excited object. When a single detection interval has both high-speed rhythm and medium-intensity tone, medium-speed rhythm and high-intensity tone, or medium-speed rhythm and medium-intensity tone, the detection interval is marked as a medium-amplitude excited object. When a single detection interval has both low-speed rhythm and medium-intensity tone, low-speed rhythm and low-intensity tone, or medium-speed rhythm and low-intensity tone, the detection interval is marked as a low-amplitude excited object. The matching analysis between other audio speaking speeds and tone intensities does not mark the amplitude of the excited objects. High-amplitude excited objects express greater excitement than medium-amplitude excited objects, that is, the emotions analyzed in the audio are more intense.

[0032] In a preferred embodiment, the steps for generating the phonetic-text difference value are:

[0033] Let the audio-text difference value be δ, and perform a formal difference analysis on the number of highly excited objects in the m detection intervals of the sample speech and the number of highly emotional objects in the n analysis intervals of the processed speech text. Specifically, let the number of detection intervals marked with highly excited objects be b, and the number of analysis intervals known to be marked with highly emotional objects be a1, and calculate the audio-text difference value δ through the company;

[0034] δ= +j, where j is the correction constant for phonetic and textual differences, a1 and b are greater than or equal to 0, and m and n are both greater than 1;

[0035] Through formula processing, it can be seen that the greater the difference between the number of detection intervals b marked as highly excited objects and the number of analysis intervals a1 marked as highly emotional objects, the greater the judgment error between the audio of the entire sample speech and the processed speech text, that is, the greater the audio-text difference value δ, the lower the authenticity of the sample speech.

[0036] In a preferred embodiment, the process of generating the phonetic-text conformity indicator is as follows:

[0037] Let the difference threshold be L, where L is an integer greater than 0. Substitute the phonetic-text difference value δ of the sample speech into the difference threshold L for comparison. When L>δ≥0, generate a phonetic-text conformance mark for the sample speech; when δ≥L, generate a phonetic-text non-conformance mark for the sample speech.

[0038] The more the audio-text mismatch mark in the sample speech indicates, the greater the judgment error between the audio of the entire sample speech and the processed speech text, while the more the audio-text match mark indicates, the smaller the judgment error between the audio of the entire sample speech and the processed speech text;.

[0039] In a preferred embodiment, the recognition target includes a forged voice target, a real voice target and a depth detection target, and the generation step of the recognition target is:

[0040] If the same sample speech contains both a speech-to-text discrepancy mark and a shallow emotion marker or a speech-to-text discrepancy mark and a balanced emotion marker, the sample speech is marked as a fake speech target; if the same sample speech contains both a speech-to-text consistency mark and a deep emotion marker or a speech-to-text consistency mark and a balanced emotion marker, the sample speech is marked as a real speech target; if the same sample speech contains both a speech-to-text discrepancy mark and a deep emotion marker or a speech-to-text consistency mark and a shallow emotion marker, the sample speech is marked as a deep detection target;

[0041] The depth detection target, i.e., the sample audio, has recognition obstacles and needs to be subsequently identified by other systems. It is not repeatedly identified in the present invention.

[0042] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0043] In this embodiment, through the matching analysis of emotional words and emotional symbols, emotional objects are generated for each sentence one by one, so as to preliminarily determine the emotional expression of the content in the sample speech, and through the generation of the excitement coefficient β, the analysis is strengthened on the basis of the matching analysis of emotional words and emotional symbols, so as to achieve in-depth sentiment analysis of the text end of the sample speech. In addition, through the dual-end analysis of the audio end and the text end, the accurate recognition function of the information-based synthesized speech is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0045] Figure 1 The present invention is a flowchart of a method for recognizing information-based synthesized speech. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] See also Figure 1 As shown, the present embodiment provides a method for recognizing information-based synthesized speech, and the method includes the following steps:

[0048] Step S100, collecting sample speech and performing preprocessing to obtain processed speech;

[0049] Step S200: extracting audio features from the processed speech using a speech content analysis model, and converting the processed speech into text to generate a processed speech text;

[0050] Step S300: extracting text content features from the processed speech text and formulaically analyzing and generating an excitement coefficient;

[0051] Step S400: setting an excitement coefficient comparison threshold, comparing the excitement coefficient comparison threshold with the excitement coefficient, and generating an emotion identifier;

[0052] Step S500: Analyze the audio features to generate a voice-text difference value, perform a threshold comparison on the voice-text difference value, and then generate a voice-text consistency indicator;

[0053] Step S600 : Integrate the emotion identifier and the audio-text consistency identifier to generate a recognition target, and mark the corresponding sample speech.

[0054] Example 1

[0055] It should be noted that preprocessing includes noise reduction, noise elimination, and audio gain adjustment. Adobe Audition: Adobe Audition is a tool for multi-directional noise reduction and audio gain adjustment. It can optimize audio quality through its built-in noise reduction algorithm and audio gain adjustment function. The specific noise reduction algorithm involved is the adaptive noise reduction algorithm, which analyzes the noise characteristics in the audio and automatically adjusts the noise reduction parameters to effectively reduce noise and retain the clarity of the original audio.

[0056] The text content features include emotion expression words and emotion symbols. The text content analysis process is as follows;

[0057] The processed speech text is divided into n analysis intervals, where one sentence represents one analysis interval. The emotion expression words and symbols in the n analysis intervals are extracted and integrated for analysis. The emotion expression words are classified according to the degree of emotional excitement. Emotions such as happiness, excitement, and anger are set as first-level emotion words. Emotions such as sadness, fear, disgust, and anxiety are set as second-level emotion words. Calm emotions are set as third-level emotion words. First-level emotion words have greater emotional reflection than second-level emotion words, and so on.

[0058] Classify "," and "." as first-level emotional symbols, "..." and "~" as second-level emotional symbols, and "?" and "!" as third-level emotional symbols. First-level emotional symbols reflect less emotion than second-level emotional symbols, and so on.

[0059] Matching analysis is performed on the emotion expression words and emotion symbols in n analysis intervals. When a single analysis interval contains both first-level emotion words and third-level emotion symbols, the analysis interval is marked as a high emotion object. When a single analysis interval contains both second-level emotion words and second-level emotion symbols, first-level emotion words and second-level emotion symbols, or second-level emotion words and third-level emotion symbols, the analysis interval is marked as a moderate emotion object. When a single analysis interval contains both second-level emotion words and first-level emotion symbols, first-level emotion words and second-level emotion symbols, or first-level emotion words and first-level emotion symbols, the analysis interval is marked as a low emotion object. The matching analysis between other emotion expression words and emotion symbols does not mark the degree of emotion objects. High emotion objects have a greater degree of emotion expression than moderate emotion objects, that is, the emotions analyzed in the text are more intense.

[0060] It should be noted that the specific emotional expression words are:

[0061] Happy: joy, happiness, excitement, happiness, pleasure, satisfaction, joy, delight, comfort, and delight, etc.

[0062] Sadness: sad, upset, melancholy, grief, disappointment, frustration, pain, sorrow, misery, tears, etc.

[0063] Angry: angry, furious, annoyed, indignant, furious, furious, resentful, furious, furious, etc.;

[0064] Fear: fear, fear, horror, panic, fright, trembling, timid, panic-stricken, etc.

[0065] Dislike: hate, boredom, aversion, disgust, disgust, disgust, dislike, contempt, hatred, impatience, etc.

[0066] Excitement: excitement, excitement, enthusiasm, excitement, passion, intensity, excitement, ecstasy, etc.

[0067] Calm: calm, cool, composed, composed, composed, peaceful, serene, and relaxed, etc.

[0068] Anxiety: anxiety, tension, uneasiness, worry, concern, irritability, restlessness, etc.

[0069] When analyzing relevant emotional expression words, a sentiment analysis model is used for emotion matching. The sentiment analysis model specifically uses Naive Bayes in the machine learning model. In addition, a sentiment recognition model is used for classification. The model can classify texts according to features such as emotional vocabulary, tone, and grammatical structure, and divide texts into different emotion categories, such as happiness, sadness, anger, fear, etc. Common emotion recognition models also include methods based on machine learning and deep learning. The training data of these two models are usually text data sets labeled with emotion or mood labels. By learning and training on a large amount of labeled data, we can learn to recognize patterns and rules of emotions and feelings, so that we can classify and analyze new texts for emotions or feelings.

[0070] The generation logic of the excitement coefficient is:

[0071] Assume that the number of analysis intervals marked as high emotion objects is a1, the number of analysis intervals marked as medium emotion objects is a2, and the number of analysis intervals marked as low emotion objects is a3. Given that the total number of analysis intervals for processing speech text is n, the excitement coefficient β is obtained by formula analysis:

[0072] β= (wherein a1, a2, and a3 are all greater than or equal to 0, and n is an integer greater than 1);

[0073] Through formula processing, it can be seen that when the proportion of the number of analysis intervals a1 containing highly emotional objects is greater, the sum of the proportion of the number of analysis intervals a2 containing moderate emotional objects and the number of analysis intervals a3 containing low emotional objects is smaller, the excitement coefficient β is larger. The larger the excitement coefficient β, the greater the degree of emotional expression in the processed speech text, and vice versa.

[0074] Emotional markers include shallow emotion markers, balanced emotion markers, and deep emotion markers. The steps for generating emotion markers are as follows:

[0075] Set the excitement coefficient comparison thresholds to qx1 and qx2, where qx1 < qx2, and both qx1 and qx2 are greater than 0, substitute the excitement coefficient β in the processed speech text into the excitement coefficient comparison thresholds qx1 and qx2 for comparison analysis. If the excitement coefficient β is less than the excitement coefficient comparison threshold qx1, generate a shallow emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx1 and less than the excitement coefficient comparison threshold qx2, generate a balanced emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx2, generate a deep emotion label for the sample speech corresponding to the processed speech text;

[0076] Among them, the processed speech text has deep emotional labels, which means that the degree of emotional expression in the processed speech text is high. The emotional labels in the processed speech text decrease in degree from deep emotional labels to shallow emotional labels. Compared with information-based synthesized speech, the degree of change of human voice is more obvious, that is, the deeper the emotional labels, the lower the degree of information-based synthesis, and so on.

[0077] By matching and analyzing emotional words and emotional symbols, we generate emotional objects for each sentence one by one, thereby preliminarily determining the emotional expression of the content in the sample speech. In addition, by further generating the excitement coefficient β, we strengthen the analysis based on the matching analysis of emotional words and emotional symbols, thereby achieving in-depth sentiment analysis of the text end of the sample speech.

[0078] Example 2

[0079] Based on the first embodiment, the present invention only analyzes the processed speech text converted from the sample speech, that is, the text end of the sample speech is analyzed, and there is a single analysis object. Based on the above situation, the audio end of the sample speech is further analyzed.

[0080] Audio features include audio speed and pitch intensity. The analysis steps for audio features are as follows:

[0081] The sample speech is divided into m detection intervals, and each sentence in the speech is represented as a single interval. It can be seen that in the m detection intervals of the sample speech and the n analysis intervals of the processed speech text, the number of detection intervals m is the same as the number of analysis intervals n;

[0082] Extract the audio speed and pitch intensity of a single detection interval in the sample speech, where the pitch is divided into high-speed rhythm, medium-speed rhythm and low-speed rhythm, and the pitch intensity can be divided into high-intensity tone, medium-intensity tone and low-intensity tone;

[0083] The audio speaking speed and tone intensity in m detection intervals are matched and analyzed. When a single detection interval has both high-speed rhythm and high-intensity tone, the detection interval is marked as a high-amplitude excited object. When a single detection interval has both high-speed rhythm and medium-intensity tone, medium-speed rhythm and high-intensity tone, or medium-speed rhythm and medium-intensity tone, the detection interval is marked as a medium-amplitude excited object. When a single detection interval has both low-speed rhythm and medium-intensity tone, low-speed rhythm and low-intensity tone, or medium-speed rhythm and low-intensity tone, the detection interval is marked as a low-amplitude excited object. The matching analysis between other audio speaking speeds and tone intensities does not mark the amplitude of the excited objects. High-amplitude excited objects express greater excitement than medium-amplitude excited objects, that is, the emotions analyzed in the audio are more intense.

[0084] It should be noted that: the significance of parameter selection;

[0085] Audio speed: Rhythm and speed can influence the perception of emotion. Fast and intense rhythms may be associated with emotions like excitement, tension, or anger, while slow and calm rhythms may be associated with emotions like calmness, sadness, or relaxation.

[0086] Pitch intensity: Pitch intensity can influence the expression of emotions. Higher pitch intensity may be associated with emotions such as excitement, anger, or tension, while lower pitch intensity may be associated with emotions such as calmness, softness, or quietness.

[0087] The steps for generating the phonetic-text difference value are:

[0088] Let the audio-text difference value be δ, and perform a formal difference analysis on the number of highly excited objects in the m detection intervals of the sample speech and the number of highly emotional objects in the n analysis intervals of the processed speech text. Specifically, let the number of detection intervals marked with highly excited objects be b, and the number of analysis intervals known to be marked with highly emotional objects be a1, and calculate the audio-text difference value δ through the company;

[0089] δ= +j, where j is the correction constant for phonetic and textual differences, a1 and b are greater than or equal to 0, and m and n are both greater than 1;

[0090] Through formula processing, it can be seen that the greater the difference between the number of detection intervals b marked as highly excited objects and the number of analysis intervals a1 marked as highly emotional objects, the greater the judgment error between the audio of the entire sample speech and the processed speech text, that is, the greater the audio-text difference value δ, the lower the authenticity of the sample speech.

[0091] The process for generating the audio-text conformity indicator is as follows:

[0092] Let the difference threshold be L, where L is an integer greater than 0. Substitute the phonetic-text difference value δ of the sample speech into the difference threshold L for comparison. When L>δ≥0, generate a phonetic-text conformance mark for the sample speech; when δ≥L, generate a phonetic-text non-conformance mark for the sample speech.

[0093] The sound-text mismatch mark in the sample speech indicates that the judgment error between the audio of the entire sample speech and the processed speech text is larger, while the sound-text match mark indicates that the judgment error between the audio of the entire sample speech and the processed speech text is smaller;

[0094] The recognition targets include fake speech targets, real speech targets, and depth detection targets. The steps for generating the recognition targets are as follows:

[0095] If the same sample speech contains both a speech-to-text discrepancy mark and a shallow emotion marker or a speech-to-text discrepancy mark and a balanced emotion marker, the sample speech is marked as a fake speech target; if the same sample speech contains both a speech-to-text consistency mark and a deep emotion marker or a speech-to-text consistency mark and a balanced emotion marker, the sample speech is marked as a real speech target; if the same sample speech contains both a speech-to-text discrepancy mark and a deep emotion marker or a speech-to-text consistency mark and a shallow emotion marker, the sample speech is marked as a deep detection target;

[0096] The deep detection target, i.e., the sample audio, has recognition obstacles and needs to be subsequently identified by other systems. In this invention, it is not repeatedly identified, thereby reducing the invalid recognition time.

[0097] In this embodiment, accurate recognition of information-based synthesized speech is achieved through dual-end analysis of the audio end and the text end.

[0098] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0099] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0100] In this application, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, "at least one of a, b, or c" can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0101] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0102] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0103] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0104] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0105] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for recognizing information-based synthesized speech, characterized in that: The method comprises the following steps: Step S100: collecting sample speech and performing preprocessing to obtain processed speech; Step S200: extracting audio features from the processed speech using a speech content analysis model, and converting the processed speech into text to generate a processed speech text; Step S300, extracting text content features from the processed speech text, and formulaically analyzing and generating an excitement coefficient; Step S400: setting an excitement coefficient comparison threshold, comparing the excitement coefficient comparison threshold with the excitement coefficient, and generating an emotion identifier; Step S500: Analyze the audio features to generate a voice-text difference value, perform a threshold comparison on the voice-text difference value, and then generate a voice-text consistency indicator; The audio features include audio speech rate and pitch intensity, and the steps of analyzing the audio features are: The sample speech is divided into m detection intervals, and each sentence in the speech is represented as a single interval. It can be seen that in the m detection intervals of the sample speech and the n analysis intervals of the processed speech text, the number of detection intervals m is the same as the number of analysis intervals n; Extract the audio speed and pitch intensity of a single detection interval in the sample speech, where the pitch is divided into high-speed rhythm, medium-speed rhythm and low-speed rhythm, and the pitch intensity is divided into high-intensity tone, medium-intensity tone and low-intensity tone; The audio speech speed and tone intensity in the m detection intervals are matched and analyzed. When a single detection interval has both high-speed rhythm and high-intensity tone, the detection interval is marked as a high-amplitude excitement object. When a single detection interval has both high-speed rhythm and medium-intensity tone, medium-speed rhythm and high-intensity tone, or medium-speed rhythm and medium-intensity tone, the detection interval is marked as a medium-amplitude excitement object. When a single detection interval has both low-speed rhythm and medium-intensity tone, low-speed rhythm and low-intensity tone, or medium-speed rhythm and low-intensity tone, the detection interval is marked as a low-amplitude excitement object. The matching analysis between other audio speech speeds and tone intensities does not mark the amplitude of the excitement object. Step S600 : Integrate the emotion identifier and the audio-text consistency identifier to generate a recognition target, and mark the corresponding sample speech.

2. A method according to claim 1, characterized in that, The text content features include emotion expression words and emotion symbols, and the text content analysis process is as follows: The processed speech text is divided into n analysis intervals, where one sentence represents one analysis interval. The emotion expression words and symbols in the n analysis intervals are extracted and integrated for analysis. The emotion expression words are classified according to the degree of emotional excitement. Happiness, excitement, and anger are set as first-level emotion words. Sadness, fear, disgust, and anxiety are set as second-level emotion words. Calmness is set as third-level emotion words. First-level emotion words have greater emotional reflection than second-level emotion words, and so on. Classify "," and "." as first-level emotional symbols, "..." and "~" as second-level emotional symbols, and "?" and "!" as third-level emotional symbols, where first-level emotional symbols reflect less emotion than second-level emotional symbols, and so on; The emotion expression words and emotion symbols in n analysis intervals are matched and analyzed. When a single analysis interval contains both first-level emotion words and third-level emotion symbols, the analysis interval is marked as a high emotion object. When a single analysis interval contains both second-level emotion words and second-level emotion symbols, first-level emotion words and second-level emotion symbols, or second-level emotion words and third-level emotion symbols, the analysis interval is marked as a moderate emotion object. When a single analysis interval contains both second-level emotion words and first-level emotion symbols, first-level emotion words and second-level emotion symbols, or first-level emotion words and first-level emotion symbols, the analysis interval is marked as a low emotion object. The matching analysis between other emotion expression words and emotion symbols does not mark the degree of emotion objects. High emotion objects have a greater degree of emotion expression than moderate emotion objects, that is, the emotions analyzed in the text are stronger.

3. A method according to claim 2, characterized in that, The generation logic of the excitement coefficient is: Assume that the number of analysis intervals marked as high emotion objects is a1, the number of analysis intervals marked as medium emotion objects is a2, and the number of analysis intervals marked as low emotion objects is a3. Given that the total number of analysis intervals for processing speech text is n, the excitement coefficient β is obtained by formula analysis: β= ； Wherein, a1, a2 and a3 are all greater than or equal to 0, and n is an integer greater than 1; Through formula processing, it can be seen that when the proportion of the number of analysis intervals a1 containing highly emotional objects is greater, the sum of the proportion of the number of analysis intervals a2 containing moderate emotional objects and the number of analysis intervals a3 containing low emotional objects is smaller, the excitement coefficient β is larger. The larger the excitement coefficient β, the greater the degree of emotional expression in the processed speech text, and vice versa.

4. A method according to claim 3, characterized in that, The emotion markers include shallow emotion markers, balanced emotion markers, and deep emotion markers. The emotion marker generation steps are: Set the excitement coefficient comparison thresholds to qx1 and qx2, where qx1 is less than qx2, and both qx1 and qx2 are greater than 0. Substitute the excitement coefficient β in the processed speech text into the excitement coefficient comparison thresholds qx1 and qx2 for comparison analysis. If the excitement coefficient β is less than the excitement coefficient comparison threshold qx1, generate a shallow emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx1 and less than the excitement coefficient comparison threshold qx2, generate a balanced emotion label for the sample speech corresponding to the processed speech text; if the excitement coefficient β is greater than the excitement coefficient comparison threshold qx2, generate a deep emotion label for the sample speech corresponding to the processed speech text.

5. A method according to claim 4, characterized in that, The steps for generating the phonetic-text difference value are: Let the audio-text difference value be δ, and perform a formulaic difference analysis on the number of highly excited objects in the m detection intervals of the sample speech and the number of highly emotional objects in the n analysis intervals of the processed speech text. Specifically, let the number of detection intervals marked with highly excited objects be b, and the number of analysis intervals known to be marked with highly emotional objects be a1, and calculate the audio-text difference value δ using the formula; δ= +j; Wherein, j is the correction constant for phonetic-to-text difference, a1 and b are greater than or equal to 0, and m and n are both greater than 1.

6. A method according to claim 5, characterized in that, The process of generating the audio-text conformity indicator is as follows: Assume that the difference threshold is L, where L is an integer greater than 0. Substitute the phonetic-text difference value δ of the sample speech into the difference threshold L for comparison. When L>δ≥0, generate a phonetic-text conformance mark for the sample speech; when δ≥L, generate a phonetic-text inconsistency mark for the sample speech.

7. A method according to claim 6, characterized in that: The recognition targets include a fake voice target, a real voice target, and a depth detection target. The generation steps of the recognition targets are: If the same sample speech contains both a speech-to-text mismatch mark and a shallow emotion marker or a speech-to-text mismatch mark and a balanced emotion marker, the sample speech is marked as a fake speech target; if the same sample speech contains both a speech-to-text mismatch mark and a deep emotion marker or a speech-to-text mismatch mark and a balanced emotion marker, the sample speech is marked as a real speech target; if the same sample speech contains both a speech-to-text mismatch mark and a deep emotion marker or a speech-to-text mismatch mark and a shallow emotion marker, the sample speech is marked as a deep detection target.

Images