Voice fraud analysis method and device, equipment and storage medium

Abstract

The invention relates to a voice fraud analysis method, which comprises the following steps of: acquiring voice signals, respectively acquiring voice data from different sources through two independently configured microphone arrays, and recording the voice data into two independent and isolated audio channels; detecting the voice data in the two audio channels by using a voice activity detection technology, identifying and marking a voice activity time period in each audio channel, performing fragmentation operation on the voice data based on the voice activity time periods, and extracting a plurality of effective voice fragments; generating a dialogue text according to the extracted effective voice segments, and screening a text part to be analyzed from the generated text; and selecting a pre-trained fraud analysis model according to the data features of the to-be-analyzed text, and determining the fraud probability of the voice data through the analysis model. The method can achieve the precise processing and analysis of voice signals, remarkably improves the quality of voice data and the accuracy of fraud detection, and is suitable for a complex voice interaction scene.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a voice fraud analysis method, device, equipment and storage medium. Background Art

[0002] Existing technologies in the field of voice fraud analysis mainly rely on traditional speech recognition and rule-based detection systems. These methods may be effective in dealing with simple and direct fraudulent behaviors, but their accuracy and real-time performance are often limited when faced with complex conversations and multi-channel data. In particular, existing systems have obvious deficiencies in distinguishing different speakers in binaural voices, understanding implicit intentions, and adapting to changing fraud strategies. In addition, most existing technologies lack intelligent analysis capabilities, have difficulty adapting to diverse fraud patterns, and perform poorly in terms of user customization and flexibility, and cannot meet the specific needs of specific industries or businesses. These defects limit the application potential of existing technologies in high-risk business scenarios, especially in the banking and insurance industries, where a more efficient and intelligent voice fraud analysis method is urgently needed.

[0003] In addition, in multiple business scenarios in banking and insurance, black and gray industries and fraud occur frequently. For example, there are professional black and gray industry teams engaged in anti-collection, insurance fraud, agency cancellation and other businesses, which have caused huge losses to banks and insurance companies. At present, the identification of such situations mainly relies on manual judgment, which is not only inefficient but also difficult to guarantee accuracy. There is an urgent need for a more intelligent method to detect and warn these fraud phenomena in a timely manner. Summary of the invention

[0004] The main purpose of the present invention is to provide a voice fraud analysis method, device, equipment and storage medium, aiming to solve the technical problems of low efficiency and accuracy of existing voice fraud analysis technology in identifying fraud scenarios.

[0005] To achieve the above object, the present invention provides a voice fraud analysis method, which comprises the following steps:

[0006] Acquire a voice signal, collect voice data from different sources from the voice signal through two independently configured microphone arrays, and record the voice data into two independent and isolated audio channels respectively;

[0007] The voice data in the two audio channels are detected by using voice activity detection technology, the voice activity time period in each audio channel is identified and marked, and the voice data is segmented based on the identified voice activity time period, so that the continuous voice data in each audio channel is segmented into multiple independent voice segments, and multiple valid voice segments are obtained;

[0008] Generate a dialogue text according to the valid voice segment, and determine the text to be analyzed in the dialogue text;

[0009] A pre-trained fraud analysis model is selected according to the data features of the text to be analyzed, and the fraud probability of the voice data is determined according to the text to be analyzed and the pre-trained fraud analysis model.

[0010] Optionally, the step of generating a dialogue text according to the valid voice segment includes:

[0011] Identify the start time and end time of the valid voice segment in the voice data, and identify the source person of the valid voice segment;

[0012] Sort the valid voice segments according to the start time and the source person to obtain time-sequence voice segments;

[0013] A dialogue text is generated according to the time-sequence speech segments.

[0014] Optionally, the step of generating a dialogue text according to the time-sequence speech segments includes:

[0015] The time-sequenced voice segments are converted into voice text using voice recognition technology, and the voice text is marked according to the source person to generate the dialogue text.

[0016] Optionally, the step of acquiring a voice signal includes:

[0017] Selecting a device that supports dual-channel recording, including a recording device with independent left-channel and right-channel recording functions;

[0018] Setting the recording parameters of the recording device, including sampling rate and quantization accuracy;

[0019] The left channel and the right channel of the recording device are used to record audio signals from two different directions or sources respectively to obtain a two-channel audio signal;

[0020] The two-channel audio signal is preprocessed, and the speech signal belonging to human voice is identified in the preprocessed two-channel audio signal.

[0021] Optionally, the step of identifying and marking the voice activity time period in each audio channel, performing a segmentation operation on the voice data based on the identified voice activity time period, and dividing the continuous voice data in each audio channel into a plurality of independent voice segments, and obtaining a plurality of valid voice segments comprises:

[0022] Constructing a speech recognition model, detecting speech activity indicators of the speech data through the speech recognition model, identifying and marking time periods containing speech activities, and generating start and end time markers for each speech segment;

[0023] Segmentation is performed based on the marked voice activity time period, and the continuous voice data in each audio channel is segmented according to the marked time period. Each segmented voice segment corresponds to an independent voice activity time period.

[0024] Exclude meaningless speech segments during segmentation, including silent segments, background noise, and irrelevant modal particles;

[0025] The valid voice segment is obtained, where the valid voice segment is a voice data segment that has been filtered and contains actual conversation content.

[0026] Optionally, the step of determining the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model includes:

[0027] Inputting the text to be analyzed into the pre-trained fraud analysis model to determine the number of fraud clauses in the text to be analyzed;

[0028] The fraud probability of the voice data is determined according to the number of the fraud clauses and a preset fraud strategy set.

[0029] Optionally, before the step of determining the fraud probability of the voice data according to the number of the fraud clauses and a preset fraud strategy set, the step further includes:

[0030] Acquire historical fraud data, and define the preset fraud strategy set according to business requirements and the historical fraud data.

[0031] Furthermore, in order to achieve the above-mentioned object, the present invention also provides a voice fraud analysis device, the voice fraud analysis device comprising:

[0032] A voice signal processing module, which obtains a voice signal, collects voice data from different sources from the voice signal through two independently configured microphone arrays, and records the voice data into two independent and isolated audio channels;

[0033] The effective segment extraction module detects the voice data in the two audio channels respectively by using the voice activity detection technology, identifies and marks the voice activity time period in each audio channel, and performs segmentation operation on the voice data based on the identified voice activity time period, so as to divide the continuous voice data in each audio channel into multiple independent voice segments, and obtain multiple effective voice segments;

[0034] A text conversion module generates a dialogue text according to the valid voice segment and determines a text to be analyzed in the dialogue text;

[0035] The fraud analysis module selects a pre-trained fraud analysis model according to the data features of the text to be analyzed, and determines the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model.

[0036] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a voice fraud analysis device, which includes a memory, a processor, and a voice fraud analysis program stored in the memory and executable on the processor, and the voice fraud analysis program, when executed by the processor, implements the steps of the voice fraud analysis method as described above.

[0037] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a storage medium, on which a voice fraud analysis program is stored, and when the voice fraud analysis program is executed by a processor, the steps of the voice fraud analysis method described above are implemented.

[0038] The present invention relates to a voice fraud analysis method, which acquires voice signals, respectively collects voice data from different sources through two independently configured microphone arrays, and records the voice data into two independent and isolated audio channels; uses voice activity detection technology to detect the voice data in the two audio channels, identifies and marks the voice activity time period in each audio channel, performs segmentation operations on the voice data based on the voice activity time period, and extracts multiple valid voice segments; generates a dialogue text based on the extracted valid voice segments, and screens out the text part to be analyzed in the generated text; selects a pre-trained fraud analysis model based on the data features of the text to be analyzed, and determines the fraud probability of the voice data through the analysis model. The present invention can realize the precise processing and analysis of voice signals, significantly improve the quality of voice data and the accuracy of fraud detection, and is suitable for complex voice interaction scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the structure of the device hardware operating environment involved in the embodiment of the voice fraud analysis device of the present invention;

[0040] Figure 2 This is a flow chart of the first embodiment of the voice fraud analysis method of the present invention;

[0041] Figure 3 Schematic diagram of functional modules of a preferred embodiment of the voice fraud analysis device of the present invention.

[0042] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0043] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0044] It should be noted that the existing technologies in the field of voice fraud analysis mainly rely on traditional speech recognition and rule-based detection systems. These methods may be effective in dealing with simple and direct fraudulent behaviors, but their accuracy and real-time performance are often limited when faced with complex conversations and multi-channel data. In particular, existing systems have obvious deficiencies in distinguishing different speakers in binaural voices, understanding implicit intentions, and adapting to changing fraud strategies. In addition, most existing technologies lack intelligent analysis capabilities, have difficulty adapting to diverse fraud patterns, and perform poorly in terms of user customization and flexibility, and cannot meet the specific needs of specific industries or businesses. These defects limit the application potential of existing technologies in high-risk business scenarios, especially in the banking and insurance industries, where a more efficient and intelligent voice fraud analysis method is urgently needed.

[0045] In addition, in multiple business scenarios in banking and insurance, black and gray industries and fraud occur frequently. For example, there are professional black and gray industry teams engaged in anti-collection, insurance fraud, agency cancellation and other businesses, which have caused huge losses to banks and insurance companies. At present, the identification of such situations mainly relies on manual judgment, which is not only inefficient but also difficult to guarantee accuracy. There is an urgent need for a more intelligent method to detect and warn these fraud phenomena in a timely manner.

[0046] Based on the above-mentioned defects, the present invention provides a voice fraud analysis device, referring to Figure 1 , Figure 1 This is a schematic diagram of the structure of the device hardware operating environment involved in the embodiment of the voice fraud analysis device of the present invention.

[0047] like Figure 1 As shown, the voice fraud analysis device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0048] Those skilled in the art will understand that Figure 1 The hardware structure of the voice fraud analysis device shown in the figure does not constitute a limitation of the voice fraud analysis device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0049] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module and a voice fraud analysis program. The operating system is a program for managing and controlling the voice fraud analysis device and software resources, and supports the operation of the network communication module, the user interface module, the voice fraud analysis program and other programs or software; the network communication module is used to manage and control the network interface 1004; the user interface module is used to manage and control the user interface 1003.

[0050] exist Figure 1 In the hardware structure of the voice fraud analysis device shown, the network interface 1004 is mainly used to connect to the backend server and perform data communication with the backend server; the user interface 1003 is mainly used to connect to the client and perform data communication with the client; the processor 1001 can call the voice fraud analysis program stored in the memory 1005 and perform the following operations:

[0051] Acquire a voice signal, collect voice data from different sources from the voice signal through two independently configured microphone arrays, and record the voice data into two independent and isolated audio channels respectively;

[0052] The voice data in the two audio channels are detected by using voice activity detection technology, the voice activity time period in each audio channel is identified and marked, and the voice data is segmented based on the identified voice activity time period, so that the continuous voice data in each audio channel is segmented into multiple independent voice segments, and multiple valid voice segments are obtained;

[0053] Generate a dialogue text according to the valid voice segment, and determine the text to be analyzed in the dialogue text;

[0054] A pre-trained fraud analysis model is selected according to the data features of the text to be analyzed, and the fraud probability of the voice data is determined according to the text to be analyzed and the pre-trained fraud analysis model.

[0055] Furthermore, the step of generating a dialogue text according to the valid voice segment includes:

[0056] Identify the start time and end time of the valid voice segment in the voice data, and identify the source person of the valid voice segment;

[0057] Sort the valid voice segments according to the start time and the source person to obtain time-sequence voice segments;

[0058] A dialogue text is generated according to the time-sequence speech segments.

[0059] Furthermore, the step of generating a dialogue text according to the time-sequence speech segments includes:

[0060] The time-sequenced voice segments are converted into voice text using voice recognition technology, and the voice text is marked according to the source person to generate the dialogue text.

[0061] Furthermore, the step of acquiring the voice signal includes:

[0062] Selecting a device that supports dual-channel recording, including a recording device with independent left-channel and right-channel recording functions;

[0063] Setting the recording parameters of the recording device, including sampling rate and quantization accuracy;

[0064] The left channel and the right channel of the recording device are used to record audio signals from two different directions or sources respectively to obtain a two-channel audio signal;

[0065] The two-channel audio signal is preprocessed, and the speech signal belonging to human voice is identified in the preprocessed two-channel audio signal.

[0066] Furthermore, the step of identifying and marking the voice activity time period in each audio channel, performing a segmentation operation on the voice data based on the identified voice activity time period, and dividing the continuous voice data in each audio channel into a plurality of independent voice segments, and obtaining a plurality of valid voice segments includes:

[0067] Constructing a speech recognition model, detecting speech activity indicators of the speech data through the speech recognition model, identifying and marking time periods containing speech activities, and generating start and end time markers for each speech segment;

[0068] Segmentation is performed based on the marked voice activity time period, and the continuous voice data in each audio channel is segmented according to the marked time period. Each segmented voice segment corresponds to an independent voice activity time period.

[0069] Exclude meaningless speech segments during segmentation, including silent segments, background noise, and irrelevant modal particles;

[0070] The valid voice segment is obtained, where the valid voice segment is a voice data segment that has been filtered and contains actual conversation content.

[0071] Furthermore, the step of determining the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model includes:

[0072] Inputting the text to be analyzed into the pre-trained fraud analysis model to determine the number of fraud clauses in the text to be analyzed;

[0073] The fraud probability of the voice data is determined according to the number of the fraud clauses and a preset fraud strategy set.

[0074] Furthermore, before the step of determining the fraud probability of the voice data according to the number of the fraud clauses and the preset fraud strategy set, it also includes:

[0075] Acquire historical fraud data, and define the preset fraud strategy set according to business requirements and the historical fraud data.

[0076] The specific implementation of the voice fraud analysis device of the present invention is basically the same as the various embodiments of the voice fraud analysis method described below, and will not be repeated here.

[0077] The present invention also provides a voice fraud analysis method based on the voice fraud analysis device.

[0078] Reference Figure 2 , Figure 2 Schematic diagram of the flow of the first embodiment of the voice fraud analysis method of the present invention.

[0079] The embodiment of the present invention provides an embodiment of a voice fraud analysis method. It should be noted that although a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in an order different from that shown here.

[0080] In various embodiments of the voice fraud analysis method, the execution entity is a voice fraud analysis device.

[0081] The voice fraud analysis method is applied to a voice fraud analysis device, and the voice fraud analysis method comprises:

[0082] Step S10, acquiring a voice signal, collecting voice data from different sources from the voice signal through two independently configured microphone arrays, and recording the voice data into two independent and isolated audio channels respectively;

[0083] In this embodiment, during a call, voice signals are acquired through two independently configured microphone arrays. These microphone arrays can capture voice data from different directions or sources respectively, and record these data into two independent and isolated audio channels. This step ensures that the voice signals recorded in each channel are clear and do not interfere with each other.

[0084] The two-channel audio signals captured by the microphone array naturally have the characteristic of spatial separation, that is, the left channel corresponds to the voice signal from one source (such as the user's voice), and the right channel corresponds to the voice signal from another source (such as the other party's voice). This spatial separation is achieved through the independent configuration of the microphone array, ensuring that the voice data of the two channels can be processed and analyzed independently, without the need for complex signal separation processing later.

[0085] In the separated left and right channel audio data, voice activity detection (VAD) technology is used to identify the time period of voice activity. Through VAD technology, the time period that actually contains voice content in each channel can be accurately located, invalid segments (such as silence, noise) can be excluded, and valid voice segments can be extracted to provide high-quality input for subsequent processing.

[0086] Step S20, using voice activity detection technology to detect the voice data in the two audio channels respectively, identify and mark the voice activity time period in each audio channel, and perform a segmentation operation on the voice data based on the identified voice activity time period, so as to segment the continuous voice data in each audio channel into multiple independent voice segments, and obtain multiple valid voice segments;

[0087] In this embodiment, the voice activity segment refers to the portion of the audio signal that actually contains voice. These time periods are identified by voice activity detection technology (VAD), which is usually based on the characteristics of the audio signal, such as energy level, spectrum characteristics, zero crossing rate, etc. VAD is a technology used to identify and distinguish voice signals from non-voice signals (such as silence, background noise, etc.). By analyzing the energy, spectrum, zero crossing rate and other characteristics of the audio signal, VAD technology can detect the portion of the audio signal that actually contains voice activity and mark these time periods. VAD is usually used in real-time communication, voice recording and voice processing systems to help reduce the interference of non-voice data, thereby improving the processing efficiency and accuracy of the system.

[0088] Each channel (left channel and right channel) in a two-channel recording may record speech signals from different sources (such as two different speakers). In order to ensure the validity and accuracy of the speech data in each channel, VAD technology is applied to the two audio channels separately.

[0089] In each channel, VAD runs independently, identifies the time periods in that channel that actually contain speech, and marks these active segments. This allows the speech data of the two channels to be processed independently to avoid interfering with each other.

[0090] Segmentation refers to dividing a continuous audio signal into multiple independent voice segments based on the identified voice activity time period. Each segment corresponds to a voice activity time period, ensuring that the segmented segments are logically independent and complete voice data. Based on the identified voice activity, corresponding time tags are generated, which record the start and end time of the voice activity. Based on these time tags, the audio signal is segmented into several independent voice segments. The segmentation process is performed through a signal processing algorithm, which can be accurate to the millisecond level to ensure the accuracy of the segment boundaries.

[0091] Valid voice segments refer to segments that actually contain voice content after segmentation and filtering operations. These segments remove meaningless parts (such as silence, background noise, and irrelevant modal particles) and only retain valuable data related to voice activity. After the segmentation operation is completed, the system further filters out segments that do not contain actual conversation content. Filtering can be based on the intensity of voice activity, segment length, voice features, etc. The retained valid voice segments are high-quality voice data that can be used for subsequent text generation and voice analysis.

[0092] The process of identifying valid voice segments is to ensure that useful conversation content is extracted from the recorded data while excluding background noise and meaningless filler words. Detect and mark valid voice segments in each channel in real time, exclude meaningless human voices (such as "um", "ah", etc.), and use edge computing to achieve local real-time voice activity detection, reduce latency and improve efficiency. Steps may include:

[0093] First, feature extraction is performed, such as Mel-frequency cepstral coefficients (MFCCs), spectrogram, energy, and zero-crossing rate. These features can help the model distinguish between speech and meaningless noise.

[0094] Mel-frequency cepstral coefficients (MFCCs): MFCCs are features extracted from speech signals and can effectively represent the spectral characteristics of speech.

[0095] Spectrogram: Spectrogram represents the energy distribution of speech signal at different frequencies.

[0096] Energy: The energy characteristic of a speech signal, which can be used to distinguish speech from silence.

[0097] Zero Crossing Rate: Indicates the number of times the signal passes through the zero value and can be used to detect voice activity.

[0098] Then, the valid speech segments in the speech data can be identified by voice activity detection. Voice activity detection is used to distinguish speech signals from silent parts and identify valid speech segments. The steps may include:

[0099] Select a suitable VAD algorithm, such as simple VAD based on energy, VAD based on statistical models, or deep learning models (such as CNN, RNN) for detection. Deep learning models can be trained to more accurately identify voice activities.

[0100] The extracted speech features (such as MFCCs, spectrograms, etc.) are input into the VAD model. The model outputs the probability or binary result (speech / non-speech) of whether each time frame contains speech activity. The detection results are smoothed to reduce misjudgments caused by noise. Methods such as sliding windows or low-pass filters can be used. Meaningless filler words and background noise are filtered out to retain valid conversation content.

[0101] The separated and extracted valid voice segments are converted into text to generate call records. Then, the text parts that need further analysis are screened and marked for semantic analysis or fraud detection.

[0102] Step S30, generating a dialogue text according to the valid voice segment, and determining a text to be analyzed in the dialogue text;

[0103] In this embodiment, after valid voice segments are identified, these segments need to be converted into dialogue texts, and further the texts to be analyzed for fraud analysis are determined. The steps may include:

[0104] First, determine the speech recognition model. You can use one or more speech recognition models, such as deep neural network (DNN), convolutional neural network (CNN), Transformer model, etc., and improve the accuracy of transcription through ensemble learning. Multi-model integration can effectively reduce the error of a single model and improve the overall recognition effect.

[0105] At the same time, the speech recognition model can be trained according to user needs to adapt it to the professional terms and expressions in different fields. For example, the terminology of the financial industry is different from that of the insurance industry, and the model needs to be trained and adjusted separately.

[0106] Moreover, the speech recognition model can be trained to support the recognition of multiple languages ​​and dialects, increasing its applicability in multilingual environments. For example, by adding training data for commonly used dialects, the application effect can be improved in different regions.

[0107] When the speech recognition model is used to transcribe valid speech segments into text, the timestamp information of each speech segment is maintained to ensure the synchronization of the generated text with the original speech. Timestamp synchronization helps to quickly locate specific speech segments during subsequent analysis.

[0108] Integrate the transcribed text fragments in chronological order to generate a complete conversation text, ensuring the natural flow and logic of the conversation for subsequent analysis.

[0109] The speaker identity (such as customer service or client) of each sentence in the conversation text is marked to ensure that the speech of different roles can be distinguished during analysis. Speaker annotation can be distinguished through voiceprint recognition or dual-channel recording data.

[0110] Clean up the conversation text and remove noise words and meaningless filler words, such as "um", "ah", "oh", etc. You can use pre-trained language models (such as GPT-4, BERT) for semantic analysis to identify and remove meaningless content.

[0111] Finally, determine the text to be analyzed for possible fraud detection from the generated conversation text. The steps may include:

[0112] One implementation method is to select a text of a specific source person from the conversation text as the text to be analyzed. For example, in the conversation text between the customer service and the customer, the customer's conversation text content is selected as the text to be analyzed.

[0113] Another implementation method may be to extract key sentences and potential fraudulent behavior descriptions from the generated conversation text. Natural language processing (NLP) techniques, such as keyword extraction and topic modeling, may be used to identify important information in the conversation. The conversation content may also be understood based on the context, and key paragraphs and sentences in the conversation may be determined, and the key paragraphs and sentences may be used as the text to be analyzed. For example, through semantic analysis and context association, important information such as questions raised by customers and answers from customer service may be identified.

[0114] Step 40: Select a pre-trained fraud analysis model based on the data features of the text to be analyzed, and determine the fraud probability of the voice data based on the text to be analyzed and the pre-trained fraud analysis model.

[0115] In this embodiment, data features refer to information extracted from the text to be analyzed, including but not limited to keywords, semantic patterns, grammatical structures, contextual relationships, sentiment analysis results, etc. These features reflect the nature of the text content and provide a basis for model selection and analysis.

[0116] Use Natural Language Processing (NLP) techniques to extract valuable features from text. This may include Bag-of-Words, Word Embeddings, TF-IDF values, sentence structure analysis, Named Entity Recognition, etc.

[0117] Pre-trained models refer to machine learning or deep learning models that have been preliminarily trained on a large corpus. These models already have strong language understanding capabilities and can be adapted to specific tasks through fine-tuning.

[0118] Based on the extracted text features, select the pre-trained model that is most suitable for analyzing the text to be analyzed. The selection criteria may include the model's architecture (such as BERT, GPT-3, RoBERTa), the model's performance on similar tasks, and the model's ability to handle terminology in a specific field. By comparing the text features with the model's capabilities, the optimal model is determined. For example, a text containing a large amount of financial terms may require the selection of a model pre-trained in the financial field.

[0119] Fraud probability refers to the possibility of fraud in text content after model analysis. This probability value is usually expressed as a decimal between 0 and 1, where the closer to 1, the more likely it is that fraud exists. The text to be analyzed is input into the selected pre-trained model, and the model analyzes the text features through its inherent parameters and structure, and outputs a fraud probability. This process may involve the calculation of multi-layer neural networks, matching of feature vectors, and the application of risk assessment algorithms.

[0120] The present invention relates to a voice fraud analysis method, which acquires voice signals, respectively collects voice data from different sources through two independently configured microphone arrays, and records the voice data into two independent and isolated audio channels; uses voice activity detection technology to detect the voice data in the two audio channels, identifies and marks the voice activity time period in each audio channel, performs segmentation operations on the voice data based on the voice activity time period, and extracts multiple valid voice segments; generates a dialogue text based on the extracted valid voice segments, and screens out the text part to be analyzed in the generated text; selects a pre-trained fraud analysis model based on the data features of the text to be analyzed, and determines the fraud probability of the voice data through the analysis model. The present invention can realize the precise processing and analysis of voice signals, significantly improve the quality of voice data and the accuracy of fraud detection, and is suitable for complex voice interaction scenarios.

[0121] Furthermore, a second embodiment of the voice fraud analysis method of the present invention is proposed. In the above step S20, the step of generating a dialogue text according to the valid voice segment includes:

[0122] Step a1, identifying the start time and end time of the valid voice segment in the voice data, and identifying the source person of the valid voice segment;

[0123] Step a2, sorting the valid voice segments according to the start time and the source person to obtain time-sequenced voice segments;

[0124] Step a3, generating a dialogue text according to the time-sequence speech segments.

[0125] In this embodiment, the start time and end time of each valid voice segment in the voice data are identified. Timestamp information is the key to ensuring the timing and synchronization of the conversation. The source person (such as customer service or client) of each valid voice segment is determined. The voices of different speakers can be distinguished through dual-channel recording data, voiceprint recognition technology or other recognition methods.

[0126] According to the start time and source person of the recognition, all valid voice segments are sorted to obtain time-sequenced voice segments, ensuring the natural fluency and logic of the conversation.

[0127] Convert the sorted time-sequenced speech segments into text to generate a complete conversation text. Ensure the timing and coherence of the conversation text for subsequent analysis.

[0128] Label the speaker identity of each sentence in the conversation text (such as customer service or customer) to ensure that the speeches of different roles can be distinguished during analysis.

[0129] For example:

[0130] Segment 1: Start time - 00:00:05, End time - 00:00:10;

[0131] Clip 2: Start time - 00:00:12, End time - 00:00:15;

[0132] Segment 3: Start time - 00:00:08, End time - 00:00:11.

[0133] Identify the source of the voice clip.

[0134] Clip 1: Source - Customer Service;

[0135] Fragment 2: Source person-customer;

[0136] Clip 3: Source person-customer.

[0137] According to the start time and source person of the recognition, all valid voice segments are sorted to obtain time-sequenced voice segments, ensuring the natural fluency and logic of the conversation.

[0138] Clip 1: Start time - 00:00:05, End time - 00:00:10, Source - Customer Service;

[0139] Clip 3: Start time - 00:00:08, End time - 00:00:11, Source - Customer;

[0140] Clip 2: Start time – 00:00:12, End time – 00:00:15, Source – Customer.

[0141] Convert the sorted time-sequence speech segments into text to generate a complete conversation text.

[0142] Clip 1 (Customer Service): Hello, this is the bank’s customer service center. How can I help you?

[0143] Snippet 3 (Customer): I have some questions about my credit card bill.

[0144] Clip 2 (Customer): There was an expense on the bill that I don't remember.

[0145] Clean up the conversation text and remove noise words and meaningless filler words, such as "um", "ah", "oh", etc. And mark the speaker identity of each sentence in the conversation text (such as customer service or customer) to ensure that the speeches of different roles can be distinguished during analysis.

[0146] Customer Service: Hello, this is the bank’s customer service center. How can I help you?

[0147] Customer: I have some questions about my credit card bill.

[0148] Customer: There is an expense on the bill that I don't remember.

[0149] Through the above steps, this embodiment can improve the natural fluency and logic of the conversation and provide high-quality input data for subsequent fraud analysis.

[0150] Furthermore, a third embodiment of the voice fraud analysis method of the present invention is proposed, wherein the above step a3 comprises:

[0151] Step b1, using speech recognition technology to convert the time-series speech segments into speech text, marking the speech text according to the source person, and generating the dialogue text.

[0152] In this embodiment, after valid voice segments are identified, these segments need to be converted into dialogue texts, and further the texts to be analyzed for fraud analysis need to be determined. The steps of generating dialogue texts based on time-series voice segments may include:

[0153] Using speech recognition technology to convert time-series speech segments into speech text, you can use a variety of speech recognition models, such as deep neural network (DNN), convolutional neural network (CNN), Transformer model, etc., to improve the accuracy of transcription through ensemble learning. Multi-model integration can effectively reduce the error of a single model and improve the overall recognition effect. Label the speaker identity of each sentence in the conversation text (such as customer service or client) to ensure that the speech of different roles can be distinguished during analysis.

[0154] For example:

[0155] Clip 1: Start time - 00:00:05, End time - 00:00:10, Source - Customer Service;

[0156] Clip 3: Start time - 00:00:08, End time - 00:00:11, Source - Customer;

[0157] Clip 2: Start time – 00:00:12, End time – 00:00:15, Source – Customer.

[0158] Convert the sorted time-sequence speech segments into text to generate a complete conversation text.

[0159] Clip 1 (Customer Service): Hello, this is the bank’s customer service center. How can I help you?

[0160] Snippet 3 (Customer): I have some questions about my credit card bill.

[0161] Clip 2 (Customer): There was an expense on the bill that I don't remember.

[0162] Clean up the conversation text and remove noise words and meaningless filler words, such as "um", "ah", "oh", etc. And mark the speaker identity of each sentence in the conversation text (such as customer service or customer) to ensure that the speeches of different roles can be distinguished during analysis.

[0163] Customer Service: Hello, this is the bank’s customer service center. How can I help you?

[0164] Customer: I have some questions about my credit card bill.

[0165] Customer: There is an expense on the bill that I don't remember.

[0166] Through the above method, this embodiment can accurately convert valid voice segments into dialogue texts, and can mark the source of the dialogue content, so as to filter the texts to be analyzed that need to be detected according to the dialogue objects.

[0167] Furthermore, a fourth embodiment of the voice fraud analysis method of the present invention is proposed. In the above step S10, the step of acquiring the voice signal includes:

[0168] Step c1, selecting a device that supports dual-channel recording, wherein the device includes a recording device with independent left-channel and right-channel recording functions;

[0169] Step c2, setting the recording parameters of the recording device, including sampling rate and quantization accuracy;

[0170] Step c3, recording audio signals from two different directions or sources through the left channel and the right channel of the recording device respectively to obtain a two-channel audio signal;

[0171] Step c4, preprocessing the two-channel audio signal, and identifying the speech signal belonging to human voice in the preprocessed two-channel audio signal.

[0172] In this embodiment, a device capable of recording a two-channel audio signal is selected. The two-channel recording device has independent left and right channels, and can capture sound signals from two different directions or sources at the same time. This ensures that the recorded audio signal has spatial separation characteristics, so that subsequent processing can better identify and distinguish voice signals from different sources. Set key parameters of the recording device, such as sampling rate and quantization accuracy. The sampling rate determines the number of audio samples recorded per second, and a high sampling rate helps to retain more audio details. The quantization accuracy determines the dynamic range and signal-to-noise ratio of the audio signal, and a higher quantization accuracy can improve the clarity and sound quality of the recording. The microphone arrays of the left and right channels record audio signals from two different directions or sources respectively. This step ensures that the recorded audio signal can accurately reflect the sounds from different sources and maintain the independence between the channels, which is helpful for subsequent signal processing and analysis. After the recording is completed, the two-channel audio signal is preprocessed.

[0173] Generally speaking, in addition to the conversation voice, the call content also includes various audio such as environmental noise. It is necessary to extract the conversation voice data belonging to the human voice from the messy audio. Among them, high-fidelity stereo recording equipment can be used to obtain two-channel audio data, including the two-channel audio data of the customer service and customer conversation. These audio data can come from call center recordings, telephone customer service recordings, etc.

[0174] The collected audio data is preprocessed, including: using noise suppression algorithms to reduce the interference of background noise; ensuring that the volume of different sound sources is consistent to improve the accuracy of speech recognition; reducing the impact of echo on voice data to ensure audio quality.

[0175] Identify human voice from pre-processed two-channel audio data. Use feature extraction and voice activity detection (VAD) technology to distinguish human voice from background noise.

[0176] Feature extraction can be done in the following ways:

[0177] Extract Mel-frequency cepstral coefficients (MFCCs) and extract 13 MFCCs as the spectral features of the speech signal. The specific steps are as follows: Pre-emphasize the audio signal to enhance the high-frequency components. Divide the signal into frames, and each frame length is usually 20-40ms. Window each frame of the signal, such as using a Hamming window. Calculate the fast Fourier transform (FFT) of each frame of the signal to obtain the power spectrum. Apply a Mel filter bank and perform a weighted summation of the power spectrum. Take the logarithm of the Mel power spectrum and calculate the discrete cosine transform (DCT) to obtain the MFCCs.

[0178] Extract the spectrogram and use the short-time Fourier transform (STFT) to generate the spectrogram of the audio signal. Decompose the audio signal into multiple time frames, perform Fourier transform on each frame, and obtain the frequency distribution.

[0179] Energy extraction: Calculate the energy of each frame of audio signal as a feature to distinguish speech from silence. Frames with higher energy usually contain speech signals.

[0180] Extract the zero crossing rate (Zero Crossing Rate) and calculate the zero crossing rate of each frame signal, that is, the number of times the signal passes through the zero value. The zero crossing rate can be used to detect voice activity because the zero crossing rate of the voice signal is relatively high.

[0181] Then, choose a suitable VAD algorithm, which can be an energy-based method, a statistical model-based method, or a deep learning model (such as CNN, RNN). The extracted MFCCs, spectrograms, energy, and zero-crossing rate features are input into the VAD model. The VAD model is trained with a large amount of speech data and can accurately distinguish between speech and non-speech parts.

[0182] Through the above steps, this embodiment accurately identifies the voice data belonging to human voice from the pre-processed two-channel audio data, ensuring that the subsequent speech-to-text conversion and fraud analysis have high accuracy and reliability.

[0183] Furthermore, a fifth embodiment of the voice fraud analysis method of the present invention is proposed. In the above step S10, the steps of identifying and marking the voice activity time period in each audio channel, performing a segmentation operation on the voice data based on the identified voice activity time period, and segmenting the continuous voice data in each audio channel into multiple independent voice segments, and obtaining multiple valid voice segments include:

[0184] Step d1, constructing a speech recognition model, detecting speech activity indicators of the speech data through the speech recognition model, identifying and marking time periods containing speech activities, and generating start and end time markers for each speech segment;

[0185] Step d2, segmenting based on the marked voice activity time period, segmenting the continuous voice data in each audio channel according to the marked time period, and each segmented voice segment corresponds to an independent voice activity time period;

[0186] Step d3, excluding meaningless speech segments during the segmentation process, including silent segments, background noise, and irrelevant modal particles;

[0187] Step d4, obtaining the valid voice segment, wherein the valid voice segment is a voice data segment that has been filtered and contains actual conversation content.

[0188] In this embodiment, a model specifically used for voice activity detection is constructed. The model identifies which time periods contain actual voice activity by analyzing voice activity indicators in voice data. Voice activity indicators may include energy levels, spectral features, zero crossing rates, phoneme recognition, etc. of audio signals.

[0189] The model determines the presence of speech activity by analyzing various indicators in the speech signal. For example, when the energy of the audio signal exceeds a certain threshold, it can be considered that there is speech activity during that time period. The model generates the start and end time markers of each speech segment based on these indicators.

[0190] The model identifies the time periods with speech activity based on the detected speech activity indicators. The model generates corresponding time tags for these time periods, which record the start and end time of the speech activity. Time tags are precise locations of speech activity time periods in the audio signal, used to indicate the start and end points of each speech segment. These tags will guide subsequent segmentation operations.

[0191] After identifying and marking the speech activity time periods, the continuous speech data in each audio channel is segmented according to these marked time periods. Each segmented speech segment corresponds to an independent speech activity time period. The segmentation operation ensures that each segment is logically independent, meaning that each segment represents a complete speech activity time period. This independence makes subsequent processing and analysis more convenient and accurate.

[0192] The model can be implemented using a convolutional neural network (CNN), a recurrent neural network (RNN), or a Transformer-based architecture. To select a suitable speech recognition model, you can use an energy-based approach, a statistical model-based approach, or a deep learning model (such as a convolutional neural network CNN, a recurrent neural network RNN). Deep learning models are usually trained with a large amount of data and can more accurately identify speech activities. The main function of the model is to analyze the speech data in each audio channel and identify the time periods that contain speech activities. These time periods may include the moment when the speaker speaks, while excluding silence and noise interference.

[0193] Definition of meaningless speech segments: Meaningless speech segments include silent segments, background noise, and interjections that do not involve the actual content of the conversation (such as "um", "ah", etc.). These segments usually do not contain valuable information and may interfere with subsequent speech recognition and analysis.

[0194] Through preprocessing steps and speech recognition models, the system can automatically detect and exclude these meaningless segments. Silence segments and background noise can be identified and filtered out through energy detection or spectrum analysis, while modal particles can be excluded through semantic analysis or specific keyword filtering mechanisms.

[0195] Among them, one implementation method that can exclude meaningless human voices may be:

[0196] To filter meaningless human voice based on rules, first, a meaningless human voice vocabulary is defined, for example, a set of common meaningless words (such as "um", "ah", "oh", "um hum", etc.) is defined.

[0197] After the voice data is converted to text, it is filtered by matching these predefined nonsense words. These nonsense words can be removed from the text using regular expressions or string matching methods.

[0198] For example, suppose a conversation is transcribed to text and contains nonsense words:

[0199] Original text: "Um, I want to consult, um, I have a problem with my bill."

[0200] Filtered text: "I'd like to ask about my bill."

[0201] Another implementation could be:

[0202] Semantic analysis based on NLP models uses pre-trained language models (such as BERT and GPT) for semantic analysis. Pre-trained models can understand the context and identify meaningless content in the conversation. The pre-trained model performs semantic analysis on each sentence to identify meaningless content such as filler words and modal particles in the conversation. Filter out the identified meaningless content and only retain meaningful voice clips.

[0203] For example, use a pre-trained model to perform semantic analysis on text, identify and filter meaningless words:

[0204] Original text: "Ah, I want to know how much this service costs?"

[0205] Text recognized and filtered by the model: "I would like to know how much this service costs?"

[0206] Definition of valid speech segments: Valid speech segments refer to speech data segments that actually contain meaningful conversation content after being processed by the speech recognition model. These segments remove silence, noise, and irrelevant content, ensuring the high quality of speech data.

[0207] After identifying and eliminating meaningless segments, the remaining voice data are valid voice segments, which will be used for subsequent speech-to-text processing, text analysis, and fraud detection.

[0208] Through the above steps, this embodiment can effectively exclude meaningless human voices in the voice data and retain meaningful voice segments, thereby ensuring that subsequent voice-to-text conversion and fraud analysis have high accuracy and reliability.

[0209] Furthermore, a sixth embodiment of the voice fraud analysis method of the present invention is proposed. In the above step S30, the step of determining the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model includes:

[0210] Step e1, inputting the text to be analyzed into the pre-trained fraud analysis model to determine the number of fraud clauses in the text to be analyzed;

[0211] Step e2, determining the fraud probability of the voice data according to the number of the fraud clauses and a preset fraud strategy set.

[0212] In this embodiment, in order to determine the fraud probability of voice data, it is necessary to input the text to be analyzed into the analysis model, identify the number of fraud clauses, and determine the fraud probability according to the preset fraud strategy set. Specific methods may include:

[0213] Perform basic preprocessing on the text to be analyzed, such as removing stop words, punctuation marks, etc. Input the preprocessed text to be analyzed into a pre-trained analysis model. The model can be a language model based on deep learning (such as BERT, GPT), which has been fine-tuned and trained with a large amount of fraud-related data.

[0214] The analysis model performs semantic analysis on the input text and identifies fraudulent terms in the text. Fraudulent terms refer to key phrases or sentences in the text that are related to fraudulent behavior. The number of identified fraudulent terms is counted. Each identified fraudulent term represents a potential fraudulent behavior.

[0215] A fraud strategy set is preset, which contains the fraud probability range corresponding to different numbers of fraud clauses. These strategies can be set through statistical analysis of historical data and expert experience.

[0216] Based on the number of identified fraud clauses, the corresponding fraud probability range is found in the preset fraud strategy set. Probability calculation can be performed using a rule engine or a logical reasoning system.

[0217] For example, suppose that after the valid speech segments extracted from the speech data are converted into text, the text to be analyzed is determined as follows:

[0218] "Hi, my credit card was used unauthorizedly. Can you cancel this transaction for me?";

[0219] "Of course, we will process your request immediately. Please provide your card number and ID information.

[0220] Preprocess the text to be analyzed to remove irrelevant words and punctuation marks:

[0221] "Hi, my credit card has been stolen. Can you help me cancel this transaction?"

[0222] "Of course we will process your request immediately. Please provide your card number and ID information."

[0223] Feed the preprocessed text into the pre-trained fraud analysis model.

[0224] The analytical model identified the fraudulent terms in the text:

[0225] Fraud clause 1: "Credit card was stolen";

[0226] Fraudulent clause 2: "Cancel this transaction";

[0227] Fraud Clause 3: "Provide your card number and ID information."

[0228] Number of fraudulent clauses identified: 3.

[0229] Preset a fraud strategy set, for example:

[0230] 0-1 fraudulent clauses: low fraud probability (0-20%);

[0231] 2-3 fraudulent clauses: medium fraud probability (20-60%);

[0232] 4 or more fraudulent clauses: high probability of fraud (60-100%).

[0233] Based on the three identified fraud clauses, the fraud probability range corresponding to the preset fraud strategy set is found and determined to be a medium fraud probability (20-60%).

[0234] Through the above steps, this embodiment can effectively identify and count fraud clauses in the text to be analyzed, and determine the fraud probability of voice data according to a preset fraud strategy set, thereby ensuring the accuracy and reliability of fraud analysis.

[0235] Furthermore, a seventh embodiment of the voice fraud analysis method of the present invention is proposed, and before the above step e2, it also includes:

[0236] Step f1, obtaining historical fraud data, and defining the preset fraud strategy set according to business requirements and the historical fraud data.

[0237] In this embodiment, in order to define a preset fraud strategy set, it is necessary to obtain historical fraud data and set specific strategies based on business needs and these data analyses.

[0238] First, historical fraud data can be obtained. The data source can come from internal databases, such as internal corporate transaction records, customer service conversation records, customer complaint records, etc.; it can also come from industry reports, such as fraud analysis reports and research papers in the financial industry; or it can come from public cases, such as fraud cases and news reports obtained through the Internet and public resources.

[0239] Data preprocessing can include cleaning data, removing invalid records, duplicate data and missing values ​​to ensure data accuracy and completeness; format standardization, unifying the data formats from different sources to facilitate subsequent processing and analysis; data labeling, marking data to distinguish between normal transactions and fraudulent transactions. This can be done through manual labeling or automated classification tools.

[0240] Store the cleaned and standardized data in a database or data warehouse to facilitate subsequent analysis and model training.

[0241] Then, perform data analysis, the steps may include:

[0242] Fraud clause extraction uses natural language processing (NLP) technology to extract common fraud clauses from historical fraud data. For example, key phrases and sentences in the text are identified through methods such as keyword extraction and entity recognition.

[0243] Fraud pattern recognition: Use machine learning algorithms to identify fraud patterns in historical data. Classification models (such as decision trees, random forests, support vector machines, etc.) can be used to identify common features of fraudulent behavior. The classification model is used for recognition as follows: extract features from the original data, such as transaction amount, transaction time, transaction location, customer behavior, etc.; use labeled historical data to train the classification model to learn the pattern of fraudulent behavior; evaluate the model's accuracy, recall rate, F1 score, etc. to ensure the reliability of the model.

[0244] Perform statistical analysis on historical fraud data to calculate the frequency and combination of different fraud clauses. For example, count the number of times each fraud clause appears in all fraudulent transactions, and the probability of multiple fraud clauses appearing at the same time.

[0245] According to the data analysis results and business needs, set specific rules for the preset fraud strategy set. The strategy set should include the fraud probability range corresponding to different numbers of fraud clauses. Specific steps may include: Based on the statistical analysis results, set the risk weights of different fraud clauses. For example, clauses with high frequency of occurrence can be assigned higher risk weights. Set the fraud probability when multiple fraud clauses appear at the same time. For example, when certain clauses appear together, it may mean a higher risk of fraud.

[0246] In addition, the specific parameters of the policy set can be adjusted by combining identification analysis with business needs. For example, for high-risk transactions (such as international transactions and large transfers), the threshold for triggering fraud detection can be lowered; for low-risk transactions (such as local small-amount consumption), the threshold can be increased.

[0247] Finally, you can verify the strategy set and use some historical data to verify the set strategy set to ensure the rationality and effectiveness of the strategy set. You can verify it through methods such as cross-validation or A / B testing, and adjust the strategy set parameters based on the verification results.

[0248] One embodiment is as follows:

[0249] Collect credit card transaction records from the bank's internal database for the past five years, including both fraudulent and normal transactions. Obtain common fraud cases and patterns from research reports on the financial industry. Clean the data, remove invalid records and duplicate data, and ensure data accuracy. Unify the data formats from different sources for subsequent processing.

[0250] According to the bank's internal audit results and customer reporting records, fraud and normal labels are added to transaction data. Common fraud terms such as "credit card stolen" and "unauthorized transaction" are extracted from transaction notes and customer conversation records using NLP technology. Decision tree models are used to identify fraud patterns in historical data and extract features such as transaction amount, transaction location, time, device information, etc.

[0251] Train the model and evaluate its accuracy, recall, and F1 score to ensure that the model can accurately identify fraudulent behavior. Count the frequency of each fraud term in historical fraudulent transactions, as well as the combination of different terms. For example, count the probability of "unauthorized transaction" and "credit card fraud" appearing at the same time.

[0252] According to the statistical analysis results, set the preset fraud strategy set:

[0253] 0-1 fraudulent clauses: low fraud probability (0-20%);

[0254] 2-3 fraudulent clauses: medium fraud probability (20-60%);

[0255] 4 or more fraudulent clauses: high probability of fraud (60-100%).

[0256] For high-risk international transactions, set stricter fraud detection thresholds. For example:

[0257] 1 fraud clause: medium fraud probability (20-60%);

[0258] 2 or more fraudulent terms: high probability of fraud (60-100%).

[0259] The strategy set is validated using the past year’s trading data, and the strategy set parameters are adjusted through cross-validation methods to ensure the accuracy and reliability of the strategy set.

[0260] Through the above steps, this embodiment can define a reasonable preset fraud strategy set based on historical fraud data and business needs, and ensure the scientificity and accuracy of the strategy set based on historical data and statistical analysis, ensure that fraud can be accurately identified and prevented in actual applications, and improve business security.

[0261] The invention also provides a voice fraud analysis device.

[0262] Reference Figure 3 , Figure 3 This is a functional module diagram of the first embodiment of the voice fraud analysis device of the present invention. The voice fraud analysis device comprises:

[0263] A voice signal processing module, which obtains a voice signal, collects voice data from different sources from the voice signal through two independently configured microphone arrays, and records the voice data into two independent and isolated audio channels;

[0264] The effective segment extraction module detects the voice data in the two audio channels respectively by using the voice activity detection technology, identifies and marks the voice activity time period in each audio channel, and performs segmentation operation on the voice data based on the identified voice activity time period, so as to divide the continuous voice data in each audio channel into multiple independent voice segments, and obtain multiple effective voice segments;

[0265] A text conversion module generates a dialogue text according to the valid voice segment and determines a text to be analyzed in the dialogue text;

[0266] The fraud analysis module selects a pre-trained fraud analysis model according to the data features of the text to be analyzed, and determines the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model.

[0267] In addition, an embodiment of the present invention further provides a storage medium.

[0268] The storage medium stores a voice fraud analysis program, which implements the steps of the voice fraud analysis method described above when executed by the processor.

[0269] The specific implementation of the storage medium of the present invention is basically the same as the above-mentioned embodiments of the voice fraud analysis method, and will not be repeated here.

Claims

1. A voice fraud analysis method, characterized in that: The voice fraud analysis method comprises the following steps: Acquire a voice signal, collect voice data from different sources from the voice signal through two independently configured microphone arrays, and record the voice data into two independent and isolated audio channels respectively; The voice data in the two audio channels are detected by using voice activity detection technology, the voice activity time period in each audio channel is identified and marked, and the voice data is segmented based on the identified voice activity time period, so that the continuous voice data in each audio channel is segmented into multiple independent voice segments, and multiple valid voice segments are obtained; Generate a dialogue text according to the valid voice segment, and determine the text to be analyzed in the dialogue text; A pre-trained fraud analysis model is selected according to the data features of the text to be analyzed, and the fraud probability of the voice data is determined according to the text to be analyzed and the pre-trained fraud analysis model.

2. The voice fraud analysis method according to claim 1, characterized in that: The step of generating a dialogue text according to the valid voice segment comprises: Identify the start time and end time of the valid voice segment in the voice data, and identify the source person of the valid voice segment; Sort the valid voice segments according to the start time and the source person to obtain time-sequence voice segments; A dialogue text is generated according to the time-sequence speech segments.

3. The voice fraud analysis method as claimed in claim 2, characterized in that: The step of generating a dialogue text according to the time-sequence speech segments comprises: The time-sequenced voice segments are converted into voice text using voice recognition technology, and the voice text is marked according to the source person to generate the dialogue text.

4. The voice fraud analysis method according to claim 1, characterized in that: The step of acquiring the voice signal comprises: Selecting a device that supports dual-channel recording, including a recording device with independent left-channel and right-channel recording functions; Setting the recording parameters of the recording device, including sampling rate and quantization accuracy; The left channel and the right channel of the recording device are used to record audio signals from two different directions or sources respectively to obtain a two-channel audio signal; The two-channel audio signal is preprocessed, and the speech signal belonging to human voice is identified in the preprocessed two-channel audio signal.

5. The voice fraud analysis method according to claim 1, characterized in that: The steps of identifying and marking the voice activity time period in each audio channel, performing a segmentation operation on the voice data based on the identified voice activity time period, and dividing the continuous voice data in each audio channel into multiple independent voice segments, and obtaining multiple valid voice segments include: Constructing a speech recognition model, detecting speech activity indicators of the speech data through the speech recognition model, identifying and marking time periods containing speech activities, and generating start and end time markers for each speech segment; Segmentation is performed based on the marked voice activity time period, and the continuous voice data in each audio channel is segmented according to the marked time period. Each segmented voice segment corresponds to an independent voice activity time period. Exclude meaningless speech segments during segmentation, including silent segments, background noise, and irrelevant modal particles; The valid voice segment is obtained, where the valid voice segment is a voice data segment that has been filtered and contains actual conversation content.

6. The voice fraud analysis method according to claim 1, characterized in that: The step of determining the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model comprises: Inputting the text to be analyzed into the pre-trained fraud analysis model to determine the number of fraud clauses in the text to be analyzed; The fraud probability of the voice data is determined according to the number of the fraud clauses and a preset fraud strategy set.

7. The voice fraud analysis method according to claim 6, characterized in that: Before the step of determining the fraud probability of the voice data according to the number of the fraud clauses and the preset fraud strategy set, the method further includes: Acquire historical fraud data, and define the preset fraud strategy set according to business requirements and the historical fraud data.

8. A voice fraud analysis device, characterized in that: The voice fraud analysis device comprises: A voice signal processing module, which obtains a voice signal, collects voice data from different sources from the voice signal through two independently configured microphone arrays, and records the voice data into two independent and isolated audio channels; The effective segment extraction module detects the voice data in the two audio channels respectively by using the voice activity detection technology, identifies and marks the voice activity time period in each audio channel, and performs segmentation operation on the voice data based on the identified voice activity time period, so as to divide the continuous voice data in each audio channel into multiple independent voice segments, and obtain multiple effective voice segments; A text conversion module generates a dialogue text according to the valid voice segment and determines a text to be analyzed in the dialogue text; The fraud analysis module selects a pre-trained fraud analysis model according to the data features of the text to be analyzed, and determines the fraud probability of the voice data according to the text to be analyzed and the pre-trained fraud analysis model.

9. A voice fraud analysis device, characterized in that: The voice fraud analysis device comprises a memory, a processor, and a voice fraud analysis program stored in the memory and executable on the processor. When the voice fraud analysis program is executed by the processor, the steps of the voice fraud analysis method according to any one of claims 1 to 7 are implemented.

10. A storage medium, characterized in that: The storage medium stores a voice fraud analysis program, which, when executed by a processor, implements the steps of the voice fraud analysis method according to any one of claims 1 to 7.

Images