Communication clue analysis reminding method based on voiceprint recognition

Abstract

The invention relates to the technical field of electrical digital processing, and discloses a voiceprint recognition-based communication clue analysis reminding method, which comprises the following steps of: acquiring incoming call information of a caller in a communication process; noise reduction is carried out on the audio recording information, then feature extraction is carried out, and a corresponding waterfall spectrogram is obtained; carrying out image enhancement on the waterfall spectrogram and then sending a human voice recognition model; acquiring identity information of the caller output by the voice recognition model; obtaining clue description text data of clues provided by the caller according to the audio recording information; judging whether the clue description text data are repeated clues or not; if so, marking the caller and sending a related reminding notice, otherwise, carrying out communication clue analysis; the acquired waterfall spectrogram is subjected to image enhancement, so that the recognition accuracy of the human voice recognition model trained based on the neural network model can be improved, the reliability of judging whether repeated clues are made by the same person or not is improved, and excessive occupation and waste of limited public resources are reduced.

Description

Technical Field

[0001] The present application relates to the technical field of electronic digital processing, and in particular to a communication clue analysis and reminder method based on voiceprint recognition. Background Art

[0002] Voiceprint recognition technology extracts identity information and voiceprint features, converts sound signals into electrical signals, and then uses computers to compare and identify through relevant algorithms. It mainly relies on the individual characteristics of each person's voice, such as pronunciation habits, resonance methods, voice purity, average pitch, etc. These characteristics can remain relatively stable for a long time after adulthood. The voiceprint recognition system determines whether the speaker of two voices is the same person by comparing the voice of the speaker on the same phoneme, thereby realizing the function of "recognizing people by hearing the voice".

[0003] At present, voiceprint recognition technology has many application scenarios, such as public security, finance, social security, smart hardware, etc. Taking the public security field as an example, a voiceprint database can be established to monitor voiceprint information in real time under specific circumstances and issue early warnings; voiceprint recognition technology can be used for massive screening to narrow the scope of investigation and improve case handling efficiency; voiceprint identification technology can also be used to identify the voices involved in telecommunications fraud cases and determine the identity of the suspect; or the caller and the communication clues they provide can be verified to reduce the probability of successful malicious repeated reports.

[0004] However, when identifying the identities of people related to communication clues, since the voice characteristics of some people have not been recorded in advance, it is necessary to obtain enough samples as much as possible within the limited call time. Currently, in the process of providing communication clues from unknown people, it is necessary to improve the recognition accuracy and training efficiency of the relevant voice recognition model as soon as possible. Summary of the invention

[0005] In order to quickly improve the recognition accuracy and training efficiency of the voice recognition model under the condition of limited samples, this application provides a communication clue analysis reminder method based on voiceprint recognition, which adopts the following technical solutions:

[0006] A communication clue analysis and reminder method based on voiceprint recognition, comprising:

[0007] Step S1, obtaining the caller's caller information during the communication process; the caller information includes audio recording information, call time information and call location information;

[0008] Step S2, performing feature extraction on the audio recording information after noise reduction to obtain a corresponding waterfall spectrogram;

[0009] Step S3, performing image enhancement on the waterfall spectrogram and then sending it to a human voice recognition model;

[0010] Step S4, obtaining the identity information of the caller output by the voice recognition model;

[0011] Step S5, obtaining clue description text data of the clue provided by the caller according to the audio recording information;

[0012] Step S6, determining whether the clue description text data is a repeated clue;

[0013] If it is a repeated clue, the caller is marked and a related reminder notification is sent, otherwise the communication clue analysis is performed based on the current clue description text data and historical call information.

[0014] By adopting the above technical solution, image enhancement is performed on the waterfall spectrogram, which is beneficial to improving the recognition accuracy of the human voice recognition model trained based on the neural network model, thereby improving the reliability of subsequent judgments on whether repeated clues are provided by the same person, thereby reducing the rate of answering malicious calls from callers and reducing excessive occupation and waste of limited public resources.

[0015] Optionally, the process of performing image enhancement on the waterfall spectrogram includes:

[0016] Step S31, extracting and storing the Y component in the waterfall spectrogram;

[0017] Step S32, performing image segmentation on the waterfall spectrogram according to the brightness of the Y component to obtain corresponding Y component block images;

[0018] Step S33, performing regional contrast parallel enhancement on the Y component block image;

[0019] Step S34, performing overall equalization on the Y component block image;

[0020] Step S35, adaptively compensating the UV component image of the Y component block image.

[0021] By adopting the above technical solution, through the steps of parallel enhancement of regional contrast based on brightness, overall equalization of Y component image, etc., the contrast of the rectangular sub-image is firstly adaptively enhanced, so as to improve the image details. Then, the seams between the molecular block images after contrast enhancement can be eliminated by overall equalization of the Y component image, so as to further improve the image quality after adjustment. Finally, the UV component is adaptively compensated, so as to make the details such as edges and textures in the image clearer, so as to achieve the image quality enhancement of the waterfall spectrogram with large resolution.

[0022] Optionally, the process of extracting and storing the Y component in the waterfall spectrogram in step S31 includes:

[0023] Converting the waterfall spectrogram into a YUV format;

[0024] The Y component of the waterfall spectrogram is extracted and stored in YUV format.

[0025] Optionally, the process of performing image segmentation on the waterfall spectrogram according to the brightness of the Y component to obtain corresponding Y component block images in S32 includes:

[0026] Preset at least three brightness threshold intervals of brightness gradients;

[0027] Dividing the waterfall sound spectrogram into regions according to the brightness threshold interval;

[0028] The waterfall spectrogram is segmented according to the region division result and the minimum rectangle to obtain the Y component block image; the region where the minimum rectangle is located is a rectangular sub-image of the Y component block image.

[0029] By adopting the above technical solution, when configuring the brightness threshold interval according to the brightness, the three brightness threshold intervals can be delineated by dividing the minimum brightness and maximum brightness intervals into three equal parts; if the call time is short, that is, when few original data samples can be obtained, it can be divided into four or five equal parts, thereby greatly expanding the number of training samples and providing different types of disturbances, thereby fully improving the recognition accuracy, robustness and training efficiency of the human voice recognition model and greatly reducing the training cost.

[0030] Optionally, the process of performing parallel regional contrast enhancement on the Y component block image in step S33 includes:

[0031] Using a thread library to perform multi-threaded parallel processing on the rectangular sub-image, and enhancing the rectangular sub-images in the brightness threshold intervals of three brightness gradients at the same time;

[0032] Performing smoothing processing on each of the rectangular sub-images to obtain a corresponding low-frequency portion;

[0033] Subtracting the low-frequency part from the original image to obtain a high-frequency part;

[0034] The low-frequency part and the high-frequency part are recombined according to a preset enhanced combination algorithm.

[0035] By adopting the above technical solution, these sub-images are processed in multi-threaded parallel by using a thread library, and different enhancement methods are respectively adopted for over-bright target, normal brightness target and darker target sub-images, namely, a brightness-based regional contrast parallel enhancement method, which suppresses strong light of over-bright target sub-images, reduces target pixel values, enhances contrast, adaptively enhances contrast of normal brightness target sub-images, enhances low illumination of darker target sub-images, increases target pixel values, enhances contrast, and then adaptively offsets and compensates according to the specific brightness of the actual ambient light, readjusts the pixel value size of the image, and enhances target details. After that, the output result diagram of the algorithm is obtained by recombining the low-frequency and high-frequency images, and the processed sub-images can be adaptively contrast enhanced and averaged, so that image details can be enhanced.

[0036] Optionally, the process of performing overall equalization on the Y component block image in step S34 includes:

[0037] The Y component image is operated by using an equalization method to make the Y component image overall equalized, thereby eliminating or weakening the effect of obvious gaps between each image block.

[0038] By adopting the above technical solution, the Y component image is operated by using an equalization method in combination with a grayscale image. At this time, the Y component image is overall equalized, which can eliminate or weaken the influence of the seams, thereby improving the image quality after the adjustment, and further improving the clarity and quality of the large-resolution image.

[0039] Optionally, the process of calculating the similarity between the clue description text data and the historical clue description text data provided by the caller includes:

[0040] For a clue description text data, extract the feature description words corresponding to the preset attributes; the preset attributes have at least n;

[0041] Evaluate the similarity between the feature description words and the feature description words of the same attribute of the historical clue description text data:

[0042]

[0043] in, is an estimate of the similarity between the data numbered B in the historical clue description text data and the current clue description text data A, Represents the cosine value between the corresponding text vectors of the same attribute text data, Represents the Euclidean distance between corresponding text vectors of the same attribute text data;

[0044] The maximum value in the similarity evaluation result is taken as the similarity comparison value ,like , it is judged that it is not a repeated clue, otherwise it is a repeated clue.

[0045] In summary, the present application includes at least one of the following beneficial technical effects:

[0046] 1. Image enhancement of the waterfall spectrogram is beneficial to improving the recognition accuracy of the voice recognition model trained based on the neural network model, which is beneficial to improving the reliability of subsequent judgments on whether repeated clues are from the same person, thereby reducing the rate of answering malicious calls from callers and reducing excessive occupation and waste of limited public resources.

[0047] 2. When configuring the brightness threshold interval according to the brightness, the three brightness threshold intervals can be defined by dividing the minimum brightness and maximum brightness intervals into three equal parts; the core here is that if the call time is short, that is, when few original data samples can be obtained, it can be divided into four or five equal parts, so as to greatly expand the number of training samples and provide different types of disturbances, thereby fully improving the recognition accuracy, robustness and training efficiency of the voice recognition model and greatly reducing the training cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the principle flow of the communication clue analysis and reminder method based on voiceprint recognition in the present invention;

[0049] Figure 2 It is a flow chart of the principle of image enhancement of waterfall spectrogram in the present invention. DETAILED DESCRIPTION

[0050] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0051] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0052] The present application embodiment discloses a communication clue analysis and reminder method based on voiceprint recognition, referring to Figure 1 ,include:

[0053] Step S1, obtaining the caller's caller information during the communication process; the caller information includes audio recording information, call time information and call location information;

[0054] Step S2, performing feature extraction on the audio recording information after noise reduction to obtain a corresponding waterfall spectrogram;

[0055] The waterfall spectrogram adds a time dimension to the spectrogram, which can show the complete process of the signal spectrum changing over time. In the waterfall spectrogram, the vertical axis represents frequency (in hertz), the horizontal axis represents time (in seconds), and the signal strength or amplitude is usually represented by the brightness or depth of the color. In this way, over time, the change of the signal spectrum flows from right to left like a waterfall, hence the name "waterfall spectrogram".

[0056] Specifically, the digitally sampled data is divided into a number of blocks (usually overlapping) in the time domain, and then each block is Fourier transformed to calculate its spectral amplitude. Each block corresponds to a vertical line in the image, and the amplitude and frequency are measured at a specific time point (the midpoint of the block). These spectra or time graphs are "placed side by side" to form an image or three-dimensional surface, thereby generating a waterfall spectrogram.

[0057] In voice identification, waterfall spectrogram is used to extract and analyze the voiceprint features of the speaker. Voiceprint features refer to the speech parameters carried in human voice that can reflect the physiological and behavioral characteristics of the speaker, such as fundamental frequency, resonance peak, spectrum envelope, etc. These features are presented in the waterfall spectrogram as a specific spectrum change pattern, providing a reliable basis for voiceprint identification.

[0058] At present, waterfall spectrograms can be used as input features of neural network models, which can intuitively show the changes of signal spectrum over time. These changes contain rich audio feature information. By using these feature information as input of neural network models, neural network models can be equipped with the ability to recognize and analyze audio signals. However, when using waterfall spectrograms for analysis, the measurement accuracy of the waterfall chart and the time window selection of the waterfall chart will affect the analysis results. This requires that both the training samples and the test samples input into the neural network model must have a considerable resolution and sufficient clarity, and can reliably improve the robustness of the neural network model during training.

[0059] However, in the application environment of the present invention, the call time and call duration of the caller are unpredictable, and the number of samples that can be collected or used to participate in the training of the neural network model is very limited. Therefore, it is difficult to ensure that after receiving the call information of the caller for the first time, the neural network model can be quickly and effectively supplemented with training, so that the correct voice identity recognition can be achieved earlier when the caller calls subsequently.

[0060] In this embodiment of the present invention, step S3 can be used to perform image enhancement on the waterfall spectrogram and then send it to the human voice recognition model; in this way, the clarity of the waterfall spectrogram can be enhanced through image enhancement, and the same waterfall spectrogram can also be artificially disturbed, thereby achieving the doubling of the waterfall spectrogram, which is beneficial to the subsequent training of the human voice recognition model, improves the recognition accuracy and training efficiency of the human voice recognition model, and is beneficial to reducing the answering rate of malicious calls from the corresponding callers, and reducing the excessive occupation and waste of limited public resources.

[0061] Step S4, obtaining the identity information of the caller output by the voice recognition model;

[0062] Step S5, obtaining clue description text data of the clue provided by the caller according to the audio recording information;

[0063] Step S6, determining whether the clue description text data is a repeated clue;

[0064] If it is a repeated clue, the caller is marked and a relevant reminder notification is sent. Otherwise, communication clue analysis is performed based on the current clue description text data and historical call information. Communication clue analysis can use frequency analysis, time period analysis, space analysis and other methods to conduct in-depth analysis of communication records to reveal the laws and characteristics of communication behavior.

[0065] In this embodiment of the present invention, the process of performing image enhancement on the waterfall spectrogram specifically includes:

[0066] Step S31, extracting and storing the Y component in the waterfall spectrogram;

[0067] Step S32, performing image segmentation on the waterfall spectrogram according to the brightness of the Y component to obtain corresponding Y component block images;

[0068] Step S33, performing regional contrast parallel enhancement on the Y component block image;

[0069] Step S34, performing overall equalization on the Y component block image;

[0070] Step S35, adaptively compensating the UV component image of the Y component block image.

[0071] Through the steps of parallel enhancement of regional contrast based on brightness and overall equalization of Y component image, the contrast of rectangular sub-image is firstly adaptively enhanced to improve image details. Then, the seams between the molecular block images after contrast enhancement can be eliminated by overall equalization of Y component image, and the image quality after adjustment can be further improved. Finally, the UV component is adaptively compensated to make the details such as edges and textures in the image clearer, thereby achieving image quality enhancement of large-resolution waterfall spectrogram.

[0072] The step S31 specifically includes:

[0073] Converting the waterfall spectrogram into a YUV format;

[0074] The Y component of the waterfall spectrogram is extracted and stored in YUV format.

[0075] In this embodiment, the NV12 image format in the YUV420 format image is selected, which belongs to the YUV420SP type, that is, a variant of the YUV 4:2:0 Subsampled Planar format. All Y components (brightness information) are stored first, and then the interleaved UV components (chrominance information) are stored. For each pixel, there is a corresponding Y value.

[0076] In the NV12 format, the storage order of Y components, U components, and V components is: YYYYYYYY UVUVUVUV..., that is, all Y components are stored continuously first, and then U and V components are stored alternately. Every two adjacent pixels share a U component and a V component, but they are stored interleaved, that is, the first byte is the U component, the second byte is the V component, and then the next U and V components. Every four Y components share a set of UV components, which can minimize storage space and bandwidth requirements while ensuring image quality.

[0077] The step S32 specifically includes:

[0078] Preset at least three brightness threshold intervals of brightness gradients; divide the waterfall spectrogram into regions according to the brightness threshold intervals;

[0079] The waterfall spectrogram is segmented according to the region division result and the minimum rectangle to obtain the Y component block image; the region where the minimum rectangle is located is a rectangular sub-image of the Y component block image.

[0080] When configuring the brightness threshold interval according to the brightness, the three brightness threshold intervals can be defined by dividing the minimum brightness and maximum brightness intervals into three equal parts. The key point here is that if the call time is short, that is, when the available raw data samples are few, the interval can be divided into four or five equal parts, so as to greatly expand the number of training samples and provide different types of disturbances, thereby fully improving the recognition accuracy, robustness and training efficiency of the human voice recognition model and greatly reducing the training cost.

[0081] Taking the three-equal division as an example, since the Y component represents the brightness of the corresponding pixel position of the Y component block image, the waterfall spectrogram after the Y component is extracted can be divided into over-bright targets, normal brightness targets and darker targets according to the size of the Y component at the corresponding pixel; for example, taking the pixel value 0-255 as an example, the pixel value 200-255 is divided into over-bright targets, the pixel value 100-200 is divided into normal brightness targets, and the pixel value 0-100 is divided into darker targets. In the subsequent processing, different enhancement methods are used for over-bright targets, normal brightness targets and darker targets to prevent over-processing of different types of targets from over-compensation and distortion, and also reduce the amount of processing calculations. Among them, the size of the minimum rectangle can be set manually, and the brightness median area corresponding to the minimum rectangle is the brightness classification of the minimum rectangle.

[0082] The step S33 specifically includes:

[0083] Using a thread library to perform multi-threaded parallel processing on the rectangular sub-image, and enhancing the rectangular sub-images in the brightness threshold intervals of three brightness gradients at the same time;

[0084] Performing smoothing processing on each of the rectangular sub-images to obtain a corresponding low-frequency portion;

[0085] Subtracting the low-frequency part from the original image to obtain a high-frequency part;

[0086] The low-frequency part and the high-frequency part are recombined according to a preset enhanced combination algorithm.

[0087] Specifically, for a rectangular sub-image, according to Calculate the pixel value of the processed image;

[0088] Among them, i and j are the horizontal and vertical coordinates of any point in the rectangular sub-image, respectively. is the pixel value of the original image at the corresponding position of the rectangular sub-image, and are preset constant coefficients. It can represent the scaling factor of the image pixel value. The type is floating point. The default value is 1.0, which means no scaling. 0.25 means the pixel value is reduced by 4 times. 4.0 means the pixel value is enlarged by 4 times. Represents the offset constant of the image pixel value. The type is floating point. The default value is 0.0, which means no offset. -20.0 means the pixel value decreases by 20, making the image darker. 20.0 means the pixel value increases by 20, making the image brighter. Ensure that the processed value is normalized to 0-255, and then adaptively enhance the contrast of over-bright targets, normal brightness targets, and darker target sub-images at the same time.

[0089] Adaptive contrast enhancement uses local histogram information to map data, and uses unsharp masking technology to effectively enhance the gradient of low-contrast images. First, the image is divided into two parts. The first part is the low-frequency part obtained by smoothing the original image; the second part is the image obtained by subtracting the first part from the original image to obtain the high-frequency part. Then, the image obtained from the second part is enhanced, and the images of the first and second parts are recombined to obtain the output result of the algorithm. Specifically, it includes the following two steps:

[0090] First, set the template size to ,according to The low-frequency part is obtained by using the method of finding the local average value;

[0091]

[0092] in, For location The size is The average pixel value of the window, m is a positive integer, and , ; and get the standard deviation of the pixel value of the window ;

[0093] After that, the high frequency part is enhanced and the standard deviation is used as the gain value. It is roughly considered to be the background part. That is, the high-frequency detail part, the gain product is performed on the high frequency;

[0094] Calculate the location Pixel value The pixel value after enhancing the high-frequency part ;

[0095]

[0096] Through such settings, these sub-images are processed in multi-threaded parallel by using the thread library, and different enhancement methods are used for over-bright target, normal brightness target and darker target sub-images, namely, the brightness-based regional contrast parallel enhancement method, which suppresses the strong light of the over-bright target sub-image, reduces the target pixel value, and enhances the contrast, adaptively enhances the contrast of the normal brightness target sub-image, enhances the low illumination of the darker target sub-image, increases the target pixel value, and enhances the contrast, and then adaptively offsets and compensates according to the specific brightness of the actual ambient light, readjusts the pixel value of the image, and enhances the target details. After that, the output result diagram of the algorithm is obtained by recombining the low-frequency part and the high-frequency part of the image, and the processed sub-image can be adaptively contrast enhanced and averaged, so that the image details can be enhanced.

[0097] The step S34 specifically includes:

[0098] The Y component image is operated by using an equalization method to make the Y component image overall equalized, thereby eliminating or weakening the effect of obvious gaps between each image block.

[0099] Specifically, first count the number of pixels of each gray level of the original image of the Y component image , the value range of G is [0, max-1], max is the number of gray levels in the original image, and the number of gray levels is related to the image bit depth. If the image is 8-bit deep, there are 2^8=256 gray levels in total;

[0100] Then through the formula Calculate the occurrence probability of pixels with grayscale G;

[0101] Where Mz is the number of pixels in the image;

[0102] Then through the formula Calculate the cumulative distribution function obtained;

[0103] Finally, the mixed part is averaged and calculated:

[0104] Through the above technical solution, the Y component image is operated by using an equalization method in combination with the grayscale image. At this time, the Y component image is overall equalized, which can eliminate or weaken the influence of the seams, thereby improving the image quality after the adjustment, and further improving the clarity and quality of the large-resolution image.

[0105] The step S35 specifically includes:

[0106] By formula:

[0107]

[0108]

[0109] Calculate the position separately Enhanced U component pixel value And the V component pixel value after position enhancement ;

[0110] in, For location The position within the n*n domain The enhanced Y component image, For location The position of the n*n field range Y component image before enhancement;

[0111] Through such settings, through the processing of UV components, the color saturation, hue and other attributes of the image can be accurately adjusted to make the image color more realistic and natural, significantly improve the color performance of the image, and make the edge details in the image clearer. The neural network model's recognition of the image requires the depth of the color to identify the height of the data, thereby improving the recognition and readability of the image to ensure the training accuracy of the neural network model.

[0112] Optionally, the process of calculating the similarity between the clue description text data and the historical clue description text data provided by the caller includes:

[0113] For a clue description text data, extract the feature description words corresponding to the preset attributes; the preset attributes have at least n;

[0114] Evaluate the similarity between the feature description words and the feature description words of the same attribute of the historical clue description text data:

[0115]

[0116] in, is an estimate of the similarity between the data numbered B in the historical clue description text data and the current clue description text data A, Represents the cosine value between the corresponding text vectors of the same attribute text data, Represents the Euclidean distance between corresponding text vectors of the same attribute text data;

[0117] The maximum value in the similarity evaluation result is taken as the similarity comparison value ,like , it is judged that it is not a repeated clue, otherwise it is a repeated clue.

[0118] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A communication clue analysis and reminder method based on voiceprint recognition, characterized in that: include: Step S1, obtaining the caller's call information during the communication process; The incoming call information includes audio recording information, incoming call time information and incoming call location information; Step S2, performing feature extraction on the audio recording information after noise reduction to obtain a corresponding waterfall spectrogram; Step S3, performing image enhancement on the waterfall spectrogram and then sending it to a human voice recognition model; Step S4, obtaining the identity information of the caller output by the voice recognition model; Step S5, obtaining clue description text data of the clue provided by the caller according to the audio recording information; Step S6, determining whether the clue description text data is a repeated clue; If it is a repeated clue, the caller is marked and a related reminder notification is sent, otherwise the communication clue analysis is performed based on the current clue description text data and historical call information.

2. The communication clue analysis and reminder method based on voiceprint recognition according to claim 1 is characterized in that: The process of performing image enhancement on the waterfall spectrogram comprises: Step S31, extracting and storing the Y component in the waterfall spectrogram; Step S32, performing image segmentation on the waterfall spectrogram according to the brightness of the Y component to obtain corresponding Y component block images; Step S33, performing regional contrast parallel enhancement on the Y component block image; Step S34, performing overall equalization on the Y component block image; Step S35, adaptively compensating the UV component image of the Y component block image.

3. The communication clue analysis and reminder method based on voiceprint recognition according to claim 2 is characterized in that: The process of step S31, extracting and storing the Y component in the waterfall spectrogram, comprises: Converting the waterfall spectrogram into a YUV format; The Y component of the waterfall spectrogram is extracted and stored in YUV format.

4. The communication clue analysis and reminder method based on voiceprint recognition according to claim 2 is characterized in that: The process of performing image segmentation on the waterfall spectrogram according to the brightness of the Y component to obtain the corresponding Y component block image in S32 includes: Preset at least three brightness threshold intervals of brightness gradients; Dividing the waterfall sound spectrogram into regions according to the brightness threshold interval; The waterfall spectrogram is segmented according to the region division result and the minimum rectangle to obtain the Y component block image; the region where the minimum rectangle is located is a rectangular sub-image of the Y component block image.

5. The communication clue analysis and reminder method based on voiceprint recognition according to claim 4 is characterized in that: The process of performing parallel regional contrast enhancement on the Y component block image in step S33 includes: Using a thread library to perform multi-threaded parallel processing on the rectangular sub-image, and enhancing the rectangular sub-images in the brightness threshold intervals of three brightness gradients at the same time; Performing smoothing processing on each of the rectangular sub-images to obtain a corresponding low-frequency portion; Subtracting the low-frequency part from the original image to obtain a high-frequency part; The low-frequency part and the high-frequency part are recombined according to a preset enhanced combination algorithm.

6. The communication clue analysis and reminder method based on voiceprint recognition according to claim 2 is characterized in that: The process of performing overall equalization on the Y component block image in step S34 includes: The Y component image is operated by using an equalization method to make the Y component image overall equalized, thereby eliminating or weakening the effect of obvious gaps between each image block.

7. The communication clue analysis and reminder method based on voiceprint recognition according to claim 1 is characterized in that: The process of calculating the similarity between the clue description text data and the historical clue description text data provided by the caller includes: For a clue description text data, extract the feature description words corresponding to the preset attributes; the preset attributes have at least n; Evaluate the similarity between the feature description words and the feature description words of the same attribute of the historical clue description text data: in, is an estimate of the similarity between the data numbered B in the historical clue description text data and the current clue description text data A, Represents the cosine value between the corresponding text vectors of the same attribute text data, Represents the Euclidean distance between corresponding text vectors of the same attribute text data; The maximum value in the similarity evaluation result is taken as the similarity comparison value ,like , it is judged that it is not a repeated clue, otherwise it is a repeated clue.

Images