Voice signal processing method and system for fraud signal identification

Abstract

The present invention relates to the field of speech signal processing technology, and in particular to a speech signal processing method and system for fraud signal identification, by converting a speech signal into multiple IMF components; performing data segmentation on each IMF component, analyzing the fluctuation characteristics of the data segment to obtain a fluctuation coefficient; obtaining a suspected noise data segment according to the fluctuation coefficient combined with the fluctuation characteristics of noise data; analyzing the proportion of the suspected noise data segment in the IMF component to obtain a noise contribution; adjusting the wavelet coefficient threshold of the IMF component in wavelet denoising according to the noise contribution; performing wavelet transform on the IMF component according to the wavelet coefficient threshold to remove noise data, and superimposing and reconstructing the denoised IMF component to obtain a denoised speech simulation signal. The scheme of the present invention can more accurately remove possible noise, and ultimately provide high-quality speech audio for accurate identification of fraud signals.

Description

Technical Field

[0001] The present invention relates to the technical field of voice signal processing, and in particular to a voice signal processing method and system for fraud signal recognition. Background Art

[0002] With the rapid popularization of mobile devices such as mobile phones, personal identity information is exposed on the Internet. After the phone number is leaked, telephone harassment is carried out through communication forms such as calls and text messages to deceive users' trust, causing serious damage to the normal communication environment.

[0003] The AI ​​speech recognition algorithm model can effectively identify fraudulent and harassing calls quickly and curb the frequency of bad phone communications. However, the existing speech recognition algorithm model is limited by the quality of the audio data samples of fraudulent and harassing calls, and its recognition accuracy fluctuates greatly in noisy scenarios. In the process of denoising the telephone audio signal, denoising is performed by setting a fixed wavelet threshold, which leads to greater noise interference in the telephone audio signal, greatly affecting the accurate recognition of fraudulent audio. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a speech signal processing method and system for fraud signal recognition.

[0005] According to the first aspect of the embodiment of the present application, the speech signal processing method for fraud signal identification provided by the present application adopts the following technical solution: the method comprises:

[0006] Collecting a voice signal input by a user, converting the voice signal into a voice analog signal, performing EMD decomposition on the voice analog signal, and obtaining a plurality of IMF components;

[0007] Segmenting each of the IMF components to obtain data segments, analyzing the fluctuation characteristics of all the data segments in the IMF components to obtain the fluctuation coefficients of the data segments;

[0008] According to the fluctuation coefficient and in combination with the fluctuation characteristics of the noise data, obtaining the suspected noise data segment in the IMF component;

[0009] Analyze the proportion of the suspected noise data segment in the IMF component to obtain the noise contribution of the IMF component;

[0010] According to the noise contribution, adjusting the wavelet coefficient threshold of the IMF component in wavelet denoising to determine an adaptive wavelet coefficient threshold of the IMF component;

[0011] According to the adaptive wavelet coefficient threshold, the IMF component is subjected to wavelet transformation to remove noise data, and the denoised IMF components are superimposed and reconstructed to obtain a denoised speech simulation signal.

[0012] In some embodiments of the present invention, analyzing the fluctuation characteristics of all data segments in the IMF component to obtain the fluctuation coefficient of the data segment includes:

[0013] Analyze the difference characteristics of the feature values ​​of all adjacent feature points in the data segment to obtain the average difference of the adjacent feature points;

[0014] Analyze the difference between the characteristic values ​​of all characteristic points in the data segment and the mean of the characteristic values ​​of all characteristic points to obtain the degree of dispersion of the characteristic points in the data segment;

[0015] Performing STL residual decomposition on all feature points in the data segment to obtain the STL residual mean of all feature points in the data segment;

[0016] The fluctuation coefficient of the data segment is obtained according to the average difference of the adjacent feature points, the degree of dispersion, and the mean of the STL residuals.

[0017] In some embodiments of the present invention, the calculation formula of the fluctuation coefficient is:

[0018]

[0019] In the formula, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component The number of feature points in a data segment, Indicates IMF component The mean of the eigenvalues ​​of all feature points in a data segment, Indicates IMF component In the data segment The STL residual value of feature points, Represents the normalization function.

[0020] In some embodiments of the present invention, obtaining the suspected noise data segment in the IMF component according to the fluctuation coefficient in combination with the fluctuation characteristics of the noise data includes:

[0021] Analyzing the difference characteristics between the fluctuation coefficient of a data segment and the fluctuation coefficients of all other data segments in the IMF component to obtain the fluctuation difference of the data segment;

[0022] Detecting extreme points in the data segment by using the AMPD method, counting the number of the extreme points, and obtaining the frequency of fluctuation of the data segment based on the number of the extreme points and the difference between the duration of the data segment and the minimum value of the duration of all data segments in the corresponding IMF component;

[0023] Analyze the relationship between the range of the data segment and the mean range of all data segments in the corresponding IMF component to obtain the relative range of the data segment;

[0024] Obtaining the noise possibility of the data segment according to the fluctuation difference, the fluctuation frequency, and the relative range;

[0025] According to the noise possibility, a suspected noise data segment in the IMF component is obtained.

[0026] In some embodiments of the present invention, the calculation formula of the noise probability is:

[0027]

[0028] In the formula, Indicates IMF component The noise probability of a data segment, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates The data segment is located in In the other data segments of the IMF component The coefficient of fluctuation of the data segment, Indicates The number of data segments present in an IMF component, Indicates IMF component The number of extreme points in a data segment, Indicates IMF component The duration of a data segment, Indicates The minimum duration of all data segments in an IMF component, Indicates IMF component The range size of each data segment, Indicates The average value of the range of all data segments in the IMF component, Represents the normalization function.

[0029] In some embodiments of the present invention, the range of the data segment is the difference between the maximum value and the minimum value of the characteristic values ​​of all characteristic points in the data segment.

[0030] In some embodiments of the present invention, the noise contribution of the IMF component is: the ratio of the sum of the durations of all suspected noise data segments in the IMF component to the total duration of the IMF component.

[0031] In some embodiments of the present invention, the adaptive wavelet coefficient threshold is determined by taking the noise contribution of the IMF component as the weight value of the wavelet coefficient difference according to the difference between the largest wavelet coefficient and the smallest wavelet coefficient in the IMF component.

[0032] According to a second aspect of an embodiment of the present application, a voice signal processing system for fraud signal recognition is provided, the system comprising: a memory and a processor, wherein:

[0033] The memory is used to store program codes;

[0034] The processor is used to read the program code stored in the memory and execute any one of the methods described.

[0035] In some embodiments of the present invention, the processor comprises:

[0036] A speech recognition module is used to collect a speech signal input by a user, convert the speech signal into a speech simulation signal, and perform EMD decomposition on the speech simulation signal to obtain multiple IMF components;

[0037] A fluctuation coefficient module, used for performing data segmentation on each of the IMF components to obtain data segments, analyzing the fluctuation characteristics of all data segments in the IMF components to obtain the fluctuation coefficient of the data segments;

[0038] A suspected noise data segment module, used to obtain the suspected noise data segment in the IMF component according to the fluctuation coefficient and the fluctuation characteristics of the noise data;

[0039] A noise contribution module, used to analyze the proportion of the suspected noise data segment in the IMF component to obtain the noise contribution of the IMF component;

[0040] An adaptive wavelet coefficient threshold module, used to adjust the wavelet coefficient threshold of the IMF component in wavelet denoising according to the noise contribution, and determine the adaptive wavelet coefficient threshold of the IMF component;

[0041] The denoised speech analog signal module is used to perform wavelet transform on the IMF component to remove noise data according to the adaptive wavelet coefficient threshold, and to superimpose and reconstruct the denoised IMF component to obtain a denoised speech analog signal.

[0042] Compared with the prior art, the voice signal processing method and system for fraud signal identification provided by the present invention have the following beneficial effects:

[0043] The present invention collects a voice signal input by a user, converts the voice signal into a voice analog signal, performs EMD decomposition on the voice analog signal, and obtains a plurality of IMF components; performs data segmentation on each of the IMF components to obtain data segments, analyzes the fluctuation characteristics of all data segments in the IMF components, and obtains the fluctuation coefficients of the data segments; obtains suspected noise data segments in the IMF components according to the fluctuation coefficients and in combination with the fluctuation characteristics of noise data; analyzes the proportion of the suspected noise data segments in the IMF components to obtain the noise contribution of the IMF components; adjusts the wavelet coefficient threshold of the IMF components in wavelet denoising according to the noise contribution, and determines the adaptive wavelet coefficient threshold of the IMF components; performs wavelet transform on the IMF components according to the adaptive wavelet coefficient threshold to remove noise data, and superimposes and reconstructs the denoised IMF components to obtain denoised voice analog signals. The present invention optimizes the problem of inaccurate noise removal due to inappropriate threshold selection in the existing fixed wavelet threshold denoising. By performing EMD decomposition on the analog signal to obtain multiple components, the noise signal is more intuitively displayed in each component. Then, the data fluctuation of each component is analyzed to determine the possible noise data and calculate the noise contribution. The threshold of the wavelet coefficient is adaptively determined according to the size of the noise contribution, thereby more accurately removing possible noise and providing high-quality voice signals for the accurate identification of fraudulent calls. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are 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.

[0045] Figure 1 A basic flow chart of a voice signal processing method for fraud signal identification provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of a basic flow chart of a method for obtaining a fluctuation coefficient of a data segment provided by an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of a basic flow chart of a method for obtaining a suspected noise data segment provided by an embodiment of the present invention;

[0048] Figure 4 A schematic diagram of the basic components of a speech signal processing system for fraud signal recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the voice signal processing method and system for fraud signal recognition proposed by the present invention, its specific implementation, structure, features and effects, in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0050] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0051] The specific scheme of the speech signal processing method and system for fraud signal recognition provided by the present invention is described in detail below with reference to the accompanying drawings.

[0052] See also Figure 1 , which shows the basic process of a speech signal processing method for fraud signal identification provided by an embodiment of the present invention.

[0053] like Figure 1 As shown, a voice signal processing method for fraud signal recognition provided by an embodiment of the present invention specifically includes:

[0054] S100: Collect a voice signal input by a user, convert the voice signal into a voice analog signal, perform EMD decomposition on the voice analog signal, and obtain multiple IMF components.

[0055] Since this embodiment is suitable for improving the accuracy of fraud signal recognition by processing voice signals, it is necessary to receive voice signals through a voice recognition module, and then convert the voice signals into voice simulation signals for subsequent analysis. At the same time, since EMD decomposition can effectively adapt to dynamic changes in data when processing signals with complex fluctuations and instantaneous characteristics, thereby well preserving these dynamic change characteristics and also removing irrelevant noise, this embodiment performs EMD decomposition on the above-mentioned voice simulation signal to obtain multiple IMF components and residual res components.

[0056] So far, multiple IMF components of the analog signal are obtained.

[0057] After EMD decomposition, each noise in the speech simulation signal is decomposed into IMF components of different frequencies, making the noise data more obvious and intuitive. Therefore, the wavelet denoising method is used to analyze the noise in each IMF component, and the wavelet coefficient threshold is adaptively adjusted according to the proportion of noise to achieve the purpose of removing noise, and then the pure speech simulation signal is reconstructed by superimposing the denoised IMF components. Specifically, it includes steps S200~S700.

[0058] S200: Segmenting each of the IMF components to obtain data segments, analyzing the fluctuation characteristics of all the data segments in the IMF components, and obtaining the fluctuation coefficients of the data segments.

[0059] The ISODATA algorithm is an adaptive clustering method that can automatically determine the number of clusters based on the distribution of data and cluster data with similar fluctuations into a data cluster. First, use the ISODATA algorithm to cluster each IMF component. Then the real speech simulation signal data and noise data will be divided into different clusters due to different fluctuations. Finally, the entire IMF component data will be divided into multiple data segments with similarities, so that subsequent analysis based on data segments can be more convenient.

[0060] Then, the fluctuation characteristics of all data segments in each IMF component are analyzed, and the fluctuation coefficient is calculated to measure the fluctuation. Since there is a certain difference between the fluctuation of real speech simulation signal data and noise data, this step calculates the fluctuation coefficient of each data segment to help the subsequent screening of noise data.

[0061] See also Figure 2 , which shows the basic process of a method for obtaining the fluctuation coefficient of a data segment provided by an embodiment of the present invention.

[0062] Based on the above analysis, further Figure 2 As shown, a method for obtaining a fluctuation coefficient of a data segment provided by an embodiment of the present invention includes:

[0063] S201: Analyze the difference characteristics of the feature values ​​of all adjacent feature points in the data segment to obtain the average difference of the adjacent feature points.

[0064] First IMF component Take the data segment as an example. IMF component The difference characteristics of the eigenvalues ​​of all adjacent feature points in the data segment, and then IMF component All adjacent feature points in the data segment ( The feature points and feature points) feature values ​​( and ) and average the absolute values ​​of the differences to obtain IMF component The average difference between adjacent feature points in a data segment is .

[0065] S202: Analyze the difference between the characteristic values ​​of all characteristic points in the data segment and the average of the characteristic values ​​of all characteristic points to obtain the degree of dispersion of the characteristic points in the data segment.

[0066] First IMF component Take the data segment as an example. IMF component The feature values ​​of all feature points in the data segment The mean of the eigenvalues ​​of all feature points The absolute values ​​of the differences are summed and averaged to obtain the IMF component The discrete degree of feature points in a data segment is .

[0067] S203: Perform STL residual decomposition on all feature points in the data segment to obtain the STL residual mean of all feature points in the data segment.

[0068] First IMF component For example, IMF component All feature points in the data segment are decomposed by STL residuals to obtain the STL residuals of all feature points in the data segment. , and then the STL residuals of all feature points Sum and average to obtain IMF component The mean STL residual of all feature points in the data segment .

[0069] S204: Obtain the fluctuation coefficient of the data segment according to the average difference of adjacent feature points, the degree of dispersion, and the mean of the STL residual.

[0070] According to IMF component The average difference between adjacent feature points in the data segment , Discreteness , STL residual mean , obtain the IMF component The coefficient of fluctuation in the data segment Furthermore, the fluctuation coefficient of the data segment The calculation formula is:

[0071]

[0072] In the formula, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component The number of feature points in a data segment, Indicates IMF component The mean of the eigenvalues ​​of all feature points in a data segment, Indicates IMF component In the data segment The STL residual value of feature points, Represents the normalization function.

[0073] Represents the IMF component The average difference of the feature values ​​of any two adjacent feature points in a data segment, if The larger the value, the IMF component The greater the difference between all feature points in a data segment, the greater the The greater the overall fluctuation of a data segment, the greater the coefficient of fluctuation. Represents the IMF component The difference between the eigenvalues ​​of all feature points in a data segment and the average eigenvalue reflects the degree of dispersion of the eigenvalue data of the feature points relative to the mean eigenvalue. The larger the value, the The more discrete the eigenvalues ​​of all feature points in a data segment, the greater its fluctuation coefficient can be considered. Represents the IMF component The STL residual mean of all feature points in a data segment. Since the residual term of STL reflects the degree of data fluctuation to a certain extent, the larger the residual mean, the greater the overall fluctuation level in the corresponding data segment, and the larger the corresponding fluctuation coefficient.

[0074] Last Use The fluctuation coefficient result data is normalized to the range of [0,1], and the corresponding fluctuation coefficient is calculated for all data segments in all components.

[0075] S300: Obtaining suspected noise data segments in the IMF component according to the fluctuation coefficient and in combination with the fluctuation characteristics of the noise data.

[0076] Since the frequency distribution of real speech simulation signal data is more regular and the waveform is similar. On the contrary, noise data is usually random, and its fluctuation has the characteristics of local frequentness and irregular fluctuation, so there will be differences in the degree of fluctuation between the two; and since the fluctuation degree of noise data is more irregular than that of normal speech signal data, the possible noise data can be determined by analyzing the difference in the fluctuation coefficient value corresponding to the data segment and the range of the characteristic value of the characteristic point in the data segment; at the same time, noise data has the characteristics of local frequent fluctuations and the duration is generally short, so all extreme points in the data segment can be detected and counted by the AMPD method. If the fluctuation is more frequent, there will be more extreme points in the same time period. Therefore, possible noise data segments are screened based on the above characteristics.

[0077] See also Figure 3 , which shows the basic process of a method for obtaining suspected noise data segments provided by an embodiment of the present invention.

[0078] Based on the above analysis, further Figure 3 As shown, a method for obtaining a suspected noise data segment provided by an embodiment of the present invention includes:

[0079] S301: Analyze the difference characteristics between the fluctuation coefficient of a data segment and the fluctuation coefficients of all other data segments in the IMF component to obtain the fluctuation difference of the data segments.

[0080] First IMF component Take the data segment as an example. Calculate the IMF component The coefficient of fluctuation of the data segment The absolute value of the difference between the coefficient of fluctuation of all other data segments, and then calculate the sum of the absolute values ​​of all differences to obtain the fluctuation difference of the data segment .

[0081] S302: Detect extreme points in the data segment by using the AMPD method, count the number of extreme points, and obtain the frequency of fluctuation of the data segment based on the number of extreme points and the difference between the duration of the data segment and the minimum duration of all data segments in the corresponding IMF component.

[0082] First IMF component Take the data segment as an example. Using the AMPD method to detect the IMF component The number of extreme value points in the data segment is counted as , according to the number of extreme points , combined with Duration of a data segment The corresponding The minimum duration of all data segments in an IMF component The difference between the two values ​​is used to obtain the fluctuation frequency of the data segment. ,in This is to prevent the denominator from being zero.

[0083] S303: Analyze the relationship between the range of the data segment and the mean range of all data segments in the corresponding IMF component to obtain the relative range of the data segment.

[0084] First IMF component Take the data segment as an example. IMF component The range of the data segment The corresponding The mean of the range of all data segments in the IMF component The relationship between the two is used to obtain the relative range of the data segment. Among them, IMF component The range of the data segment is the maximum value of the eigenvalues ​​of all feature points in the data segment With minimum The difference between .

[0085] S304: Obtain the noise possibility of the data segment according to the fluctuation difference, fluctuation frequency, and relative range.

[0086] According to IMF component The fluctuation difference of the data segment , Fluctuation frequency , relative range , get the noise probability of the data segment Furthermore, the noise possibility The calculation formula is:

[0087]

[0088] In the formula, Indicates IMF component The noise probability of a data segment, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates The data segment is located in In the other data segments of the IMF component The coefficient of fluctuation of the data segment, Indicates The number of data segments present in an IMF component, Indicates IMF component The number of extreme points in a data segment, Indicates IMF component The duration of a data segment, Indicates The minimum duration of all data segments in an IMF component, Indicates IMF component The range size of each data segment, Indicates The average value of the range of all data segments in the IMF component, Represents the normalization function.

[0089] Representative IMF component The difference in fluctuation degree between a data segment and all other data segments. Since the fluctuation degree of the noise data segment is more abnormal and the proportion of noise data in the entire component is usually small, the larger the statistical value of the difference in fluctuation degree relative to other data segments, the greater the The larger the value of The more likely the data segment is a noise data segment, that is, The greater the noise possibility of each data segment. Represents the IMF component The frequency of fluctuations in each data segment (the denominator is increased by 1 to prevent the denominator from being 0). The more extreme points a data segment has during its duration, the The larger the value, the The shorter the duration of a data segment, the The smaller the value of The larger the value, the The more likely the data segment is a noise data segment, that is, The more noise there is in each data segment, the greater the possibility of noise. Represents the IMF component The relative range of each data segment is larger or smaller than that of the normal data segment, and the average range value is Closer to the range of normal data segment, so The larger the value of The greater the deviation of the range of the first data segment from the range of the normal data segment, the greater the deviation of the first data segment The more likely the data segment is a noise data segment, that is, The greater the noise probability of the data segment, the greater the noise probability of the data segment. Finally, based on the above characteristics, the corresponding noise probability value is calculated for each data segment in all IMF components. The function normalizes the result data to the range [0,1].

[0090] S305: Obtain suspected noise data segments in the IMF component according to the noise possibility.

[0091] Step S304 obtains the noise probability of each data segment in all IMF components. According to the noise probability, the suspected noise data segment in the IMF component is obtained, that is, the data segment corresponding to the noise probability value greater than or equal to the noise threshold is selected as the suspected noise data segment. In this embodiment, the noise threshold can be 0.88, that is, the suspected noise data segment is selected. The corresponding data segment is a suspected noise data segment.

[0092] At this point, the suspected noise data segments existing in each IMF component are screened out.

[0093] S400: Analyze the proportion of suspected noise data segments in the IMF component to obtain the noise contribution of the IMF component.

[0094] Since the ultimate goal is to remove noise data through the wavelet denoising method, and the noise data content in each IMF component is different, it is necessary to determine the appropriate wavelet coefficient threshold for each IMF component. If there is more noise in a certain IMF component, then the threshold of the wavelet coefficient is increased accordingly so that it can remove more noise data, otherwise the threshold of the wavelet coefficient is reduced. Therefore, the problem is transformed into calculating the proportion of noise data in each IMF component. The larger the proportion, the more noise there is. The above steps obtain the suspected noise data segment in each IMF component, so the existence of noise data can be determined by calculating the proportion of suspected noise data in the entire IMF component.

[0095] Therefore, analyzing the The sum of the durations of all suspected noise data segments in the IMF component is equal to the The noise contribution of the IMF component is obtained by calculating the ratio of the total duration of the IMF components. That is, the calculation formula for the noise contribution of the IMF component is:

[0096]

[0097] In the formula, Indicates The noise contribution of each IMF component is Indicates The total duration of the IMF components, Indicates IMF component The duration of the suspected noise data segment, Indicates The number of suspected noise data segments in an IMF component.

[0098] The size represents the The existence degree of noise data in each IMF component. Then the corresponding noise contribution is calculated for all IMF components.

[0099] At this point, each IMF component is analyzed, and suspected noise data is screened out according to the characteristic differences between the real speech simulation signal data and the noise data. Then, the noise contribution of this IMF component is calculated according to the proportion of suspected noise data in each IMF component, so as to determine different wavelet coefficient thresholds according to the noise contribution of each IMF component.

[0100] S500: adjusting the wavelet coefficient threshold of the IMF component in wavelet denoising according to the noise contribution, and determining an adaptive wavelet coefficient threshold of the IMF component.

[0101] In the wavelet domain, noise data is usually expressed as small-amplitude wavelet coefficients, while the important features of speech analog signals correspond to larger wavelet coefficients. Therefore, wavelet denoising sets a threshold for wavelet coefficients. Coefficients below this threshold can be regarded as noise and set to zero or reduced, while coefficients greater than the threshold are retained. For each IMF component, if there is more noise data, that is, the greater the noise contribution, in order to more accurately and completely remove the noise, the threshold of the wavelet coefficient needs to be increased. Conversely, the smaller the noise contribution, the correspondingly reduced threshold of the wavelet coefficient. Therefore, the threshold of the wavelet coefficient can be adjusted with the noise contribution as the weight.

[0102] Based on the above analysis, according to the noise contribution, the wavelet coefficient threshold of the IMF component in wavelet denoising is adjusted to determine the adaptive wavelet coefficient threshold of the IMF component. Further, the adaptive wavelet coefficient threshold is determined by taking the noise contribution of the IMF component as the weight value of the wavelet coefficient difference according to the difference between the largest wavelet coefficient and the smallest wavelet coefficient in the IMF component. That is, the calculation formula of the adaptive wavelet coefficient threshold is:

[0103]

[0104] In the formula, Indicates The adaptive wavelet coefficient threshold corresponding to the IMF component, and Indicates The largest wavelet coefficient and the smallest wavelet coefficient in the IMF component are Indicates The noise contribution corresponding to each IMF component.

[0105] That is, in The maximum span of the wavelet coefficients of the IMF components Based on the noise contribution As the weight, the threshold of the wavelet coefficient is adaptively adjusted. Then, for each IMF component, the corresponding threshold can be adaptively obtained through the above method, thereby achieving the purpose of adaptive and accurate denoising.

[0106] S600: According to the adaptive wavelet coefficient threshold, the IMF component is subjected to wavelet transform to remove noise data, and the denoised IMF components are superimposed and reconstructed to obtain a denoised speech analog signal.

[0107] According to the adaptive wavelet coefficient threshold of each IMF component obtained in step S500, wavelet transform is performed on each IMF component to remove noise data, and then all the denoised IMF components are superimposed and reconstructed to obtain a denoised speech simulation signal.

[0108] See also Figure 4 , which shows the basic composition of a speech signal processing system for fraud signal recognition provided by an embodiment of the present invention.

[0109] like Figure 4 As shown, the speech signal processing system includes: a memory 10 and a processor 20, wherein:

[0110] A memory 10, used for storing program codes;

[0111] The processor 20 is used to read the program code stored in the memory 10, and execute the voice signal input by the user, convert the voice signal into a voice analog signal, perform EMD decomposition on the voice analog signal, and obtain multiple IMF components; segment each IMF component to obtain data segments, analyze the fluctuation characteristics of all data segments in the IMF component, and obtain the fluctuation coefficient of the data segment; obtain the suspected noise data segment in the IMF component according to the fluctuation coefficient and the fluctuation characteristics of the noise data; analyze the proportion of the suspected noise data segment in the IMF component to obtain the noise contribution of the IMF component; adjust the wavelet coefficient threshold of the IMF component in wavelet denoising according to the noise contribution, and determine the adaptive wavelet coefficient threshold of the IMF component; according to the adaptive wavelet coefficient threshold, perform wavelet transform on the IMF component to remove noise data, and superimpose and reconstruct the denoised IMF component to obtain a denoised voice analog signal.

[0112] The processor 20 includes:

[0113] The speech recognition module 21 is configured to collect a speech signal input by a user, convert the speech signal into a speech analog signal, and perform EMD decomposition on the speech analog signal to obtain a plurality of IMF components;

[0114] The fluctuation coefficient module 22 is configured to segment the data of each IMF component to obtain data segments, analyze the fluctuation characteristics of all data segments in the IMF component, and obtain the fluctuation coefficient of the data segment;

[0115] The suspected noise data segment module 23 is configured to obtain the suspected noise data segment in the IMF component according to the fluctuation coefficient and the fluctuation characteristics of the noise data;

[0116] The noise contribution module 24 is configured to analyze the proportion of the suspected noise data segments in the IMF component to obtain the noise contribution of the IMF component;

[0117] The adaptive wavelet coefficient threshold module 25 is configured to adjust the wavelet coefficient threshold of the IMF component in the wavelet denoising according to the noise contribution, and determine the adaptive wavelet coefficient threshold of the IMF component;

[0118] The denoised speech analog signal module 26 is configured to perform wavelet transform on the IMF component to remove noise data according to the adaptive wavelet coefficient threshold, and to superimpose and reconstruct the denoised IMF component to obtain a denoised speech analog signal;

[0119] The speech signal processing method and system for fraud signal recognition provided by this solution optimizes the problem of inaccurate noise removal caused by inappropriate threshold setting when denoising speech simulation signals through fixed wavelet coefficient thresholds. The noise contribution is determined by analyzing the data fluctuation characteristics of each IMF component, and each component is denoised according to the threshold of the adaptive wavelet coefficient according to the noise contribution, so that the noise removal is more targeted, thereby improving the accuracy of fraud speech recognition.

[0120] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0121] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A speech signal processing method for fraud signal identification, characterized in that: The method comprises: Collecting a voice signal input by a user, converting the voice signal into a voice analog signal, performing EMD decomposition on the voice analog signal, and obtaining a plurality of IMF components; Segmenting each of the IMF components to obtain data segments, analyzing the fluctuation characteristics of all the data segments in the IMF components to obtain the fluctuation coefficients of the data segments; According to the fluctuation coefficient and in combination with the fluctuation characteristics of the noise data, obtaining the suspected noise data segment in the IMF component; Analyze the proportion of the suspected noise data segment in the IMF component to obtain the noise contribution of the IMF component; According to the noise contribution, adjusting the wavelet coefficient threshold of the IMF component in wavelet denoising to determine an adaptive wavelet coefficient threshold of the IMF component; According to the adaptive wavelet coefficient threshold, the IMF component is subjected to wavelet transform to remove noise data, and the denoised IMF component is superimposed and reconstructed to obtain a denoised speech simulation signal; The method for obtaining the coefficient of fluctuation of the data segment is as follows: analyzing the difference characteristics of the characteristic values ​​of all adjacent characteristic points in the data segment to obtain the average difference of adjacent characteristic points; analyzing the difference between the characteristic values ​​of all characteristic points in the data segment and the average of the characteristic values ​​of all characteristic points to obtain the degree of dispersion of the characteristic points in the data segment; performing STL residual decomposition on all characteristic points in the data segment to obtain the mean value of the STL residuals of all characteristic points in the data segment; obtaining the coefficient of fluctuation of the data segment according to the average difference of adjacent characteristic points, the degree of dispersion and the mean value of the STL residuals; The calculation formula of the volatility coefficient is: In the formula, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component In the data segment The eigenvalue size of the feature points, Indicates IMF component The number of feature points in a data segment, Indicates IMF component The mean of the eigenvalues ​​of all feature points in a data segment, Indicates IMF component In the data segment The STL residual value of feature points, represents the normalization function; Wherein, obtaining the suspected noise data segment in the IMF component according to the fluctuation coefficient and in combination with the fluctuation characteristics of the noise data includes: Analyzing the difference characteristics between the fluctuation coefficient of a data segment and the fluctuation coefficients of all other data segments in the IMF component to obtain the fluctuation difference of the data segment; Detect extreme points in the data segment by using the AMPD method, count the number of extreme points, and obtain the fluctuation frequency of the data segment based on the number of extreme points and the difference between the duration of the data segment and the minimum value of the duration of all data segments in the corresponding IMF component; Analyze the relationship between the range of the data segment and the mean range of all data segments in the corresponding IMF component to obtain the relative range of the data segment; Obtaining the noise possibility of the data segment according to the fluctuation difference, the fluctuation frequency, and the relative range; According to the noise possibility, a suspected noise data segment in the IMF component is obtained.

2. The method for processing speech signals for fraud signal identification according to claim 1, characterized in that: The calculation formula of the noise probability is: In the formula, Indicates IMF component The noise probability of a data segment, Indicates IMF component The coefficient of fluctuation of the data segment, Indicates The data segment is located in In the other data segments of the IMF component The coefficient of fluctuation of the data segment, Indicates The number of data segments present in an IMF component, Indicates IMF component The number of extreme points in a data segment, Indicates IMF component The duration of a data segment, Indicates The minimum duration of all data segments in an IMF component, Indicates IMF component The range size of each data segment, Indicates The average value of the range of all data segments in the IMF component, Represents the normalization function.

3. The method for processing speech signals for fraud signal identification according to claim 1, characterized in that: The range of the data segment is the difference between the maximum and minimum values ​​of the characteristic values ​​of all characteristic points in the data segment.

4. The voice signal processing method for fraud signal identification according to claim 1, characterized in that: The noise contribution of the IMF component is: the ratio of the sum of the durations of all suspected noise data segments in the IMF component to the total duration of the IMF component.

5. The voice signal processing method for fraud signal identification according to claim 1, characterized in that: The adaptive wavelet coefficient threshold is determined by taking the noise contribution of the IMF component as the weight value of the wavelet coefficient difference according to the difference between the largest wavelet coefficient and the smallest wavelet coefficient in the IMF component.

6. A speech signal processing system for fraud signal identification, characterized in that: The system comprises: a memory and a processor, wherein: The memory is used to store program code; The processor is configured to read the program code stored in the memory and execute the method according to any one of claims 1 to 5.

7. The voice signal processing system for fraud signal identification according to claim 6, characterized in that: The processor comprises: A speech recognition module is used to collect a speech signal input by a user, convert the speech signal into a speech simulation signal, and perform EMD decomposition on the speech simulation signal to obtain multiple IMF components; A fluctuation coefficient module, used for performing data segmentation on each of the IMF components to obtain data segments, analyzing the fluctuation characteristics of all data segments in the IMF components to obtain the fluctuation coefficient of the data segments; A suspected noise data segment module, used to obtain the suspected noise data segment in the IMF component according to the fluctuation coefficient and the fluctuation characteristics of the noise data; A noise contribution module, used to analyze the proportion of the suspected noise data segment in the IMF component to obtain the noise contribution of the IMF component; An adaptive wavelet coefficient threshold module, used to adjust the wavelet coefficient threshold of the IMF component in wavelet denoising according to the noise contribution, and determine the adaptive wavelet coefficient threshold of the IMF component; The denoised speech analog signal module is used to perform wavelet transform on the IMF component to remove noise data according to the adaptive wavelet coefficient threshold, and to superimpose and reconstruct the denoised IMF component to obtain a denoised speech analog signal.

Images