A telephone voice bad content detection system based on natural language processing technology
By using a system based on natural language processing technology to optimize voice data processing and real-time parameter adjustment, the problems of low recognition rate and delay caused by fast speech speed and low-quality voice are solved, and efficient detection and timely processing of telephone voice content are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI IFLY DIGITAL TECH CO LTD
- Filing Date
- 2025-02-11
- Publication Date
- 2026-06-26
AI Technical Summary
Existing telephone voice content detection systems suffer from low speech recognition conversion rates and delayed detection strategy adjustments when faced with fast-paced and low-quality speech, resulting in untimely identification of inappropriate calls.
The system employs natural language processing technology to optimize voice content detection by adjusting call parameters in real time through data acquisition, noise reduction, voice data analysis, conversion rate comparison, and the optimal audio encoder bitrate regression model.
It improved the conversion rate of speech to text, reduced detection delay, and enabled timely identification and disconnection of inappropriate call content.
Smart Images
Figure CN120071938B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of telephone voice inappropriate content detection system based on natural language processing technology. Background Technology
[0002] With the advancement of communication technology, telephone voice communication has become an indispensable part of people's lives and business activities. With the development of communication technology, more and more harmful information has appeared in telephone communication. Traditional telephone voice harmful content detection systems usually use fixed models and algorithms to identify and filter harmful content, such as malicious calls and fraudulent calls. However, these systems are often unable to adapt to the ever-changing voice environment and new telephone fraud methods, resulting in poor detection results. They still do not automatically hang up when harmful content is detected in the future.
[0003] In existing technologies, most methods perform preliminary nuisance call feature analysis and issue alerts based on the correlation between the IP address and the phone number of the incoming caller. The phone number correlation is determined by whether the incoming number is recorded in the communication host's address book. If it is not identified as a nuisance call, the next step is to analyze the voice content. Existing technologies mostly use natural language processing (NLP) technology for call content detection. For example, Chinese patent CN116456028B proposes a method and system for preventing nuisance calls, and Chinese patent CN113301210B proposes a method, device, and electronic device for preventing nuisance calls based on neural networks. While combining these methods and existing technologies can detect inappropriate content in telephone voice messages, in practical applications, it has been found that these methods and existing technologies have at least the following shortcomings:
[0004] (1) During the call detection process, the conversion rate of speech recognition to text is determined by many factors, such as the speech rate and voice quality of incoming call voice. Faster speech rate and lower voice quality will affect the conversion rate of speech recognition to text, resulting in inaccurate voice content analysis, which in turn affects the user's call experience.
[0005] (2) In addition, when the speech is too fast or the speech quality is poor and the speech content cannot be recognized, the speech detection strategy cannot be adjusted in time, resulting in a delay in the control of the speech detection strategy. This leads to the situation that bad calls cannot be recognized in subsequent calls. Summary of the Invention
[0006] The purpose of this invention is to provide a telephone voice message inappropriate content detection system based on natural language processing technology to address the shortcomings of the prior art.
[0007] To achieve the above objectives, the telephone voice inappropriate content detection system based on natural language processing technology includes the following steps:
[0008] The system includes:
[0009] The data acquisition module is used to acquire the voice data of the telephone, perform noise reduction processing on the voice data to obtain denoised voice data, and convert the denoised voice data into first text data.
[0010] The first analysis module is used to input the denoised speech data, speech tone, and speech volume into a preset speech data analysis model to generate the second text data.
[0011] The comparison module is used to compare the conversion rate of the second text data with that of the first text data and generate conversion rate data.
[0012] The second analysis module is used to acquire telephone identification data, input the identification data and conversion rate data into the preset optimal audio encoder bit rate regression model to obtain the optimal audio encoder bit rate;
[0013] The control module adjusts the real-time call parameters based on the optimal audio encoder bitrate and applies the adjusted real-time call parameters to the subsequent call content detection process.
[0014] In a preferred embodiment, the voice data includes call voice data and background noise data, wherein the background noise data includes ambient sound and interference sound.
[0015] In a preferred embodiment, the noise reduction process includes the following steps:
[0016] Collect multiple silent signals of background noise data, perform statistical analysis on the multiple silent signals, and calculate the mean and variance of the silent signals;
[0017] Input the mean and variance of the silence signal into a preset noise processing model to output noise suppression data;
[0018] The noise suppression data is used to perform denoising processing to obtain denoised speech data.
[0019] In a preferred embodiment, the generation logic of the noise processing model is as follows:
[0020] Obtain historical noise data and divide the historical silence data into a first training set and a first test set. The historical noise data includes the mean of the silence signal, the variance of the silence signal, and the corresponding noise suppression data.
[0021] Construct a first regression network, using the mean and variance of the silence signal in the first training set as inputs to the first regression network, and the noise suppression data in the first training set as outputs to obtain the initial noise processing network;
[0022] The initial noise processing network after training is validated based on the first test set, and the initial noise processing network whose output is less than or equal to the preset first test error is used as the noise processing model.
[0023] In a preferred embodiment, the logic for generating the noise suppression data is as follows:
[0024] S1, acquire the noisy signal from the background noise data, and perform convolution calculation on the noisy signal and the filter's weight factors to obtain the filter's output signal. The convolution calculation formula is: ;in, Let t be the output signal of the filter, and t be the discrete time. The weighting factor of the filter, For the filter's first Each weighting factor The length of the signal, For noisy signals At the present moment And before The value at each moment, , , It is an integer greater than 0;
[0025] S2, compare the output signal of the filter with the preset original signal to calculate the error signal. The calculation formula for the error signal is as follows: ,in, For error signals, For the original signal at time The value of , For the output signal at time The value of N is the same as that of N in step S1;
[0026] S3, perform gradient descent processing based on the error signal and the noisy signal to obtain the updated weight coefficients of the filter. The formula for calculating the weight coefficients is as follows: ,in, In time Update weighting factors;
[0027] S4, Repeat steps S1 to S3 until the error between the filter's output signal and the original signal is less than or equal to the preset comparison error, then set the time... Update weight factor Output as noise suppression data for, It is an integer greater than 0.
[0028] In a preferred embodiment, the generation logic of the speech data analysis model is as follows:
[0029] Acquire historical speech data and divide it into a second training set and a second test set. The historical speech data includes denoised speech data, speech pitch, speech volume, and corresponding second text data.
[0030] Construct a second regression network, using the denoised speech data, speech pitch, and speech volume from the second training set as input to the second regression network, and the second text data from the second training set as output to obtain the initial speech data analysis network;
[0031] The initial speech data analysis network after training is validated using a second test set. The output of the initial speech data analysis network with an error less than or equal to the preset second test error is used as the speech data analysis model.
[0032] In a preferred embodiment, the logic for comparing the conversion rates of the second text data and the first text data is as follows:
[0033] Valid information is obtained from the second text data and the first text data respectively. The valid information includes the number of characters, the number of words, and the number of stop words. The valid information from the second text data is compared and analyzed with the valid information from the first text data to calculate the conversion rate data. The formula for calculating the conversion rate data is as follows: = ;in, For conversion rate data, , , These represent the number of characters, words, and stop words in the first text data. , , These represent the number of characters, words, and stop words in the second text data, respectively.
[0034] In a preferred embodiment, the identification data includes a spectrogram and an acoustic feature vector.
[0035] In a preferred embodiment, the generation logic of the optimal audio encoder bitrate regression model is as follows:
[0036] Acquire call history data and divide the call history data into a third training set and a third test set. The call history data includes recognition data, conversion rate data and the corresponding optimal audio encoder bit rate.
[0037] Construct a third regression network, using the recognition data and conversion rate from the third training set as input to the third regression network, and the optimal audio encoder bit rate from the third training set as output to obtain the initial optimal audio encoder bit rate regression network.
[0038] The initial optimal audio encoder bitrate regression network after training is validated using the third test set. The initial optimal audio encoder bitrate regression network whose output is less than or equal to the preset third test error is used as the optimal audio encoder bitrate regression model.
[0039] In a preferred embodiment, the logic for generating the optimal audio encoder bit rate is as follows:
[0040] S1: Set the initial audio encoder bit rate of the phone. , The initial value is 8, and the unit is kbps;
[0041] S2: Adjust the bit rate of the initial audio encoder, making... = + To obtain the adjusted audio encoder bit rate , and Integers greater than zero;
[0042] S3: Adjusting the audio encoder bitrate Next, the voice clarity of the telephone conversation is obtained. If the actual voice clarity is greater than or equal to the preset standard voice clarity threshold, the bit rate of the audio encoder will be adjusted. As the optimal audio encoder bit rate; if the actual clarity is less than the preset standard clarity threshold, then let = + And return to step S2, Integers greater than zero;
[0043] S4: Repeat steps S2 to S3 until the optimal audio encoder bit rate is obtained, then end the loop.
[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:
[0045] This invention discloses a system for detecting inappropriate content in telephone voice messages based on natural language processing technology, comprising:
[0046] The system acquires voice data from a phone call, performs denoising processing to obtain denoised voice data, and converts the denoised voice data into first text data. The denoised voice data, voice tone, and voice volume are input into a preset voice data analysis model to generate second text data. The conversion rate of the second text data is compared with that of the first text data to generate conversion rate data. The system acquires the phone call recognition data, inputs the recognition data and conversion rate data into a preset optimal audio encoder bitrate regression model to obtain the optimal audio encoder bitrate. Based on the optimal audio encoder bitrate, the system adjusts the real-time call parameters and applies the adjusted real-time call parameters to subsequent incoming calls. In the good content detection process; based on the above process, the present invention is beneficial to improve the accuracy of the conversion rate data of subsequent calls by optimizing the preprocessing of speech-to-text during the call detection process, thereby judging whether the voice content detection of the call is accurate. When the voice content detection is inaccurate, the voice detection strategy is adjusted in real time. That is, the optimal audio encoder bit rate is generated by using the call recognition data and conversion rate data, and the real-time audio encoder bit rate is adjusted according to the optimal audio encoder bit rate, thereby reducing the latency of voice detection strategy adjustment and realizing timely detection and disconnection of bad call content in subsequent calls. Attached Figure Description
[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0048] Figure 1 This is a flowchart of a telephone voice inappropriate content detection system based on natural language processing technology according to the present invention. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] Example 1
[0051] Please see Figure 1 As shown in the figure, this embodiment discloses a telephone voice inappropriate content detection system based on natural language processing technology. The system includes:
[0052] The data acquisition module 101 is used to acquire the voice data of the telephone, perform noise reduction processing on the voice data to obtain noise-reduced voice data, and convert the noise-reduced voice data into first text data.
[0053] Specifically, the voice data includes call voice data and background noise data, and the background noise data includes ambient sound and interference sound;
[0054] Specifically, the noise reduction process includes the following steps:
[0055] Collect multiple silent signals of background noise data, perform statistical analysis on the multiple silent signals, and calculate the mean and variance of the silent signals;
[0056] For example, if the collected multiple silent signals are: [-0.02, 0.03, -0.01, 0.00, 0.02, 0.01, -0.03, -0.01, 0.02, 0.01], then the mean is 0.005 and the variance is 0.0014.
[0057] Input the mean and variance of the silence signal into a preset noise processing model to output noise suppression data;
[0058] The generation logic of the noise processing model is as follows:
[0059] Obtain historical noise data and divide the historical silence data into a first training set and a first test set. The historical noise data includes the mean of the silence signal, the variance of the silence signal, and the corresponding noise suppression data.
[0060] Specifically, the logic for generating the noise suppression data is as follows:
[0061] S1, acquire the noisy signal from the background noise data, and perform convolution calculation on the noisy signal and the filter's weight factors to obtain the filter's output signal. The convolution calculation formula is: ;in, Let t be the output signal of the filter, and t be the discrete time. The weighting factor of the filter, For the filter's first Each weighting factor The length of the signal, For noisy signals At the present moment And before The value at each moment, , , It is an integer greater than 0;
[0062] It should be noted that the filter has been initialized before use. Noise signals usually refer to random, useless signals, such as white noise, Gaussian noise, colored noise, etc. Their characteristic is that the amplitude of the signal changes randomly over time, without obvious periodicity or regularity.
[0063] A noisy signal refers to a signal whose original signal has been altered by noise. This signal may be a useful information signal, but it has been interfered with by noise during transmission or acquisition, resulting in a deterioration in signal quality. For example, voice signals and image signals can be affected by environmental noise or transmission noise.
[0064] S2, compare the output signal of the filter with the preset original signal to calculate the error signal. The calculation formula for the error signal is as follows: ,in, For error signals, For the original signal at time The value of , For the output signal at time The value of N is the same as that of N in step S1;
[0065] It should be noted that the raw signal is acquired through a sensor.
[0066] S3, perform gradient descent processing based on the error signal and the noisy signal to obtain the updated weight coefficients of the filter. The formula for calculating the weight coefficients is as follows: ,in, In time Update weighting factors;
[0067] S4, Repeat steps S1 to S3 until the error between the filter's output signal and the original signal is less than or equal to the preset comparison error, then set the time... Update weight factor Output as noise suppression data for, It is an integer greater than 0;
[0068] It is important to understand that by repeatedly generating updated weighting factors, the output signal of the filter can be made closer to the original signal, thereby achieving the effects of noise suppression and signal enhancement.
[0069] Construct a first regression network, using the mean and variance of the silence signal in the first training set as inputs to the first regression network, and the noise suppression data in the first training set as outputs to obtain the initial noise processing network;
[0070] The initial noise processing network after training is validated based on the first test set, and the initial noise processing network whose output is less than or equal to the preset first test error is used as the noise processing model.
[0071] Noise suppression processing is performed on the noise suppression data to obtain denoised speech data.
[0072] It should be noted that the conversion process of the first text data is existing technology and will not be described in detail.
[0073] The first analysis module 102 is used to input the denoised speech data, speech tone, and speech volume into a preset speech data analysis model to generate the second text data.
[0074] Specifically, the generation logic of the speech data analysis model is as follows:
[0075] Acquire historical speech data and divide it into a second training set and a second test set. The historical speech data includes denoised speech data, speech pitch, speech volume, and corresponding second text data.
[0076] Construct a second regression network, using the denoised speech data, speech pitch, and speech volume from the second training set as input to the second regression network, and the second text data from the second training set as output to obtain the initial speech data analysis network;
[0077] The initial speech data analysis network after training is validated using a second test set. The output of the initial speech data analysis network with an error less than or equal to the preset second test error is used as the speech data analysis model.
[0078] The comparison module 103 is used to compare the conversion rate of the second text data with the first text data and generate conversion rate data.
[0079] Specifically, the logic for comparing the conversion rates of the second text data and the first text data is as follows:
[0080] Valid information is obtained from the second text data and the first text data respectively. The valid information includes the number of characters, the number of words, and the number of stop words. The valid information from the second text data is compared and analyzed with the valid information from the first text data to calculate the conversion rate data. The formula for calculating the conversion rate data is as follows: = ;in, For conversion rate data, , , These represent the number of characters, words, and stop words in the first text data. , , They are the number of characters, the number of words, and the number of stop words in the second text data respectively;
[0081] It should be understood that: the conversion rate data The closer it is to 1, the higher the conversion efficiency during the speech-to-text conversion process at this time, and the lower the situation of abnormal speech recognition. The situation of abnormal speech recognition includes incorrect recognition or non-recognition of speech due to too fast speech speed, blurred speech, or low speech volume.
[0082] It should be noted that the number of characters is the total number of characters in the text data, the number of words is the total number of words in the text data, and the stop words are words that often appear in natural language processing but have little significance for text analysis. For example, "the", "is", "at", etc. in English, and "的", "了", "在", etc. in Chinese. The units of the number of characters, the number of words, and the number of stop words are the same.
[0083] This embodiment is beneficial to improving the accurate analysis of the conversion rate data of subsequent calls by optimizing the preprocessing during the speech-to-text conversion in the process of call detection, so as to judge whether the speech content detection of the call is accurate. [[ID=Useful acoustic features are extracted from the spectrogram, such as Mel frequency cepstral coefficients (MFCC), sound intensity, and pitch, which are then used as acoustic feature vectors.
[0092] Specifically, the logic for generating the optimal audio encoder bit rate is as follows:
[0093] S1: Set the initial audio encoder bit rate of the phone. , The initial value is 8, and the unit is kbps;
[0094] S2: Adjust the bit rate of the initial audio encoder, making... = + To obtain the adjusted audio encoder bit rate , and Integers greater than zero;
[0095] S3: Adjusting the audio encoder bitrate Next, the voice clarity of the telephone conversation is obtained. If the actual voice clarity is greater than or equal to the preset standard voice clarity threshold, the bit rate of the audio encoder will be adjusted. As the optimal audio encoder bit rate; if the actual clarity is less than the preset standard clarity threshold, then let = + And return to step S2, Integers greater than zero;
[0096] It should be noted that the criteria for judging the clarity of the telephone voice is as follows:
[0097] Based on whether the call signal-to-noise ratio, spectral smoothness, voice distortion, frequency response, and time domain characteristics reach the corresponding set thresholds, the phone's voice can be made to be smooth, fluent, and distortion-free.
[0098] S4: Repeat steps S2 to S3 until the optimal audio encoder bit rate is obtained, then end the loop.
[0099] Construct a third regression network, using the recognition data and conversion rate from the third training set as input to the third regression network, and the optimal audio encoder bit rate from the third training set as output to obtain the initial optimal audio encoder bit rate regression network.
[0100] The initial optimal audio encoder bitrate regression network after training is validated using the third test set. The initial optimal audio encoder bitrate regression network whose output is less than or equal to the preset third test error is used as the optimal audio encoder bitrate regression model.
[0101] The control module 105 is used to adjust the bit rate of the real-time audio encoder according to the optimal audio encoder bit rate, and apply the adjusted real-time audio encoder bit rate to the subsequent caller ID process for detecting unwanted content.
[0102] In one specific embodiment, the adjustment step is as follows:
[0103] Obtain the optimal audio encoder bitrate;
[0104] The real-time audio encoder bitrate is adjusted based on the optimal audio encoder bitrate.
[0105] In this embodiment, the optimal audio encoder bitrate is generated by using telephone recognition data and conversion rate data. The real-time audio encoder bitrate is then adjusted based on the optimal audio encoder bitrate, thereby reducing the latency of voice detection strategy control and enabling timely detection and disconnection of inappropriate telephone content in subsequent calls.
[0106] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.
[0107] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0108] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0109] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0110] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0111] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0112] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0113] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A system for detecting inappropriate content in telephone voice messages based on natural language processing technology, characterized in that, The system includes: The data acquisition module is used to acquire the voice data of the telephone, perform noise reduction processing on the voice data to obtain denoised voice data, and convert the denoised voice data into first text data. The first analysis module is used to input the denoised speech data, speech tone, and speech volume into a preset speech data analysis model to generate the second text data. The comparison module is used to compare the conversion rate of the second text data with that of the first text data and generate conversion rate data. The second analysis module is used to acquire telephone identification data, input the identification data and conversion rate data into the preset optimal audio encoder bit rate regression model to obtain the optimal audio encoder bit rate; The control module adjusts the real-time call parameters based on the optimal audio encoder bitrate and applies the adjusted real-time call parameters to the subsequent call content detection process.
2. The telephone voice inappropriate content detection system based on natural language processing technology according to claim 1, characterized in that, The voice data includes call voice data and background noise data, and the background noise data includes ambient sound and interference sound.
3. The telephone voice inappropriate content detection system based on natural language processing technology according to claim 2, characterized in that, The noise reduction process is as follows: Collect multiple silent signals of background noise data, perform statistical analysis on the multiple silent signals, and calculate the mean and variance of the silent signals; Input the mean and variance of the silence signal into a preset noise processing model to output noise suppression data; The noise suppression data is used to perform denoising processing to obtain denoised speech data.
4. The telephone voice inappropriate content detection system based on natural language processing technology according to claim 3, characterized in that, The generation logic of the noise processing model is as follows: Obtain historical noise data and divide the historical silence data into a first training set and a first test set. The historical noise data includes the mean of the silence signal, the variance of the silence signal, and the corresponding noise suppression data. Construct a first regression network, using the mean and variance of the silence signal in the first training set as inputs to the first regression network, and the noise suppression data in the first training set as outputs to obtain the initial noise processing network; The initial noise processing network after training is validated based on the first test set, and the initial noise processing network whose output is less than or equal to the preset first test error is used as the noise processing model.
5. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 4, characterized in that, The logic for generating the noise suppression data is as follows: S1, acquire the noisy signal from the background noise data, and perform convolution calculation on the noisy signal and the filter's weight factors to obtain the filter's output signal. The convolution calculation formula is: ;in, Let t be the output signal of the filter, and t be the discrete time. The weighting factor of the filter, For the filter's first Each weighting factor The length of the signal, For noisy signals At the present moment And before The value at each moment, , , It is an integer greater than 0; S2, compare the output signal of the filter with the preset original signal to calculate the error signal. The calculation formula for the error signal is as follows: ,in, For error signals, For the original signal at time The value of , For the output signal at time The value of N is the same as that of N in step S1; S3, perform gradient descent processing based on the error signal and the noisy signal to obtain the updated weight coefficients of the filter. The formula for calculating the weight coefficients is as follows: ,in, In time Update weighting factors; S4, Repeat steps S1 to S3 until the error between the filter's output signal and the original signal is less than or equal to the preset comparison error, then set the time... Update weight factor Output as noise suppression data for, It is an integer greater than 0.
6. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 5, characterized in that, The generation logic of the speech data analysis model is as follows: Acquire historical speech data and divide it into a second training set and a second test set. The historical speech data includes denoised speech data, speech pitch, speech volume, and corresponding second text data. Construct a second regression network, using the denoised speech data, speech pitch, and speech volume from the second training set as input to the second regression network, and the second text data from the second training set as output to obtain the initial speech data analysis network; The initial speech data analysis network after training is validated using a second test set. The output of the initial speech data analysis network with an error less than or equal to the preset second test error is used as the speech data analysis model.
7. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 6, characterized in that, The logic for comparing the conversion rates of the second text data and the first text data is as follows: Valid information is obtained from the second text data and the first text data respectively. The valid information includes the number of characters, the number of words, and the number of stop words. The valid information from the second text data is compared and analyzed with the valid information from the first text data to calculate the conversion rate data. The formula for calculating the conversion rate data is as follows: = ;in, For conversion rate data, , , These represent the number of characters, words, and stop words in the first text data. , , These represent the number of characters, words, and stop words in the second text data, respectively.
8. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 1, characterized in that, The identification data includes spectrograms and acoustic feature vectors.
9. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 8, characterized in that, The generation logic of the optimal audio encoder bitrate regression model is as follows: Acquire call history data and divide the call history data into a third training set and a third test set. The call history data includes recognition data, conversion rate data and the corresponding optimal audio encoder bit rate. Construct a third regression network, using the recognition data and conversion rate from the third training set as input to the third regression network, and the optimal audio encoder bit rate from the third training set as output to obtain the initial optimal audio encoder bit rate regression network. The initial optimal audio encoder bitrate regression network after training is validated using the third test set. The initial optimal audio encoder bitrate regression network whose output is less than or equal to the preset third test error is used as the optimal audio encoder bitrate regression model.
10. A telephone voice inappropriate content detection system based on natural language processing technology according to claim 9, characterized in that, The logic for generating the optimal audio encoder bitrate is as follows: S1: Set the initial audio encoder bit rate of the phone. , The initial value is 8, and the unit is kbps; S2: Adjust the bit rate of the initial audio encoder, making... = + To obtain the adjusted audio encoder bit rate , and Integers greater than zero; S3: Adjusting the audio encoder bitrate Next, the voice clarity of the telephone conversation is obtained. If the actual voice clarity is greater than or equal to the preset standard voice clarity threshold, the bit rate of the audio encoder will be adjusted. As the optimal audio encoder bit rate; if the actual clarity is less than the preset standard clarity threshold, then let = + And return to step S2, Integers greater than zero; S4: Repeat steps S2 to S3 until the optimal audio encoder bit rate is obtained, then end the loop.