Potential user mining method and system based on voice analysis

Abstract

The invention provides a potential user mining method and system based on voice analysis, and relates to the technical field of voice analysis. Voice data published by a client on a platform is preprocessed, and the preprocessed voice data is analyzed from the aspects of emotional tendency and emotional abundance; an emotional abundant coefficient and an emotional tendency coefficient are respectively generated, a comprehensive emotional tendency coefficient of each section of effective voice is calculated through the emotional abundant coefficient and the emotional tendency coefficient, and the comprehensive emotional tendency coefficient is further analyzed to generate an emotional recommendation coefficient and a comprehensive recommendation coefficient; and evaluating potential mining values of the users according to the emotion recommendation coefficient and the comprehensive recommendation coefficient, and classifying the users. According to the method, the emotional abundance degree of the language is analyzed, the defect of tone analysis in traditional language analysis is overcome, evaluation of the mining value of the client is more accurate, pushing resources are saved, and high-quality clients are more easily mined.

Description

Technical Field

[0001] The present invention relates to the technical field of speech analysis, and in particular to a method and system for mining potential users based on speech analysis. Background Art

[0002] With the widespread application of voice data in customer service and market analysis, mining potential users through voice analysis has become an important means of corporate competition. However, existing user mining methods are mostly limited to the analysis of text information, with less use of voice data, and cannot fully reflect the emotional characteristics of users and their potential needs. At the same time, the existing technology for sentiment analysis of voice data usually lacks accurate multi-dimensional comprehensive indicators and cannot effectively quantify the emotional richness and tendency of users. In addition, the results of emotional recommendation are judged only by a single indicator, ignoring the statistical correlation between different voice data, resulting in low accuracy of user classification and affecting the effectiveness of potential user mining.

[0003] In the prior art, the publication number CN116303978B discloses a potential user mining method based on voice analysis, through S1, obtaining the overall mining evaluation of each historical product and the specific cases fed back by the customer, dividing the overall mining evaluation into a positive evaluation set and a negative evaluation set, and calculating the scores of different voice answer paths in the voice analysis model according to the distribution of the positive evaluation set and the negative evaluation set on each voice answer path and the voice answer path corresponding to the specific case, the voice answer path is the whole process of answering from the beginning of the outbound call to the end of the outbound call; S2, when promoting the new product of the customer to the outbound user, the voice answer of the outbound user is scored according to the updated voice analysis model to obtain the answer intention score, and the outbound user including the answer intention score is pushed to the customer. However, the prior art still has shortcomings. In addition to being affected by the content of the expression itself, the tone of voice is also an important influencing factor. The prior art only analyzes the content of the language and ignores the analysis of the tone. This analysis is incomplete, which will lead to inaccurate results of the voice analysis, thereby affecting the evaluation of the customer mining value, and then leading to the waste of push resources and the loss of high-quality customers.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for mining potential users based on speech analysis to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A potential user mining method based on speech analysis, the specific steps include:

[0008] Step 1: Obtain the voice data published by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate;

[0009] Step 2: extracting acoustic features of each valid speech segment of each speech data, obtaining acoustic feature data of each valid speech segment, and generating an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstrum coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment;

[0010] Step 3: extracting semantic features from each valid speech segment of each speech data, obtaining semantic feature data of each valid speech segment, and generating a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio;

[0011] Step 4: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; perform statistical processing on the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient;

[0012] Step 5: Sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-category potential users, second-category potential users, and third-category potential users by comparing the emotional recommendation coefficient and the comprehensive recommendation coefficient with the corresponding emotional recommendation threshold and the comprehensive recommendation threshold.

[0013] If the emotional recommendation coefficient of a voice data is greater than the emotional recommendation threshold, the user is defined as a first-class potential user. If the user's comprehensive recommendation coefficient is greater than the comprehensive threshold, the user is defined as a second-class potential user.

[0014] Furthermore, the specific logic for extracting effective speech is as follows: the speech signal is divided into 25ms time windows, each window contains N sampling points, and slides with a fixed step size of 10ms to form time frame speech data, the short-time energy and zero-crossing rate of the time frame speech data are calculated, the short-time energy threshold and the zero-crossing rate threshold are preset, and the speech with short-time energy greater than the short-time energy threshold and the zero-crossing rate less than the zero-crossing rate threshold is determined as effective speech, and each section of effective speech is extracted;

[0015] The specific formula for calculating short-time energy is:

[0016]

[0017] in, is the short-time energy of the time frame speech data, is the number of sampling points, For the The sound signal, is the sound signal index, ;

[0018] The specific formula for calculating the zero-crossing rate is:

[0019]

[0020]

[0021] in, is the zero-crossing rate of the time frame speech data.

[0022] Furthermore, the specific logic for generating the emotional abundance coefficient is as follows: calculating the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; generating the emotional abundance coefficient by using the fundamental frequency, spectrum centroid, power spectrum density, the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; the specific formula for generating the emotional abundance coefficient is as follows:

[0023]

[0024] in, For the emotional fullness coefficient, is the distribution variance of the 3rd to 13th Mel frequency cepstral coefficients, is the spectrum centroid, is the fundamental frequency, is the average of the first two Mel-frequency cepstral coefficients, is the power spectral density.

[0025] Furthermore, the specific logic for extracting semantic features is as follows: converting effective speech into text data, using natural language processing tools to extract words from the text, and obtaining the number of keyword occurrences, the number of positive words, and the number of negative words; the keywords include the category name and brand name of the product or service; analyzing the number of keyword occurrences to obtain the subject word occurrence rate, and adding 1 to the number of negative words to obtain the positive-negative word ratio; the specific formula for generating the subject word occurrence rate is as follows:

[0026]

[0027] in, is the occurrence rate of the subject word, The number of times the category name of the product or service appears. is the number of times the brand name appears, is the total number of words extracted from each valid speech segment, , is the weight coefficient, and , .

[0028] Furthermore, the specific logic for generating the sentiment tendency coefficient is: generating the sentiment tendency coefficient according to the occurrence rate of the subject word and the positive and negative word ratio; the specific formula for generating the sentiment tendency coefficient is:

[0029]

[0030] in, To generate the sentiment tendency coefficient, is the positive-negative vocabulary ratio, is the occurrence rate of the subject word.

[0031] Furthermore, the specific logic for generating the emotional recommendation coefficient is as follows: The specific formula for generating the emotional recommendation coefficient is as follows: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; Calculate the average value of the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient; The specific formula for generating the comprehensive emotional tendency coefficient is as follows:

[0032]

[0033] in, is the comprehensive sentiment tendency coefficient, To generate the sentiment tendency coefficient, is the emotional abundance coefficient; the specific formula for generating the emotional recommendation coefficient is:

[0034]

[0035] in, is the sentiment recommendation coefficient, For the same voice data The comprehensive emotional tendency coefficient of the effective speech segment, is a valid voice index in the same voice data; is the total number of valid speech segments in the same speech data. Emotional recommendation coefficient Further, the specific logic for generating the comprehensive recommendation coefficient is: summing the emotional recommendation coefficients of each speech data to obtain the comprehensive recommendation coefficient, and the specific formula for generating the comprehensive recommendation coefficient is:

[0036]

[0037] in, is the comprehensive recommendation factor, For the same customer The emotional recommendation coefficient of the voice data; The number of voice data for the same customer. An index of voice data for the same customer;

[0038] The specific logic for classifying users is as follows: condition A is defined as that the customer has at least one emotional recommendation coefficient greater than the recommended emotional threshold, and condition B is defined as that the comprehensive recommendation coefficient is greater than the comprehensive recommendation threshold. If the user meets both conditions A and B, the user is defined as a third-category potential user. If the user only meets condition A and does not meet time B, the user is defined as a second-category potential user. If the user only meets condition B and does not meet event A, the user is defined as a first-category potential user. If the user does not meet both conditions A and B, the user is defined as a non-potential user.

[0039] The present invention further provides a potential user mining system based on speech analysis, the system is used to implement the potential user mining method based on speech analysis, specifically comprising:

[0040] The effective segmentation module is used to obtain the voice data released by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate;

[0041] The emotional abundance analysis module is used to extract acoustic features of each valid speech segment of each speech data, obtain acoustic feature data of each valid speech segment, and generate an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstral coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment;

[0042] The tendency analysis module is used to extract semantic features from each valid speech segment of each speech data, obtain semantic feature data of each valid speech segment, and generate a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio;

[0043] A comprehensive analysis module is used to calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; the comprehensive emotional tendency coefficient of each valid speech segment of each speech data is statistically processed to obtain the emotional recommendation coefficient;

[0044] The customer classification module is used to sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-class potential users, second-class potential users, and third-class potential users by comparing the emotional recommendation coefficients and the comprehensive recommendation coefficients with the corresponding emotional recommendation thresholds and the comprehensive recommendation thresholds.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention analyzes the language to generate an emotion richness coefficient reflecting the emotion richness of the language, and further analyzes it in combination with the language content to determine the potential mining value of the customer. By analyzing the emotion richness, the deficiency of tone analysis in traditional language analysis is compensated, so that the assessment of the mining value of the customer is more accurate, pushing resources are saved, and it is easier to mine high-quality customers. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the overall method flow of the present invention.

[0048] Figure 2 It is a schematic diagram of the overall system structure of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0050] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0051] Example:

[0052] See also Figure 1 , the present invention provides a technical solution:

[0053] A potential user mining method based on speech analysis, the specific steps include:

[0054] Step 1: Obtain the voice data published by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate;

[0055] The specific logic for extracting effective speech is as follows: the speech signal is divided into 25ms time windows, each window contains N sampling points, and slides with a fixed step size of 10ms to form time frame speech data, the short-time energy and zero-crossing rate of the time frame speech data are calculated, the short-time energy threshold and the zero-crossing rate threshold are preset, and the speech with short-time energy greater than the short-time energy threshold and the zero-crossing rate less than the zero-crossing rate threshold is determined as effective speech, and each section of effective speech is extracted;

[0056] The short-time energy is set to be greater than the short-time energy threshold to ensure that the signal has sufficient energy, and the zero-crossing rate is set to be less than the zero-crossing rate threshold to ensure that the frequency characteristics of the signal are consistent with human speech. Meeting these two conditions at the same time ensures that human voice is extracted.

[0057] The specific formula for calculating short-time energy is:

[0058]

[0059] in, is the short-time energy of the time frame speech data, is the number of sampling points, For the The sound signal, is the sound signal index, ;

[0060] The specific formula for calculating the zero-crossing rate is:

[0061]

[0062]

[0063] in, is the zero-crossing rate of the time frame speech data.

[0064] Step 2: extracting acoustic features of each valid speech segment of each speech data, obtaining acoustic feature data of each valid speech segment, and generating an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstrum coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment;

[0065] The fundamental frequency is obtained by analyzing the autocorrelation function of the speech signal, specifically, obtaining the speech signal delay when the autocorrelation of the speech signal reaches the maximum, that is, the speech signal delay when the autocorrelation function takes the maximum value. At this time, the inverse of the speech signal delay is the fundamental frequency.

[0066] The calculation method of the spectrum centroid and power spectrum density is: obtain the spectrum data of the speech signal by performing Fourier transform on the speech signal, and input the spectrum data into MATLAB to obtain the spectrum centroid and power spectrum density.

[0067] The first 13 Mel-frequency cepstral coefficients are obtained by performing a short-time Fourier transform on each speech signal to obtain the spectrum information of the speech. The spectrum is mapped to the Mel scale, the logarithm of the Mel spectrum is taken, the discrete cosine transform is calculated, the energy of the Mel spectrum is compressed, a set of Mel-frequency cepstral coefficients is generated, and the first 13 Mel-frequency cepstral coefficients are taken. These are all conventional technical means in the art and will not be described in detail here.

[0068] The specific logic for generating the emotional abundance coefficient is as follows: calculate the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; generate the emotional abundance coefficient by taking the fundamental frequency, spectrum centroid, power spectrum density, the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; the specific formula for generating the emotional abundance coefficient is as follows:

[0069]

[0070] in, For the emotional fullness coefficient, is the distribution variance of the 3rd to 13th Mel frequency cepstral coefficients, is the spectrum centroid, is the fundamental frequency, is the average of the first two Mel-frequency cepstral coefficients, is the power spectrum density. The emotional abundance coefficient reflects the emotional abundance of speech. The larger its value, the more emotional the speech is. This coefficient quantifies the emotional expressiveness of speech from multiple dimensions such as pitch, timbre, frequency brightness and overall energy, ensuring the comprehensiveness and accuracy of the emotional abundance coefficient. The first two Mel-frequency cepstral coefficients reflect the overall energy distribution of the speech signal and the approximate shape of the spectrum envelope, including the pitch of the speech. The average value of the two reflects the stability of the speech pitch. The larger its value, the worse the pitch stability and the more passionate the speech. The fundamental frequency directly describes the frequency of vocal cord vibration and is the main indicator of the pitch of speech. A higher fundamental frequency is usually associated with emotional states such as excitement and agitation, while a lower fundamental frequency corresponds to calm and low emotional expressions. The average value of the first two Mel-frequency cepstral coefficients reflecting the pitch is highly correlated with the fundamental frequency of the pitch, both describing the overall tone of the speech and the intensity of the emotion. Therefore, they are combined to describe the intensity of the speech. The spectral centroid represents the energy center of the speech spectrum. When the spectral centroid is higher, the timbre of the speech is brighter and the emotional abundance is higher. The 3-13th Mel-frequency cepstral coefficients can reflect the timbre of speech. The distribution variance of the 3-13th Mel-frequency cepstral coefficients reflects the energy change in the high-frequency region of mid-speech, which is closely related to the timbre and emotional richness of speech. A larger variance indicates that the timbre of speech is more complex and the emotional expression is richer. The centroid of the spectrum and the distribution variance of the 3-13th Mel-frequency cepstral coefficients together describe the emotional expressiveness and complexity of speech. Therefore, they are combined to quantify the emotional expressiveness and complexity of this speech. The power spectral density reflects the overall energy distribution of the speech signal. The higher the energy, the greater the emotional intensity expressed by the speech. The function itself is monotonically decreasing, and if it is used as the denominator, the overall value is monotonically increasing. The dynamic range of energy is large, and taking the logarithm can compress the data to prevent a certain parameter from excessively dominating the result.

[0071] Step 3: extracting semantic features from each valid speech segment of each speech data, obtaining semantic feature data of each valid speech segment, and generating a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio;

[0072] The specific logic for extracting semantic features is: converting valid speech into text data, using natural language processing tools such as (LDA topic model and F-IDF topic model) to extract words from the text, and obtaining the number of keyword occurrences, the number of positive words, and the number of negative words; the keywords include the category name and brand name of the product or service; the positive words and negative words are identified through the existing language sentiment word library.

[0073] The number of times keywords appear is analyzed to get the occurrence rate of the subject words, and the positive word count plus the negative word count plus 1 is used to get the positive-negative word ratio; adding 1 is to prevent the positive word count ratio from being meaningless due to the lack of negative words;

[0074] The specific formula for generating the occurrence rate of subject words is:

[0075]

[0076] in, is the occurrence rate of the subject word, The number of times the category name of the product or service appears. is the number of times the brand name appears, is the total number of words extracted from each valid speech segment, , is the weight coefficient, and , The reason why the weight of the product or service category name occurrence rate is set to be more than three times the weight of the brand name occurrence rate is that when customers consume, they pay much more attention to their needs than to the brand.

[0077] The specific logic for generating the sentiment tendency coefficient is: generating the sentiment tendency coefficient according to the occurrence rate of the subject word and the positive and negative word ratio; the specific formula for generating the sentiment tendency coefficient is:

[0078]

[0079] in, To generate the sentiment tendency coefficient, is the positive-negative vocabulary ratio, is the occurrence rate of the keyword. The sentiment coefficient reflects the customer's sentiment towards the product or service. The larger the value, the greater the customer's positive sentiment towards the product or service. The generation of this coefficient can predict the customer's sentiment towards the recommended product and provide an important basis for mining potential customers. The positive and negative word ratio reflects the customer's evaluation of the product or service. The larger the value, the higher the customer's evaluation of the product or service.

[0080] The appearance rate of the subject word reflects the degree of demand for the pushed product or service. The larger the value is, the higher the customer's demand for the pushed product or service is. In the formula, the larger the value of the subject word appearance rate and the positive and negative word ratio is, the greater the emotional tendency is. If the subject word appearance rate is 0, it means that the user has demand for the product or service, so the overall emotional tendency is also 0. However, if the positive and negative word ratio is 0, although the customer evaluation is poor, the customer still has demand, so it is added by 1 to prevent the situation where the demand is ignored and potential customers are lost.

[0081] Step 4: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; perform statistical processing on the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient;

[0082] The specific logic for generating the emotional recommendation coefficient is as follows: The specific formula for generating the emotional recommendation coefficient is as follows: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; Calculate the average value of the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient; The specific formula for generating the comprehensive emotional tendency coefficient is as follows:

[0083]

[0084] in, is the comprehensive sentiment tendency coefficient, To generate the sentiment tendency coefficient, It is the emotion abundance coefficient; the comprehensive emotion tendency coefficient reflects the customer's comprehensive positive emotion tendency towards a product. The larger its value is, the greater the comprehensive emotion tendency is. This coefficient analyzes the customer's comprehensive tendency towards a product or service from the perspective of the customer's emotion abundance and emotion tendency. The larger the emotion tendency coefficient is, the stronger the customer's emotion towards the product or service is. The larger the emotion tendency coefficient is, the greater the customer's positive emotion tendency is. Combining the two, the greater the customer's inclination towards the product or service and the more intense the emotion is, the greater the customer's comprehensive positive emotion tendency is. This coefficient quantifies the customer's comprehensive emotion tendency towards the product or service, and can provide an important basis for the exploration of potential customers.

[0085] The specific formula for generating the sentiment recommendation coefficient is:

[0086]

[0087] in, is the sentiment recommendation coefficient, For the same voice data The comprehensive emotional tendency coefficient of the effective speech segment, is a valid voice index in the same voice data; The total number of valid voice segments in the same voice data. The emotional recommendation coefficient is simply an average value, which is used to reflect the comprehensive emotional tendency of customers towards a product or service at a time. Based on this emotional tendency, it can be determined whether to recommend a product or service to a customer and the degree of recommendation. It can provide an important basis for mining potential customers.

[0088] Step 5: Sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-category potential users, second-category potential users, and third-category potential users by comparing the emotional recommendation coefficient and the comprehensive recommendation coefficient with the corresponding emotional recommendation threshold and the comprehensive recommendation threshold.

[0089] The specific logic for generating the comprehensive recommendation coefficient is: summing up the emotional recommendation coefficients of each voice data to obtain the comprehensive recommendation coefficient. The specific formula for generating the comprehensive recommendation coefficient is:

[0090]

[0091] in, is the comprehensive recommendation factor, For the same customer The emotional recommendation coefficient of the voice data; The number of voice data for the same customer. It is the index of voice data of the same customer. This coefficient comprehensively analyzes the customer's feedback on multiple products, and then analyzes the customer's mining value. The larger the value, the greater the customer's potential mining value. The generation of this coefficient can provide an important basis for mining potential customers.

[0092] The specific logic for classifying users is as follows: Condition A is defined as the customer having at least one emotional recommendation coefficient greater than the recommended emotional threshold, and Condition B is defined as the comprehensive recommendation coefficient greater than the comprehensive recommendation threshold. If the user meets both Condition A and Condition B, the user is defined as the third type of potential user. This type of customer has a higher mining value and is recommended to be pushed first.

[0093] If the user only meets condition A, but not time B, then the user is defined as the second type of potential user. This type of customer has relatively low demand, but has a good response to high-quality products and services. If there are high-quality products and services, they will be pushed first;

[0094] If the user only meets condition B but not time A, then the user is defined as the first type of potential user; this type of customer service has stable needs, but the response to the product is relatively average, so push notifications are not given priority, but push notifications are recommended;

[0095] If a user does not meet both conditions A and B, the user is defined as a non-potential user. Such users have no mining value and are not recommended for push.

[0096] The present invention further provides a potential user mining system based on speech analysis, the system is used to implement the potential user mining method based on speech analysis, specifically comprising:

[0097] The effective segmentation module is used to obtain the voice data released by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate;

[0098] The emotional abundance analysis module is used to extract acoustic features of each valid speech segment of each speech data, obtain acoustic feature data of each valid speech segment, and generate an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstral coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment;

[0099] The tendency analysis module is used to extract semantic features from each valid speech segment of each speech data, obtain semantic feature data of each valid speech segment, and generate a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio;

[0100] A comprehensive analysis module is used to calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; the comprehensive emotional tendency coefficient of each valid speech segment of each speech data is statistically processed to obtain the emotional recommendation coefficient;

[0101] The customer classification module is used to sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-class potential users, second-class potential users, and third-class potential users by comparing the emotional recommendation coefficients and the comprehensive recommendation coefficients with the corresponding emotional recommendation thresholds and the comprehensive recommendation thresholds.

[0102] 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.

[0103] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0104] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0105] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the 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.

Claims

1. A potential user mining method based on speech analysis, characterized in that: The specific steps include: Step 1: Obtain the voice data published by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate; Step 2: extracting acoustic features of each valid speech segment of each speech data, obtaining acoustic feature data of each valid speech segment, and generating an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstrum coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment; Step 3: extracting semantic features from each valid speech segment of each speech data, obtaining semantic feature data of each valid speech segment, and generating a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio; Step 4: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; perform statistical processing on the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient; Step 5: Sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-category potential users, second-category potential users, and third-category potential users by comparing the emotional recommendation coefficient and the comprehensive recommendation coefficient with the corresponding emotional recommendation threshold and the comprehensive recommendation threshold.

2. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for extracting effective speech is as follows: the speech signal is divided into 25ms time windows, each window contains N sampling points, and slides with a fixed step size of 10ms to form time frame speech data, the short-time energy and zero-crossing rate of the time frame speech data are calculated, the short-time energy threshold and the zero-crossing rate threshold are preset, and the speech with short-time energy greater than the short-time energy threshold and the zero-crossing rate less than the zero-crossing rate threshold is determined as effective speech, and each section of effective speech is extracted; The specific formula for calculating short-time energy is: in, is the short-time energy of the time frame speech data, is the number of sampling points, For the The sound signal, is the sound signal index, ; The specific formula for calculating the zero-crossing rate is: in, is the zero-crossing rate of the time frame speech data.

3. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for generating the emotional abundance coefficient is as follows: calculate the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; generate the emotional abundance coefficient by taking the fundamental frequency, spectrum centroid, power spectrum density, the average of the first two Mel-frequency cepstral coefficients and the distribution variance of the 3rd to 13th Mel-frequency cepstral coefficients; the specific formula for generating the emotional abundance coefficient is as follows: in, For the emotional fullness coefficient, is the distribution variance of the 3rd to 13th Mel frequency cepstral coefficients, is the spectrum centroid, is the fundamental frequency, is the average of the first two Mel-frequency cepstral coefficients, is the power spectral density.

4. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for extracting semantic features is: converting valid speech into text data, using natural language processing tools to extract words from the text, and obtaining the number of keyword occurrences, the number of positive words, and the number of negative words; the keywords include the category name and brand name of the product or service; The occurrence frequency of keywords is analyzed to obtain the occurrence rate of the subject words, and the positive word ratio is obtained by adding 1 to the number of negative words. The specific formula for generating the occurrence rate of the subject words is: in, is the occurrence rate of the subject word, The number of times the category name of the product or service appears. is the number of times the brand name appears, is the total number of words extracted from each valid speech segment, , is the weight coefficient, and , .

5. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for generating the sentiment tendency coefficient is: generating the sentiment tendency coefficient according to the occurrence rate of the subject word and the positive and negative word ratio; the specific formula for generating the sentiment tendency coefficient is: in, To generate the sentiment tendency coefficient, is the positive-negative vocabulary ratio, is the occurrence rate of the subject word.

6. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for generating the emotional recommendation coefficient is as follows: The specific formula for generating the emotional recommendation coefficient is as follows: Calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; Calculate the average value of the comprehensive emotional tendency coefficient of each valid speech segment of each speech data to obtain the emotional recommendation coefficient; The specific formula for generating the comprehensive emotional tendency coefficient is as follows: in, is the comprehensive sentiment tendency coefficient, To generate the sentiment tendency coefficient, is the emotional abundance coefficient; the specific formula for generating the emotional recommendation coefficient is: in, is the sentiment recommendation coefficient, For the same voice data The comprehensive emotional tendency coefficient of the effective speech segment, is a valid voice index in the same voice data; It is the total number of valid speech segments in the same speech data.

7. The method for mining potential users based on speech analysis according to claim 1, characterized in that: The specific logic for generating the comprehensive recommendation coefficient is: summing up the emotional recommendation coefficients of each voice data to obtain the comprehensive recommendation coefficient. The specific formula for generating the comprehensive recommendation coefficient is: in, is the comprehensive recommendation factor, For the same customer The emotional recommendation coefficient of the voice data; The number of voice data for the same customer. An index of voice data for the same customer; The specific logic for classifying users is as follows: condition A is defined as that the customer has at least one emotional recommendation coefficient greater than the recommended emotional threshold, and condition B is defined as that the comprehensive recommendation coefficient is greater than the comprehensive recommendation threshold. If the user meets both conditions A and B, the user is defined as a third-category potential user. If the user only meets condition A and does not meet time B, the user is defined as a second-category potential user. If the user only meets condition B and does not meet event A, the user is defined as a first-category potential user. If the user does not meet both conditions A and B, the user is defined as a non-potential user.

8. A potential user mining system based on speech analysis, characterized by: The system is used to implement the potential user mining method based on speech analysis according to any one of claims 1 to 7, specifically comprising: The effective segmentation module is used to obtain the voice data released by each customer on the platform within half a year, and process the voice data to obtain time frame voice data; calculate the short-time energy and zero-crossing rate of the time frame voice data, and extract the effective voice of each voice data through the short-time energy and zero-crossing rate; The emotional abundance analysis module is used to extract acoustic features of each valid speech segment of each speech data, obtain acoustic feature data of each valid speech segment, and generate an emotional abundance coefficient through the acoustic feature data of each valid speech segment; the acoustic feature data includes the first 13 Mel frequency cepstral coefficients, fundamental frequency, spectrum centroid and power spectrum density of each valid speech segment; The tendency analysis module is used to extract semantic features from each valid speech segment of each speech data, obtain semantic feature data of each valid speech segment, and generate a sentiment tendency coefficient through the semantic feature data of each valid speech segment; the semantic feature data includes the occurrence rate of the subject word and the positive and negative word ratio; A comprehensive analysis module is used to calculate the comprehensive emotional tendency coefficient of each valid speech segment through the emotional abundance coefficient and the emotional tendency coefficient; the comprehensive emotional tendency coefficient of each valid speech segment of each speech data is statistically processed to obtain the emotional recommendation coefficient; The customer classification module is used to sum the emotional recommendation coefficients of each voice data to obtain a comprehensive recommendation coefficient, preset an emotional recommendation threshold and a comprehensive recommendation threshold, and classify users into non-potential users, first-class potential users, second-class potential users, and third-class potential users by comparing the emotional recommendation coefficients and the comprehensive recommendation coefficients with the corresponding emotional recommendation thresholds and the comprehensive recommendation thresholds.

Images