AI voice anti-bulley cloud platform application system fused with intelligent algorithm

Abstract

The invention discloses an AI voice anti-bullingering cloud platform application system fused with an intelligent algorithm, and particularly relates to the technical field of intelligent algorithms, the AI voice anti-bullingering cloud platform application system comprises a voice acquisition module, a preprocessing module, a voice recognition module, a natural language processing module, an intelligent algorithm analysis module and an alarm module, the voice recognition module extracts voice features to analyze bullying hidden dangers, the natural language processing module converts voice data into texts and recognizes the hidden dangers, the intelligent algorithm analysis module conducts deep learning verification on voice segments with the hidden dangers, and finally the alarm module decides whether to send out alarm signals or not according to verification results. According to the method, potential bullying behaviors in voice can be captured in time, response is made immediately, the risk of delay and missing report is remarkably reduced, the system can capture more complex and hidden bullying behaviors through combination of multi-modal data, and the method is suitable for various voice communication scenes and has wide application prospects and social benefits.

Description

Technical Field

[0001] The present invention relates to the field of intelligent algorithm technology, and more specifically, to an AI voice anti-bullying cloud platform application system integrating intelligent algorithms. Background Art

[0002] With the rapid development of the Internet and communication technology, voice communication has become an indispensable part of people's daily lives. Voice communication is widely used in various fields such as social networks, online education, and game communication. However, with the popularization of voice communication, the phenomenon of voice bullying has become more and more serious, posing a serious threat to the physical and mental health of individuals, especially young people.

[0003] Traditional anti-bullying technology mainly relies on text analysis, such as sentiment analysis and keyword recognition of text on social media. However, voice bullying is real-time and complex, and victims often find it difficult to identify and respond in the first place. In addition, the emotions, tone and other information implied in voice are more complex than text, which makes it difficult for traditional text analysis technology to be directly applied to the identification of voice bullying. Moreover, in most cases, existing technologies require manual review or delayed analysis to judge voice bullying behavior, and cannot achieve real-time automatic identification and intervention. Therefore, an AI voice anti-bullying cloud platform application system that integrates intelligent algorithms is proposed here to solve the above problems. Summary of the invention

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The AI ​​voice anti-bullying cloud platform application system that integrates intelligent algorithms includes voice acquisition module, preprocessing module, voice recognition module, natural language processing module, intelligent algorithm analysis module, and alarm module;

[0006] The voice collection module is used to collect the user's voice data in real time, obtain the user's voice data set, and then transmit the user's voice data set to the preprocessing module;

[0007] The preprocessing module is used to perform denoising and segmentation on the user's speech data set to obtain multiple speech audio segments, and then transmit the multiple speech audio segments to the speech recognition module and the natural language processing module respectively;

[0008] The speech recognition module is used to extract and analyze speech features of multiple speech audio segments to identify whether the target speech audio segment contains bullying risks;

[0009] The natural language processing module is used to convert the speech data corresponding to the multiple speech audio segments into multiple text paragraphs and analyze them to identify whether the target text paragraphs contain bullying risks;

[0010] The intelligent algorithm analysis module is used to verify the audio and video segments with bullying risks through deep learning algorithms, and then transmit the verification results to the alarm module;

[0011] In a preferred embodiment, the AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms also includes an alarm module for deciding whether to issue or not to issue an alarm signal based on the verification result.

[0012] In a preferred embodiment, the data recording module is used to record all collected voice data, analysis results of the voice recognition module and the natural language processing module, verification results of the intelligent algorithm analysis module and alarm records of the alarm module.

[0013] In a preferred embodiment, the preprocessing module is used to segment the user's speech data set, which includes:

[0014] The speech data set is divided into multiple speech segments in the time-frequency domain by short-time Fourier transform.

[0015] In a preferred embodiment, the speech recognition module is used to extract and analyze speech features of multiple speech audio segments respectively, which means:

[0016] Extracting speech features of the target speech segment includes Mel frequency cepstral coefficients, fundamental frequency, energy envelope, and formant frequency;

[0017] The extracted speech features of the target speech segment are combined into a speech vector and transmitted to a pre-trained convolutional neural network. The output data of the pre-trained convolutional neural network is the probability of bullying.

[0018] In a preferred embodiment, identifying whether the target audio segment has a bullying risk refers to:

[0019] The probability of bullying in the target audio frequency segment is obtained, and then the probability of bullying is compared with a preset hidden danger threshold value 1. If the probability of bullying is greater than or equal to the preset hidden danger threshold value 1, it is judged that there is a bullying hidden danger; if the probability of bullying is less than the preset hidden danger threshold value 1, it is judged that there is no bullying hidden danger.

[0020] In a preferred embodiment, the natural language processing module is used to convert the speech data corresponding to the multiple speech segments into multiple text segments and analyze them, which means:

[0021] Get the target speech segment, then convert the frequency domain signal back to the time domain signal through inverse short-time Fourier transform and get the speech data of the corresponding time segment, then convert the speech data of the time segment into a text paragraph, then use the pre-trained sentiment analysis model to perform sentiment analysis on the text paragraph, the pre-trained sentiment analysis model outputs the probability of each preset sentiment category, then use the preset bullying vocabulary list to extract the bullying words in the text paragraph and count the total number of times all the extracted bullying words appear in the text paragraph, and finally calculate the bullying occurrence coefficient:

[0022] Qi represents the probability of each preset emotion category i, wi represents the preset proportion coefficient corresponding to each emotion category i, k represents the total number of preset emotion categories, N represents the total number of times all bullying words appear in the text paragraph, L represents the total number of words in the text paragraph, and BIC represents the bullying occurrence coefficient.

[0023] In a preferred embodiment, identifying whether a target text paragraph has a bullying risk refers to:

[0024] Obtain the bullying occurrence coefficient of the target text paragraph, and then compare the bullying occurrence coefficient with the preset hidden danger threshold value one. If the bullying occurrence coefficient is greater than or equal to the preset hidden danger threshold value two, it is judged that there is a bullying hidden danger. If the bullying occurrence coefficient is less than the preset hidden danger threshold value two, it is judged that there is no bullying hidden danger.

[0025] In a preferred embodiment, the intelligent algorithm analysis module is used to verify the audio and video segments with potential bullying risks through a deep learning algorithm, which means:

[0026] Obtain the bullying risk recognition result of the voice segment corresponding to the voice data of the time segment, and the bullying risk recognition result of the corresponding text segment. If at least one of the recognition results indicates that there is a bullying risk, perform a verification operation on the voice data of the time segment. If both recognition results indicate that there is no bullying risk, do not perform a verification operation on the voice data of the time segment.

[0027] The verification operation refers to:

[0028] The extracted speech features of the speech segment, the probability of bullying occurrence corresponding to the speech segment, each preset emotion category of the text paragraph and its probability, the frequency of bullying words, and the bullying occurrence coefficient are combined to form a comprehensive feature vector;

[0029] The comprehensive feature vector is input into the pre-trained deep learning model, and the deep learning model outputs one of two preset classification labels, indicating whether the comprehensive feature vector represents the existence of bullying behavior.

[0030] In a preferred embodiment, the alarm module is used to determine whether to issue an alarm signal according to the verification result, which means:

[0031] The total number of bullying labels of all classification labels indicating the existence of bullying behavior is obtained. If the total number of bullying labels is zero, no alarm signal is issued. If the total number of bullying labels is not zero, the total number of bullying labels is divided by the preset signal level division value and the result is rounded up, and an alarm signal of the level corresponding to the rounded integer value is issued.

[0032] Technical effects and advantages of the present invention:

[0033] The present invention realizes real-time monitoring and accurate identification of voice bullying behavior by integrating speech recognition, natural language processing and deep learning algorithms. Compared with traditional text analysis methods, the present invention can timely capture potential bullying behavior in voice and respond immediately, significantly reducing the risk of delays and omissions. The present invention makes comprehensive use of speech features and text features, and further improves the accuracy of bullying behavior detection by performing dual analysis of speech signals and converted text paragraphs. The combination of multimodal data enables the system to capture more complex and hidden bullying behaviors.

[0034] The alarm module of the present invention can dynamically adjust the alarm level according to the number and severity of bullying tags identified. This flexible alarm mechanism ensures that the system responds appropriately to bullying behaviors of different intensities, effectively preventing the flooding or neglect of alarms. The present invention is applicable to a variety of voice communication scenarios, such as social networks, online education, online games, etc., and can effectively prevent psychological and social problems caused by bullying behaviors in voice communication, and has broad application prospects and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0036] Figure 1 This is the schematic diagram of the AI ​​voice anti-bullying cloud platform application system that integrates intelligent algorithms. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Example 1

[0039] The AI ​​voice anti-bullying cloud platform application system that integrates intelligent algorithms includes voice acquisition module, preprocessing module, voice recognition module, natural language processing module, intelligent algorithm analysis module, and alarm module;

[0040] The voice acquisition module is used to collect the user's voice data in real time, obtain the user's voice data set, and then transmit the user's voice data set to the preprocessing module; use a microphone or other voice input device to collect voice data in real time, convert the voice data stream into a digital signal, and use sampling technology (such as 16kHz sampling rate) to obtain the voice waveform.

[0041] Take the campus application environment as an example: voice collection modules should be distributed throughout the campus, especially in key areas where bullying may occur, such as toilets, dormitories, corners of playgrounds, corridors, etc. This ensures real-time collection of sounds on campus, and the voice collection module should have wireless communication capabilities to achieve connection with the system. This ensures real-time transmission and reception of audio data, thereby achieving real-time analysis and processing of audio data. The voice collection module should have a built-in high-sensitivity microphone to ensure clear capture and transmission of surrounding sounds. The voice collection module can collect voice data through, for example, smoke detectors with built-in high-sensitivity microphones and wireless communication capabilities that are arranged on campus.

[0042] The preprocessing module is used to perform denoising and segmentation on the user's voice data set to obtain multiple voice audio segments, and then transmit the multiple voice audio segments to the voice recognition module and the natural language processing module respectively; denoising: use filters (such as low-pass filters, band-pass filters) or adaptive filtering algorithms (such as LMS algorithms) to eliminate background noise. The preprocessing module is used to segment the user's voice data set, which means: the voice data set is divided into multiple voice audio segments in the time-frequency domain through short-time Fourier transform.

[0043] X(m,k) is the result of short-time Fourier transform, which is the spectrum value corresponding to time frame m and frequency index k. It is the time-frequency representation obtained after the short-time Fourier transform (STFT) decomposes the signal. x[n] is the nth sampling point of the input time domain signal (i.e., the original speech signal). It is a discrete signal obtained from the time domain. w[n-mR] is a window function, usually a window such as a Hanning window or a Hamming window, which is used to intercept a part of the speech signal. m is the frame index, and R is the frame shift (i.e., the distance between two adjacent windows). This window function is used to limit the range of the signal in time, and the frequency components can be analyzed within a period of time. e -j2πkn / Nis the basis function of the Fourier transform, which is used to convert the signal from the time domain to the frequency domain. Here k is the frequency index and N is the size of the window. This part performs frequency analysis and converts the signal into its spectral representation. n is the index of the sampling point, from 0 to N-1, representing the discrete points of the signal in the current window. m is the index of the window, which represents the time frame of the current analysis. Each m corresponds to a specific time window. k is the frequency index, which represents the frequency component in the STFT result. Each k corresponds to a specific frequency. STFT is an extension of the Fourier transform, which can decompose the signal into the frequency domain while retaining the time information. This is particularly useful for processing non-stationary signals (such as speech signals), and can capture the frequency changes of the signal at different time points. Specific benefits include: Time-frequency analysis: STFT allows you to observe the frequency components of the signal at different time points. Speech signals are dynamic and contain speech information that changes over time, so time-frequency analysis can provide richer information than simple frequency analysis. Capture local frequency characteristics: The window function can perform frequency analysis in a local area of ​​the signal, thereby capturing instantaneous frequency changes and characteristics in a short period of time.

[0044] The speech recognition module is used to extract and analyze speech features of multiple speech audio segments respectively, which means: extracting speech features of target speech audio segments including Mel frequency cepstral coefficients, fundamental frequency, energy envelope, and formant frequency;

[0045] MFCC n Represents the nth Mel-frequency cepstral coefficient. This is a feature value extracted from the speech signal, usually used to represent the sound quality characteristics of the speech. X(k) is the spectrum value after the Mel filter bank is processed, which represents the energy of the frequency components of the speech signal after passing through the Mel-frequency filter bank. k is the filter index, which represents the kth filter in the Mel filter bank. Each filter corresponds to a specific frequency range. K is the number of filters, that is, the total number of filters in the Mel filter bank. Usually this value is pre-set, such as 20 or 40 filters. n is the coefficient index of MFCC, usually starting from 1 and increasing. Generally speaking, the first 12 to 13 MFCC coefficients are used for speech feature analysis.

[0046] f0 is the fundamental frequency, which represents the main frequency component of the speech signal (usually the speaker's vocal cord vibration frequency) and is an important sound quality feature of speech. T0 is the fundamental period, that is, the duration of the periodic repetition of the speech signal (in seconds). This value is usually estimated by autocorrelation or other period detection methods.

[0047] E mis the energy of the mth frame. It represents the total energy of the speech signal in a certain frame (time period), and is usually used to represent the loudness or intensity of the speech signal. x[n] is the amplitude of the input speech signal at the nth sampling point. It is a discrete representation of the speech signal, usually a digitized signal. n is the index of the sampling point, indicating the position of the nth sampling point in the current frame. Usually, a frame contains multiple sampling points. N is the number of sampling points per frame. It determines the length of time contained in each frame, usually between 20 and 40 milliseconds.

[0048] A(z) is a polynomial representing the linear predictive coding (LPC) polynomial used to estimate the formants of the speech signal. Formants are the frequency components of the speech signal. These frequencies correspond to the resonant frequencies of the vocal tract, a k is the linear prediction coefficient (LPC coefficient), which indicates the k-th order LPC coefficient. The LPC coefficient is obtained by analyzing the autocorrelation function of the speech signal and is used to predict the value of the current sample point. p is the order of LPC, which indicates how many past sample points are used to predict the current sample point. The higher the order, the higher the prediction accuracy may be, but it is also more prone to overfitting. -k is a delay operator, which indicates a time delay of k sampling points. It is used to construct a prediction polynomial to represent the influence of past information of the speech signal on the current information.

[0049] The extracted speech features of the target speech segment are combined into a speech vector and transmitted to a pre-trained convolutional neural network. The output data of the pre-trained convolutional neural network is the probability of bullying.

[0050] The extracted features (MFCC, fundamental frequency, energy envelope, formant frequency) are combined into a feature vector as the input of the convolutional neural network (CNN). The architecture of CNN is: Input layer: The input layer accepts the speech feature vector. Convolution layer: A series of convolution layers perform convolution operations on the input features to extract higher-level features. Pooling layer: Usually after the convolution layer, the pooling layer is used for downsampling to reduce the dimension of the features. Fully connected layer: The feature map after convolution and pooling is converted to output probability through the fully connected layer. Output layer: The output layer uses the sigmoid activation function to output the probability of bullying. Training process: Forward propagation: The input feature vector is passed through the network to calculate the output probability. Loss function: The cross entropy loss function is used to measure the difference between the predicted result and the true label.

[0051] y i represents the true label, Represents the predicted probability, N is the number of samples, Back propagation: The gradient of the loss function with respect to each parameter in the network is calculated through the back propagation algorithm. Gradient descent: Use the gradient descent optimization algorithm (such as Adam or SGD) to update the weights and biases in the CNN and gradually minimize the loss function. Iteration: The process of gradually minimizing the loss function is iterated continuously until the loss function converges or the preset number of training rounds is reached.

[0052] Identifying whether there is a bullying risk in the target audio frequency band refers to: obtaining the probability of bullying in the target audio frequency band, and then comparing the probability of bullying with a preset risk threshold value one; if the probability of bullying is greater than or equal to the preset risk threshold value one, it is determined that there is a bullying risk; if the probability of bullying is less than the preset risk threshold value one, it is determined that there is no bullying risk.

[0053] The natural language processing module is used to convert the speech data corresponding to the multiple speech audio segments into multiple text paragraphs and analyze them to identify whether the target text paragraphs contain bullying risks;

[0054] The target speech segment is obtained, and then the frequency domain signal is converted back to the time domain signal through the inverse short-time Fourier transform to obtain the speech data of the corresponding time segment, and then the speech data of the time segment is converted into a text paragraph, and then the pre-trained sentiment analysis model is used to perform sentiment analysis on the text paragraph. The pre-trained sentiment analysis model outputs the probability of each preset sentiment category, and then the preset bullying vocabulary list is used to extract the bullying words in the text paragraph and the total number of times all the extracted bullying words appear in the text paragraph is counted, and finally the bullying occurrence coefficient is calculated. The bullying occurrence coefficient can be calculated based on the emotional tendency of the text and the number of occurrences of the bullying words. This coefficient can be used to measure the overall bullying level of the paragraph.

[0055] L represents the total number of words in the text paragraph, and BIC represents the bullying occurrence coefficient. N is the total number of times all bullying words extracted from the text paragraph appear in the text, indicating the frequency of words involving bullying behavior in the text paragraph. L is the total number of words in the text paragraph. This parameter is used for normalization to prevent deviations caused by different text lengths. Qi is the probability of the i-th preset emotion category, that is, the emotion probability output by the model. Different emotion categories may contribute differently to the BIC value. wi is the weight coefficient associated with the i-th emotion category. The weight coefficient is used to adjust the impact of different emotion categories on BIC. For example, negative emotions may be given greater weight than neutral emotions. k is the total number of preset emotion categories. For example, if the model can recognize positive, negative, and neutral emotions, k=3. The part is used to comprehensively consider the emotional distribution of the text paragraph. The weight coefficients wi of different emotions can be set according to specific needs to reflect the impact of emotions on the possibility of bullying. is the vocabulary density, which is used to calculate the density of bullying words in a text paragraph. A high density indicates that there is a greater possibility of bullying in the text paragraph. The size of the BIC value indicates the potential bullying risk in the text paragraph. The larger the value, the higher the possibility of bullying in the text paragraph.

[0056] The intelligent algorithm analysis module is used to verify the audio and video segments with bullying risks through deep learning algorithms, and then transmit the verification results to the alarm module;

[0057] Obtain the bullying risk recognition results of the audio segment corresponding to the voice data of the time segment, and the bullying risk recognition results of the corresponding text paragraph. If at least one of the recognition results is that there is a bullying risk, the voice data of the time segment is verified. If both recognition results are that there is no bullying risk, the voice data of the time segment is not verified. The two preliminary recognition results are used to determine whether further verification operations are required. By combining the analysis of both voice and text, the accuracy of recognition can be improved and false positives or negatives can be reduced. The conditions for the verification operation are: if any of the audio segment or text paragraph identifies the existence of bullying risks, the verification operation is triggered. If neither of the two identifies bullying risks, the verification is skipped. It ensures that further in-depth analysis is only performed when there are obvious signs of risk, optimizes the efficiency of the system, and avoids unnecessary computing overhead.

[0058] The verification operation refers to: combining the extracted speech features of the speech segment, the probability of bullying corresponding to the speech segment, each preset emotion category and its probability in the text paragraph, the frequency of bullying words, and the bullying occurrence coefficient to form a comprehensive feature vector; extracting speech features: including Mel-frequency cepstral coefficients (MFCC), fundamental frequency, energy envelope, and formant frequency. Obtaining the probability of bullying: using the probability of bullying in the speech segment calculated by the previous model. Emotion category and its probability: each preset emotion category and its probability corresponding to the text paragraph (such as positive, negative, and neutral). Bullying word frequency: counting the frequency of bullying words in the text paragraph. Bullying occurrence coefficient: the bullying occurrence coefficient of the text paragraph calculated according to the formula. The comprehensive feature vector combines the multi-dimensional information of speech and text, providing a comprehensive input to the deep learning model, enabling the model to more accurately determine whether there is bullying behavior.

[0059] The comprehensive feature vector is input into the pre-trained deep learning model, and the deep learning model outputs one of the two preset classification labels, indicating whether the comprehensive feature vector represents the existence of bullying behavior. The trained deep learning model is used for the final verification of bullying risks. Model architecture: Multilayer perceptron (MLP), convolutional neural network (CNN) or other suitable deep learning architectures can be used to process the comprehensive feature vector. Taking the multilayer perceptron (MLP) as an example, the input layer: receives the comprehensive feature vector. Hidden layer: multiple hidden layers process features and extract complex nonlinear relationships. Output layer: uses softmax or sigmoid activation function to output one of the two classification labels (for example: bullying risk exists / no bullying risk exists). Training process: Dataset preparation: Use a labeled dataset, in which each sample contains voice and text features, as well as the corresponding label (whether there is a bullying risk). Loss function: Use the cross entropy loss function to measure the difference between the predicted result and the true label. Optimization algorithm: Use optimization algorithms such as Adam or SGD to optimize the network parameters through the back propagation algorithm. Model training: Through multiple rounds of training, the model gradually learns the association between the comprehensive feature vector and bullying behavior. Reasoning process: The comprehensive feature vector of the new sample is input into the trained model. The model outputs the classification result, indicating whether the sample has the potential for bullying. The deep learning model can make a more accurate judgment on potential bullying behavior by comprehensively considering the multi-dimensional features of speech and text. This model not only improves the recognition accuracy of the system, but also reduces the possibility of misjudgment.

[0060] Suppose that after the speech data of a time segment is analyzed: Speech segment analysis: The probability of bullying is found to be 0.8 (high). Text paragraph analysis: The coefficient of bullying is calculated to be 0.6 (medium), and the probability of negative emotion = 0.7. Since at least one result (speech analysis) shows that there is a risk of bullying, the verification operation is triggered. Comprehensive feature vector construction: The extracted speech features (MFCC, fundamental frequency, etc.), the probability of bullying, the probability of emotion, the frequency of bullying words, and the coefficient of bullying are combined into a vector. Input deep learning model: Input the vector, and the model outputs the classification result "there is a risk of bullying".

[0061] The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms also includes an alarm module for deciding whether to issue or not to issue an alarm signal according to the verification result. The total number of bullying labels of all classification labels indicating the existence of bullying behavior is obtained. If the total number of bullying labels is zero, no alarm signal is issued. If the total number of bullying labels is not zero, the total number of bullying labels is divided by the preset signal level division value and the result is rounded up, and an alarm signal of the level corresponding to the rounded integer value is issued.

[0062] The specific steps are as follows: During the verification operation, the deep learning model will output a classification label indicating the presence or absence of bullying behavior. The number of all labels indicating the presence of bullying behavior is added up to obtain the total number of bullying labels. The total number of bullying labels reflects the intensity and frequency of bullying behaviors identified by the system. The larger this number is, the more bullying behaviors are detected in the analyzed voice clips or text paragraphs, and the higher the risk.

[0063] Decide whether to issue an alarm signal: If the total number of bullying tags is zero, it means that no bullying behavior has been detected in all the analyzed clips, and the system does not issue an alarm signal. If the total number of bullying tags is not zero, it means that there is a certain degree of bullying behavior, and the system will continue to calculate the alarm level. It ensures that the alarm is only issued when bullying behavior is detected, avoiding unnecessary alarm signals and reducing false alarm rates. Divide the total number of bullying tags by a preset signal level division value D. This division value is usually set according to the actual application scenario, indicating the minimum number of bullying tags required for each alarm level, and set T to the total number of bullying tags.

[0064] The integer value after rounding up represents the level of the alarm, and the system issues a corresponding alarm signal based on this level. This grading mechanism allows the system to dynamically adjust the alarm level according to the severity of the bullying behavior, providing a flexible response method. For example, a low-level alarm is issued when the bullying behavior is relatively minor, while severe bullying behavior will trigger a high-level alarm.

[0065] For example: Assume that the signal level division value D set by the system is 5. Total number of bullying tags: In a certain analysis, the total number of bullying tags detected by the system is T, which is 12. First, calculate the division result T / D=2.4, round the result up, and get the alarm level 3, and the system issues a level 3 alarm signal. In this example, the system detects 12 tags indicating bullying behavior and calculates that a level 3 alarm should be issued. Level 3 alarms indicate that there is moderate bullying behavior in the voice data or text, which may require the attention and intervention of system administrators or relevant personnel. The alarm module can effectively decide whether to issue an alarm signal based on the intensity of bullying behavior detected by the system, and ensure the accuracy and adaptability of the alarm by calculating the alarm level. This mechanism can flexibly respond to different degrees of bullying behavior, thereby providing users with more accurate and targeted protection measures.

[0066] Data recording module, the data recording module is used to record all collected voice data, the analysis results of the voice recognition module and the natural language processing module, the verification results of the intelligent algorithm analysis module and the alarm records of the alarm module.

[0067] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0068] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0069] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0070] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0071] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms is characterized by: It includes voice acquisition module, preprocessing module, voice recognition module, natural language processing module, intelligent algorithm analysis module and alarm module; The voice collection module is used to collect the user's voice data in real time, obtain the user's voice data set, and then transmit the user's voice data set to the preprocessing module; The preprocessing module is used to perform denoising and segmentation on the user's speech data set to obtain multiple speech audio segments, and then transmit the multiple speech audio segments to the speech recognition module and the natural language processing module respectively; The speech recognition module is used to extract and analyze speech features of multiple speech audio segments to identify whether the target speech audio segment contains bullying risks; The natural language processing module is used to convert the speech data corresponding to the multiple speech audio segments into multiple text paragraphs and analyze them to identify whether the target text paragraphs contain bullying risks; The intelligent algorithm analysis module is used to verify the audio and video segments with bullying risks through deep learning algorithms, and then transmit the verification results to the alarm module; The alarm module is used to decide whether to issue an alarm signal according to the verification result.

2. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 1 is characterized in that: It also includes a data recording module, which is used to record all collected voice data, analysis results of the voice recognition module and the natural language processing module, verification results of the intelligent algorithm analysis module and alarm records of the alarm module.

3. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 1 is characterized in that: The preprocessing module is used to segment the user's speech data set, which refers to: The speech data set is divided into multiple speech segments in the time-frequency domain by short-time Fourier transform.

4. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 3 is characterized in that: The speech recognition module is used to extract and analyze speech features of multiple speech audio segments respectively, which means: Extracting speech features of the target speech segment includes Mel frequency cepstral coefficients, fundamental frequency, energy envelope, and formant frequency; The extracted speech features of the target speech segment are combined into a speech vector and transmitted to a pre-trained convolutional neural network. The output data of the pre-trained convolutional neural network is the probability of bullying.

5. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 4 is characterized in that: Identifying whether the target audio segment has bullying risks means: The probability of bullying in the target audio frequency segment is obtained, and then the probability of bullying is compared with a preset hidden danger threshold value 1. If the probability of bullying is greater than or equal to the preset hidden danger threshold value 1, it is judged that there is a bullying hidden danger; if the probability of bullying is less than the preset hidden danger threshold value 1, it is judged that there is no bullying hidden danger.

6. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 5 is characterized in that: The natural language processing module is used to convert the speech data corresponding to the multiple speech segments into multiple text segments and analyze them. Get the target speech segment, then convert the frequency domain signal back to the time domain signal through inverse short-time Fourier transform and get the speech data of the corresponding time segment, then convert the speech data of the time segment into a text paragraph, then use the pre-trained sentiment analysis model to perform sentiment analysis on the text paragraph, the pre-trained sentiment analysis model outputs the probability of each preset sentiment category, then use the preset bullying vocabulary list to extract the bullying words in the text paragraph and count the total number of times all the extracted bullying words appear in the text paragraph, and finally calculate the bullying occurrence coefficient: Qi represents the probability of each preset emotion category i, wi represents the preset proportion coefficient corresponding to each emotion category i, k represents the total number of preset emotion categories, N represents the total number of times all bullying words appear in the text paragraph, L represents the total number of words in the text paragraph, and BIC represents the bullying occurrence coefficient.

7. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 6 is characterized in that: Identifying whether a target text paragraph contains bullying risks means: Obtain the bullying occurrence coefficient of the target text paragraph, and then compare the bullying occurrence coefficient with the preset hidden danger threshold value one. If the bullying occurrence coefficient is greater than or equal to the preset hidden danger threshold value two, it is judged that there is a bullying hidden danger. If the bullying occurrence coefficient is less than the preset hidden danger threshold value two, it is judged that there is no bullying hidden danger.

8. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 7 is characterized in that: The intelligent algorithm analysis module is used to verify the audio and video segments with bullying risks through deep learning algorithms. It refers to: Obtain the bullying risk recognition result of the voice segment corresponding to the voice data of the time segment, and the bullying risk recognition result of the corresponding text segment. If at least one of the recognition results indicates that there is a bullying risk, perform a verification operation on the voice data of the time segment. If both recognition results indicate that there is no bullying risk, do not perform a verification operation on the voice data of the time segment. The verification operation refers to: The extracted speech features of the speech segment, the probability of bullying occurrence corresponding to the speech segment, each preset emotion category of the text paragraph and its probability, the frequency of bullying words, and the bullying occurrence coefficient are combined to form a comprehensive feature vector; The comprehensive feature vector is input into the pre-trained deep learning model, and the deep learning model outputs one of two preset classification labels, indicating whether the comprehensive feature vector represents the existence of bullying behavior.

9. The AI ​​voice anti-bullying cloud platform application system integrating intelligent algorithms according to claim 8 is characterized in that: The alarm module is used to decide whether to issue an alarm signal based on the verification results: The total number of bullying labels of all classification labels indicating the existence of bullying behavior is obtained. If the total number of bullying labels is zero, no alarm signal is issued. If the total number of bullying labels is not zero, the total number of bullying labels is divided by the preset signal level division value and the result is rounded up, and an alarm signal of the level corresponding to the rounded integer value is issued.

Images