Inplanatable voice psychological crisis recognition method, medium and equipment

Abstract

The invention discloses an interpretable voice psychological crisis recognition method, a medium and equipment, and relates to the technical field of psychological crisis intervention.The interpretable voice psychological crisis recognition method mainly comprises the steps that according to a hotline real record, a suicide risk scale is utilized to obtain a crisis intervention hotline suicide risk voice data set, and a crisis intervention hotline suicide risk voice data set is obtained; carrying out sublingual feature extraction by utilizing a Wilkokson symbol rank test method and a Man-Huitinib U test method to obtain a training data set; constructing a multi-task crisis detection framework by using a deep learning model; training and verifying the multi-task crisis detection framework by using a cross validation method according to the training data set; by implementing the interpretable voice psychological crisis recognition method, the medium and the equipment provided by the invention, the accuracy of psychological crisis recognition can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of psychological crisis intervention, and more specifically to an interpretable voice psychological crisis recognition method, medium and device. Background Art

[0002] The existing speech recognition and analysis methods are mainly: 1) Speech psychological crisis recognition: Speech psychological crisis recognition technology identifies the psychological state of individuals by analyzing speech signals, especially those who may be in psychological crisis. This technology has broad application prospects in psychological counseling, medical health monitoring, intelligent customer service, driving safety monitoring and other fields. With the development of deep learning technology, emotional speech recognition technology has been rapidly developed. Speech recognition technology based on deep learning models such as convolutional neural network (CNN), recurrent neural network (RNN) and long short-term memory network (LSTM) has achieved remarkable results. However, emotional speech recognition technology still faces many challenges in practical applications. People's emotional expression is affected by many factors, such as cultural background, personal experience, language habits, etc., which makes it very difficult to accurately identify and understand people's emotional state. In addition, noise interference and environmental changes, data privacy and security, cross-language and cross-cultural research, real-time and other issues are also key issues that need to be solved in this field. In order to address these challenges, future research may focus more on the following directions: combining deep learning technology, using more complex neural network structures to extract speech features and improve the accuracy of emotion recognition; multimodal emotion recognition, combining information from multiple modalities such as facial expressions and body language for emotion recognition; strengthening privacy protection and data security to ensure user privacy and data security; cross-language and cross-cultural research to improve the generalization ability of emotional speech recognition; and improving the real-time and robustness of emotional speech recognition. At present, in order to solve the problem of lack of labeled data, researchers have proposed a variety of methods, including semi-supervised learning and cross-modal knowledge transfer. Semi-supervised learning improves learning effects by combining a small amount of labeled data with a large amount of unlabeled data, while cross-modal knowledge transfer uses the labeled information of one modality to enhance the learning effect of another modality. These methods have alleviated the problem of data shortage to a certain extent and provided new possibilities for the development of emotional speech recognition technology. 2) Research on interpretable algorithms: Research on interpretable algorithms is an increasingly important research direction in the field of artificial intelligence. With the widespread application of machine learning models, especially deep learning models, in various industries, people are paying more and more attention to the interpretability of models, that is, whether the decision-making process and results of the models can be understood and trusted by humans. This demand is particularly important in high-risk fields such as medicine, finance, and justice, because decisions in these fields are often accompanied by significant social impacts and legal responsibilities. The research on interpretable algorithms aims to develop and improve algorithms to make the internal working mechanism of the model more transparent, thereby improving the credibility of the model and user acceptance. Research methods can be mainly divided into three categories: data-based interpretability, model-based interpretability, and result-based interpretability.Data-based methods provide explanations by analyzing the relationship between input data and model output; model-based methods aim to improve the model structure to make it inherently interpretable; and result-based methods explain the behavior of the model through external analysis after the model makes a decision. At present, the challenges faced by the research on interpretable algorithms include algorithm maturity, computing power cost, and data sparsity. In terms of algorithm maturity, although some interpretable algorithms have been proposed, their stability and accuracy in practical applications still need to be further improved. In terms of computing power cost, some algorithms such as LIME and SHAP require huge computing resources when processing large-scale data. In terms of data sparsity, in order to train an interpretable model, a large amount of labeled data is usually required, which may be difficult to obtain in practical applications. Future research trends may focus on improving the balance between algorithm interpretability and performance, reducing the computational cost of the algorithm, and developing new data-efficient learning methods. In addition, with the improvement of ethical and legal requirements for artificial intelligence, the research on interpretable algorithms will also pay more attention to the fairness and privacy protection of the model. In practical applications, interpretable algorithms can help users with non-technical backgrounds understand the decision logic of the model, thereby improving the transparency of the model and the trust of users. For example, in medical diagnosis, explainable AI models can explain the basis of their diagnosis to doctors, helping them make more accurate judgments. In the financial field, explainable models can clearly show users the basis of credit assessment and enhance users' trust in services. 3) Language behavior feature recognition algorithm: Language is a complex naturally acquired human motor ability. Adults are characterized by making 14 different sounds per second through the coordinated movement of about 100 muscles. Speaker recognition refers to the ability of software or hardware to receive voice signals, identify the speakers appearing in the voice signals, and then identify the speakers. Feature extraction is achieved by converting the speech waveform into a parameter representation at a relatively minimum data rate for subsequent processing and analysis. Therefore, acceptable classification is derived from good and high-quality features. The process of speech recognition technology includes three aspects: preprocessing of speech signals, feature extraction, pattern matching and judgment. First, the system receives speech input, and after signal preprocessing and feature extraction, the speech data of the target information is obtained to form a speech information pattern library, and then the input speech signal and the pattern library are matched to obtain the result of information matching between the two. One is the Mel-Frequency Cepstral Coefficients (MFCC) extraction algorithm. Mel frequency is based on the auditory characteristics of the human ear, and it has a nonlinear correspondence with the Hz frequency. MFCC uses this relationship between them to calculate the Hz spectrum features. MFCC was originally suggested to be used to identify monosyllabic words in continuous spoken sentences, but not for speaker recognition.MFCC calculation is a replication of the human auditory system, assuming that the human ear is a reliable speaker identifier to artificially implement the working principle of the human ear. MFCC features are derived from the difference in the critical bandwidth of the human ear. Frequency filters with linear spacing at low frequencies and logarithmic spacing at high frequencies are used to preserve the phonetic important characteristics of the speech signal. Speech signals usually contain tones of different frequencies, each with an actual frequency, f (Hz), and the subjective pitch is calculated on the Mel scale. The Mel frequency scale has linear frequency spacing below 1000Hz and logarithmic frequency spacing above 1000Hz. 1 kHz pitch, 40 dB above the perceptual audible threshold, is defined as 1000 mels, as a reference point. MFCC is achieved with the help of filter banks to decompose the signal. MFCC gives the discrete cosine transform (DCT) of the real logarithm of the short-term energy displayed on the Mel frequency scale. MFCC is used to identify flight reservations, phone numbers, and security of speech recognition systems. In order to obtain better robustness, some modifications have been made to the basic MFCC algorithm, such as boosting the log-mel amplitude to an appropriate power (about 2 or 3) before applying dct0 and reducing the influence of the low-energy part. The second is the hidden Markov model HMM. The hidden Markov model is a statistical model that is used to describe a Markov process with hidden unknown parameters. The difficulty is to determine the hidden parameters of the process from the observable parameters. These parameters are then used for further analysis, such as pattern recognition. For example, we may get a string of numbers (rolling the dice 10 times): 1, 6, 3, 5, 2, 7, 3, 5, 2, 4. This string of numbers is called a visible state chain. But in the hidden Markov model, we not only have such a visible state chain, but also a hidden state chain. In this example, this hidden state chain is the sequence of dice you use. For example, the hidden state chain may be: D6, D8, D8, D6, D4, D8, D6, D6, D4, D8. The third is Linear Prediction Cepstral Coefficients (LPCC), which are cepstral coefficients obtained from the spectrum envelope calculated by Linear Prediction Coefficients (LPC), and are the coefficients of the Fourier transform of the LPC logarithmic amplitude spectrum. Cepstral analysis is a commonly used analysis method in the field of speech processing because it can perfectly characterize speech waveforms and features with limited features.

[0003] "A Psychological Crisis Intervention System and Method" with publication number CN114121193A provides a psychological crisis intervention system and method. By setting a self-evaluation module, a peer-evaluation module, a crisis reporting module, a re-evaluation module, a crisis approval module, an intervention module and a status monitoring module, the comprehensive psychological crisis situation of the person being evaluated is obtained through self-evaluation and peer-evaluation, so that it is reported when it is judged that a psychological crisis exists, and the psychological crisis state of the person being evaluated is further confirmed through the re-evaluation module, and the approval is carried out through the crisis approval module to obtain the approval result. According to a pre-set intervention model, an intervention conclusion is generated according to the crisis approval result, and the mental health condition of the person being evaluated is counseled or diagnosed to help him restore his mental health and reduce the risk of hurting others and self-harm caused by psychological crisis problems. By setting a status monitoring module, the intervention process of the person being evaluated is monitored and recorded, so that the intervention process can be further adjusted according to the intervention situation to improve the intervention effect.

[0004] "A method and system for identifying and intervening in psychological crises" with publication number CN110993061A discloses a method and system for identifying and intervening in psychological crises. By establishing a psychological crisis assessment model with five psychological crisis factor dimensions and scoring the five psychological crisis factors of the person being assessed separately according to the psychological crisis factor score table, the health level of the person being assessed in each psychological crisis factor can be understood. By weighted scoring the psychological crisis status of the person being assessed according to the weight of each psychological crisis factor, a general evaluation reflecting the overall mental health status of the person being assessed can be obtained. By judging the mental health and suicidal tendency of the person being assessed, a psychological crisis file is given, which enables counselors, psychological consultants, etc. to grasp the changes in the mental health and intervention of the person being assessed. By providing intervention process navigation, the process guidance and monitoring of the psychological crisis intervention of the person being assessed are carried out, the mental health status of the person being assessed is further improved, and a psychological crisis prevention and rapid response mechanism is improved.

[0005] In the current field of psychological crisis intervention, traditional speech recognition and analysis methods face many challenges.

[0006] First, existing crisis hotline voice data sources are limited and the data quality is uneven, which limits the accuracy and reliability of speech recognition systems.

[0007] Secondly, due to compression in recording equipment and transmission, speech signals often lose a lot of important acoustic information, resulting in inaccurate analysis results. In addition, traditional machine learning methods rely on manual feature selection, which not only increases the workload but also may introduce human bias.

[0008] Furthermore, people’s psychological states are complex and dynamically changing, and existing models often find it difficult to capture contextual information in speech, which is crucial for accurately identifying psychological crises.

[0009] Finally, deep learning models are often viewed as “black boxes” and lack interpretability, which limits their generalization in practical applications. Summary of the invention

[0010] The purpose of the present invention is to provide an interpretable voice psychological crisis recognition method, medium and device, which can improve the accuracy of psychological crisis recognition.

[0011] The present invention provides an interpretable voice psychological crisis recognition method, comprising the following steps: S1: obtaining a crisis intervention hotline suicide risk voice data set by using a suicide risk scale according to a real hotline recording; S2: extracting paralinguistic features by using a Wilcoxon signed rank test method and a Mann-Whitney U test method according to the crisis intervention hotline suicide risk voice data set to obtain a training data set; S3: constructing a multi-task crisis detection framework by using a deep learning model; S4: training and verifying the multi-task crisis detection framework by using a cross-validation method according to the training data set to obtain a trained multi-task crisis detection framework.

[0012] Furthermore, step S1 specifically includes: S11: recording the crisis intervention hotline call to obtain a real hotline recording; S12: cross-annotating the real hotline recording using a suicide risk scale to obtain annotated audio; S13: segmenting the annotated audio according to the voice start timestamps of the caller and the operator to obtain caller audio segments; S14: merging the caller audio segments using an audio feature extraction tool to obtain a complete caller audio segment; S15: cropping and sampling the complete caller audio segment according to a preset duration and a preset sampling rate to obtain an audio segment; and obtaining a crisis intervention hotline suicide risk voice dataset based on the audio segment.

[0013] Furthermore, step S2 specifically includes: S21: according to the crisis intervention hotline suicide risk voice data set, using the Wilcoxon signed rank test method, obtaining the difference in voice duration between callers and operators of different categories; S22: according to the crisis intervention hotline suicide risk voice data set, using the Mann-Whitney U test method, obtaining the gender difference of callers of different categories; S23: resampling the audio segments in the crisis intervention hotline suicide risk voice data set according to a preset resampling rate to obtain resampled audio segments; S24: according to the difference in voice duration between callers and operators of different categories and the gender difference of callers of different categories, feature extraction is performed on the resampled audio segments according to a preset Fourier window size and a preset step size to obtain paralinguistic features; S25: normalizing, statistically analyzing, variance analysis and post hoc test the paralinguistic features to obtain a training data set.

[0014] Furthermore, the above-mentioned paralinguistic features include emotional intensity and dynamic range features, timbre and brightness features, speech speed and rhythm features, pitch and intonation features, and texture and content features.

[0015] Furthermore, the above-mentioned multi-task crisis detection framework includes a dual bidirectional long short-term memory network, a self-attention mechanism, a first fully connected layer, a second fully connected layer, a first normalization layer and a second normalization layer; the dual bidirectional long short-term memory network includes a first bidirectional long short-term memory network and a second bidirectional long short-term memory network.

[0016] Furthermore, the above multi-task crisis detection framework also includes a loss function, such as the formula: , , , , in, is the focal loss, is the probability that the model correctly classifies the actual label, is a class-specific weighting factor, To reduce the factor of the impact of easy-to-classify examples, Loss of mission for crisis identification, Audio feature input for crisis identification tasks, is the label corresponding to the crisis identification task, is the crisis category weight, is the loss for the gender recognition task, Audio feature input for the gender recognition auxiliary task, is the label corresponding to the gender recognition auxiliary task, is the total loss, is the weight assigned to the crisis classification task, is the weight assigned to the gender recognition auxiliary task.

[0017] Furthermore, the above cross-validation method is StratifiedKFold.

[0018] Furthermore, the above-mentioned explainable voice psychological crisis identification method also includes: using a trained multi-task crisis detection framework to perform psychological crisis identification on the voice to be analyzed to obtain an identification result.

[0019] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned interpretable voice psychological crisis identification method are implemented.

[0020] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned interpretable speech psychological crisis identification method when executing the program.

[0021] The interpretable voice psychological crisis identification method, medium and device provided by the present invention have the following beneficial effects: Aiming at the deficiencies of the existing technology in the psychological crisis intervention hotline service, the present invention is based on deep learning and multi-task learning technology, fully considers the theoretical knowledge of psychology and the existing theoretical research results, fully considers the suicide risk and psychological crisis phenomenon, and takes into account the problems of gender differences and category imbalance. The extraction algorithm corresponding to speech frequency, emotional tendency and semantics is used to calculate and process the speech feature value; the present invention combines the deep learning model of psychological knowledge to construct a gender-assisted multi-task crisis detection framework; the model adopts a bidirectional long short-term memory network, with gender recognition as an auxiliary task and crisis detection as the main task; the core architecture is characterized by a dual bidirectional long short-term memory network (Dual Bi-LSTM) (drop-out rate is 0.5) and self-attention mechanism, aiming to accurately capture and evaluate the psychological state of the help-seekers, so as to improve the accuracy and timeliness of suicide intervention; its application in crisis intervention hotlines in the future will reduce the burden and pressure of crisis intervention work, and provide strong support for the efficient operation of crisis intervention hotlines; the present invention proposes a deep learning network for crisis intervention hotlines from both theoretical and data-driven perspectives, and uses deep learning technology to automatically extract speech features, avoiding the bias that may be introduced by manual feature selection, while taking into account gender differences and category imbalance, and combining multi-task learning, which not only improves the performance of the model, but also enhances the interpretability of the model in practical applications, promotes the application of the model in real scenarios, and constructs a crisis identification method with more theoretical support and interpretability, which can achieve better results and Applicability: The present invention constructs a high-quality speech data set through a carefully designed data collection and preprocessing process, effectively solving the problems of insufficient data sources and uneven data quality in the prior art; in addition, in response to the challenges of poor recording quality and loss of acoustic information, the present invention adopts advanced signal reconstruction technology, which not only significantly improves the quality of speech signals, but also retains more effective information, providing a more accurate and reliable speech analysis basis for psychological crisis intervention, thereby improving the accuracy of psychological crisis identification; the present invention can accurately analyze and identify the signs of psychological crisis contained in the voice signals of the help-seekers, and assist in suicide identification and crisis detection through machines; the present invention can improve the accuracy of identification and the timeliness of intervention, significantly improve the efficiency and effectiveness of hotline services, and thus more effectively prevent the occurrence of suicide and psychological crisis events. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 It is a flow chart of the interpretable voice psychological crisis identification method provided by the present invention; Figure 2 It is the overall flow chart of the psychological crisis identification method provided by the present invention; Figure 3 It is a flow chart of constructing a data set provided by the present invention; Figure 4 This is the data set construction process provided by the present invention; Figure 5 It is a characteristic statistical analysis flow chart provided by the present invention; Figure 6 It is a crisis identification model framework diagram provided by the present invention; Figure 7 It is a structural block diagram of the computer device provided by the present invention. DETAILED DESCRIPTION

[0023] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0024] Figure 1 A schematic diagram of an interpretable speech psychological crisis identification method of this embodiment is shown. In this embodiment, the interpretable speech psychological crisis identification method includes the following steps: S1: Based on the real recordings of the hotline, the suicide risk scale is used to obtain the crisis intervention hotline suicide risk voice dataset; Specifically, the suicide risk scale includes a death intention section, a hopelessness section, and a psychological distress section; e.g. Figure 2 Shown is an overall flow chart of the psychological crisis identification method provided by this embodiment; Specifically, step S1 includes: S11: recording the crisis intervention hotline call to obtain the real hotline recording; S12: cross-annotating the real hotline recording using the suicide risk scale to obtain annotated audio; S13: segmenting the annotated audio according to the voice start timestamps of the caller and the operator to obtain the caller audio segment; S14: merging the caller audio segments using the audio feature extraction tool to obtain the caller audio complete segment; S15: cropping and sampling the caller audio complete segment according to the preset duration and the preset sampling rate to obtain the audio segment; according to the audio segment, the crisis intervention hotline suicide risk voice dataset is obtained; Specifically, the preset duration is 5s; the preset sampling rate is 16kHz; It should be noted that in step S1, a crisis intervention hotline suicide risk speech dataset is constructed and the speech is labeled based on SRS; in this embodiment, data collection and annotation research data comes from Wuhan "Heart to Heart" crisis intervention hotline, which collects 81 real recordings ranging from 5 to 65 minutes in length, from operators and callers, and randomly matched by gender; the recordings are used with the permission of the participants and processed to remove identifiable personal characteristics; the audio is cross-annotated in terms of death intention, despair and psychological distress according to the modified Suicide Risk Scale (SRS); the scale is divided into three main parts, each containing three questions, for a total of nine questions; the score range for each question is 1 to 5, and the score is based on the severity or frequency of the symptoms; the total score range of the scale is 9 to 45 points, and callers can be classified as non-suicidal calls, crisis calls or acute suicide calls according to the total score; the first part "Death Intent" mainly assesses the caller's intention to die, and the questions include: "How often does the caller mention suicidal thoughts?", "How detailed is the caller's discussion of suicide plans?" "How much do callers think about their future?" and "How detailed is the suicide preparation described by the caller?" The second part, "Despair", focuses on assessing the caller's hopelessness about their life and future. Questions include: "How does the caller view their future?", that is, whether they see hope or think the prospects are bleak; "How does the caller describe their current situation?", assessing their views on current living conditions, environment or events; and "How does the caller perceive their self-worth?", examining the degree to which they recognize their own value and whether they have low self-esteem or self-denial tendencies; the third part, "Psychological distress", focuses on assessing the caller's emotional stability and The severity of psychological stress, including questions like: "How stable is the caller's mood?", to see if their mood fluctuates greatly, or if they have signs of emotional breakdown; "How severe are the social difficulties caused by psychological problems?", that is, whether their psychological problems affect their social or work functions; and "How high is the stress level described by the caller?", to understand whether the stress they feel exceeds their tolerance, and whether they have persistent tension, anxiety, or helplessness; the total score is added up based on the score of each question, and the score range of the scale is 9 to 45 points; 9-21 points means that the caller has no obvious suicidal tendencies and is non-suicidal. Suicidal calls; 22-33 points indicate a crisis and may require further attention; 34-45 points indicate an acute suicide risk and require immediate intervention; the dataset was processed by segmenting the caller and operator's audio by marking the caller's and operator's voice start timestamps; the audio feature extraction tool liborsa was then used to merge the caller's audio segments into a complete segment; in order to make full use of the audio information, the caller's voice was muted, and after cutting off the silent segments greater than 1 second, the remaining audio was segmented into audio segments with a duration of 5 seconds and a sampling rate of 16kHz;The complete process of dataset construction is as follows; Figure 3 As shown, the construction process is as follows Figure 4 As shown; S2: Based on the crisis intervention hotline suicide risk speech dataset, the Wilcoxon signed rank test method and the Mann-Whitney U test method were used to extract paralinguistic features to obtain the training dataset; Specifically, step S2 includes: S21: Based on the crisis intervention hotline suicide risk voice dataset, the Wilcoxon signed-rank test method was used to obtain the differences in voice duration between different categories of callers and operators; S22: Based on the crisis intervention hotline suicide risk voice dataset, the Mann-Whitney U test method was used to obtain the gender differences of callers in different categories; S23: resampling the audio segments in the crisis intervention hotline suicide risk voice dataset according to a preset resampling rate to obtain a resampled audio segment; Specifically, the preset resampling rate is 22050Hz; S24: according to the difference in voice duration between different categories of callers and operators and the difference in gender of different categories of callers, feature extraction is performed on the resampled audio segment according to a preset Fourier window size and a preset step size to obtain paralinguistic features; Specifically, the preset Fourier window size is 2048 Hz, and the preset step size is 1024 Hz; Specifically, paralinguistic features include emotional intensity and dynamic range features, timbre and brightness features, speech rate and rhythm features, pitch and intonation features, and texture and content features; S25: normalizing, statistically analyzing, variance analysis and post hoc test the paralinguistic features to obtain the training data set; It should be noted that in step S2, the variability of speech duration between callers and high-level features of speech between different callers was studied by digitization of speech signals; the Wilcoxon sign-rank test, a non-parametric test applied to paired samples, was used to analyze the differences in speech duration between different categories of callers and operators through interaction effect tests and simple slope analysis; similarly, the Mann-Whitney-U-Test was used to test gender differences between different categories of callers, which is a non-parametric method suitable for independent sample difference tests, with interaction effect tests and simple slope analysis; when extracting features, the liborsa toolkit in Python was used to extract paralinguistic features of telephone audio; the extracted features included five paralinguistic dimensions, such as emotional intensity and dynamic range, richness and brightness, speech rate and rhythm, pitch and tone and texture, and Content analysis, a total of 178 features; specifically, including: 10 emotion intensity and dynamic range features, which are generally related to the intensity and emotional expression of speech; 9 timbre and brightness features, which can respond to the complexity of speech; 6 speech rate and rhythm features, which reflect a person's language performance; 12 pitch and intonation features, which reflect the pitch level of speech; and 141 texture and content features, which contain detailed information about subtle changes in speech and responses to speech; Due to the low sampling rate of telephone audio, this embodiment resamples the original audio to 22050Hz, and then extracts features with a Fourier window size of 2048Hz and a step size of 1024Hz; Finally, 178 features of 5 categories of 79 time frames are extracted from 5 seconds of audio, and the features are normalized on each time frame; The features are then statistically analyzed, and variance analysis and post hoc tests are performed on the features of different crisis categories; The process of feature statistical analysis is as follows Figure 5 As shown; S3: Using deep learning models to build a multi-task crisis detection framework; Specifically, the multi-task crisis detection framework includes a dual bidirectional long short-term memory network, a self-attention mechanism, a first fully connected layer, a second fully connected layer, a first normalized layer, and a second normalized layer; the dual bidirectional long short-term memory network includes a first bidirectional long short-term memory network and a second bidirectional long short-term memory network; Specifically, the dropout rate of the dual bidirectional LSTM network is 0.5; Specifically, the multi-task crisis detection framework also includes a loss function, such as the formula: , , , , in, is the focal loss, is the probability that the model correctly classifies the actual label, is a class-specific weighting factor, To reduce the factor of the impact of easy-to-classify examples, Loss of mission for crisis identification, Audio feature input for crisis identification tasks, is the label corresponding to the crisis identification task, is the crisis category weight, is the loss for the gender recognition task, Audio feature input for the gender recognition auxiliary task, is the label corresponding to the gender recognition auxiliary task, is the total loss, is the weight assigned to the crisis classification task, is the weight assigned to the gender recognition auxiliary task; It should be noted that in step S3, a deep learning model combined with psychology knowledge is applied to build a gender-assisted multi-task crisis detection framework; the model adopts a bidirectional long short-term memory (Bi-LSTM) network, gender recognition is an auxiliary task, and crisis detection is the main task; the core architecture is characterized by a dual bidirectional long short-term memory network with a random dropout rate of 0.5 to prevent overfitting and a self-attention mechanism to enhance the temporal dynamics and feature weights of audio data; it includes two fully connected layers and two normalization layers to improve generalization of different data; due to the significant imbalance of crisis types in the dataset, a FocalLoss with custom weights is used; the loss function is defined as (1) in, p is the probability that the model correctly classifies the actual label, is a factor that reduces the impact of easy-to-classify examples, is a class-specific weighting factor; the loss for crisis and gender recognition tasks is calculated as follows: (2) (3) The total loss is the weighted sum of the two task losses: (4) in, and Represent the weights assigned to crisis classification and gender identification tasks respectively; the crisis identification model framework is as follows Figure 6 As shown; S4: Based on the training data set, the multi-task crisis detection framework is trained and verified using a cross-validation method to obtain a trained multi-task crisis detection framework; Specifically, the cross-validation method is StratifiedKFold; It should be noted that in step S4, StratifiedKFold was used to perform five cross-validations to ensure the effectiveness and generalization of the model. An early stopping mechanism was set to stop training when the performance indicators of the validation set converged without improvement for 20 consecutive epochs to avoid overfitting; Specifically, in step S4, the model considers both precision and recall, and also focuses on a few crisis categories, so the performance indicators mainly use macro Score; the first step is to calculate the precision and recall accuracy of each category as follows: (5) (6) in TP For authenticity, FP is a false positive, FN It is a false negative; The formula for calculating a class of F1 based on precision and recall is: (7) For each task, the macro F1 score is calculated as: (8) in, is the number of categories in the task; Combining the crisis identification and gender identification tasks, the weighted F1 score is calculated as follows: (9) Among them, the comparison of model weighted F1 scores on existing datasets is shown in Table 1; Table 1: Weighted F1 score comparison table

[0025] Optionally, the above-mentioned explainable voice psychological crisis identification method also includes: using a trained multi-task crisis detection framework to perform psychological crisis identification on the voice to be analyzed to obtain an identification result.

[0026] This embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned interpretable voice psychological crisis identification method are implemented. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk drive (HDD) or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above-mentioned types of memory.

[0027] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned interpretable speech psychological crisis identification method are implemented.

[0028] like Figure 7 As shown, the computer device may include: at least one processor 121, such as a central processing unit (CPU), at least one communication interface 123, a memory 124, and at least one communication bus 122. The communication bus 122 is used to realize the connection and communication between these components. The communication interface 123 may include a display screen (Display), a keyboard (Keyboard), and the optional communication interface 123 may also include a standard wired interface and a wireless interface. The memory 124 may be a random access memory (Random Access Memory, RAM), or a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 124 may also be at least one storage device located away from the aforementioned processor 121. The memory 124 stores an application program, and the processor 121 calls the program code stored in the memory 124 to perform any of the above method steps. The communication bus 122 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7Only one line is used to represent it, but it does not mean that there is only one bus or one type of bus. Among them, the memory 124 may include a volatile memory (volatile memory), such as a random-access memory (random-access memory, RAM); the memory may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a hard disk drive (hard disk drive, HDD) or a solid-state drive (solid-state drive, SSD); the memory 124 may also include a combination of the above-mentioned types of memory. Among them, the processor 121 may be a central processing unit (central processing unit, CPU), a network processor (network processor, NP) or a combination of CPU and NP. Among them, the processor 121 may further include a hardware chip. The above-mentioned hardware chip may be an application-specific integrated circuit (application-specific integrated circuit, ASIC), a programmable logic device (programmable logic device, PLD) or a combination thereof. The above-mentioned PLD may be a complex programmable logic device (complex programmable logic device, CPLD), a field-programmable gate array (field-programmable gate array, FPGA), a generic array logic (generic array logic, GAL) or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call the program instructions to implement the interpretable voice psychological crisis identification method of this embodiment.

[0029] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. An interpretable voice psychological crisis identification method, characterized in that: The following steps are involved: S1: Based on the real recordings of the hotline, the suicide risk scale is used to obtain the crisis intervention hotline suicide risk voice dataset; S2: Based on the crisis intervention hotline suicide risk speech dataset, the Wilcoxon signed rank test method and the Mann-Whitney U test method are used to extract paralinguistic features to obtain a training dataset; S3: Using deep learning models to build a multi-task crisis detection framework; S4: According to the training data set, the multi-task crisis detection framework is trained and verified by using a cross-validation method to obtain a trained multi-task crisis detection framework.

2. The interpretable speech psychological crisis identification method according to claim 1, characterized in that: Step S1 specifically includes: S11: Record the crisis intervention hotline calls to obtain the real recording of the hotline; S12: cross-annotating the real recording of the hotline using the suicide risk scale to obtain annotated audio; S13: segmenting the annotated audio according to the voice start timestamps of the caller and the operator to obtain the caller audio segment; S14: merging the caller audio segments using an audio feature extraction tool to obtain a complete caller audio segment; S15: According to a preset duration and a preset sampling rate, the complete audio segment of the caller is cropped and sampled to obtain an audio segment; and a crisis intervention hotline suicide risk voice dataset is obtained based on the audio segment.

3. The interpretable speech psychological crisis identification method according to claim 1, characterized in that: Step S2 specifically includes: S21: Based on the crisis intervention hotline suicide risk voice dataset, the Wilcoxon signed-rank test method is used to obtain the differences in voice durations of different categories of callers and operators; S22: Based on the crisis intervention hotline suicide risk voice dataset, the Mann-Whitney U test method is used to obtain the gender differences of callers in different categories; S23: resampling the audio segments in the crisis intervention hotline suicide risk voice dataset according to a preset resampling rate to obtain resampled audio segments; S24: according to the difference in voice duration between the callers of different categories and the operator and the gender difference of the callers of different categories, feature extraction is performed on the resampled audio segment according to a preset Fourier window size and a preset step size to obtain paralinguistic features; S25: performing normalization processing, statistical analysis, variance analysis and post hoc test on the paralinguistic features to obtain a training data set.

4. The interpretable speech psychological crisis identification method according to claim 3 is characterized in that: The paralinguistic features include emotional intensity and dynamic range features, timbre and brightness features, speech speed and rhythm features, pitch and intonation features, and texture and content features.

5. The interpretable speech psychological crisis identification method according to claim 1, characterized in that: The multi-task crisis detection framework includes a dual bidirectional long short-term memory network, a self-attention mechanism, a first fully connected layer, a second fully connected layer, a first normalization layer and a second normalization layer; the dual bidirectional long short-term memory network includes a first bidirectional long short-term memory network and a second bidirectional long short-term memory network.

6. The interpretable speech psychological crisis identification method according to claim 5, characterized in that: The multi-task crisis detection framework also includes a loss function, such as the formula: ， ， ， ， in, is the focal loss, is the probability that the model correctly classifies the actual label, is a class-specific weighting factor, To reduce the factor of the impact of easy-to-classify examples, Loss of mission for crisis identification, Audio feature input for crisis identification tasks, is the label corresponding to the crisis identification task, is the crisis category weight, is the loss for the gender recognition task, Audio feature input for the gender recognition auxiliary task, is the label corresponding to the gender recognition auxiliary task, is the total loss, is the weight assigned to the crisis classification task, is the weight assigned to the gender recognition auxiliary task.

7. The interpretable speech psychological crisis identification method according to claim 1, characterized in that: The cross validation method is StratifiedKFold.

8. The interpretable voice psychological crisis identification method according to claim 1, characterized in that: The explainable voice psychological crisis identification method also includes: using the trained multi-task crisis detection framework to perform psychological crisis identification on the voice to be analyzed to obtain an identification result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the interpretable speech psychological crisis identification method as described in any one of claims 1-8 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the interpretable speech psychological crisis identification method as described in any one of claims 1-8 are implemented.

Images