An artificial intelligence-based adolescent emotional health intelligent evaluation method

Abstract

The present application relates to the technical field of emotional health assessment, in particular to a kind of adolescent emotional health intelligent evaluation method based on artificial intelligence.Method comprising: different abnormal emotion corresponding different voice feature rule model is built, the voice feature rule model is called based on the voice feature value of current user answering question, the voice emotion abnormal index of current user answering question on different voice feature is output, and the emotional consistency degree of the text content and voice of current user answering question is obtained, the text content emotional abnormal index of current user answering question with abnormal emotion dimension is weighted and fused based on emotional consistency degree, and the fusion abnormal score of current user on different abnormal emotion is obtained, and the emotional health assessment result of current user is obtained based on fusion abnormal score.The present application can significantly improve the accuracy of adolescent emotional health intelligent evaluation.

Description

Technical Field

[0001] This invention relates to the field of emotional health assessment technology, specifically to an artificial intelligence-based intelligent assessment method for adolescent emotional health. Background Technology

[0002] Because adolescents are in a period of rapid psychological and cognitive development, their emotional state is easily affected by various factors such as academic pressure, interpersonal relationships, and online information, leading to emotional problems such as anxiety and depression. However, adolescents' emotions are often highly concealed, and they generally tend to be reluctant to seek help. This makes it difficult to detect and intervene in their emotional health problems in a timely manner, which can trigger a chain of negative consequences, such as long-term cognitive impairment, decreased problem-solving abilities, and impaired future social adaptation; physical symptoms such as insomnia and headaches; and even self-harming behaviors or suicidal ideation. Therefore, timely and accurate assessment of adolescents' emotional health has become an important and urgent developmental need in the field of mental health.

[0003] Traditional assessments of adolescent emotional health primarily rely on psychological scales and human interviews. However, psychological scales are easily influenced by the respondent's subjective consciousness, social expectations, and attempts to conceal emotions, while human interviews are inefficient and fail to meet the needs of large-scale, continuous monitoring of adolescent emotional health. Therefore, with the development of artificial intelligence, AI-based assessment methods have been introduced. However, existing AI-based methods typically analyze adolescent emotions solely based on characteristics such as speech rate and volume, neglecting the fact that the user's own state (e.g., fatigue, tension) significantly affects their vocal characteristics, leading to misidentification. For example, normal fluctuations in state are easily misjudged as emotional abnormalities, and the consistency between the dialogue text and the actual vocal state is not thoroughly analyzed. Consequently, it is difficult to effectively identify the conflict between the expressed content and the inner feelings that adolescents may be trying to hide, or the conflict between the dialogue content and the true emotions, thus affecting the accuracy of adolescent emotional health assessment results. Therefore, how to effectively identify the consistency between dialogue content and vocal characteristics to improve the accuracy of adolescent emotional health assessment has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an artificial intelligence-based intelligent assessment method for adolescent emotional health, the specific technical solution of which is as follows:

[0005] One embodiment of the present invention provides an artificial intelligence-based intelligent assessment method for adolescent emotional health, comprising the following steps:

[0006] When conducting psychological scale assessments on users, the audio and text of the user's answers to each question, the average heart rate, and the emotional abnormality index of the text content are obtained. Based on the audio and text of the user's answers to each question, the voice feature values ​​of the user's answers to the corresponding questions are extracted. The voice features include speech rate, energy changes, and pauses.

[0007] Based on the emotional abnormality index of the text content of the historical users' answers to questions, the voice feature value and the average heart rate, the probability of abnormal emotions of each historical user sample is calculated, and a different voice feature pattern model corresponding to different abnormal emotions is constructed. The abnormal emotions include depression, anxiety and stress.

[0008] Based on the voice feature values ​​of the current user's answer to the question, the voice feature pattern model is invoked to output the voice emotion abnormality index of the current user's answer to the question on different voice features. The voice emotion abnormality index is then corrected based on the probability of the current user's abnormal emotion and compared with the emotion abnormality index of the text content of the current user's answer to the question to obtain the degree of consistency between the emotion of the text content and the voice of the current user's answer to the question. Based on the degree of emotional consistency, the emotion abnormality index of the text content of the current user's answer to the question on the same abnormal emotion dimension is weighted and fused to obtain the fusion abnormality score of the current user on different abnormal emotions. Based on the fusion abnormality score, the emotional health assessment result of the current user is obtained.

[0009] Beneficial Effects: This invention first acquires the audio, text, average heart rate, and emotional abnormality index of each user's answer to a question, and then extracts the voice feature values ​​of the user's answers to the corresponding questions based on the audio and text of each answer. Next, based on the emotional abnormality index of the text content of historical user answers, voice feature values, and average heart rate, the probability of abnormal emotions for each historical user sample is calculated, and a different voice feature pattern model corresponding to different abnormal emotions is constructed. Then, based on the voice feature values ​​of the current user's answer, the voice feature pattern model is invoked to output the voice emotional abnormality index of the current user's answer on different voice features. The voice emotional abnormality index is then corrected based on the probability of the current user's abnormal emotions and compared with the emotional abnormality index of the text content of the current user's answer to obtain the degree of emotional consistency between the text content and voice of the current user's answer. Based on the degree of emotional consistency, the emotional abnormality index of the text content of the current user's answers to questions of the same abnormal emotion dimension is weighted and fused to obtain the fused abnormality score of the current user on different abnormal emotions. Based on the fused abnormality score, the emotional health assessment result of the current user is obtained. Furthermore, by analyzing the consistency between text content and speech features and establishing a multimodal fusion mechanism that considers data credibility, this invention can significantly improve the accuracy of intelligent assessment of adolescent emotional health, or in other words, effectively overcome the misjudgment problems caused by emotional masquerading, individual differences, and temporary state interference, thereby providing more reliable and accurate emotional health assessment results. Detailed Implementation

[0010] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the protection scope of the embodiments of the present invention.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art.

[0012] This embodiment provides an artificial intelligence-based intelligent assessment method for adolescent emotional health, detailed as follows:

[0013] Step S001: When conducting psychological scale assessments on users, obtain the audio and text of the user's answers to each question, the average heart rate, and the emotional abnormality index of the text content. Based on the audio and text of the user's answers to each question, extract the voice feature values ​​of the user's answers to the corresponding questions. The voice features include speech rate, energy changes, and pauses.

[0014] Traditional AI-based assessment methods typically analyze adolescents' emotions solely based on characteristics such as speech rate and volume, neglecting the significant impact of the user's own state (e.g., fatigue, tension) on their speech characteristics. This leads to misinterpreting normal fluctuations as emotional abnormalities and fails to analyze the consistency between the dialogue text and the user's actual speech state. Consequently, it struggles to effectively identify inconsistencies between dialogue content and genuine emotions. Furthermore, adolescents often mask their true emotions due to societal expectations, resulting in responses that don't accurately reflect their feelings, thus affecting the accuracy of adolescent emotional health assessments. This embodiment aims to effectively identify inconsistencies between dialogue content and genuine emotions to improve the accuracy of adolescent emotional health assessments. It combines dialogue text and speech feature consistency analysis to enhance the accuracy of adolescent emotional health assessments. Specifically, this embodiment first collects and acquires dialogue text... The text, voice, and heart rate data collected during the process provide a data foundation for consistency analysis. Then, voice features such as speech rate, energy changes, and pauses are extracted. Combined with historical data (such as the emotional abnormality index of the text content of sample users' answers to questions, voice feature values, and average heart rate), the system learns the different voice feature change patterns under different abnormal emotions (depression, anxiety, and high stress). Then, the system obtains the emotional state reflected by the current user's voice features through the obtained voice feature change patterns, which is the voice emotional abnormality index. The system compares the emotional state reflected by the current user's voice features with the emotional abnormality index of the text content to calculate the consistency between the text and voice states in each dialogue. That is, it calculates the emotional consistency between the text content and voice of each question answered by the current user. Using the consistency degree as a weight, the system merges the text emotional index to assess the user's merged abnormality score in the three aspects of depression, anxiety, and high stress. Finally, the system outputs the emotional health level based on the severity grading scoring table.

[0015] Based on the above analysis, this embodiment assesses the user's emotional health using a psychological scale or engages in dialogue with the user based on questions in the psychological scale. Audio data of the user's answers to each question is captured and recorded as the corresponding user's answer audio. That is, in the psychological scale assessment process, the audio segment captured and stored in real time when the user verbally answers each specific question in the scale is the user's answer audio for that question. Furthermore, the answer audio for any given user to any question refers to the audio recorded from the moment the user begins answering the question until the end of the answer. In specific applications, the implementer can set the audio sampling frequency according to the actual situation. For example, in this embodiment, the audio sampling frequency can be set to 16kHz, where kHz represents kilohertz. The psychological scale used is the DASS-21 scale, also known as the Depression-Anxiety-Stress Scale. The DASS-21 contains 21 items, categorized as depression, anxiety, and stress. Each item corresponds to one question. Specifically, the DASS-21 includes 7 items in the depression dimension, 7 in the anxiety dimension, and 7 in the stress dimension. Therefore, the aforementioned psychological scale assessment refers to the process of a user interacting with AI and answering questions based on the DASS-21 scale. Audio data was acquired to facilitate the identification of whether users were faking their true emotions during the responses. Then, an automatic speech recognition model (such as the Whisper speech recognition model) was used to perform text recognition on the audio of each user's answers to each question. The results of the text recognition for each user's answers to each question were recorded as the corresponding answer text. Then, heart rate monitoring devices (such as smart bracelets) are used to obtain the average heart rate of each user when answering each question. That is, for a certain user, the average of all heart rate data collected during the time period from when the user starts answering a question to when the user finishes answering the question is recorded as the average heart rate of the user when answering the question.

[0016] For historical users in the sample, in order to obtain the emotional abnormality index reflected in each answer, users need to fill out the DASS-21 scale on the same day to obtain the abnormality score for each question in their historical answers. Since the scores for each question on the DASS-21 scale are 0, 1, 2, and 3 respectively, this embodiment needs to unify the scores to the range of 0 to 1 for easier calculation. That is, the abnormality scores for each question in the historical user answer scale are linearly normalized, and the linear normalization result is recorded as the normalized score for the corresponding question. Furthermore, for the sample... The minimum and maximum values ​​used to normalize the scores of historical users' answers to each question on the scale are 0 and 3, respectively. Therefore, if the original score is 0, the normalized score after normalization is still 0; if the original score is 1, the normalized score after normalization is 1 / 3 or 0.33; if the original score is 2, the normalized score after normalization is 2 / 3 or 0.67; and if the original score is 3, the normalized score after normalization is 1. After evaluating the user using the DASS-21 scale, the normalized score of the corresponding user's answer to each question can be obtained.

[0017] After obtaining the answer texts and normalized scores of historical users' answers to questions, the answer texts and normalized scores of all historical users' answers to questions are used as training samples to train the BERT model. Specifically, the answer text and normalized score of each question answered by historical users are paired to form a data pair, which is then used as input to train the BERT model. The normalized score of each question answered by historical users represents the sentiment abnormality index reflected in the corresponding historical user's answer to that question, while the normalized score of each question answered by historical users serves as a label for supervising the BERT model training. In specific applications, implementers need to set the training sample size according to actual conditions such as model accuracy and computing resources, as in this embodiment, it is optional. The answer texts and normalized scores of 3,000 adolescent users who participated in the DASS-21 scale assessment were used as training samples. That is, the number of historical users in this embodiment can be set to 3,000. Since the training samples are known, the process of training the BERT model is a well-known technique. In this embodiment, the trained BERT model is used as the emotional abnormality model of the answer text content. In addition, the emotional abnormality index of the text content of each question answered by the historical users of the sample is the normalized score of the corresponding historical user's answer to that question. In this embodiment, the historical users of the sample are all adolescents who have completed the emotional health intelligent assessment. Moreover, each user in this embodiment has completed the entire dialogue process with the AI ​​system, that is, they have all answered all the questions in the DASS-21 scale.

[0018] In this embodiment, the emotional abnormality model for the answer text content is mainly used to output the emotional abnormality index of the text content for each question answered by the current user. That is, for the current user, there is no need to select the score of their answer by filling in the DASS-21 scale, or to obtain the emotional abnormality index or abnormal score of the text content for each question answered by the current user by filling in the DASS-21 scale. Instead, the emotional abnormality index of the text content is obtained through a trained emotional abnormality model for the answer text content. In other words, the answer text of each question answered by the current user is input into the emotional abnormality model for the answer text content, and the model outputs the emotional abnormality index of the text content for the corresponding question answered by the current user. The current user refers to the adolescent individual undergoing emotional health assessment in the adolescent emotional health intelligent assessment strategy provided in this embodiment. That is, it is necessary to analyze the degree of emotional consistency between the text content and speech of each question answered by the current user, and use the degree of emotional consistency as a weight to weight and fuse the emotional index of the text content of the current user's answer questions, so as to assess the fusion abnormality score of the current user in the three emotional dimensions of depression, anxiety, and high stress.

[0019] Therefore, this embodiment can obtain the audio, text, average heart rate, and emotional abnormality index of the DASS-21 scale answers from historical users and the current user through the above process. This embodiment then needs to quantify the consistency between the emotional state reflected in the current user's voice features and the emotional abnormality index of the text content by comparing them. The emotional state reflected in the current user's voice features needs to be determined based on the different voice feature change patterns under different abnormal emotions learned from historical question text content emotional abnormality index, voice feature values, and average heart rate data. Therefore, this embodiment also needs to obtain the voice feature values ​​of the user's answers. Furthermore, since the user's speaking speed, voice changes, and pauses will change when the user is in an abnormal emotional state, and these voice changes are reflected by changes in audio signal energy, the audio signal energy... The changes are mainly characterized by energy changes in speech segments; and single features are often easily affected by individual differences among users, such as different users having different speaking speeds. Therefore, in order to analyze the user's true emotional state, this embodiment needs to extract multi-dimensional speech feature values ​​for each user's answer audio. The speech feature values ​​extracted in this embodiment include speech rate feature values, energy change feature values, and pause feature values. Since the method for obtaining the speech feature values ​​for each user's answer to each question is the same, for ease of understanding and description, this embodiment will subsequently describe the process of obtaining the speech feature values ​​for any user A's answer to the a-th question as an example. The specific process of obtaining the speech feature values ​​for user A's answer to the a-th question is as follows:

[0020] Since users may not speak continuously during the answering process, there may be pauses. That is, the actual speaking time in the answer audio may be less than the total length of the corresponding answer audio. However, the audio energy during pauses is significantly lower than the audio energy during speaking, and the sound characteristics are relatively stable within a short period. Therefore, in order to accurately extract speech feature values, this embodiment needs to first divide the answer audio based on a short time length, and then extract the speaking segments and pause segments based on the audio energy changes of each segment. The speech feature values ​​are then quantified based on the extraction results of the speaking and pause segments. Therefore, based on the above description, this embodiment needs to first construct a short time segmentation length, that is, to first construct a preset time length, and then divide the answer audio of the a-th question into non-overlapping segments based on the preset time window to obtain the audio segments of the answer audio of the a-th question. In other words, the answer to the a-th question is divided according to the time period of the preset time length as the basic unit. The audio is cut sequentially from beginning to end. Each segment of the audio obtained from the cut is independent, connected end to end, and does not overlap, ultimately forming a series of continuous, short audio segments. In addition, the last audio segment of the answer audio for question a may have a duration shorter than the preset duration. If the duration of the last audio segment of the answer audio for question a is not equal to the preset duration, then in this embodiment, the new segment obtained by merging the last two audio segments of the answer audio for question a, as well as the other audio segments before the last two audio segments of the answer audio for question a, are all recorded as audio windows of the answer audio for question a. If the duration of the last audio segment of the answer audio for question a is equal to the preset duration, then in this embodiment, each audio segment of the answer audio for question a is recorded as an audio window of the answer audio for question a. In practical applications, implementers can set a preset time length based on actual conditions such as sound characteristics and audio stability characteristics. However, the preset time length is required to be no longer than the time length of the answer audio for the a-th question. Since it is usually difficult to obtain a stable audio signal in a short time, and sound characteristics are usually relatively stable in a short time (such as when saying a word, the pitch and energy of the sound usually do not change much during the process), this embodiment can set the preset time length to 20ms, where ms represents milliseconds.

[0021] Then, the mean square energy of each audio window on the audio of the answer to the a-th question is calculated and denoted as the audio energy of the corresponding audio window; mean square energy is a fundamental parameter for accurately distinguishing speech segments from pause segments in audio, and it is also a fundamental parameter for subsequent extraction of speech features; and the expression for the audio energy of the w-th audio window on the audio of the answer to the a-th question is:

[0022]

[0023] in, The audio energy of the w-th audio window on the audio of the answer to the a-th question. Let w be the total number of sampling moments in the w-th audio window. This represents the amplitude corresponding to the i-th sampling moment in the w-th audio window; The average energy of the audio signal segment in the w-th audio window is represented by the value. The larger the value, the higher the signal energy and the more drastic the sound pressure change. The smaller the value, the lower the signal energy and the closer to silence. The number of sampling moments in the audio window is related to the audio sampling frequency mentioned above. If the audio sampling frequency is set to 16kHz, then the time interval between adjacent sampling moments is 0.0625 milliseconds.

[0024] Since different users have different speaking energy levels, directly dividing speech and pauses based on the calculated audio energy might misclassify quieter speech segments as pauses. Therefore, this embodiment, after obtaining the audio energy, also calculates the relative energy of the audio window relative to the overall average audio energy of the answer audio of the corresponding question. This relative energy is used to eliminate interference from overall volume differences between different audio segments, ensuring that energy judgment is not affected by user differences and other factors, and more accurately distinguishing between speech segments and pauses. The specific calculation process for relative energy is as follows: calculate the average audio energy of all audio windows on the answer audio of question a, and record it as the global average energy of question a; calculate the linearly normalized result of the ratio of the audio energy of each audio window on the answer audio of question a to the global average energy of question a, and record it as the relative energy of the corresponding audio window; linear normalization is also minimum-maximum normalization; taking the calculation process of the relative energy of the w-th audio window on the answer audio of question a as an example, the expression for the relative energy of the w-th audio window is: :

[0025]

[0026] in, Let be the relative energy of the w-th audio window, and norm() be the linear normalization function. Let be the global energy mean for the a-th problem. The audio energy of the w-th audio window on the audio of the answer to the a-th question; The larger the value, the higher the proportion of the energy of the w-th audio window in the audio of the answer to the a-th question, corresponding to a louder sound, and a greater likelihood that it belongs to the effective human voice region or a speech segment; conversely, the smaller the value, the higher the proportion of the energy of the w-th audio window in the audio of the answer to the a-th question. The smaller the value, the lower the proportion of the energy of the w-th audio window in the audio of the answer to the a-th question, the lower the loudness of the corresponding sound, and the more likely it is to belong to a silent segment or a background noise segment, which means it is more likely to belong to a pause segment.

[0027] In this embodiment, after obtaining the relative energy of each audio window on the audio of the answer to the a-th question, the speech window and pause window are identified based on the comparison between the relative energy of the audio window and a preset energy threshold. The specific identification process is as follows: Since the background sound energy is usually low when the user is not speaking, in this embodiment, audio windows with relative energy greater than or equal to the preset energy threshold are identified as speech windows, and audio windows with relative energy less than the preset energy threshold are identified as pause windows. In specific applications, the implementer can set the preset energy threshold according to the actual situation such as historical data distribution. For example, in this embodiment, the preset energy threshold can be obtained based on Otsu's method. That is, based on Otsu's method, the optimal threshold can be automatically calculated from the relative energy of the audio windows on the question and answer audio of all questions answered by users in all samples in history. This threshold is the preset energy threshold.

[0028] Because users' speech rate characteristics differ under different emotional states, after distinguishing between speech and pauses, the speech rate characteristic of user A's answer to question a is reflected by the total number of words in the speech segments of the audio recording of the answer to question a. Specifically, the process for obtaining the speech rate feature value of question a is as follows: The total number of words in the answer text for question a is counted, and the total duration of the speech window in the audio recording of the answer to question a is counted. The total duration of the speech window in the audio recording of the answer to question a is the sum of the durations of all speech windows in the audio recording of the answer to question a. The normalized result of comparing the total number of words in the answer text for question a with the total duration of the speech window in the audio recording of the answer to question a is denoted as the speech rate feature value of question a. The expression for the speech rate feature value of question a is: :

[0029]

[0030] in, Let be the speech rate feature value for the a-th question. The total number of words in the answer text for the a-th question. The total duration of the audio window for the answer to the a-th question; The larger and The smaller, or The larger the value, the faster user A answers the a-th question, and vice versa. The smaller the value, the slower user A speaks when answering the a-th question.

[0031] Since the volume of a user's speech is not completely stable and varies depending on their emotional state, the changes in audio energy also differ. For example, the volume of a user's speech will fluctuate naturally and smoothly under normal emotional conditions, while the volume will be lower and fluctuate less under depressed conditions, and the volume will fluctuate more drastically and may experience sudden changes under anxious conditions. Furthermore, since audio energy reflects the volume of a user's speech, this embodiment will calculate the energy change characteristic value of the a-th question based on the standard deviation of the audio energy of all speaking windows in the audio of the answer to the a-th question and the continuity or smoothness of the audio energy changes in all speaking windows. The specific calculation process for the energy change characteristic value of the a-th question is as follows: the standard deviation of the audio energy of all speaking windows in the audio of the answer to the a-th question is denoted as the standard deviation of the audio energy of the a-th question. The mean of the standard deviations of the speech audio energy of all historical questions is denoted as the historical standard deviation mean. All historical questions refer to all questions answered by all sample historical users. The linearly normalized result of comparing the standard deviation of the speech audio energy of question a with the historical standard deviation mean is denoted as the first energy change characterization value. The process of obtaining the standard deviation of the speech audio energy of any question answered by any sample historical user is the same as that of question a. The linearly normalized result of the mean of the absolute values ​​of the second-order differences of the audio energy between adjacent speech windows in the answer audio of question a is denoted as the second energy change characterization value. The second energy change characterization value reflects the continuity of audio energy changes in the speech window. The product of the first energy change characterization value and the second energy change characterization value is denoted as the energy change characteristic value of question a. The expression for the energy change characteristic value of question a is:

[0032]

[0033] in, Let be the characteristic value of the energy change for the a-th problem. Let the standard deviation of the speech audio energy for the a-th question be . The historical standard deviation is the mean. The total number of audio windows displaying the answer to the a-th question. , and These are the audio energies of the (r-1), r, and r+1th speaking windows on the audio of the answer to the a-th question, respectively. It measures the degree of fluctuation in volume of the audio response to the a-th question during the speaking process. The larger the value, the greater the volume fluctuation and the more unstable the audio segment. It characterizes the overall average volume fluctuation level of historical users in the sample, which is equivalent to a benchmark or reference value used to quantify the relative magnitude of volume fluctuation. and The ratio represents the relative magnitude of the volume fluctuation of the audio response to question a compared to the average volume fluctuation of the overall historical user volume. A larger ratio indicates that the volume fluctuation of the audio response to question a is relatively smaller compared to the average volume fluctuation of the entire sample of historical users. The stronger the intensity, the less it indicates that the fluctuation in audio volume of the answer to question a is relative to... The more stable; The absolute value of the second-order difference of the audio energy between the r-th speech window and the two speech windows before and after it in the audio of the answer to the a-th question is also the relative difference between the average energy of the r-th speech window and the average energy of the two speech windows before and after it. It characterizes the abrupt change or discontinuity of the audio energy locally. The larger the value, the more likely the volume of the r-th speech window is to have suddenly increased or decreased, and the less smooth the change is. The mean value characterizes the average degree of non-stationarity of the volume change of the audio of the answer to the a-th question throughout the entire speech process. The larger the value, the less smooth the volume changes across all speaking windows in the audio of the answer to question a. The larger and When it is larger, The larger, and The larger the value, the more drastic and unstable the energy fluctuations during the speech segment in the audio response to question a, and the more discontinuous or uneven the changes in the audio energy within the speech window. The smaller the value, the more stable and less volatile the energy during the speech segment in the audio of the answer to question a, and the more continuous and smooth the change. Additionally, there might be a situation where the total number of speech windows in the audio of the answer to question a is no more than 2. If this occurs, to ensure the above calculation formula is effective and that the energy change characteristic value takes into account the standard deviation of the speech audio energy, this embodiment does not calculate the second energy change characterization value when the total number of speech windows in the audio of the answer to question a is no more than 2. Instead, it directly assigns a very small positive value to the second energy change characterization value. For example, when the total number of speech windows in the audio of the answer to question a is no more than 2, this embodiment directly records the second energy change characterization value as 0.01, meaning it is not calculated. And replace this item with 0.01.

[0034] Since users' pauses vary depending on their emotional state—for example, pauses tend to increase and lengthen when a user is depressed, while pauses tend to decrease significantly when a user is anxious or tense—this embodiment will calculate the pause feature value for question a based on all pause windows in the audio of the answer to question a. The specific calculation process for the pause feature value of question a is as follows:

[0035] All pause windows that are consecutive in time on the audio of the answer to the a-th question are merged, and the merged result is recorded as a pause segment. If it is determined that the first c pause windows on the audio of the answer to the a-th question are consecutive in time, but the c-th pause window and the (c+1)-th pause window are not consecutive in time, then the merged result of the first c pause windows is taken as a pause segment. If it is determined that the (c+1)-th pause window and the (c+2)-th pause window are not consecutive in time, then the (c+1)-th pause window is taken as a pause segment, and the determination of whether the (c+2)-th pause window and the (c+3)-th pause window are consecutive in time is continued until all pause windows on the audio of the answer to the a-th question have been traversed, and all the obtained pause segments are counted. The absence of a speech window between pause windows indicates consecutive time. In this embodiment, one pause segment on the audio of the answer to the a-th question represents one pause, that is, the audio of the answer to the a-th question... The total number of pauses in the audio is the number of pauses user A makes when answering question a. The ratio of the number of pauses for question a to the historical maximum number of pauses is denoted as the relative number of pauses for question a. Each question answered by a user in any sample is considered a historical question, and the historical maximum number of pauses is the maximum number of pauses among all historical questions. The process for obtaining the number of pauses for any historical question is the same as that for question a. The percentage of the total duration of pauses in the audio of the answer to question a is calculated. This percentage is the ratio of the total duration of pauses in the audio of the answer to question a to the total duration of the audio of the answer to question a. The total duration of pauses in the audio of the answer to question a is the sum of the durations of all pauses in the audio of the answer to question a. The linearly normalized result of multiplying the percentage of the total duration of pauses in the audio of the answer to question a by the relative number of pauses for question a is denoted as the pause feature value for question a. The expression for the pause feature value for question a is:

[0036]

[0037] in, Let be the pause characteristic value for the a-th problem. The total audio duration of the answer to the a-th question. The total duration of pauses in the audio recording of the answer to question a is given. This represents the number of pauses user A made while answering the a-th question. This represents the maximum number of pauses on historical issues. The percentage of the pause time in answering the a-th question. The larger the value, the longer the pause. Characterizing the relative pause frequency, The larger the value, the closer the number of pauses in the audio of the response is to the highest value among all historical response audios, indicating that the speech is less fluent and coherent. The smaller the value, the fewer the pauses, and the smoother the speech. and The larger, The larger, and A larger value indicates a longer audio pause time, more pauses, or more frequent pauses in the answer to question a; conversely, a smaller value indicates a shorter pause time. The smaller the value, the shorter the audio pause time and the fewer the number of pauses in the answer to question a.

[0038] Therefore, this embodiment obtains the speech feature values ​​of the user's answer to each question through the above process, including speech rate feature values, energy change feature values, and pause feature values.

[0039] Step S002: Based on the emotional abnormality index of the text content of the historical users' answers to the questions, the voice feature value and the average heart rate, calculate the probability of abnormal emotions for each historical user and construct a model of different voice feature patterns corresponding to different abnormal emotions, including depression, anxiety and stress.

[0040] Since depression, anxiety, and stress do not manifest in identical voice features, this embodiment will analyze the different voice feature changes for each abnormal emotion separately. Furthermore, because voice features during user responses are influenced not only by emotions but also by the user's physical state—for example, voice features in a fatigued state may resemble those of depression, while voice features in a tense state may resemble those of anxiety—it is necessary to distinguish between genuine abnormal emotional expressions and normal expressions influenced by the user's state when analyzing the voice feature changes for each abnormal emotion. Moreover, the subsequent analysis in this embodiment primarily relies on the emotional abnormality index and voice features of the text content of historical user responses. This embodiment analyzes the changing patterns of different speech features for different abnormal emotions using the emotional abnormality index, speech feature value, and average heart rate of the text content of the sample historical users' answers to questions. Therefore, after obtaining the emotional abnormality index, speech feature value, and average heart rate of the text content of the sample historical users' answers to questions, it calculates the probability of abnormal emotions for each sample historical user based on the emotional abnormality index, speech feature value, and average heart rate, and constructs different speech feature pattern models corresponding to different abnormal emotions. That is, abnormal emotions include depression, anxiety, and stress, and the speech feature pattern model corresponding to any abnormal emotion includes a speech rate feature pattern model, an energy change feature pattern model, and a pause feature pattern model. The specific process of the different speech feature pattern models corresponding to different abnormal emotions is as follows:

[0041] To extract more accurate patterns of speech feature changes, the reliability of historical data used for model construction needs to be considered when building the model. The reliability of historical data is determined by comparing the emotional abnormality index of the text content of historical users' answers to questions with the probability that the corresponding user is experiencing different depressive moods. This serves as the fitting weight during model construction, effectively reducing the contamination of the model's patterns by data whose user text does not match their true emotions or is affected by temporary states. This allows for the extraction of speech feature change patterns that better represent true abnormal emotions. Based on the above description, this embodiment needs to calculate the probability of depression, anxiety, and high stress among historical users. In this embodiment, the probability of depression, anxiety, and high stress is mainly measured based on the average heart rate of users answering questions. Therefore, this embodiment will next calculate the probability of depression, anxiety, and high stress for each user based on the average heart rate of historical users' question-and-answer sessions. Since the calculation method for the probability of depression, anxiety, and high stress is the same for each user, for ease of understanding and description, this embodiment will use the calculation process of user A's probability of depression, anxiety, and high stress as an example. The specific calculation process for user A's probability of depression, anxiety, and high stress is as follows:

[0042] 1) The specific calculation process for the probability of user A's depression:

[0043] Because users in a depressed mood remain in a low-activity state, resulting in very low heart rate fluctuations throughout the conversation, it is necessary to first measure the user's overall heart rate fluctuation during the entire conversation. Specifically, based on questions from a psychological scale, the average absolute difference in heart rate between adjacent answers to user A's questions is recorded as the user A's overall heart rate fluctuation. The expression for the user A's overall heart rate fluctuation is:

[0044]

[0045] in, J0 represents the overall heart rate fluctuation level of user A, and J0 represents the total number of questions answered by user A, which is also the total number of questions on the psychological scale. In this embodiment, J0 is 21. and These are the average heart rates of user A when answering the j-th and j+1-th questions, respectively; the j-th and j+1-th questions are consecutive in the time it takes for user A to answer them. The magnitude of heart rate change for user A between answering two different questions was quantified. The larger the value, the more frequently or the greater the fluctuation in user A's heart rate throughout the conversation; This characterizes the overall heart rate fluctuation level of user A throughout the entire conversation. The larger the mean, The larger, and A higher heart rate indicates a greater overall fluctuation in user A's heart rate throughout the entire conversation, and vice versa. A smaller value indicates that user A experienced less overall heart rate fluctuation throughout the entire conversation. Furthermore, because the DASS-21 scale interweaves questions addressing different aspects, even users in a normal state will experience some heart rate fluctuations when faced with relatively sensitive questions. Users with depressive symptoms, however, exhibit decreased responsiveness to sensitive questions, resulting in smaller heart rate changes. Therefore, the smaller the overall heart rate fluctuation, the more likely they are to be experiencing depressive symptoms. The smaller the value, the greater the likelihood that user A is in a depressive mood; conversely, the smaller the value, the less likely user A is to be in a depressive mood.

[0046] Furthermore, a user's physical condition can also affect their emotional expression during the response process. For example, a user who is fatigued or sleep-deprived may exhibit low responsiveness similar to depression. However, the low responsiveness in a true depressive state is continuous, while the performance of a normal user is usually gradual. For instance, in a fatigued state, the user's speech characteristics may be relatively normal in the early stages, but gradually decrease in speech speed, energy level, and pauses in the later stages. Therefore, in order to reduce misidentification, it is necessary to analyze the trend of the user's speech speed characteristics during the entire dialogue. In other words, in order to ensure the accuracy of the analysis of the possibility of depression, this embodiment also needs to combine the trend of user A's speech speed characteristics during the entire dialogue to analyze the possibility of user A's normal emotions, and combine the analyzed possibility of user A's normal emotions with the degree of fluctuation of user A's overall heart rate to quantify the possibility of user A's depression. In other words, this embodiment first needs to obtain the probability of user A's emotional normality in speech rate performance based on the fitting residuals and fitting slope of user A's speech rate feature values ​​in answering questions; then, based on the fitting residuals and fitting slope of user A's energy feature values ​​in answering questions, it needs to obtain the probability of user A's emotional normality in energy change performance; and finally, based on the fitting residuals and fitting slope of user A's pause feature values ​​in answering questions, it needs to obtain the probability of user A's emotional normality in pause performance. Then, it calculates the mean of the probabilities of user A's emotional normality in speech rate performance, energy change performance, and pause performance, and records this as user A's emotional normality probability. , The smaller the time, the higher the likelihood that user A is experiencing abnormal emotions. The lower the likelihood of an emotion, and the more likely the expression is... , Assess the likelihood that User A's speech rate reflects normal emotional state. To determine the likelihood of user A exhibiting normal emotional responses in relation to energy changes. This involves assessing the likelihood of user A's emotions being normal based on their pauses. Since distinguishing between normal and abnormal emotions typically requires a combination of various speech features, a single feature may not be accurate. For example, some users may naturally speak slowly with minimal changes in pace, appearing continuous, but their speech energy or pauses show a trend, indicating that the user is not continuously experiencing an abnormal emotion. Therefore, this embodiment uses a comprehensive approach, based on user A's speech rate, energy changes, and pauses, to comprehensively reflect the likelihood of user A's emotions being normal, ensuring greater reliability for subsequent analyses of the likelihood of abnormal emotions.

[0047] Since the method for obtaining the probability of normal emotion in user A's different speech features is the same, for ease of understanding and description, this embodiment will take the specific process of obtaining the probability of normal emotion in user A's speech rate as an example. That is, the specific process of obtaining the probability of normal emotion in user A's speech rate is as follows:

[0048] A time-speech rate feature value coordinate system is constructed, with the horizontal axis representing time and the vertical axis representing speech rate feature values. The time when user A starts answering each question and the speech rate feature value of each answer are mapped to the time-speech rate feature value coordinate system to obtain the speech rate feature value point of each question answered by user A. That is, the horizontal and vertical coordinates of the speech rate feature value point of any question answered by user A are the start time when answering the question and the speech rate feature value of the question. The least squares method is used to perform linear fitting on the speech rate feature value points of all questions answered by user A. The straight line obtained by fitting is denoted as the speech rate feature value fitting line corresponding to user A. The larger the absolute value of the fitting slope, the more obvious the trend of the user's speech rate change.

[0049] Since a better fit, meaning the mean absolute value of the fitted residuals is closer to 0, indicates extremely high statistical significance and reliability of the fitted slope, the trend dimension has greater information value than the residual dimension. Therefore, the trend dimension should be prioritized or its weight increased to accurately identify and distinguish pseudo-anomalies caused by temporary physical conditions such as fatigue. The trend dimension refers to the fitted slope, and the residual dimension refers to the mean absolute value of the fitted residuals. Conversely, a worse fit indicates lower reference value for the fitted slope, and the residual dimension has greater value than the trend dimension. Therefore, the residual dimension should be prioritized to capture abnormal oscillations or repeated fluctuations, identifying persistent random deviations caused by emotional abnormalities. Therefore, this embodiment next calculates the characterization value of user A's speech rate fitting effect based on the fitting residuals obtained from the fitted line corresponding to user A's speech rate, and records this as the influence weight of user A's speech rate fitting effect. ;and The expression is , Let J be the speech rate feature value of user A's answer to the j-th question. This is the fitted value of the speech rate feature for user A's answer to the j-th question. The fitted value of the speech rate feature for the j-th question is the ordinate value of the point on the straight line corresponding to user A's speech rate feature value, with the x-axis being the start time when answering the j-th question. Alternatively, it can be the output result of the straight line function expression for user A's speech rate feature value, with the start time being the input. The smaller the mean absolute value of the residuals for all questions answered by user A, the better the fit; the larger the mean absolute value of the residuals for all questions answered by user A, the worse the fit. When it is larger, The larger, and The larger the value, the worse the fit, and vice versa. The smaller the value, the better the fit. Additionally, to avoid... Whether the value is too large or too small, the likelihood of a normal subsequent emotional outcome will be dominated by one aspect. This embodiment limits this. Within a certain scope, the scope of restriction can be set by the implementer according to the actual situation, such as in this embodiment. The maximum and minimum values ​​are 0.7 and 0.3 respectively, that is... When the value is greater than 0.7, Take 0.7 directly. When the value is less than 0.3, Take 0.3 directly.

[0050] After obtaining the influence weights of user A's speech rate fitting effect, the expression for the probability of user A's normal emotional expression in speech rate performance, based on the influence weights of user A's speech rate fitting effect, is as follows:

[0051]

[0052] in, The norm() linear normalization function is used to determine the probability that user A's speech rate reflects normal emotions. The weights for the fitting effect on user A's speech rate are defined, and J0 represents the total number of questions answered by user A. Let J be the speech rate feature value of user A's answer to the j-th question. K0 is the fitted value of the speech rate feature for the j-th question answered by user A, and K0 is the slope of the fitted line for the speech rate feature value corresponding to user A.

[0053] This indicates the weight of the fitting residual dimension or non-trend when analyzing the probability that the emotions reflected by user speech rate characteristics are normal. The larger the value, the higher the fitted residual dimension when judging whether the user's emotions are normal. The greater the role played, The smaller the value, the more it indicates the trend dimension, which is the slope of the fitted line. The greater the role it plays. The absolute value of the average residual of the fitting of user A’s speech rate feature reflects the degree of fluctuation or concentration of the speech rate feature values ​​of all of user A’s answers around the fitting line. The larger the value of this item, the higher the degree of deviation of the actual speech rate feature value points from the fitting line. The smaller the value of this item, the more evenly the actual data points are distributed around the fitting line, or the residual fluctuates around 0. This measure assesses the strength of the trend or the rate of change of user A's speech rate characteristics over time. A larger value indicates that user A's speech rate characteristics show a significant upward or downward trend over time, and the change has a strong trend. A significant trend may indicate that the speech rate change is caused by the user's temporary state (such as fatigue or tension) rather than a continuous abnormal emotion. That is, a smaller value indicates that user A's speech rate characteristics have almost no linear trend over time, are relatively stable, or fluctuate randomly around a certain value.

[0054] Because if a user is in a sustained abnormal emotional state, their speech rate tends to show a continuous trend rather than a gradual increase or decrease. Specifically, this manifests as a smaller absolute value of the fitted slope, with the residuals fluctuating around 0 or the absolute value of the residual mean close to 0. In other words, the smaller the absolute value of the slope K0 and the closer the absolute value of the residual mean is to 0, the better. It is again a weighted sum of the slope K0 term and the absolute value of the residual mean, therefore The smaller the value of the slope K0, the lower the probability that user A is in a normal emotional state or the higher the probability of a persistent abnormal emotional state. Conversely, the larger the absolute value of the slope K0 (indicating a clear upward or downward trend in speech rate) or the stronger the trend in speech rate characteristics over time, and the more regular or larger the residual average, the more likely this behavior is caused by the user's temporary state (such as fatigue or tension). In other words, the user is more likely to be in a normal emotional state rather than a persistent abnormal emotional state. The larger the value, the higher the probability that user A is in a normal emotional state or the lower the probability that user A is in a persistent abnormal emotional state. Additionally, it should be noted that when obtaining the probability of user A's emotional normality based on energy changes and the probability of emotional normality based on pauses, simply replace the speech rate feature value with the energy change feature value and the pause feature value.

[0055] Next, based on the probability of user A's emotions being normal and the degree of fluctuation in user A's overall heart rate, the probability of user A's depression is calculated; and the probability of user A's depression... The expression is:

[0056]

[0057] in, To assess the likelihood of user A's depression, For user A, the overall heart rate fluctuation level, The probability that user A's emotions are normal; due to The smaller the value, the higher the likelihood that user A is in a depressive state. The smaller the value, the higher the likelihood that user A is experiencing abnormal emotions. smaller and The smaller, The larger, therefore The larger the value, the higher the likelihood that user A is in a depressive state, and vice versa. The smaller the value, the lower the likelihood that user A is in a depressive state.

[0058] 2) The specific calculation process for the probability of user A's anxiety:

[0059] Anxiety is often manifested as significant heart rate fluctuations; that is, anxiety can manifest as large heart rate fluctuations across multiple issues, and recovery is relatively difficult. Therefore, when analyzing the likelihood of anxiety, it is necessary to first measure the difference in heart rate fluctuations when user A answers anxiety questions compared to when answering other questions. Specifically, in this embodiment, we will first calculate the mean absolute difference between the average heart rate of user A answering each anxiety question and the previous question corresponding to that anxiety question, and record this as the heart rate fluctuation range when user A answers anxiety questions. Anxiety questions refer to those belonging to the anxiety dimension of the scale; the same applies to stress and depression questions. The degree of heart rate fluctuation when user A answers anxiety questions... The expression is:

[0060]

[0061] in, The heart rate fluctuation level of user A when answering anxiety questions is represented by J1, which is the number of anxiety-related questions answered by user A. In this embodiment, J1 is 7. Let $\frac{ ... This is the absolute value of the difference between the average heart rate of user A when answering the first anxiety question and the heart rate of user A before answering the first question. The heart rate of user A before answering the first question is the heart rate at the moment of heart rate sampling before user A started answering the first question. The heart rate sampling frequency also needs to be set by the implementer according to the actual situation or according to the performance of the acquisition equipment. A higher value indicates a greater degree of heart rate fluctuation when user A answers anxiety questions, while... The arithmetic mean is ,therefore The larger the value, the greater the fluctuation in user A's heart rate when answering anxiety questions. The smaller the value, the smaller the heart rate fluctuation when user A answers anxiety questions.

[0062] Then, user A's anxiety level was measured in terms of heart rate fluctuations. , The expression is:

[0063]

[0064] in, Assess user A's level of anxiety regarding heart rate fluctuations. For user A, the overall heart rate fluctuation level, The degree of heart rate fluctuation when user A answers anxiety questions. This is a preset constant to prevent the denominator from being 0. In specific applications, the implementer can set the preset constant according to the actual situation. For example, the preset constant in this embodiment can be set to 0.01. The absolute difference between overall heart rate variability and heart rate variability related to anxiety reflects the difference between a user's response to anxiety questions and their overall heart rate variability. A smaller value indicates a smaller difference in heart rate fluctuations and overall condition when users experience anxiety-related issues. The higher the value, the greater the fluctuation in the user's heart rate throughout the conversation. Since anxiety often manifests as significant heart rate fluctuations across multiple issues, an anxious user not only experiences large heart rate fluctuations when facing specific anxiety-inducing questions, but also struggles to return to a calm state when answering other non-anxiety-related questions. This results in a consistently high heart rate throughout the conversation. smaller and The larger the heart rate, the more consistent the heart rate characteristics of user A with anxiety, or the more physiological characteristics support the existence of persistent anxiety, indicating a higher level of anxiety or a greater likelihood of anxiety. smaller and When it is larger, The larger, therefore The larger the value, the more consistent the user A's emotional expression is with abnormal anxiety or the higher the level of anxiety when answering anxiety questions. It also indicates that the user A is more likely to be in an anxious state. Conversely, the smaller the value, the lower the value. The smaller the value, the less likely user A's emotions are to be consistent with abnormal anxiety characteristics when answering anxiety questions, or the less likely user A is to be in an anxious emotional state.

[0065] Then, based on the probability of user A's emotions being normal and the degree of anxiety in user A's heart rate fluctuations, the probability of user A's anxiety was calculated; and the probability of user A's depression was also calculated. The expression is:

[0066]

[0067] in, The possibility of user A's anxiety; due to The larger the value, the higher the user A's anxiety level. The smaller the value, the higher the likelihood that user A is experiencing abnormal emotions. The larger and The smaller, And the larger it is, therefore The larger the value, the higher the likelihood that user A is in an anxious state, and vice versa. The smaller the value, the lower the likelihood that user A is in an anxious emotional state.

[0068] 3) The specific calculation process for the likelihood of user A experiencing high stress:

[0069] Abnormal stress often manifests as significant heart rate fluctuations. Specifically, high stress typically only leads to larger heart rate fluctuations when answering targeted questions, while the heart rate quickly returns to normal when answering other questions. Therefore, when analyzing the likelihood of high stress, it's necessary to first measure the difference in heart rate fluctuations between user A's answers to stressful questions and their answers to other questions. In this embodiment, the mean absolute difference between user A's average heart rate when answering each stressful question and the previous question in the corresponding stressful question is calculated and recorded as the heart rate fluctuation range when user A answers stressful questions. The expression is:

[0070]

[0071] in, J1 represents the degree of heart rate fluctuation when user A answers stress questions, and J2 represents the number of stress questions user A answers. In this embodiment, J2 is 7. Let $\frac{f}{f}$ be the absolute value of the difference between the average heart rate of user A when answering the $f$-th stress question and the average heart rate of user A when answering the previous stress question. If the first question answered by user A is the first stress question answered by user A, then $\frac{f}{f}$ is the absolute value of the difference between the average heart rate of user A when answering the $f$-th stress question and the average heart rate of user A when answering the previous stress question. This is the absolute value of the difference between the average heart rate of user A when answering the first stress question and the heart rate of user A before answering the first question. Here, the heart rate of user A before answering the first question is also the heart rate at the heart rate sampling moment before user A started answering the first question. A higher value indicates greater fluctuations in user A's heart rate when answering stress questions, while... The arithmetic mean is ,therefore The larger the value, the greater the fluctuation in user A's heart rate when answering questions about overall stress. The smaller the value, the smaller the heart rate fluctuation when user A answers the overall stress question.

[0072] Then, the stress level of user A in terms of heart rate fluctuation is measured from the perspective of heart rate fluctuation. User A's stress level regarding heart rate fluctuations The expression is:

[0073]

[0074] in, For user A, the level of stress related to heart rate fluctuations, For user A, the overall heart rate fluctuation level, The degree of heart rate fluctuation when user A answers stress questions. This indicates the degree of heart rate fluctuation when a user answers stress-related questions; the higher the value, the stronger the heart rate fluctuation. Rather than the overall heart rate fluctuations throughout the conversation The greater the difference, the more likely it is that a stressed user's heart rate will fluctuate significantly only when answering targeted questions, and will recover to normal levels more quickly when answering other questions. This means that a stressed user's intense heart rate fluctuations are only triggered when facing direct stressful questions, and recover more quickly when facing other questions. It will be significantly higher than Therefore The larger the value, the more closely user A matches the heart rate characteristics under stress, or the greater the likelihood that user A is experiencing stressful emotions; although stress responses are specific, a larger value... This still indicates that the user's overall heart rate fluctuation level was high during the conversation, possibly due to multiple stressors, incomplete recovery, or a strong stress response itself. It can characterize the overall intensity of the pressure, that is The larger the heart rate, the more closely it indicates that user A's heart rate characteristics match those of someone experiencing high stress, or it may also reflect a greater likelihood that user A is in a state of high stress. The larger and When it is larger, The larger, therefore The larger the value, the more consistent the user A's emotions are with abnormal stress characteristics when answering stress-related questions, or the greater the likelihood that user A is experiencing high stress. Conversely, the smaller the value, the more consistent the user A's emotions are with abnormal stress characteristics when answering stress-related questions. The smaller the value, the less consistent the user A's emotions are with the abnormal characteristics of high stress when answering stress-related questions, or the greater the likelihood that user A is experiencing high stress.

[0075] Then, based on the probability of user A's emotions being normal and the degree of stress in user A's heart rate fluctuations, the probability of user A being depressed was calculated; and the probability of user A being under great stress was also calculated. The expression is:

[0076]

[0077] in, The pressure on user A is likely high; due to The larger the value, the greater the likelihood that user A is experiencing high stress. The smaller the value, the higher the likelihood that user A is experiencing abnormal emotions. The larger and The smaller, The larger, therefore The larger the value, the greater the likelihood that user A is experiencing high stress, and vice versa. The smaller the value, the lower the likelihood that user A is in a state of high stress.

[0078] This embodiment, after obtaining the probability of a user being depressed, anxious, or stressed, quantifies the credibility of the user's answers to each question based on a comparison between the probability of depression, anxiety, and stress among historical users and the emotional abnormality index of the text content of each historical user's answers to the corresponding questions. Since the methods for obtaining the credibility of depression, anxiety, and stress in the answers to depression questions, anxiety questions, and stress questions are the same for each historical user's answers, this embodiment uses the process of obtaining the credibility of depression in the b1st depression question, the credibility of anxiety in the b2nd anxiety question, and the credibility of stress in the b3rd stress question from any historical user B as an example for ease of understanding and description. That is, the expressions for the credibility of depression in the b1st depression question, the credibility of anxiety in the b2nd anxiety question, and the credibility of stress in the b3rd stress question from historical user B are:

[0079]

[0080]

[0081]

[0082] in, The confidence level of depression in the b1th depression question answered by historical user B in the sample. The credibility of anxiety in the b2nd anxiety question answered by historical user B in the sample. The credibility of the stress response to the b3rd stress question by historical user B in the sample is high. The likelihood of depression for historical user B in the sample. The likelihood of anxiety for historical user B in the sample. The sample of historical user B is likely under great pressure. The emotional abnormality index of the text content of the b1st depression question answered by historical user B in the sample is used. The emotional abnormality index of the text content of the b2th anxiety question answered by the sample user B is used. The emotional abnormality index of the text content of the b3rd stress question answered by the sample user B is used. This is a preset constant to prevent the denominator from being 0. Measuring the likelihood of depression based on a user's voice characteristics Sentiment Abnormality Index of the Text Content of the Answer Questions The higher the consistency between them, the better. The smaller the value, the closer and more consistent the emotional state reflected in the speech and text content. Therefore, the more reliable the data points related to the b1th depression question are when analyzing emotional characteristic patterns, and consequently, the greater the weight of these data points in subsequent pattern fitting. Conversely, the smaller the value, the lower the consistency. The larger the value, the less closely and inconsistent the emotional state reflected in the speech and text content are. Therefore, the data points related to the b1th depression question are less reliable for analyzing emotional characteristic patterns, and consequently, their weight should be smaller when participating in subsequent pattern fitting. The smaller, The larger the value, the more reliable the data points related to the b1th depression problem are when used to analyze the patterns of emotional characteristics. Consequently, the data points related to the b1th depression problem also have a greater weight when participating in subsequent pattern fitting. and Similarly.

[0083] After obtaining the credibility of depression, anxiety, and stress in each question answered by each historical user sample, the credibility of depression in each depression question answered by each historical user sample was used as the fitting weight. The weighted least squares method was used to fit the text content emotion abnormality index-speech rate feature points of each depression question answered by all historical users sample, and the fitted straight line was recorded as the speech rate law model corresponding to depression. The weighted least squares method was used to fit the text content emotion abnormality index-energy change feature points of each depression question answered by all historical users sample, and the fitted straight line was recorded as the energy change law model corresponding to depression. The weighted least squares method was used to fit the text content emotion abnormality index-pause feature points of each depression question answered by all historical users sample, and the fitted straight line was recorded as the pause law model corresponding to depression. Using the anxiety credibility of each anxiety question answered by historical users in each sample as the fitting weight, the weighted least squares method is used to fit the text content emotion abnormality index-speech rate feature points of all anxiety questions answered by historical users in all samples. The fitted straight line is denoted as the speech rate law model corresponding to anxiety. The weighted least squares method is used to fit the text content emotion abnormality index-energy change feature points of each anxiety question answered by historical users in all samples. The fitted straight line is denoted as the energy change law model corresponding to anxiety. The weighted least squares method is used to fit the text content emotion abnormality index-pause feature points of each anxiety question answered by historical users in all samples. The fitted straight line is denoted as the pause law model corresponding to anxiety. Using the credibility of each stress question answered by historical users in each sample as the fitting weight, the weighted least squares method is used to fit the text content sentiment abnormality index-speech rate feature points of all stress questions answered by historical users in all samples. The fitted line is denoted as the speech rate law model corresponding to stress. The weighted least squares method is used to fit the text content sentiment abnormality index-energy change feature points of each stress question answered by historical users in all samples. The fitted line is denoted as the energy change law model corresponding to stress. The weighted least squares method is used to fit the text content sentiment abnormality index-pause feature points of each stress question answered by historical users in all samples. The fitted line is denoted as the pause law model corresponding to stress.Furthermore, when fitting the text content sentiment abnormality index-speech rate feature value points for each depression question answered by all historical users in all samples using the weighted least squares method, the weight of the text content sentiment abnormality index-speech rate feature value point for any depression question is the depression credibility of that depression question. The weights of other points involved in the fitting are similarly determined. Moreover, since the points involved in the fitting and their weights are known, the process of performing the weighted least squares fitting is a well-known technique. The text content sentiment abnormality index-speech rate feature value points are located in the text content sentiment abnormality index-speech rate feature value coordinate system. The horizontal axis of the coordinate system represents the text content emotion abnormality index, and the vertical axis represents the speech rate characteristic value. The text content emotion abnormality index-energy change characteristic value point is located in the text content emotion abnormality index-energy change characteristic value coordinate system, where the horizontal axis represents the text content emotion abnormality index, and the vertical axis represents the energy change characteristic value. The text content emotion abnormality index-pause characteristic value point is located in the text content emotion abnormality index-pause characteristic value coordinate system, where the horizontal axis represents the text content emotion abnormality index, and the vertical axis represents the pause characteristic value.

[0084] Therefore, this embodiment obtains different speech feature patterns corresponding to different abnormal emotions through the above process, namely, the speech rate pattern, energy change pattern, and pause pattern corresponding to depression, anxiety, and stress.

[0085] Step S003: Based on the voice feature values ​​of the current user's answer to the question, the voice feature pattern model is invoked to output the voice emotion abnormality index of the current user's answer to the question on different voice features. The voice emotion abnormality index is then corrected based on the probability of the current user's abnormal emotion and compared with the emotion abnormality index of the text content of the current user's answer to the question to obtain the degree of consistency between the emotion of the text content and the voice of the current user's answer to the question. Based on the degree of emotional consistency, the emotion abnormality index of the text content of the current user's answer to the question on the same abnormal emotion dimension is weighted and fused to obtain the fusion abnormality score of the current user on different abnormal emotions. Based on the fusion abnormality score, the emotional health assessment result of the current user is obtained.

[0086] This embodiment constructs different speech feature pattern models corresponding to different abnormal emotions. It then uses these models to obtain the speech emotion abnormality index for each question answered by the current user, based on different speech features. The speech emotion abnormality index is primarily used for subsequent analysis of the consistency between the text content and the emotional tone of the speech. For any depression question answered by the current user, the speech rate feature value of the depression question is input into the speech rate pattern model corresponding to depression, and the model output is recorded as the speech emotion abnormality index of the depression question in terms of speech rate. Similarly, the energy change feature value of the depression question is input into the energy change pattern model corresponding to depression, and the model output is recorded as the speech emotion abnormality index of the depression question in terms of energy change. Finally, the pause feature value of the depression question is input into the pause pattern model corresponding to depression, and the model... The output is denoted as the speech emotion abnormality index at pauses for the depression problem. The speech rate feature value v0 of any depression problem is input into the speech rate pattern model corresponding to depression. The model output is either the speech rate feature value or the x-coordinate of the point whose ordinate is the speech rate feature value v0, or the output of the function expression after the speech rate feature value v0 belongs to the speech rate pattern model corresponding to depression. The method for determining the output of other feature values ​​input into the corresponding pattern model for other types of problems is the same as the method for determining the output of the speech rate feature value input into the speech rate pattern model corresponding to depression. Furthermore, the process of obtaining the speech emotion abnormality index for anxiety or stress problems on different speech features is the same as the process of obtaining the speech emotion abnormality index for depression problems on different speech features.

[0087] This embodiment, after obtaining the speech emotion abnormality index of each question answered by the current user on different speech features, corrects the speech emotion abnormality index of the current user's answers to questions with the same abnormal emotion on different speech features based on the probability of the current user's abnormal emotion, and then compares and analyzes the result with the emotion abnormality index of the text content of the current user's answer to obtain the degree of consistency between the emotion of the current user's answer text content and speech; the specific acquisition process is as follows:

[0088] For the z1th depression question answered by the current user: the average of the current user's depression probability and the voice emotion abnormality index of the z1th depression question across various voice features is denoted as the corrected voice emotion abnormality index of the z1th depression question on the corresponding voice feature. Specifically, the average of the current user's depression probability and the voice emotion abnormality index of the z1th depression question on speech rate is the corrected voice emotion abnormality index of the z1th depression question on speech rate; the average of the current user's depression probability and the voice emotion abnormality index of the z1th depression question on energy change is the corrected voice emotion abnormality index of the z1th depression question on energy change; and the average of the current user's depression probability and the voice emotion abnormality index of the z1th depression question on pauses is the corrected voice emotion abnormality index of the z1th depression question on pauses. This correction operation is essentially an averaging operation, where the current user's voice feature values ​​(such as speech rate, energy rate, and pauses) are used to calculate the corrected voice emotion abnormality index of the z1th depression question on pauses. The emotional abnormality index obtained by substituting speed, energy change, and pauses into the speech feature law model is combined with the abnormality probability calculated based on the user's overall behavior. This utilizes the information of speech features and integrates the user's overall behavior, thereby reducing the bias of a single data source and improving the accuracy of identifying hidden emotions in adolescents. The absolute difference between the corrected speech emotional abnormality index of the z1th depression question on each speech feature and the text content emotional abnormality index of the z1th depression question is accumulated and then inversely mapped to the result of linear normalization. This result is recorded as the degree of consistency between the text content and speech emotion of the current user's answer to the z1th depression question. The inverse mapping here is to take the reciprocal. To avoid the denominator being 0, a constant needs to be added to the denominator during the inverse mapping. That is, the accumulated result is added to the preset constant and then inversely mapped to the result of linear normalization to obtain the degree of consistency between the text content and speech emotion of the current user's answer to the z1th depression question.

[0089] Furthermore, the expression for the consistency of the emotional content between the text and speech of the current user's answer to the z1th depression question is as follows:

[0090]

[0091] in, The degree of emotional consistency between the text and spoken responses of the current user to the z1th depression question. The emotional abnormality index of the text content of the current user's answer to the z1th depression question. The corrected speech emotion abnormality index is applied to the z1th depression question answered by the current user, based on the speech rate. The corrected voice emotion abnormality index is calculated based on the energy change of the z1th depression question answered by the current user. The corrected speech emotion abnormality index is set for the pauses in the z1th depression question answered by the current user. The preset constant here is also to prevent the denominator from being 0.

[0092] The larger the absolute value of the difference between the text content emotion anomaly index and the corrected speech emotion anomaly index, the greater the inconsistency between the emotional state reflected by the speech features and the emotional state expressed in the dialogue text. This also indicates a greater likelihood that the user is hiding or disguising their true emotions in the dialogue. Conversely, the smaller the absolute value of the difference between the text content emotion anomaly index and the corrected speech emotion anomaly index, the greater the consistency between the emotional state reflected by the speech features and the emotional state expressed in the dialogue text. This also indicates a greater consistency between the user's verbal expression and the emotional signals conveyed by their speech features, with no obvious hidden or conflicting behaviors. To reduce the interference of hidden emotions on the evaluation results, in subsequent fusion, the text content emotion anomaly index with a high degree of consistency should be given more importance or higher participation, while the text content emotion anomaly index with a lower degree of consistency should have lower participation. The larger, therefore A larger value indicates that when subsequently obtaining the current user's fusion anomaly score for depression or fusing the text content sentiment anomaly indices of all depression questions answered by the current user, the contribution of the text content sentiment anomaly index of the z1th depression question is greater, and vice versa. The smaller the value, the less contribution the emotional abnormality index of the text content of the z1st depression question makes when subsequently obtaining the user's fusion abnormality score for depression. Similarly, the corrected voice emotional abnormality index for the z2th anxiety question and the z3th stress question can be obtained based on speech rate, energy changes, and pauses. Following the method used to obtain the emotional consistency between the text content and speech of the user's z1st depression question, the emotional consistency between the text content and speech of each anxiety question and each stress question can be obtained.

[0093] After obtaining the degree of emotional consistency between the text content and speech of each question answered by the current user, the emotional anomalousness index of the text content of the current user's answers to questions with the same abnormal emotional dimension is weighted and fused based on the degree of emotional consistency, and the fusion anomalous score of the current user on different abnormal emotions is calculated; the specific process of obtaining the fusion anomalous score of the previous user on different abnormal emotions is as follows:

[0094] The normalized ratio of the emotional consistency between the text content and speech of all the depression questions answered by the current user is denoted as the text-speech consistency weight for the corresponding depression question. The weighted sum of the text-speech consistency weights of each depression question answered by the current user and the corresponding emotional aberration index of the text content is denoted as the current user's fusion aberration score for depression. The expression for the current user's fusion aberration score Q1 for depression is: Where S1 is the total number of depression questions answered by the current user, and in this embodiment, S1 is 7. The degree of emotional consistency between the text and speech of the z1th depression question answered by the current user. The emotional abnormality index of the text content of the z1th depression question answered by the current user; the higher the current user's fusion abnormality score for depression, the more severe the current user's depression, and vice versa. The fusion abnormality scores for anxiety and stress are obtained in the same way as those for depression, so they will not be described in detail.

[0095] After obtaining the current user's fusion abnormality scores for depression, anxiety, and stress, the current user's emotional health assessment result is obtained based on the fusion abnormality scores. That is, firstly, each value in the original severity rating table is linearly normalized to obtain the normalized values ​​as shown in Table 1. Then, the abnormal emotion level or severity of the current user is determined based on the score range of the corresponding emotion dimension column in Table 1 for the current user's fusion abnormality scores for depression, anxiety, and stress. Since the original severity rating table is constructed by multiplying the original total score by 2, the maximum score in the original severity rating table is 42 and the minimum score is 0. Therefore, the linear normalization of each value in the original severity rating table is based on the minimum value of 0 and the maximum value of 42. Table 1 is the linearly normalized severity rating table. The columns in Table 1 correspond to the three abnormal emotion dimensions to be assessed. Each row of the table represents the score range of the three abnormal emotion dimensions under the corresponding severity. The symbol / in Table 1 is equivalent to ÷.

[0096] Table 1

[0097]

[0098] Furthermore, the level of the user's fusion abnormal score in depression in the column of Table 1 represents the severity level of the user's depression; the level of the user's fusion abnormal score in anxiety in the column of Table 1 represents the severity level of the user's anxiety; and the level of the user's fusion abnormal score in stress in the column of Table 1 represents the severity level of the user's stress. If the user's fusion abnormal score in depression falls within the range of 14 / 42-20 / 42 in the column of depression, then the user's severity of depression is determined to be moderate.

[0099] Thus, this embodiment completes the intelligent assessment of adolescent emotional health. Furthermore, this embodiment, based on historical data (such as the emotional abnormality index of text content in historical user answers, voice feature values, and average heart rate), learns a model of different voice feature changes under different abnormal emotions (depression, anxiety, and high stress). This model can accurately predict the emotional abnormality index reflected in the current user's voice features. By using the consistency between the emotional abnormality index reflected in the current user's voice features and the emotional abnormality index of the text content as weights to fuse the text emotional index, it can better reflect the user's true emotional state. This improves the accuracy of the assessment of the user's scores in depression, anxiety, and high stress, thereby significantly improving the accuracy of the intelligent assessment of adolescent emotional health.

[0100] In summary, this embodiment first acquires the audio, text, average heart rate, and emotional abnormality index of the user's answers to each question, and then extracts the voice feature values ​​of the user's answers to the corresponding questions based on the audio and text of the user's answers to each question. Next, based on the emotional abnormality index of the text content of the historical user's answers, the voice feature values, and the average heart rate, the probability of abnormal emotions for each historical user sample is calculated, and different voice feature pattern models corresponding to different abnormal emotions are constructed. Then, based on the voice feature values ​​of the current user's answer, the voice feature pattern model is invoked to output the voice emotional abnormality index of the current user's answer on different voice features. The voice emotional abnormality index is then corrected based on the probability of the current user's abnormal emotions and compared with the emotional abnormality index of the text content of the current user's answer to obtain the degree of emotional consistency between the text content and the voice of the current user's answer. Based on the degree of emotional consistency, the emotional abnormality index of the text content of the current user's answers to questions of the same abnormal emotion dimension is weighted and fused to obtain the fused abnormality score of the current user on different abnormal emotions. Furthermore, by analyzing the consistency between text content and speech features and establishing a multimodal fusion mechanism that considers data credibility, this embodiment can significantly improve the accuracy of intelligent assessment of adolescent emotional health, or in other words, effectively overcome the misjudgment problem caused by emotional masquerading, individual differences, and temporary state interference, thereby providing more reliable and accurate emotional health assessment results.

[0101] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An AI-based intelligent assessment method for adolescent emotional health, characterized in that, The method includes the following steps: The process of assessing users' emotional health based on psychological scales involves obtaining the audio and text of the user's answers to each question, the average heart rate, and the emotional abnormality index of the text content. Based on the audio and text of the user's answers to each question, the voice feature values ​​of the user's answers to the corresponding questions are extracted. The voice features include speech rate, energy changes, and pauses. Based on the emotional abnormality index of the text content of the historical users' answers to questions, the voice feature value and the average heart rate, the probability of abnormal emotions of each historical user sample is calculated, and a different voice feature pattern model corresponding to different abnormal emotions is constructed. The abnormal emotions include depression, anxiety and stress. Based on the voice feature values ​​of the current user's answer to the question, the voice feature pattern model is invoked to output the voice emotion abnormality index of the current user's answer to the question on different voice features. The voice emotion abnormality index is then corrected based on the probability of the current user's abnormal emotion and compared with the emotion abnormality index of the text content of the current user's answer to the question to obtain the degree of consistency between the emotion of the text content and the voice of the current user's answer to the question. Based on the degree of emotional consistency, the emotion abnormality index of the text content of the current user's answer to the question on the same abnormal emotion dimension is weighted and fused to obtain the fusion abnormality score of the current user on different abnormal emotions. Based on the fusion abnormality score, the emotional health assessment result of the current user is obtained.

2. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, The sentiment aberration index of the current user's answers to each question is obtained by calling the sentiment aberration model of the answer text content of the current user's answers to the corresponding question. The sentiment aberration index of the text content of the historical user's answers to each question is the normalized score of the corresponding historical user's answers to the corresponding question.

3. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, Methods for obtaining speech feature values ​​include: For any user's answer to the a-th question: the normalized result of comparing the total number of words in the answer text of the a-th question with the total duration of the speech window in the audio of the answer to the a-th question is recorded as the speech rate feature value of the a-th question; the standard deviation of the audio energy of all speech windows in the audio of the answer to the a-th question is recorded as the speech audio energy standard deviation of the a-th question; the mean of the standard deviation of the speech audio energy of all questions answered by all historical users in all samples is recorded as the historical standard deviation mean; the normalized result of comparing the speech audio energy standard deviation of the a-th question with the historical standard deviation mean is recorded as the first energy change characterization value; the normalized result of the mean of the absolute values ​​of the second-order differences of the audio energy between adjacent speech windows in the audio of the answer to the a-th question is recorded as the second energy change characterization value; the product of the first energy change characterization value and the second energy change characterization value is recorded as the energy change feature value of the a-th question; the normalized result of multiplying the proportion of the total duration of pauses in the audio of the answer to the a-th question with the relative number of pauses in the a-th question is recorded as the pause feature value of the a-th question.

4. The AI-based intelligent assessment method for adolescent emotional health as described in claim 3, characterized in that, The relative energy of any audio window is the normalized result of the ratio of the audio energy of the corresponding audio window to the global average energy. The audio energy of any audio window is the mean square energy of the corresponding audio window. The global average energy is the average audio energy of all audio windows on the answer audio to which the audio window belongs. The audio windows on any answer audio are the result of dividing the corresponding answer audio into non-overlapping segments according to a preset time length. Audio windows with relative energy not less than a preset energy threshold are speech windows, and audio windows with relative energy less than the preset energy threshold are pause windows. The number of pauses for the a-th question is the total number of pause segments in the answer audio of the a-th question. The pause segments in the answer audio of the a-th question are the result of merging all time-continuous pause windows in the answer audio of the a-th question. The relative number of pauses for the a-th question is the ratio of the number of pauses for the a-th question to the maximum number of pauses for all historical questions.

5. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, Methods for assessing the likelihood of abnormal emotions include: For any user A: When user A is assessed using psychological scales, the mean absolute difference in heart rate between answers to adjacent questions is recorded as user A's overall heart rate fluctuation. The mean absolute difference in heart rate between answers to each anxiety question and the previous question is recorded as user A's heart rate fluctuation when answering anxiety questions. The mean absolute difference in heart rate between answers to each stress question and the previous question is recorded as user A's heart rate fluctuation when answering stress-related questions. Based on the fitting residuals and slopes of user A's speech rate characteristic values, the probability of user A's emotional normality in speech rate performance is obtained. Based on the fitting residuals and slopes of user A's energy change characteristic values, the probability of user A's emotional normality in energy change performance is obtained. Based on the fitting residuals and slopes of user A's pause characteristic values, the probability of user A's emotional normality in pause performance is obtained. The mean of user A's emotional normality in speech rate performance, emotional normality in energy change performance, and emotional normality in pause performance is recorded as user A's emotional normality probability. Let this be the likelihood of depression for user A. and Let's categorize these as user A's high likelihood of anxiety and stress. , norm() is a linear normalization function. For user A, the overall heart rate fluctuation level, Assuming user A's emotions are normal, For user A's anxiety level regarding heart rate, For user A, the stress level related to heart rate, The degree of heart rate fluctuation when user A answers anxiety-related questions. The degree of heart rate fluctuation when user A answers stress-related questions. The value is a preset constant; the probability of depression, the probability of anxiety, and the probability of high stress all belong to the probability of abnormal emotions.

6. The AI-based intelligent assessment method for adolescent emotional health as described in claim 5, characterized in that, The expression for calculating the probability that User A's speech rate reflects normal emotions is: ； in, The norm() linear normalization function is used to determine the probability that user A's speech rate reflects normal emotions. The weights for the impact of the speech rate fitting effect on user A are defined, and J0 represents the total number of questions answered by user A. Let J be the speech rate feature value of user A's answer to the j-th question. The speech rate feature fitting value for the j-th question answered by user A is given by K0, which is the slope of the speech rate feature fitting line corresponding to user A. The speech rate feature fitting line is obtained by linearly fitting the speech rate feature value points of all questions answered by user A. The x and y coordinates of the speech rate feature value point of any question are the start time when answering the question and the speech rate feature value of the question. The speech rate feature fitting value for the j-th question is the output result of the function expression of the speech rate feature fitting line input to the start time when answering the j-th question. The influence weight of the speech rate fitting effect of user A is the mean absolute difference between the speech rate feature values ​​of each question answered by user A and the corresponding speech rate feature fitting value.

7. The AI-based intelligent assessment method for adolescent emotional health as described in claim 5, characterized in that, Methods for constructing speech rate pattern models corresponding to different abnormal emotions include: Using the credibility of depression questions, anxiety questions, and stress questions from historical users as fitting weights, a weighted least squares method is used to fit the text content emotion abnormality index-speech rate feature points of all historical users' answers to questions of the same emotion dimension, obtaining speech rate pattern models corresponding to depression, anxiety, and stress, respectively. The x and y coordinates of the text content emotion abnormality index-speech rate feature points for any question represent the corresponding text content emotion abnormality index and speech rate feature value. The credibility of depression in the b1st depression question answered by any historical user B is... The credibility of user B's answer to the b2nd anxiety question is: User B's answer to the b3rd stress question has a high credibility of [value missing]. , , and The samples represent the high likelihood of historical user B experiencing depression, anxiety, and stress. The text content of the b1st depression question is an index of abnormal emotion. The emotional abnormality index of the text content for the second anxiety question. The emotional abnormality index of the text content for the b3rd stress question.

8. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, The methods for obtaining the abnormality index of voice emotion in the current user's answer to a question in terms of speech rate include: For any depression question answered by the current user, the speech rate feature value of the depression question is input into the speech rate pattern model corresponding to depression, and the output of the model is recorded as the speech emotion abnormality index of the depression question in terms of speech rate.

9. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, Methods for obtaining emotional consistency include: For any depression question answered by the current user, the mean of the probability of the current user's depression and the voice emotion abnormality index of the question on each voice feature is denoted as the corrected voice emotion abnormality index of the question on the corresponding voice feature. The absolute differences between the corrected voice emotion abnormality index of the question on each voice feature and the text content emotion abnormality index of the question are accumulated and then back-mapped to the result of normalization processing, which is denoted as the degree of consistency between the text content and voice emotion of the current user's answer to the question.

10. The AI-based intelligent assessment method for adolescent emotional health as described in claim 1, characterized in that, The current methods for obtaining users' fusion anomaly scores in depression include: The normalized ratio of the emotional consistency between the text content and speech of all the depression questions answered by the current user is recorded as the text-speech consistency weight of the corresponding depression question. The weighted sum of the text-speech consistency weights of each depression question answered by the current user and the emotional abnormality index of the text content of the corresponding depression question is recorded as the fusion abnormality score of the current user in depression.

Images