Scene verbal skill classification system

Abstract

The invention relates to the technical field of electronic digital data processing, in particular to a scene verbal skill classification system which comprises a verbal skill sample keyword extraction module, an audio processing module and a verbal skill screening module. The verbal skill sample keyword extraction module, the audio processing module and the verbal skill screening module are connected in sequence; the verbal skill sample keyword extraction module is used for extracting keywords in the fraud sample verbal skill and classifying and adding the keywords into a word library; the audio processing module is used for extracting keywords in the audio and classifying and adding the keywords into a word bank; the verbal skill screening module is used for evaluating the risk level of the text according to keywords in the word bank; according to the method, the service function requirements in special scenes can be met, and efficient and high-accuracy scene verbal skill classification can be ensured under the condition that few resources are occupied.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a scene speech classification system. Background Art

[0002] With the active promotion of anti-telecom fraud, how to efficiently and accurately match and classify large amounts of audio data in order to better control, intercept and trace the source, and protect users' personal information and property safety from loss, has become a relatively important task.

[0003] In the modern information society, a variety of methods are applied to this work. For example, some methods use semantic recognition combined with AI for classification, or match based on set rules. These technologies have greatly promoted the development of informatization in the direction of public security.

[0004] Although existing classification technologies have been widely used in this field, they still have some obvious limitations and shortcomings. First, most existing technologies combined with AI require large hardware resources and long-term, large-scale data training, and the accuracy of the results requires further manual intervention and review in the early stage. Secondly, like some other simple rule matching methods, it cannot meet the business function requirements in special scenarios. Summary of the invention

[0005] The purpose of the present invention is to provide a scenario speech classification system that can meet the business function requirements in special scenarios and can ensure efficient and high-accuracy scenario speech classification while occupying very few resources.

[0006] To achieve the above-mentioned purpose, the present invention provides a scene speech classification system, including a speech sample keyword extraction module, an audio processing module and a speech screening module; The speech sample keyword extraction module, the audio processing module and the speech screening module are connected in sequence; The speech sample keyword extraction module is used to extract keywords from the fraud sample speech and classify the keywords into a vocabulary; The audio processing module is used to extract keywords from the audio and classify the keywords into a vocabulary; The speech screening module is used to evaluate the risk level of the text based on the keywords in the vocabulary.

[0007] The speech sample keyword extraction module includes a speech sample keyword extraction unit and a category marking unit; the category marking unit is connected to the speech sample keyword extraction unit; The speech sample keyword extraction unit is used to extract keywords from the fraud sample speech; The category marking unit is used to classify keywords and add them into the vocabulary.

[0008] The audio processing module includes a noise reduction unit, a speech recognition unit, a speech text keyword extraction unit, a hot word library maintenance unit and a speech model adjustment unit; The noise reduction unit, the speech recognition unit, the speech text keyword extraction unit, the hot word library maintenance unit and the speech model adjustment unit are connected in sequence; The noise reduction unit is used to perform noise reduction processing on the input audio; The speech recognition unit is used to convert the noise-reduced audio into speech text using a speech recognition model; The speech text keyword extraction unit is used to extract keywords from the speech text and classify the keywords into a hot word library; The hot word library maintenance unit is used to continuously maintain and update the hot word library according to specific application scenarios and user needs; The speech model adjustment unit is used to increase the weight of hot words in the speech recognition model to improve the recognition accuracy of key terms.

[0009] The speech screening module includes a keyword filtering and classification unit, a risk level classification unit, a data output unit and a new word storage unit; the keyword filtering and classification unit, the risk level classification unit, the data output unit and the new word storage unit are connected in sequence; The keyword filtering and classification unit is used to match the text to be tested with a mixed regular expression and filter it according to the set risk level; The risk level classification unit is used to classify the risk level according to the set threshold value, and divide the risk level into four categories: high, medium, low, and miss; The data output unit is used to output high-risk data to the result report, and the remaining data is input into the unsupervised learning model to collect new keywords; The new word storage unit is used to manually screen and review the new words matched by the unsupervised model and add them to the word library to form a cycle.

[0010] A scenario speech classification system of the present invention, the speech sample keyword extraction module adopts the WWH method to extract sample fraud speech keywords, namely: who (who), what (what), how (how). For example, for the following fraud speech: "Hello, I am the customer service of XX Bank. There is a risk with your bank card. You need to provide your card number and password for verification." The extracted keywords are: who is he (bank customer service), what he wants to do with me (verify bank card risk), what do I need to do (provide card number and password). The audio processing module uses spectral subtraction to perform noise reduction on the input audio. This is a widely used noise reduction technology. Noise reduction is achieved by estimating the power spectrum of the noise and subtracting it from the power spectrum of the noisy frequency signal. The advantage of this method is that it is simple to implement and has a good noise reduction effect for constant noise; the audio is converted into a text mode through ASR, and keyword extraction is performed on it, and the extracted keywords are added to the hot word library. By increasing the weight of hot words in the language model, the recognition accuracy of these key terms is improved. During the recognition process, the system will give priority to hot words, making them more likely to be correctly recognized in the speech recognition output. This process can be achieved by dynamically adjusting the probability distribution of the language model, or by training a special hot word enhancement model to enhance the recognition effect. In addition, the maintenance and update of the hot word library needs to be carried out continuously according to the specific application scenarios and user needs to ensure that the system can adapt to new keywords and terms in a timely manner, thereby providing more accurate and efficient speech recognition services. The speech screening module evaluates the risk level of the text according to the keywords of the word library. By reading the text data in the file, the speech screening module checks the keywords in each record in turn, and records the hit words, categories and their risk levels after matching to a specific category. Finally, the processing results are written into a new file to facilitate further analysis and processing by users. The present invention does not use a relatively large AI model, but also takes into account the increase in labor costs caused by excessive manual intervention, and adds a visual interface result display, so that users can quickly obtain results when experiencing, and it is also more convenient to check the results. The present invention can meet the business function requirements in special scenarios, and can ensure efficient and high-accuracy scene speech classification while occupying very few resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0012] Figure 1 It is a structural diagram of a scene speech classification system of the present invention.

[0013] Figure 2 It is a flow chart of the speech sample keyword extraction module of the present invention.

[0014] Figure 3 It is a flow chart of the audio processing module of the present invention.

[0015] Figure 4 It is a flow chart of the speech screening module of the present invention.

[0016] 101- speech sample keyword extraction module, 102- audio processing module, 103- speech screening module, 104- speech sample keyword extraction unit, 105- category marking unit, 106- noise reduction unit, 107- speech recognition unit, 108- speech text keyword extraction unit, 109- hot word library maintenance unit, 110- speech model adjustment unit, 111- keyword filtering and classification unit, 112- risk level division unit, 113- data output unit, 114- new word storage unit. DETAILED DESCRIPTION

[0017] See also Figure 1-Figure 4 ,in, Figure 1 is a structural diagram of a scene speech classification system of the present invention, Figure 2 is a flow chart of a speech sample keyword extraction module of the present invention, Figure 3 is a flow chart of the audio processing module of the present invention, Figure 4 It is a flow chart of the speech screening module of the present invention.

[0018] The present invention provides a scene speech classification system, including a speech sample keyword extraction module 101, an audio processing module 102 and a speech screening module 103; the speech sample keyword extraction module 101 includes a speech sample keyword extraction unit 104 and a category marking unit 105; the audio processing module 102 includes a noise reduction unit 106, a speech recognition unit 107, a speech text keyword extraction unit 108, a hot word library maintenance unit 109 and a speech model adjustment unit 110; the speech screening module 103 includes a keyword filtering and classification unit 111, a risk level division unit 112, a data output unit 113 and a new word storage unit 114; the above-mentioned scheme can meet the business function requirements in special scenarios, and can ensure efficient and high-accuracy scene speech classification while occupying very few resources.

[0019] According to this specific implementation, the speech sample keyword extraction module 101, the audio processing module 102 and the speech screening module 103 are connected in sequence; The speech sample keyword extraction module 101 is used to extract keywords from the fraud sample speech and classify the keywords into a vocabulary; The audio processing module 102 is used to extract keywords from the audio and classify the keywords into a vocabulary; The speech screening module 103 is used to evaluate the risk level of the text according to the keywords in the vocabulary.

[0020] Keyword extraction of sample speech is the first step of this classification system. The speech template provided is the basis for subsequent system classification. The system must make judgments according to the provided speech template and category. Most speech templates are relatively lengthy. Therefore, key information must be extracted as the basis for subsequent matching to reduce the system burden and increase coverage. The speech sample keyword extraction module 101 adopts the WWH method to extract sample fraud speech keywords, namely: who (who), what (what), how (how). For example, for the following fraud speech: "Hello, I am a customer service of XX Bank. There is a risk with your bank card. You need to provide your card number and password for verification." The extracted keywords are: who is he (bank customer service), what he wants to do with me (verify bank card risks), and what I need to do (provide card number and password). The audio processing module 102 uses spectral subtraction to perform noise reduction on the input audio. This is a widely used noise reduction technology. Noise reduction is achieved by estimating the power spectrum of the noise and subtracting it from the power spectrum of the noisy frequency signal. The advantage of this method is that it is simple to implement and has a good noise reduction effect on constant noise; the audio is converted into a text mode through ASR, and keywords are extracted, and the extracted keywords are added to the hot word library. By increasing the weight of the hot words in the language model, the recognition accuracy of these key terms is improved. During the recognition process, the system will give priority to hot words, making them more likely to be correctly recognized in the speech recognition output. This process can be achieved by dynamically adjusting the probability distribution of the language model, or by training a special hot word enhancement model to enhance the recognition effect. In addition, the maintenance and update of the hot word library needs to be continuously carried out according to the specific application scenario and user needs to ensure that the system can adapt to new keywords and terms in a timely manner, thereby providing more accurate and efficient speech recognition services. The speech screening module 103 evaluates the risk level of the text according to the keywords in the vocabulary. By reading the text data in the file, the speech screening module 103 checks the keywords in each record in turn, and records the hit words, categories and their risk levels after matching to a specific category. Finally, the processing results are written into a new file to facilitate further analysis and processing by the user.

[0021] The category marking unit 105 is connected to the speech sample keyword extraction unit 104; The speech sample keyword extraction unit 104 is used to extract keywords from the fraud sample speech; The category marking unit 105 is used to classify keywords and add them into the vocabulary.

[0022] The speech sample keyword extraction unit 104 uses the WWH method to extract sample fraud speech keywords, namely: who (who), what (what), how (how). For example, for the following fraud speech: "Hello, I am the customer service of XX Bank. There is a risk with your bank card. You need to provide your card number and password for verification." The extracted keywords are: who is he (bank customer service), what does he want me to do (verify bank card risks), what do I need to do (provide card number and password). The category marking unit 105 adds the keyword classification to the word library, which refers to sensitive words related to fraud, such as JD Finance, transfer, etc.

[0023] Secondly, the noise reduction unit 106, the speech recognition unit 107, the speech text keyword extraction unit 108, the hot word library maintenance unit 109 and the speech model adjustment unit 110 are connected in sequence; The noise reduction unit 106 is used to perform noise reduction processing on the input audio; The speech recognition unit 107 is used to convert the noise-reduced audio into speech text using a speech recognition model; The speech text keyword extraction unit 108 is used to extract keywords from the speech text and classify the keywords into a hot word library; The hot word library maintenance unit 109 is used to continuously maintain and update the hot word library according to specific application scenarios and user needs; The speech model adjustment unit 110 is used to increase the weight of hot words in the speech recognition model to improve the recognition accuracy of key terms.

[0024] The noise reduction unit 106 uses spectral subtraction to perform noise reduction on the input audio, which is a widely used noise reduction technology. Noise reduction is achieved by estimating the power spectrum of the noise and subtracting it from the power spectrum of the noisy frequency signal. The advantage of this method is that it is simple to implement and has a good noise reduction effect on constant noise; the speech recognition unit 107 uses a speech recognition model to convert the denoised audio into a text mode. The speech recognition model uses an ASR model. The speech text keyword extraction unit 108 extracts keywords and adds the extracted keywords to the hot word library. The speech model adjustment unit 110 increases the weight of hot words in the language model to improve the recognition accuracy of these key terms. During the recognition process, the system will give priority to hot words so that they are more likely to be correctly recognized in the speech recognition output. This process can be achieved by dynamically adjusting the probability distribution of the language model, or by training a special hot word enhancement model to enhance the recognition effect. Specifically, when the ASR model is converting speech to text, if it encounters polyphonic words or characters, it will be converted into common words for this type of speech. For example, when the ASR model encounters the pronunciation of xinzhenke during speech to text conversion, I would prefer it to be translated into the Criminal Investigation Department rather than the Administrative Department. Therefore, the Criminal Investigation Department is added to the hot word library of the ASR model to facilitate accurate text conversion and subsequent keyword matching. In addition, the maintenance and update of the hot word library needs to be carried out continuously according to specific application scenarios and user needs to ensure that the system can adapt to new keywords and terms in a timely manner, thereby providing more accurate and efficient speech recognition services.

[0025] At the same time, the keyword filtering and classification unit 111, the risk level classification unit 112, the data output unit 113 and the new word storage unit 114 are connected in sequence; The keyword filtering and classification unit 111 is used to match the text to be tested with a mixed regular expression and filter it according to the set risk level; The risk level classification unit 112 is used to classify the risk level according to the set threshold value, and divide the risk level into four categories: high, medium, low, and miss; The data output unit 113 is used to output high-risk data to the result report, and the remaining data is input into the unsupervised learning model to collect new keywords; The new word storage unit 114 is used to manually screen and review the new words matched by the unsupervised model and add them to the word library to form a cycle.

[0026] Using hybrid regular expressions to match text can quickly and accurately capture keywords and phrases, thereby effectively filtering and classifying text. This method avoids the tedious and time-consuming traditional manual processing and improves processing efficiency.

[0027] A scene speech classification system of the present invention utilizes the technology combining AI with traditional models and manual processing, such as spectral subtraction noise reduction, ASR hot word processing, and the combination of regularization and weight division. These technologies can not only improve the accuracy of audio classification, but also enhance the adaptability and robustness of the system. For example, the audio preprocessing module uses spectral subtraction for noise reduction processing. This method has good adaptability to different types of noise and can effectively reduce the noise level under different environmental conditions, so that the system can also ensure the accurate output of the ASR model in a complex audio environment. In addition, the unsupervised learning model also ensures that the vocabulary can be updated in real time without excessive manual intervention to achieve a cyclic effect. The present invention does not use a relatively large AI model, but also takes into account the increase in labor costs caused by excessive manual intervention, and adds a visual interface result display, so that users can quickly obtain results when experiencing, and can also check the results more conveniently.

[0028] What is disclosed above is only one or more preferred embodiments of the present application, and cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A scene speech classification system, characterized in that: It includes a speech sample keyword extraction module, an audio processing module and a speech screening module; The speech sample keyword extraction module, the audio processing module and the speech screening module are connected in sequence; The speech sample keyword extraction module is used to extract keywords from the fraud sample speech and classify the keywords into a vocabulary; The audio processing module is used to extract keywords from the audio and classify the keywords into a vocabulary; The speech screening module is used to evaluate the risk level of the text based on the keywords in the vocabulary.

2. A scene speech classification system as claimed in claim 1, characterized in that: The speech sample keyword extraction module includes a speech sample keyword extraction unit and a category marking unit; the category marking unit is connected to the speech sample keyword extraction unit; The speech sample keyword extraction unit is used to extract keywords from the fraud sample speech; The category marking unit is used to classify keywords and add them into the vocabulary.

3. A scene speech classification system as claimed in claim 2, characterized in that: The audio processing module includes a noise reduction unit, a speech recognition unit, a speech text keyword extraction unit, a hot word library maintenance unit and a speech model adjustment unit; The noise reduction unit, the speech recognition unit, the speech text keyword extraction unit, the hot word library maintenance unit and the speech model adjustment unit are connected in sequence; The noise reduction unit is used to perform noise reduction processing on the input audio; The speech recognition unit is used to convert the noise-reduced audio into speech text using a speech recognition model; The speech text keyword extraction unit is used to extract keywords from the speech text and classify the keywords into a hot word library; The hot word library maintenance unit is used to continuously maintain and update the hot word library according to specific application scenarios and user needs; The speech model adjustment unit is used to increase the weight of hot words in the speech recognition model to improve the recognition accuracy of key terms.

4. A scene speech classification system as claimed in claim 3, characterized in that: The speech screening module includes a keyword filtering and classification unit, a risk level classification unit, a data output unit and a new word storage unit; the keyword filtering and classification unit, the risk level classification unit, the data output unit and the new word storage unit are connected in sequence; The keyword filtering and classification unit is used to match the text to be tested with a mixed regular expression and filter it according to the set risk level; The risk level classification unit is used to classify the risk level according to the set threshold value, and divide the risk level into four categories: high, medium, low, and miss; The data output unit is used to output high-risk data to the result report, and the remaining data is input into the unsupervised learning model to collect new keywords; The new word storage unit is used to manually screen and review the new words matched by the unsupervised model and add them to the word library to form a cycle.

Images