Online voice anti-harassment intelligent recognition method and system

Abstract

The present invention discloses an online voice anti-harassment intelligent recognition method and system, comprising: predicting the language type that a user who is about to make a call may use, and pre-selecting a language emotion judgment model according to the language type that the customer may use; using the pre-selected language emotion judgment model to perform real-time emotion analysis in real time during the call, and outputting a preset emotion analysis result; obtaining audio of the user's online voice for a certain duration at the beginning of the call, and converting the audio into text, judging the user's language type through the audio and text, and simultaneously performing dual real-time emotion analysis to obtain a real-time emotion analysis result; judging whether the twice selected language emotion judgment models are the same, and outputting the final emotion analysis result after the judgment. The present application combines the predicted language emotion judgment model with the real-time language emotion judgment model to perform online voice anti-harassment intelligent recognition, and the recognition result is more accurate, which can improve the quality of customer after-sales maintenance service.

Description

Technical Field

[0001] The present invention relates to the field of online voice anti-harassment intelligent recognition, and in particular to an online voice anti-harassment intelligent recognition method and system. Background Art

[0002] At present, in order to better serve customers, major platforms, enterprises and brands pay attention to customer after-sales return visits. After-sales return visits are an important part of the communication between enterprises and customers, and have many meanings:

[0003] 1. It can improve customer satisfaction: Through after-sales return visits, companies can understand the customer's satisfaction with the product or service, and promptly resolve customer questions and dissatisfaction, thereby improving customer satisfaction. Satisfied customers are more likely to become loyal customers of the company;

[0004] 2. Collect feedback information: After-sales return visits are an effective way to collect customer feedback. These feedbacks can help companies discover problems in products or services and provide a basis for product improvement and service optimization;

[0005] 3. It can enhance customer loyalty: Regular after-sales return visits can make customers feel the care and attention of the company, which helps to establish long-term and stable customer relationships and enhance customer loyalty;

[0006] 4. It can tap into potential needs: During the return visit, the company can discover the potential needs of customers through communication with them, provide them with more personalized products or services, and also provide a direction for the company to develop new products or services;

[0007] 5. It can promote secondary sales: Good after-sales return visits can not only maintain existing customer relationships, but also promote customers' secondary purchases and even attract new customers through word-of-mouth marketing;

[0008] 6. Can enhance brand image: Proactive after-sales service can enhance the company's image in the market and increase the brand's trust and reputation;

[0009] 7. Prevent and solve problems: Through after-sales follow-up visits, companies can learn about possible problems in advance and take measures to prevent or solve them, reduce customer dissatisfaction and complaints, and avoid negative reviews.

[0010] To sum up, after-sales follow-up visits can be used to collect customer service experience and push new products, thereby adjusting products, business models and service models, maintaining customer stickiness and supporting the long-term development of the company.

[0011] For after-sales follow-up, manual follow-up is costly, slow and inefficient, so intelligent calling systems came into being. However, conventional intelligent calling systems are not accurate enough in recognizing the emotions of the person being called. The recognition results output by the system lead to incorrect judgments about the customer's intentions, which goes against the original intention of after-sales follow-up and is unable to accurately identify potential customers and collect feedback. Summary of the invention

[0012] In view of the problems in the related art, the present invention proposes an online voice anti-disturbance intelligent recognition method and system to overcome the technical problems existing in the existing technology.

[0013] To this end, the specific technical solution adopted by the present invention is as follows:

[0014] An online voice anti-harassment intelligent recognition method comprises the following steps:

[0015] S1. Predict the language type that the user is likely to use, and pre-select a language emotion judgment model based on the language type that the customer is likely to use. The language types include Mandarin, dialects, and non-standard Mandarin.

[0016] S2, during the call, a pre-selected language emotion judgment model is used to perform real-time emotion analysis and output preset emotion analysis results;

[0017] S3: When the call starts, the audio of the user's online voice for a certain period of time is obtained, and the audio is converted into text. The user's language type is determined by the audio and text, and dual real-time sentiment analysis is performed simultaneously to obtain real-time sentiment analysis results;

[0018] S4, judging whether the two selected language emotion judgment models are the same, if they are the same, taking the preset emotion analysis result as the final emotion analysis result, if they are different, assigning weights to the preset emotion analysis result and the real-time emotion analysis result respectively based on the call duration and the time node of starting the dual real-time emotion analysis, and outputting the final emotion analysis result based on the weights;

[0019] S5. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions and requires manual secondary judgment.

[0020] This application obtains preset emotion analysis results by pre-selecting a language emotion judgment model for real-time emotion analysis, and combines the preset emotion analysis results with the real-time emotion analysis results obtained by performing double real-time emotion analysis based on the call audio during the call, and uses the language type of the caller as a judgment factor for selecting the language emotion judgment model. It considers multiple factors and combines them for emotion analysis. The emotion analysis recognition results are more accurate, which can improve the quality of customer after-sales maintenance services.

[0021] Furthermore, the language type that the user who is about to make a call may use includes:

[0022] S1.1. Obtain and analyze the user's identity information and historical delivery information to identify the user's registered residence, previous residence, and permanent residence;

[0023] S1.2, obtaining the user's registered dialect type based on the user's registered place of residence;

[0024] S1.3. Statistically analyze the length of residence and language type of the user's former and permanent residences;

[0025] S1.4. Based on the influence of the user's previous residence and permanent residence on the user's registered dialect, predict the type of language that the user is likely to use in the upcoming call.

[0026] Furthermore, the language emotion judgment model includes a Mandarin emotion judgment model, a dialect emotion judgment model and a non-standard Mandarin emotion judgment model. The dialect emotion judgment model and the non-standard Mandarin emotion judgment model are obtained by adjusting parameters and adding feature parameters on the basis of the Mandarin emotion judgment model. The Mandarin emotion judgment model performs text field analysis based on a large language model to realize emotion judgment.

[0027] Furthermore, adjusting parameters includes: adjusting parameters of sensitive words, commendatory words, and emotional expression words in the Mandarin emotion judgment model according to the word usage habits of dialects in different provinces and cities, so as to adapt to the language habits of different provinces and cities.

[0028] Furthermore, additional feature parameters include: adding pitch, volume, speaking speed, intonation and ending sound habit features to the Mandarin emotion judgment model according to the emotion expression habits of dialect speech in different provinces and cities, so as to adapt to the emotion expression habits of dialect speech in different provinces and cities.

[0029] Furthermore, in step S3, the audio duration of the user's online voice obtained at the beginning of the call for conversion into text is determined based on the average call duration of historical customer service return visits and the duration of the language type recognized by the large language model.

[0030] Furthermore, in step S4, if the language emotion judgment models selected twice are different, the total call duration T and the analysis duration T1 after the real-time language emotion judgment model is selected are obtained, and the ratio of T1 to T is calculated.

[0031] B1: When T is within the lowest period, the low segment is directly output, and the user is judged to have negative emotions and thinks he is being harassed;

[0032] B2: When T is in the analyzable period, if the ratio of T1 to T is greater than 0.8, the preset sentiment analysis result is used as the final sentiment analysis result. If the ratio of T1 to T is between 0.5 and 0.8, the preset sentiment analysis result score accounts for 70-80% of the total score, and the real-time sentiment analysis result score accounts for 20-30% of the total score. If the ratio of T1 to T is below 0.5, the preset sentiment analysis result score accounts for 60-90% of the total score, and the real-time sentiment analysis result score accounts for 10-40% of the total score.

[0033] B3: When T is greater than the analyzable period and the real-time language emotion judgment model can be identified, a high score segment is directly output to determine that the user has positive emotions and is willing to continue communicating.

[0034] According to another aspect of the present invention, there is also provided an online voice anti-harassment intelligent recognition system, comprising:

[0035] Language emotion judgment model pre-selection module: used to predict the language type that the user who is about to call may use, and pre-select a language emotion judgment model according to the language type that the customer may use. The language types include Mandarin, dialects and non-standard Mandarin;

[0036] Preset sentiment analysis module: used to perform real-time sentiment analysis using a pre-selected language sentiment judgment model during a call, and output preset sentiment analysis results;

[0037] Real-time sentiment analysis module: It is used to obtain the audio of the user's online voice for a certain period of time at the beginning of the call, convert the audio into text, determine the user's language type through the audio and text, and simultaneously perform dual real-time sentiment analysis to obtain real-time sentiment analysis results;

[0038] Sentiment analysis selection module: used to determine whether the two selected language emotion judgment models are the same. If they are the same, the preset sentiment analysis result is used as the final sentiment analysis result. If they are different, the preset sentiment analysis result and the real-time sentiment analysis result are weighted based on the call duration and the time point when the dual real-time sentiment analysis starts, and the final sentiment analysis result is output based on the weight.

[0039] Sentiment analysis output module: used to output the final sentiment analysis results. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions and requires manual secondary judgment.

[0040] The beneficial effects of the present invention are:

[0041] (1) This application combines the preset emotion analysis results obtained by pre-selecting a language emotion judgment model for real-time emotion analysis with the real-time emotion analysis results obtained by performing dual real-time emotion analysis based on the call audio during the call, and takes the language type of the caller as a judgment factor for selecting the language emotion judgment model. It considers multiple factors and combines them for emotion analysis. The emotion analysis recognition results are more accurate, which can improve the quality of customer after-sales maintenance services.

[0042] (2) This application identifies the user's registered residence, former residence, and permanent residence and determines the user's language type based on the length of time the user has lived there. This takes into account the impact of language habits in different regions on the user's language type within the range of the judgment factor for selecting the language type, and can more accurately determine the user's language type.

[0043] (3) The language emotion judgment model of this application includes Mandarin emotion judgment model, dialect emotion judgment model and non-standard Mandarin emotion judgment model. The appropriate language emotion judgment model is selected based on the emotional differentiation of different language types, and the emotion judgment is more accurate;

[0044] (4) The dialect emotion judgment model and non-standard Mandarin emotion judgment model of the present application are obtained by adjusting parameters and adding feature parameters on the basis of the Mandarin emotion judgment model. The model is adjusted by combining the word usage habits of dialects in different provinces and cities and the voice emotion expression habits of dialects in different provinces and cities, so that the language emotion judgment model can recognize emotions more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 This is a flow chart of an online voice anti-harassment intelligent recognition method according to the first embodiment of the present invention;

[0047] Figure 2 This is a module diagram of an online voice anti-harassment intelligent recognition system according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0048] 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.

[0049] According to an embodiment of the present invention, an online voice anti-disturbance intelligent recognition method and system are provided. Embodiment 1

[0050] like Figure 1 As shown, an online voice anti-harassment intelligent recognition method includes the following steps:

[0051] S1. First, before the call, predict the language type that the user is likely to use. Specifically:

[0052] S1.1. Obtain and analyze the user's identity information and historical delivery information to identify the user's registered residence, previous residence, and permanent residence;

[0053] S1.2, obtaining the user's registered dialect type based on the user's registered place of residence;

[0054] S1.3. Statistically analyze the length of residence and language type of the user's former and permanent residences;

[0055] S1.4. Based on the influence of the user's previous residence and permanent residence on the user's registered dialect, predict the type of language that the user is likely to use in the upcoming call.

[0056] Then, a language emotion judgment model is pre-selected according to the language type that the customer may use. The language types include Mandarin, dialects and non-standard Mandarin. The language emotion judgment model is designed based on the language type, including the Mandarin emotion judgment model, the dialect emotion judgment model and the non-standard Mandarin emotion judgment model. The dialect emotion judgment model and the non-standard Mandarin emotion judgment model are obtained by adjusting parameters and adding feature parameters on the basis of the Mandarin emotion judgment model. The Mandarin emotion judgment model performs text field analysis based on a large language model to realize emotion judgment.

[0057] Among them, the adjustment parameters include: adjusting the parameters of the emotional expression differences of sensitive words, commendatory words, and emotional expression words in the Mandarin emotion judgment model according to the word usage habits of dialects in different provinces and cities, so as to adapt to the language habits of different provinces and cities; the additional feature parameters include: adding pitch, volume, speaking speed, intonation, and tail sound habit features to the Mandarin emotion judgment model according to the emotional expression habits of dialect voices in different provinces and cities, so as to adapt to the emotional expression habits of language dialect voices in different provinces and cities.

[0058] S2. After the call starts, a pre-selected language emotion judgment model is used to perform real-time emotion analysis during the call, and a preset emotion analysis result is output.

[0059] S3. At the beginning of the call, the audio of the user's online voice for a certain duration is obtained, and the audio is converted into text. The user's language type is determined by the audio and text, and dual real-time sentiment analysis is performed simultaneously to obtain real-time sentiment analysis results. Specifically, the audio duration of the user's online voice obtained at the beginning of the call for conversion into text is determined based on the average call duration of historical customer service return visits and the duration of the language type recognized by the large language model.

[0060] S4. Then determine whether the two selected language emotion judgment models are the same. If they are the same, the preset emotion analysis result is used as the final emotion analysis result. If they are different, weights are assigned to the preset emotion analysis result and the real-time emotion analysis result based on the call duration and the time point when the dual real-time emotion analysis is started, and the final emotion analysis result is output based on the weights. Specifically:

[0061] If the language emotion judgment models selected twice are different, obtain the total call duration T and the analysis duration T1 after selecting the real-time language emotion judgment model, and calculate the ratio of T1 to T.

[0062] B1: When T is within the lowest period, the low segment is directly output, and the user is judged to have negative emotions and thinks he is being harassed;

[0063] B2: When T is in the analyzable period, if the ratio of T1 to T is greater than 0.8, the preset sentiment analysis result is used as the final sentiment analysis result. If the ratio of T1 to T is between 0.5 and 0.8, the preset sentiment analysis result score accounts for 70-80% of the total score, and the real-time sentiment analysis result score accounts for 20-30% of the total score. If the ratio of T1 to T is below 0.5, the preset sentiment analysis result score accounts for 60-90% of the total score, and the real-time sentiment analysis result score accounts for 10-40% of the total score.

[0064] S5. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions. The sentiment analysis score output by step S4 is judged to be any one of the low, medium and high segments to obtain the sentiment analysis result. Embodiment 2

[0065] When performing online voice anti-harassment intelligent recognition based on the method of Example 1, first, the registered permanent residence, former residence and permanent residence of user A are obtained based on the user's delivery information. It is assumed that the registered permanent residence of user A is in the Northeast region, the former residence and permanent residence are in the Northeast and Hangzhou respectively, and he has lived in Hangzhou for more than 5 consecutive years. Based on the characteristics of the Northeast dialect and the influence of the standard Mandarin commonly used language in Hangzhou on the Northeast dialect, it is predicted that the language type of user A is non-standard Mandarin, and the language emotional habits tend to be Northeast characteristics. Therefore, it is predicted that the language emotion judgment model suitable for user A is a non-standard Mandarin emotion judgment model, and the pre-selected non-standard Mandarin emotion judgment model is used in real time during the call to perform real-time emotion analysis, and output a preset emotion analysis result.

[0066] Since the predicted language emotion judgment model may have errors, the accuracy of emotion recognition using only the predicted language emotion judgment model is poor, so it is necessary to combine the real-time audio to re-judge and select the language emotion judgment model, which requires a certain length of recording for real-time analysis. The length of the recording needs to be sufficient to identify the language emotion judgment model, and it must also be selected based on the average call duration of historical customer service follow-up visits. For example, the average call duration of historical customer service follow-up visits is 5s, but the time required to identify the language emotion judgment model exceeds 5s. It may be difficult to identify the language emotion judgment model in real time. Therefore, if this happens, a model with a faster recognition speed must be selected. The language emotion judgment models used in this application include Mandarin emotion judgment models, dialect emotion judgment models, and non-standard Mandarin emotion judgment models. The dialect emotion judgment models and non-standard Mandarin emotion judgment models are obtained by adjusting parameters and adding feature parameters based on the Mandarin emotion judgment model. The Mandarin emotion judgment model performs text field analysis based on a large language model to achieve emotion judgment, and has the fastest analysis speed. For the parameter adjustment of the dialect emotion judgment model and the non-standard Mandarin emotion judgment model, it is necessary to adjust the parameters of the Mandarin emotion judgment model based on the emotional expression differences of sensitive words, commendatory words, and emotional expression words. At the same time, it is necessary to add the pitch, volume, speaking speed, intonation, and tail sound habit features to the Mandarin emotion judgment model to adapt to the language dialect speech emotion expression habits of different provinces and cities. The above parameters are arranged and combined to form a plurality of adjustment parameter sets with different numbers of adjustment parameters to adjust the Mandarin emotion judgment model to obtain the dialect emotion judgment model and the non-standard Mandarin emotion judgment model. For example, only the tail sound and commendatory words are used as parameters to adjust the parameters of the model, or sensitive words, commendatory words, emotional expression words, as well as pitch, volume, speaking speed, intonation, and tail sound habits are used as parameters to adjust the parameters of the model. The more parameters there are, the slower the recognition speed. Therefore, compared with the Mandarin emotion judgment model, the recognition speed of the dialect emotion judgment model and the non-standard Mandarin emotion judgment model is slower. Therefore, it is necessary to match the language emotion judgment model after appropriate parameter adjustment based on the average call duration of historical customer service return visits.

[0067] Assume that the average call duration of historical customer service follow-up calls is 20 seconds, and sensitive words, commendatory words, emotional expression words, as well as pitch, volume, speaking speed, intonation, and ending sound habits are used as parameters to adjust the model parameters. The recognition language emotion judgment model requires an audio duration of 3 seconds, and the recognition time is 2 seconds. Therefore, it is judged that the non-standard Mandarin emotion judgment model can be selected for real-time emotion recognition.

[0068] After the call starts, the audio of the user's online voice is read for 3 seconds and converted into text. The user's language type is determined by the audio and text. That is, after 5 seconds, the dual real-time non-standard Mandarin emotion judgment model is used to simultaneously perform real-time emotion analysis to obtain real-time emotion analysis results.

[0069] After the call is over, the preset sentiment analysis results and the real-time sentiment analysis results are obtained. At this time, it is necessary to determine whether the two selected language sentiment judgment models are the same. If they are the same, it means that the predicted language sentiment judgment model is accurate, and the preset sentiment analysis result is used as the final sentiment analysis result. If they are different, it is necessary to assign weights to the preset sentiment analysis result and the real-time sentiment analysis result according to the call duration to make the root result more accurate. Assuming that the total call duration is 28s, the analysis duration after selecting the real-time language sentiment judgment model is 20s, the ratio of 20s to 28s is greater than 0.8, and 28s is in the analyzable period, so the preset sentiment analysis result is selected as the final sentiment analysis result. Assuming that the sentiment analysis result score is 86 points, the low segment is set to 0-40 points, the middle segment is 40-60 points, and the high segment is 60-100 points. The low segment represents negative emotions and feels harassed, the high segment represents positive emotions and is willing to continue communicating, and the middle segment represents stable emotions. The output sentiment analysis result is positive emotions and willingness to continue communicating. The user can be recorded as an intended customer and continue to provide accurate services later. Embodiment 3

[0070] An online voice anti-harassment intelligent recognition system, used to implement an online voice anti-harassment intelligent recognition method of embodiment 1 or 2, comprising:

[0071] Language emotion judgment model pre-selection module: used to predict the language type that the user who is about to call may use, and pre-select a language emotion judgment model according to the language type that the customer may use. The language types include Mandarin, dialects and non-standard Mandarin;

[0072] Preset sentiment analysis module: used to perform real-time sentiment analysis using a pre-selected language sentiment judgment model during a call, and output preset sentiment analysis results;

[0073] Real-time sentiment analysis module: It is used to obtain the audio of the user's online voice for a certain period of time at the beginning of the call, convert the audio into text, determine the user's language type through the audio and text, and simultaneously perform dual real-time sentiment analysis to obtain real-time sentiment analysis results;

[0074] Sentiment analysis selection module: used to determine whether the two selected language emotion judgment models are the same. If they are the same, the preset sentiment analysis result is used as the final sentiment analysis result. If they are different, the preset sentiment analysis result and the real-time sentiment analysis result are weighted based on the call duration and the time point when the dual real-time sentiment analysis starts, and the final sentiment analysis result is output based on the weight.

[0075] Sentiment analysis output module: used to output the final sentiment analysis results. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions and requires manual secondary judgment.

[0076] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An online voice anti-harassment intelligent recognition method, characterized in that: The following steps are involved: S1. Predict the language type that the user is likely to use, and pre-select a language emotion judgment model based on the language type that the customer is likely to use. The language types include Mandarin, dialects, and non-standard Mandarin. S2, during the call, a pre-selected language emotion judgment model is used to perform real-time emotion analysis and output preset emotion analysis results; S3: When the call starts, the audio of the user's online voice for a certain period of time is obtained, and the audio is converted into text. The user's language type is determined by the audio and text, and dual real-time sentiment analysis is performed simultaneously to obtain real-time sentiment analysis results; S4, judging whether the two selected language emotion judgment models are the same, if they are the same, taking the preset emotion analysis result as the final emotion analysis result, if they are different, assigning weights to the preset emotion analysis result and the real-time emotion analysis result respectively based on the call duration and the time node of starting the dual real-time emotion analysis, and outputting the final emotion analysis result based on the weights; S5. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions and requires manual secondary judgment.

2. According to claim 1, an online voice anti-harassment intelligent recognition method is characterized in that: Predict the language types that the user might use in the upcoming call including: S1.

1. Obtain and analyze the user's identity information and historical delivery information to identify the user's registered residence, previous residence, and permanent residence; S1.2, obtaining the user's registered dialect type based on the user's registered place of residence; S1.

3. Statistically analyze the length of residence and language type of the user's former and permanent residences; S1.

4. Based on the influence of the user's previous residence and permanent residence on the user's registered dialect, predict the type of language that the user is likely to use in the upcoming call.

3. According to claim 1, an online voice anti-harassment intelligent recognition method is characterized in that: The language emotion judgment model includes the Mandarin emotion judgment model, the dialect emotion judgment model and the non-standard Mandarin emotion judgment model. The dialect emotion judgment model and the non-standard Mandarin emotion judgment model are obtained by adjusting parameters and adding feature parameters on the basis of the Mandarin emotion judgment model. The Mandarin emotion judgment model performs text field analysis based on a large language model to realize emotion judgment.

4. According to claim 3, an online voice anti-harassment intelligent recognition method is characterized in that: The adjustment parameters include: adjusting the parameters of sensitive words, commendatory words, and emotional expression words in the Mandarin emotion judgment model according to the word usage habits of dialects in different provinces and cities, so as to adapt to the language habits of different provinces and cities.

5. According to claim 3, an online voice anti-harassment intelligent recognition method is characterized in that: Additional feature parameters include: adding pitch, volume, speaking speed, intonation and ending sound habit features to the Mandarin emotion judgment model according to the emotion expression habits of dialect speech in different provinces and cities, so as to adapt to the emotion expression habits of dialect speech in different provinces and cities.

6. According to claim 1, an online voice anti-harassment intelligent recognition method is characterized in that: In step S3, the audio duration of the user's online voice that is obtained for conversion into text at the beginning of the call is determined based on the average call duration of historical customer service return visits and the duration of the language type recognized by the large language model.

7. The online voice anti-harassment intelligent recognition method according to claim 6 is characterized in that: In step S4, if the language emotion judgment models selected twice are different, the total call duration T and the analysis duration T1 after the real-time language emotion judgment model is selected are obtained, and the ratio of T1 to T is calculated. B1: When T is within the lowest period, the low segment is directly output, and the user is judged to have negative emotions and thinks he is being harassed; B2: When T is in the analyzable period, if the ratio of T1 to T is greater than 0.8, the preset sentiment analysis result is used as the final sentiment analysis result. If the ratio of T1 to T is between 0.5 and 0.8, the preset sentiment analysis result score accounts for 70-80% of the total score, and the real-time sentiment analysis result score accounts for 20-30% of the total score. If the ratio of T1 to T is below 0.5, the preset sentiment analysis result score accounts for 60-90% of the total score, and the real-time sentiment analysis result score accounts for 10-40% of the total score. B3: When T is greater than the analyzable period and the real-time language emotion judgment model can be identified, a high score segment is directly output to determine that the user has positive emotions and is willing to continue communicating.

8. An online voice anti-harassment intelligent recognition system, used to implement an online voice anti-harassment intelligent recognition method according to any one of claims 1 to 7, characterized in that: include: Language emotion judgment model pre-selection module: used to predict the language type that the user who is about to call may use, and pre-select a language emotion judgment model according to the language type that the customer may use. The language types include Mandarin, dialects and non-standard Mandarin; Preset sentiment analysis module: used to perform real-time sentiment analysis using a pre-selected language sentiment judgment model during a call, and output preset sentiment analysis results; Real-time sentiment analysis module: It is used to obtain the audio of the user's online voice for a certain period of time at the beginning of the call, convert the audio into text, determine the user's language type through the audio and text, and simultaneously perform dual real-time sentiment analysis to obtain real-time sentiment analysis results; Sentiment analysis selection module: used to determine whether the two selected language emotion judgment models are the same. If they are the same, the preset sentiment analysis result is used as the final sentiment analysis result. If they are different, the preset sentiment analysis result and the real-time sentiment analysis result are weighted based on the call duration and the time point when the dual real-time sentiment analysis starts, and the final sentiment analysis result is output based on the weight. Sentiment analysis output module: used to output the final sentiment analysis results. The final sentiment analysis results include low, medium and high segments. The low segment represents negative emotions and the person feels harassed. The high segment represents positive emotions and the person is willing to continue communicating. The medium segment represents stable emotions and requires manual secondary judgment.

Images