Anti-fraud method based on ASR speech recognition

Abstract

The present invention discloses an anti-fraud method based on ASR voice recognition, relates to the technical field of anti-fraud of ASR voice recognition, and solves the technical problem that the prior art uses a fraud database technology to perform fraud identification, but has poor fraud effect and low efficiency; the present invention performs fraud behavior analysis on the user's call voice once every fixed time period; extracts a voice signal to be analyzed, and converts the voice signal to be analyzed into voice text; identifies the voice features of the voice signal to be analyzed; calculates the fraud score of the voice signal to be analyzed based on the voice text and the voice features, and evaluates the fraud behavior according to the fraud score; and performs a comprehensive analysis in combination with the voice text content and the voice features, so as to more comprehensively evaluate the suspicious degree of the call, thereby improving the accuracy and efficiency of fraud identification, which is of great significance for protecting the communication security of individuals or enterprises.

Description

Technical Field

[0001] The present invention belongs to the field of anti-fraud of ASR speech recognition, and specifically relates to an anti-fraud method based on ASR speech recognition. Background Art

[0002] With the continuous development of the service industry and the diversification of marketing methods, more and more companies are using telemarketing to serve their customers and promote their products. Often, these companies will choose a call platform to facilitate their business. This puts higher demands on call platform companies, such as how to prevent telephone fraud, online chat fraud and other criminal activities.

[0003] In the prior art, the preventive measures against telephone fraud are mainly to establish a fraud number database through the telephone numbers commonly used by mobile phone fraudsters, and then identify whether the incoming call number is in the fraud number database. If the current incoming call number is in the database, the user answering the call is prompted that the current call is a fraud call. However, in the prior art, the fraudster can continue to commit fraud by changing the telephone number, so the anti-fraud effect is poor, and if the fraud number database is continuously updated, it is cumbersome and the anti-fraud efficiency is low. Therefore, the present invention provides an anti-fraud method based on ASR voice recognition. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an anti-fraud method based on ASR speech recognition, which is used to solve the problem that the prior art uses fraud database technology to identify fraud, but has poor fraud effect and low efficiency.

[0005] To achieve the above object, the first aspect of the present invention provides an anti-fraud method based on ASR speech recognition, comprising the following steps:

[0006] Step 1: Extract the speech signal to be analyzed and convert it into speech text;

[0007] Step 2: Identify the speech features of the speech signal to be analyzed; wherein the speech features include speech intensity and speech speed;

[0008] Step 3: Based on the voice text and voice information, calculate the fraud score of the voice signal to be analyzed, and analyze whether there is fraudulent behavior; if yes, evaluate the fraud level; if not, do nothing.

[0009] Preferably, the extracting the voice signal to be analyzed and converting the voice signal to be analyzed into voice text includes:

[0010] The user's call voice is marked as a voice signal to be analyzed every N time period, and the voice signal to be analyzed is converted into voice text in sequence using ASR technology.

[0011] By converting voice signals into text in real time, the present invention enables the system to timely analyze the content of the conversation, identify key words or phrases, and thus quickly respond or take action; this is particularly important for scenarios such as sentiment analysis in customer service centers, key information extraction in emergency call centers, and real-time meeting minutes generation; ASR technology can automatically and continuously convert voice into text, greatly improving work efficiency compared to manual dictation.

[0012] Preferably, identifying the speech intensity of the speech signal to be analyzed includes:

[0013] The speech signal to be analyzed is converted into digital form to obtain the digital audio signal to be analyzed; the digital audio signal to be analyzed is divided into several audio time frames, the average energy of each audio time frame is calculated, and the average energy of each audio time frame is calculated by the formula The speech intensity Lp of the speech signal to be analyzed is calculated; wherein Pi is the average energy of each audio time frame; P0 is the reference power; i is the number of each audio time frame, i=0, 1, ..., N, N is a positive integer.

[0014] The present invention takes advantage of the characteristics of digital signals that are highly anti-interference and easy to store and transmit, which is convenient for subsequent processing and analysis; voice signals of different sources and formats are uniformly converted into digital signals, which is convenient for standardized processing and algorithm application; the digital audio signal is divided into a plurality of audio time frames of fixed length in chronological order, the average energy of each audio time frame is analyzed and summed up to obtain the strength of the voice signal to be analyzed, thereby laying a foundation for the subsequent analysis of the fraud score of the voice signal.

[0015] Preferably, the average energy of each audio time frame is the mean square value of the amplitude of the digital signal in each audio time frame.

[0016] Preferably, identifying the speech rate of the speech signal to be analyzed includes:

[0017] According to the speech text of the speech signal to be analyzed, the number of syllables of the speech text is counted, and the speaking rate YV of the speech signal to be analyzed is calculated by the formula YV=YJS / T; wherein YJS is the number of syllables, that is, each Chinese character in the speech text is a syllable; T is the duration of the speech signal to be analyzed.

[0018] The present invention expresses the speaking speed as the number of syllables per unit time, providing a standardized measurement method, making it possible to compare the speaking speeds of different voice samples; the calculation is based on the actual content and duration of the voice signal, reducing the influence of subjective judgment and providing a more objective speaking speed evaluation index; and research has found that changes in speaking speed are related to the individual's emotional state and psychological stress level, which in turn indirectly reflects the possibility of fraud in the call.

[0019] Preferably, the step of calculating the fraud score of the voice signal to be analyzed based on the voice text and the voice features includes:

[0020] Inputting the speech text and speech features of the speech signal to be analyzed into the fraud model, and outputting the number of abnormal speech items in the speech signal to be analyzed; wherein the fraud model is constructed based on the MLP model; and,

[0021] By formula The fraud score ZPF of the voice signal to be analyzed is calculated; wherein SC is the score of a single abnormal voice item; M is the number of abnormal voice items in the voice signal to be analyzed; abnormal voice items include sensitive words, abnormal voice intensity and abnormal speaking speed; abnormal voice intensity refers to the voice intensity exceeding the preset intensity; abnormal speaking speed refers to the speaking speed exceeding the preset speaking speed.

[0022] The present invention comprehensively considers sensitive words and voice features to ensure comprehensive identification of potential fraudulent behaviors; utilizes the efficient computing characteristics of the MLP (multi-layer perceptron) model to complete the analysis of a large amount of voice data in a short time and detect abnormal items in a timely manner; by calculating the total score of abnormal voice items, it directly reflects the fraud possibility of the voice signal to be analyzed, facilitates intelligent decision-making, helps to detect fraudulent behaviors in a timely manner, and contributes to building a safe digital society.

[0023] Preferably, the fraud model is constructed based on the MLP model, including:

[0024] Extracting speech texts and speech features of several speech signals from historical data, as well as the number of abnormal speech items that exist;

[0025] The speech text and speech features of the speech signal are used as input data of the MLP model; the number of abnormal speech items is used as output data of the MLP model; the MLP model is trained based on the input data and output data to obtain a fraud model.

[0026] Through a large amount of historical data, the model of the present invention can learn patterns and rules related to fraud, so as to more accurately identify abnormal voice items; the sensitive words and voice features in the historical data provide rich training materials for the model, ensuring the effectiveness and reliability of its recognition ability; the model can automatically analyze the abnormal voice items in the input voice signal without manual intervention, greatly improving the detection efficiency.

[0027] Preferably, based on the fraud score, the analysis of whether there is fraudulent behavior includes:

[0028] The fraud score is compared with a preset score range. If the fraud score is lower than the preset score range, there is no fraud in the voice signal to be analyzed;

[0029] If the fraud score is within the preset score range, the fraud level of the voice signal to be analyzed is level one, which means there is a slight fraud phenomenon;

[0030] If the fraud score is greater than the preset score range, the fraud level of the voice signal to be analyzed is level two, which means that there is serious fraud.

[0031] By mapping fraud scores to different preset ranges, the present invention can clearly distinguish the mild and severe degrees of fraud, provide a basis for subsequent treatment measures, and convert abstract fraud possibilities into specific numerical values, making the evaluation process more objective and quantifiable, facilitating unified standards and cross-departmental collaboration; and according to different levels of fraud, the response strategy can be dynamically adjusted, such as monitoring minor fraud and taking immediate legal measures for serious fraud.

[0032] Preferably, when the fraud level is level one, a fraud warning signal is sent to the user terminal; when the fraud level is level two or the fraud level is level one for several consecutive times, an on-hook operation will be performed.

[0033] Preferably, when there is no fraud in several consecutive voice signals to be analyzed, the fraud scores of the several voice signals to be analyzed are accumulated, the accumulated fraud scores are compared with a preset score range, the fraud level is analyzed, and measures are taken.

[0034] Even if the fraud score is low in a single conversation, if minor anomalies occur frequently over a period of time, the accumulated score may reveal potential fraud patterns or attempts; further identifying fraud behaviors that attempt to evade immediate detection through scattered, low-frequency attempts, enhancing the long-term monitoring effectiveness of the system, and improving the timeliness and accuracy of alarms.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention extracts the voice signal to be analyzed and converts it into voice text; identifies the voice features of the voice signal to be analyzed; calculates the fraud score of the voice signal to be analyzed based on the voice text and the voice features, and analyzes whether there is fraud; if yes, the fraud level is evaluated; if not, there is no fraud; by comprehensively analyzing the three dimensions of voice text content, voice intensity and speech speed, the normality of the voice signal can be more comprehensively evaluated, and the accuracy of identifying fraud can be effectively improved;

[0037] 2. When there is no fraud in a number of consecutive voice signals to be analyzed, the present invention accumulates the fraud scores of the voice signals to be analyzed, compares the accumulated fraud scores with a preset score range, and analyzes whether there is fraud; it is used to identify potential fraud to avoid the situation where some fraudsters may not directly expose their intentions in a single call, but gradually build trust through multiple contacts and then commit fraud; by accumulating fraud scores, the system can capture this progressive strategy to prevent long-term potential threats from being ignored due to a single analysis failing to reach a threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] 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 or the description of the prior art 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.

[0039] Figure 1 It is a schematic diagram of the process of the present invention;

[0040] Figure 2 Schematic diagram of the speech intensity analysis method of the present invention;

[0041] Figure 3 It is a schematic diagram of the fraud score analysis method of the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. 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.

[0043] See also Figure 1 The first aspect of the present invention provides an anti-fraud method based on ASR speech recognition, comprising the following steps:

[0044] Step 1: Extract the speech signal to be analyzed and convert it into speech text;

[0045] The user's call voice is marked as a voice signal to be analyzed every N time period, and the voice signal to be analyzed is converted into voice text in sequence using ASR technology.

[0046] Specifically, step S1: the SIP switch freeswitch (a soft voice switch) transmits the 8K audio stream to the "audio acquisition" service in real time through the media_bug method;

[0047] Step S2: The "audio acquisition service" receives the 8K audio stream converted into a 16K audio stream in freeswitch, performs relevant noise reduction and VAD processing, and transmits the processed audio stream to the ASR service;

[0048] Step S3: The ASR service receives the 16K audio stream, recognizes the text in real time based on the context, and returns it to the "audio acquisition service".

[0049] Step S4: The "audio acquisition service" receives the text recognized by the ASR and transmits the text-related information to the MQ.

[0050] Step S5: Business processing: The business service receives the text through MQ and performs relevant processing according to the business.

[0051] See also Figure 2 ,Step 2: Identify the speech features of the speech signal to be analyzed; wherein the speech features include speech intensity and speech speed;

[0052] Perform analog-to-digital conversion on the speech signal to be analyzed to obtain the digital audio signal to be analyzed; divide the digital audio signal to be analyzed into several audio time frames, calculate the average energy of each audio time frame, and use the formula The speech intensity Lp of the speech signal to be analyzed is calculated; wherein Pi is the average energy of each audio time frame; P0 is the reference power; i is the number of each audio time frame, i=0, 1, ..., N, N is a positive integer.

[0053] It should be noted that the average energy of each audio time frame is the mean square value of the digital signal amplitude in each audio time frame; P0 is generally taken as 10 -12 , because this value is close to the lowest sound intensity that the human ear can perceive.

[0054] Specifically, first, an analog voice signal is converted into a digital audio signal through analog-to-digital conversion; analog-to-digital conversion includes three steps: sampling, quantization, and encoding, which converts the continuous analog signal into a discrete digital signal for computer processing;

[0055] Secondly, divide the digital audio signal into a series of continuous time segments of equal length, each of which is called a frame; assuming that the audio signal sampling rate is 44100 Hz and the frame length is 20 milliseconds, each frame will contain 882 sample points (44100×0.02=882).

[0056] For each frame of audio signal, the mean square value of the amplitude of all digital signals in the frame is the average energy of the audio signal in the frame. For example, if the sample points of the i-th frame are x1, x2, ..., xN, then the average energy of the frame Pi=[(x1-x0) ^2+(x2-x0) ^2 +…+(xN-x0) ^2 ] / N; and based on the reference power, the speech intensity of the speech signal, that is, the audio decibel size, is analyzed and obtained.

[0057] The present invention evaluates the audio decibel of the speech signal to be analyzed by speech intensity; the audio decibel is mainly related to the average energy of each frame of audio in the speech signal to be analyzed, and the reference power is a set value. If the average energy of each frame of audio is large and the reference power is a fixed value, it indicates that the energy of the speech signal to be analyzed is large, then the audio decibel of the speech signal is large, that is, the speech intensity is large.

[0058] According to the speech text of the speech signal to be analyzed, the number of syllables in the speech text is counted, and the speaking rate YV of the speech signal to be analyzed is calculated by the formula YV=YJS / T; where YJS is the number of syllables, that is, each Chinese character in the speech text is a syllable; T is the duration of the speech signal to be analyzed.

[0059] The present invention evaluates the speed of the speech signal to be analyzed by the speaking rate; the speed of the speech signal to be analyzed is mainly closely related to the number of syllables in the speech text of the speech signal to be analyzed and the duration of the speech signal to be analyzed. If the number of syllables in the speech text is large and the duration is short, indicating that more syllables are output in a shorter time, the speaking rate of the speech signal to be analyzed is fast, that is, the speaking rate value is large.

[0060] See also Figure 3 ,Step 3: Based on the voice text and voice information, calculate the fraud score of the voice signal to be analyzed, and analyze whether there is fraud; if yes, evaluate the fraud level; if not, do nothing.

[0061] Input the speech text and speech features of the speech signal to be analyzed into the fraud model, and output the number of abnormal speech items in the speech signal to be analyzed; wherein the fraud model is constructed based on the MLP model;

[0062] Among them, the fraud model construction process is to extract the voice text and voice features of several voice signals from historical data, as well as the number of abnormal voice items that exist;

[0063] The speech text and speech features of the speech signal are used as input data of the MLP model; the number of abnormal speech items is used as output data of the MLP model; the MLP model is trained based on the input data and output data to obtain a fraud model.

[0064] For example: collect a large amount of historical voice data, which should include known normal calls and fraudulent call samples, and each voice signal includes at least the following information (i.e., included voice items): voice text: text transcription of the call; voice features: extract features from the audio, such as speech speed and volume changes; set the judgment rules for abnormal voice items: such as sensitive word lists (such as "transfer", "Alipay", "WeChat", etc.), preset intensity (limited value of speaking volume) and preset speed (limited value of speaking speed); when sensitive words appear in the voice text or the intensity of the voice exceeds the preset intensity or the speaking speed exceeds the preset speed, the corresponding voice item is judged to be abnormal and marked as 1; if the corresponding voice item is normal, it is marked as 0;

[0065] The prepared data is used to train the MLP model, which consists of an input layer, one or more hidden layers, and an output layer. The input layer receives preprocessed speech text and speech features, and the output layer counts and outputs the number of abnormal speech items. During the training process, the weights of the model are adjusted to minimize the difference between the predicted value and the actual number of abnormal speech items.

[0066] And, through the formula The fraud score ZPF of the voice signal to be analyzed is calculated; wherein SC is the score of a single abnormal voice item; M is the number of abnormal voice items in the voice signal to be analyzed; abnormal voice items include sensitive words, abnormal voice intensity and abnormal speaking speed; sensitive words include "payment", "WeChat" and "Alipay"; abnormal voice intensity refers to the voice intensity exceeding the preset intensity; abnormal speaking speed refers to the speaking speed exceeding the preset speaking speed.

[0067] It should be noted that the score of a single abnormal speech item, that is, the score of each abnormal speech item, and the corresponding scores are set the same.

[0068] The present invention evaluates the possibility that the voice signal to be analyzed involves fraud through a fraud score; the possibility of fraud is mainly closely related to the number of abnormal voice items and the score of a single abnormal voice item; when the number of abnormal voice items is more and the score of a single abnormal voice item is larger, it indicates that there is more fraud information involved in the voice signal to be analyzed, and the possibility of fraud is greater, that is, the fraud score is larger.

[0069] The fraud score is compared with a preset score range. If the fraud score is lower than the preset score range, the voice signal to be analyzed does not contain fraud.

[0070] If the fraud score is within the preset score range, the fraud level of the voice signal to be analyzed is level one, which means there is a slight fraud phenomenon;

[0071] If the fraud score is greater than the preset score range, the fraud level of the voice signal to be analyzed is level two, which means that there is serious fraud.

[0072] Among them, when there is no fraud phenomenon in several consecutive voice signals to be analyzed, the fraud scores of several voice signals to be analyzed are accumulated, the accumulated fraud scores are compared with the preset score range, the fraud level is analyzed, and measures are taken.

[0073] When the fraud level is level one, a fraud warning signal will be sent to the user end; when the fraud level is level two or the fraud level is level one for several consecutive times, the on-hook operation will be performed.

[0074] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0075] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. Anti-fraud method based on ASR speech recognition, characterized in that: The following steps are involved: Step 1: Extract the speech signal to be analyzed and convert it into speech text; Step 2: Identify the speech features of the speech signal to be analyzed; wherein the speech features include speech intensity and speech speed; Step 3: Based on the voice text and voice features, calculate the fraud score of the voice signal to be analyzed, and analyze whether there is fraud; if yes, evaluate the fraud level; if not, do nothing; The fraud score calculation process of the voice signal is as follows: Inputting the speech text and speech features of the speech signal to be analyzed into the fraud model, and outputting the number of abnormal speech items in the speech signal to be analyzed; wherein the fraud model is constructed based on the MLP model; and, By formula Calculate the fraud score of the voice signal to be analyzed ;in, The score of a single abnormal speech item; M is the number of abnormal speech items in the speech signal to be analyzed; abnormal speech items include sensitive words, abnormal speech intensity and abnormal speaking speed; abnormal speech intensity refers to speech intensity exceeding the preset intensity; abnormal speaking speed refers to speaking speed exceeding the preset speaking speed.

2. The anti-fraud method based on ASR speech recognition according to claim 1, characterized in that: The extracting of the speech signal to be analyzed and converting the speech signal to be analyzed into speech text comprises: The user's call voice is marked as a voice signal to be analyzed every N time period, and the voice signal to be analyzed is converted into voice text in sequence using ASR technology.

3. The anti-fraud method based on ASR speech recognition according to claim 1, characterized in that: Identify the speech intensity of the speech signal to be analyzed, including: The speech signal to be analyzed is converted into digital form to obtain the digital audio signal to be analyzed; the digital audio signal to be analyzed is divided into several audio time frames, the average energy of each audio time frame is calculated, and the average energy of each audio time frame is calculated by the formula Calculate the speech intensity of the speech signal to be analyzed ;in, is the average energy of each audio time frame; is the reference power; i is the number of each audio time frame, i=0, 1,…, N, N is a positive integer.

4. The anti-fraud method based on ASR speech recognition according to claim 3 is characterized in that: The average energy of each audio time frame is the mean square value of the digital signal amplitude in each audio time frame.

5. The anti-fraud method based on ASR speech recognition according to claim 1, characterized in that: Identify the speech rate of the speech signal to be analyzed, including: According to the speech text of the speech signal to be analyzed, the number of syllables of the speech text is counted, and the formula Calculate the speech rate of the speech signal to be analyzed ;in, is the number of syllables, that is, each Chinese character in the speech text is a syllable; T is the duration of the speech signal to be analyzed.

6. The anti-fraud method based on ASR speech recognition according to claim 1, characterized in that: The fraud model is built based on the MLP model, including: Extracting speech texts and speech features of several speech signals from historical data, as well as the number of abnormal speech items that exist; The speech text and speech features of the speech signal are used as input data of the MLP model; the number of abnormal speech items is used as output data of the MLP model; the MLP model is trained based on the input data and output data to obtain a fraud model.

7. The anti-fraud method based on ASR speech recognition according to claim 1, characterized in that: Based on the fraud score, the analysis determines whether there is fraudulent behavior, including: The fraud score is compared with a preset score range. If the fraud score is lower than the preset score range, there is no fraud in the voice signal to be analyzed; If the fraud score is within the preset score range, the fraud level of the voice signal to be analyzed is level one, which means there is a slight fraud phenomenon; If the fraud score is greater than the preset score range, the fraud level of the voice signal to be analyzed is level two, which means that there is serious fraud.

8. The anti-fraud method based on ASR speech recognition according to claim 7, characterized in that: When the fraud level is level one, a fraud warning signal will be sent to the user end; when the fraud level is level two or the fraud level is level one for several consecutive times, the on-hook operation will be performed.

9. The anti-fraud method based on ASR speech recognition according to claim 7, characterized in that: When there is no fraud phenomenon in a number of consecutive voice signals to be analyzed, the fraud scores of the voice signals to be analyzed are accumulated, the accumulated fraud scores are compared with the preset score range, the fraud level is analyzed, and measures are taken.

Images