Processing method and device for fraud-related call, electronic equipment and readable medium

Abstract

The invention relates to a processing method and device for a fraud-related call, electronic equipment and a readable medium, and the method comprises the steps: collecting an incoming call audio under the condition that a target object receives the fraud-related call, and dividing the incoming call audio into a plurality of audio segments; generating a to-be-arranged text of the current round according to the audio segment; a scene exclusive prompt is injected into the to-be-arranged text, a prompt word is obtained, and the scene exclusive prompt comprises a human-set template of a fraud-related scene, a rigid security boundary rule and a response strategy; generating a target reply text corresponding to the cue word according to a preset dialogue generation model; and generating a reply voice matched with the target object according to the target reply text, and returning the reply voice to the calling party of the fraud-related call. The problem of how to broadcast the safe and controllable voice conforming to the characteristics of the victim in the fraud-related call is solved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method, apparatus, electronic device, and readable medium for processing fraudulent calls. Background Technology

[0002] In today's world of rapid information technology development and people's constant use of mobile phones, more and more fraudsters are using mobile phones, telephones, the internet and other communication tools to carry out contactless telecommunications fraud. Telephone fraud is a major form of telecommunications and internet fraud.

[0003] Current fraudulent calls are characterized by highly manipulative scripts, multiple rounds of entanglement, and frequent requests for sensitive information. Existing intelligent call-calling systems only support custom basic scripts and general interactions, and have not built a response mechanism applicable to fraudulent scenarios. As a result, when dealing with the inducements of fraudsters, they cannot maintain a stable response logic that matches the characteristics of the victim in real-time interactions, and it is difficult to effectively avoid the risk of sensitive information leakage.

[0004] Therefore, how to broadcast safe and controllable voice messages that match the characteristics of the victim during fraudulent calls has become an urgent problem to be solved.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a method, apparatus, electronic device, and readable medium for processing fraudulent calls, in order to solve the aforementioned technical problem of "how to broadcast secure and controllable voice that conforms to the characteristics of the victim in fraudulent calls".

[0007] According to one aspect of the embodiments of this application, this application provides a method for processing fraudulent calls, including: when a target receives a fraudulent call, collecting the incoming call audio and dividing the incoming call audio into multiple audio segments; generating a text to be arranged for the current round based on the audio segments; injecting scene-specific prompts into the text to be arranged to obtain prompt words, wherein the scene-specific prompts include a persona template for the fraudulent scenario, rigid security boundary rules, and response strategies; generating a target reply text corresponding to the prompt words based on a preset dialogue generation model; generating a reply voice matching the target target based on the target reply text, and sending the reply voice back to the caller in the fraudulent call.

[0008] Optionally, generating the text to be arranged for the current round based on the audio segment includes: converting the audio segment into a text segment in the target format through speech recognition; validating the text segment; and if the text segment passes the validity check, concatenating the text segment with the historical conversation text to obtain the text to be arranged for the current round.

[0009] Optionally, after validating the validity of the text segment, the method further includes: if the text segment fails the validity check, obtaining a secure transition text; generating a secure transition voice matching the target object based on the secure transition text; and sending the secure transition voice back to the caller in the fraudulent call.

[0010] Optionally, scenario-specific prompts are injected into the text to be arranged to obtain prompt words, including: determining the current conversation stage based on the historical conversation text in the text to be arranged; calling the fraud scenario strategy library and extracting the response strategy from the fraud scenario strategy library according to the current conversation stage; retrieving the persona template and rigid security boundary rules, and determining the response strategy, persona template, and rigid security boundary rules as scenario-specific prompts; and injecting the scenario-specific prompts into the text to be arranged to obtain prompt words.

[0011] Optionally, generating a target response text corresponding to the prompt word according to a preset dialogue generation model includes: inputting the prompt word into the preset dialogue generation model and obtaining candidate response texts output by the preset dialogue generation model; performing sensitivity rule verification on the candidate response texts to obtain a risk level, wherein the risk level includes low risk, medium risk, and high risk; if the risk level is medium risk, optimizing the candidate response texts to obtain the target response text; if the risk level is low risk, determining the candidate response text as the target response text; and if the risk level is high risk, ending the current round of the call.

[0012] Optionally, generating a response speech matching the target object based on the target response text includes: obtaining a reference timbre resource package of the target object, wherein the reference timbre resource includes reference timbre audio and reference text corresponding to the reference timbre audio, and the reference text is in the target format; inputting the target response text, the reference timbre audio, and the reference text into a preset synthesis model to obtain the response speech output by the synthesis model, wherein the preset synthesis model is used to obtain speech prosody based on the semantic association between the target response text and the reference text, and to extract speech timbre based on the reference timbre audio, so as to generate response speech based on speech prosody and speech timbre.

[0013] Optionally, the method further includes structurally reconstructing the multi-turn dialogue of the case in the following manner: obtaining case descriptions of multiple cases, wherein the case descriptions are obtained from the call logs of the fraudulent calls; traversing each case description; during the traversal, generating input instructions for the current case, wherein the input instructions include multiple scenario constraints; inputting the input instructions into a preset dialogue generation model to obtain candidate multi-turn dialogue texts output by the preset dialogue generation model; sequentially parsing and processing the candidate multi-turn dialogue texts to obtain the target multi-turn dialogue text; and storing the target multi-turn dialogue text of the current case to obtain structured dialogue data.

[0014] According to another aspect of the embodiments of this application, this application provides a processing device for fraudulent calls, including: a collection module, used to collect incoming call audio when a target object receives a fraudulent call, and divide the incoming call audio into multiple audio segments; a first generation module, used to generate text to be arranged for the current round based on the audio segments; an injection module, used to inject scene-specific prompts into the text to be arranged to obtain prompt words, wherein the scene-specific prompts include a persona template for the fraudulent scene, rigid security boundary rules, and response strategies; a second generation module, used to generate target reply text corresponding to the prompt words according to a preset dialogue generation model; and a third generation module, used to generate reply voice matching the target object based on the target reply text, and send the reply voice back to the caller in the fraudulent call.

[0015] According to another aspect of the embodiments of this application, this application provides an electronic device, including a memory, a processor, a communication interface and a communication bus. The memory stores a computer program that can run on the processor. The memory and the processor communicate with each other through the communication bus and the communication interface. When the processor executes the computer program, it implements the steps of the above method.

[0016] According to another aspect of the embodiments of this application, this application also provides a computer-readable medium having processor-executable non-volatile program code that causes the processor to perform the above-described method.

[0017] Compared with related technologies, the technical solutions provided in this application have the following advantages: This application provides a method for processing fraudulent calls, including: when a target receives a fraudulent call, collecting the incoming call audio and dividing it into multiple audio segments; generating a text to be arranged for the current round based on the audio segments; injecting scenario-specific prompts into the text to be arranged to obtain prompt words, wherein the scenario-specific prompts include a persona template for the fraudulent scenario, rigid security boundary rules, and response strategies; generating a target response text corresponding to the prompt words based on a preset dialogue generation model; generating a response voice matching the target target based on the target response text, and sending the response voice back to the caller in the fraudulent call. By segmenting the incoming call audio and generating a text to be arranged, injecting prompt words containing a victim persona, security boundaries, and response strategies, then generating a target response text using a preset model, and finally generating a voice response matching the victim based on the target response text, a safe and controllable voice broadcast that conforms to the victim's characteristics is achieved, solving the problem of how to broadcast safe and controllable voice that conforms to the victim's characteristics in fraudulent calls. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the hardware environment for an optional method for processing fraudulent calls according to an embodiment of this application; Figure 2 A flowchart illustrating an optional method for processing fraudulent calls according to an embodiment of this application; Figure 3 This is a schematic diagram of an optional processing system for fraudulent calls provided according to an embodiment of this application; Figure 4 A timing diagram of an optional real-world call forwarding method for fraudulent calls provided according to an embodiment of this application; Figure 5 This is a flowchart illustrating an optional real-time single-turn interaction for fraudulent call handling, according to an embodiment of this application. Figure 6 This is a block diagram of an optional processing apparatus for fraudulent calls according to an embodiment of this application; Figure 7 This is a schematic diagram of an optional electronic device structure provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustration and has no specific meaning in itself. Therefore, "module" and "part" may be used interchangeably.

[0023] In today's world of rapid information technology development and people's constant use of mobile phones, more and more fraudsters are using mobile phones, telephones, the internet and other communication tools to carry out contactless telecommunications fraud. Telephone fraud is a major form of telecommunications and internet fraud.

[0024] Current fraudulent calls are characterized by highly manipulative scripts, multiple rounds of entanglement, and frequent requests for sensitive information. Existing intelligent call-calling systems only support custom basic scripts and general interactions, and have not built a response mechanism applicable to fraudulent scenarios. As a result, when dealing with the inducements of fraudsters, they cannot maintain a stable response logic that matches the characteristics of the victim in real-time interactions, and it is difficult to effectively avoid the risk of sensitive information leakage.

[0025] Therefore, how to broadcast safe and controllable voice messages that match the characteristics of the victim during fraudulent calls has become an urgent problem to be solved.

[0026] In order to address the problems mentioned in the background art, according to one aspect of the embodiments of this application, an embodiment of a method for processing fraudulent calls is provided.

[0027] Optionally, in the embodiments of this application, the above-described method for processing fraudulent calls can be applied to, for example... Figure 1 The hardware environment shown consists of terminal 101 and server 103. Figure 1 As shown, server 103 is connected to terminal 101 via a network and can be used to provide services to the terminal or clients installed on the terminal. Database 105 can be set up on the server or independently of the server to provide data storage services for server 103. The network mentioned above includes, but is not limited to, wide area network, metropolitan area network or local area network. Terminal 101 includes, but is not limited to, PC, mobile phone, tablet computer, etc.

[0028] A method for handling fraudulent calls in this application embodiment can be executed by server 103, or it can be jointly executed by server 103 and terminal 101, such as... Figure 2 As shown, it includes: Step 201: When the target receives a fraudulent call, collect the incoming call audio and divide the audio into multiple audio segments; Step 202: Generate the text to be arranged for the current round based on the audio segment; Step 203: Inject scenario-specific prompts into the text to be arranged to obtain prompt words. Scenario-specific prompts include persona templates for fraud scenarios, rigid security boundary rules, and response strategies. Step 204: Generate the target response text corresponding to the prompt words based on the preset dialogue generation model; Step 205: Generate a response voice matching the target object based on the target response text, and send the response voice back to the caller in the fraudulent call.

[0029] This application is applicable to scenarios involving the interception of telecommunications network fraud. When a target (victim) receives a fraudulent call, the system automatically answers the call on their behalf, simulating the victim's response to achieve a safe deception. This avoids direct exposure to fraudulent tactics by the user and effectively prevents the leakage of sensitive information.

[0030] Figure 3 The schematic diagram of the fraud-related call processing system provided in this application is shown in the figure. The system consists of the following functional units connected in series in sequence (adjacent units are connected by data flow and control flow): (1) Voice access and conversation orchestration unit: interface with telephone gateway or file input, and complete media stream decoding, voice activity detection (or fixed time window) segmentation, round boundary determination, abnormal / silent processing and timing coordination; (2) ASR (Automatic Speech Recognition) unit: provides file-based or streaming recognition interface, transcribes audio segments into text, and returns JSON in the form of {"text":"…"}; (3) Prompt word and strategy arrangement unit: splice ASR text with conversation history, inject the persona and security boundaries of the present invention into the system prompts, and select the speech strategy according to the stage; (4) LLM (Large Language Model) dialogue generation unit: receives parameters and returns the victim's spoken response text; (5) Text review and security inspection unit: Remove explanatory prefixes and suffixes, template headers and irrelevant tags, retain the broadcastable text, and perform sensitive action / sensitive field rule verification (rewrite or reject when triggered); (6) TTS (Text-to-Speech) synthesis unit: using reference audio timbre and its transcribed text as joint cues to perform zero-sample synthesis, outputting speech segments that can be directly played back; (7) Playback and timing control unit: Send the synthesized speech back to the other end with low latency, and insert appropriate transitional phrases to smooth the waiting when the next round of processing is not ready; (8) Audit and Log Unit: Records the time spent, generated text, sensitive triggers, fault-tolerant branches and round counts for each round of ASR / LLM / TTS; (9) Case reconstruction unit: construct strong constraint prompts for each case in the input case JSON list and call LLM to output a multi-turn dialogue with a fixed JSON structure.

[0031] This application first collects and segments audio of fraudulent calls, then converts the audio to text and splices it with the conversation history to obtain the text to be arranged. Next, it injects exclusive prompts containing victim profiles, security boundaries, and response strategies into the text, generates safe and compliant reply text using a preset model, and finally synthesizes the text into a voice matching the user's voice and sends it back to the caller, thus achieving safe call forwarding and handling of fraudulent calls.

[0032] As an optional embodiment, generating the text to be arranged for the current round based on the audio segment includes: converting the audio segment into a text segment in the target format through speech recognition; validating the text segment; and if the text segment passes the validity check, concatenating the text segment with the historical conversation text to obtain the text to be arranged for the current round.

[0033] The ASR unit is invoked. The ASR unit converts the received audio segment into text information through a file-based or streaming recognition interface, generating a text segment in the target format. The text segment format is consistent with the output standard of the ASR unit. Specifically, the target format is a JSON format containing a "text" field (storing the recognized text content) and a "segment_id" field (a unique identifier associated with the corresponding audio segment), for example: {"text":"What is the matter with the account freeze you mentioned?", "segment_id":"call_001_segment_02"}.

[0034] By converting ASR and standardizing the format, we can ensure that the text segment can be directly adapted to the processing requirements of the special prompts for subsequent injection scenarios, thus avoiding process delays caused by format incompatibility.

[0035] The text segment in the target format output by the ASR unit is validated. The validation includes whether the "text" field is empty (i.e. the recognition result has no content) and whether the confidence of the recognized text segment is lower than the preset threshold.

[0036] The preset threshold can be set based on the accuracy requirements of fraud-related dialogue recognition, for example, the preset threshold can be set to 0.6.

[0037] If the "text" field is not empty and the confidence level meets the standard, the text segment is considered to have passed the validity check; otherwise, it is considered to have failed the validity check.

[0038] Text with low confidence due to noise or dialect ambiguity is considered invalid. By filtering out invalid text through verification, the reliability of the text to be edited can be ensured.

[0039] If the text segment passes the validity check, the historical conversation text is retrieved, and the content of the "text" field of the current text segment is concatenated with the historical conversation text in chronological order to obtain the text to be arranged for the current round.

[0040] The historical conversation text consists of text segments that have passed validation in previous rounds and their corresponding response texts, which together form the conversation context.

[0041] By piecing together historical conversation texts to form a complete context, the text to be arranged can reflect the logic of multiple rounds of dialogue, laying the foundation for generating responses that fit the conversation process and conform to the victim's response logic, and further ensuring the authenticity and rationality of the broadcast voice.

[0042] As an optional embodiment, after validating the validity of the text segment, the method further includes: if the text segment fails the validity check, obtaining a secure transition text; generating a secure transition voice matching the target object based on the secure transition text; and sending the secure transition voice back to the caller in the fraudulent call.

[0043] If the text segment fails validation, the secure transition text acquisition process is triggered. The validation results quickly identify invalid text scenarios, preventing subsequent conversations from being interrupted due to invalid text and ensuring the continuity of fraudulent call interactions.

[0044] The system directly retrieves a pre-set secure transition text library and selects text that matches the normal response logic of a victim in a fraud scenario. For example, "I don't quite understand what you mean, could you please repeat it?" or "I didn't hear you clearly, could you please explain in more detail?"

[0045] Based on the typical responses of victims to ambiguous information in fraud scenarios, pre-set texts are designed to ensure that transition texts fit the persona while avoiding the introduction of sensitive content, thus ensuring the security of responses from the source.

[0046] The acquired safe transition text is input into the TTS synthesis unit. The TTS synthesis unit uses the reference timbre audio and its transcribed text as joint cues to generate a safe transition speech that matches the timbre and tone of the target object in zero-sample synthesis mode.

[0047] Zero-shot synthesis is an efficient synthesis paradigm in the field of text-to-speech. It refers to the ability of a model to generate speech that is highly similar to the timbre and intonation of the target speaker, without the need to obtain specialized training data of the target speaker (such as without recording a large amount of the target speaker's speech for model fine-tuning). It only requires a small amount (usually a few seconds to tens of seconds) of the target speaker's reference speech (or combined with its transcribed text) as a prompt. The core is that the model captures key acoustic features in the reference speech through generalization ability, thereby achieving auditory learning.

[0048] By leveraging the zero-sample timbre alignment capability of the TTS unit, the transitional speech is made to match the victim's characteristics, avoiding arousing suspicion from fraudsters due to abrupt changes in speech style, while ensuring that the generated speech conforms to the system's unified timbre standard.

[0049] Figure 4The timing diagram of the real-world call handling method for fraudulent calls provided in this application is shown in the figure. The main process is as follows: an audio segment is generated based on the incoming call, which is transcribed into text by ASR (Automatic Speech Recognition); the text is passed to the prompt word unit to form a message, which is then input into LLM (Large Language Model) to generate candidate responses r_raw; the review and verification unit purifies r_raw to obtain playable text r, which is then synthesized into a speech segment by TTS (Text-to-Speech), and finally played back to the caller. At the same time, data from each stage is synchronously transmitted back to the corresponding unit. If a recognition error occurs in the main process, a transition phrase (such as "Please say it again") is triggered and directly synthesized by TTS before playback, while re-recognition is initiated to ensure call continuity.

[0050] The playback and timing control unit receives the secure transition voice and transmits it back to the caller (fraudster) in the fraudulent call with low latency through the telephone gateway. By utilizing the low-latency transmission capability of the playback unit, the time from the generation of the transition voice to its transmission is shortened, maintaining the natural rhythm of the call interaction, while also giving the system time to reprocess the next round of incoming call audio.

[0051] As an optional implementation, scenario-specific prompts are injected into the text to be arranged to obtain prompt words, including: determining the current conversation stage based on the historical conversation text in the text to be arranged; calling the fraud scenario strategy library and extracting the response strategy from the fraud scenario strategy library according to the current conversation stage; retrieving the persona template and rigid security boundary rules, and determining the response strategy, persona template, and rigid security boundary rules as scenario-specific prompts; and injecting the scenario-specific prompts into the text to be arranged to obtain prompt words.

[0052] The text to be arranged includes the text segment converted from the audio segment of the current round and the historical conversation text. The caller's (fraudster's) words in the historical conversation text are analyzed by the prompt words and strategy arrangement unit to determine the current conversation stage. The conversation stage includes, but is not limited to, the identity probing stage, the question confirmation stage, the pressure and urgency stage, and the guidance operation stage.

[0053] For example, if the caller only mentions "identity verification" (such as "I am a staff member of the XX public security, procuratorate and court") in the historical conversation text, it is judged as the identity probing stage; if it mentions issues such as "account freezing" or "affecting credit score", it is judged as the issue confirmation stage; if it mentions content such as "limited time operation" or "failure to cooperate will result in consequences", it is judged as the pressure and urgency stage; if it mentions content such as "transfer steps" or "provide verification code", it is judged as the guidance operation stage.

[0054] Based on the typical evolution of rhetoric in fraudulent calls—from identity spoofing to pressure and then to inducement—the current stage of the interaction is located through historical dialogue context, providing a basis for matching response strategies.

[0055] The system uses prompts and strategy orchestration units to access a pre-defined strategy library for fraud-related scenarios. This library pre-stores response patterns that match each stage of the conversation. For example, for the identity probing stage, a response strategy of "cautiously inquiring about details" is extracted, such as "What is your job number?"; for the problem confirmation stage, a response strategy of "expressing doubt but not agreement" is extracted, such as "I haven't done anything wrong, why is my account frozen?"; for the pressure and urgency stage, a response strategy of "showing hesitation" is extracted, such as "Is it so urgent? I'm a little afraid of making a mistake"; and for the guided operation stage, a response strategy of "clearly refusing sensitive actions" is extracted, such as "Transferring money involves money, I dare not operate casually." Through pre-matching of stages and strategies, the system ensures that the response content aligns with the interaction needs of different stages in a fraud-related scenario, avoiding logical disconnects caused by generic responses.

[0056] The prompt and strategy arrangement unit retrieves the system's preset "persona template" and "rigid security boundary rules." The persona template is "cautiously cooperating, slightly hesitant but not cooperating with key steps" (which matches the typical behavioral characteristics of victims in fraud scenarios). The rigid security boundary rules are "never disclosing sensitive information such as bank card numbers, ID numbers, and verification codes, and never performing dangerous operations such as transferring money or downloading specified software." The extracted response strategies, persona templates, and rigid security boundary rules are integrated to form scenario-specific prompts.

[0057] By combining multiple dimensions of rules, the generated response content is constrained from three aspects: role positioning (persona), safety bottom line (boundary rules), and interaction logic (response strategy), to ensure that the response meets the requirements for handling fraud-related scenarios.

[0058] The prompt word and strategy orchestration unit injects scenario-specific prompts into the text to be orchestrated in a fixed format. Using prefixes such as "System prompt: [Character template content]"; "Security rule: [Rigid security boundary rule content]"; and "Current stage response requirement: [Response strategy content]", it combines historical conversation text and the current round text segment from the reception orchestration text to form prompt words for use by the large-scale language model.

[0059] By using structured injection, scenario-specific rules are transformed into an input format that LLM can understand, ensuring that the response text generated by LLM simultaneously meets the requirements of persona, security, and phased strategy.

[0060] As an optional embodiment, generating target response text corresponding to the prompt words according to a preset dialogue generation model includes: inputting the prompt words into the preset dialogue generation model and obtaining candidate response texts output by the preset dialogue generation model; performing sensitivity rule verification on the candidate response texts to obtain risk levels, wherein the risk levels include low risk, medium risk, and high risk; if the risk level is medium risk, optimizing the candidate response texts to obtain the target response text; if the risk level is low risk, determining the candidate response text as the target response text; and if the risk level is high risk, ending the current round of the call.

[0061] The prompts are input into a preset dialogue generation model, namely LLM. Based on the persona constraints, safety boundary rules, and conversation context in the prompts, LLM generates candidate response texts that conform to the victim's spoken logic.

[0062] The sensitive rule verification function of the text review and security inspection unit is invoked to detect candidate response texts. The specific detection is based on the preset sensitive action library and sensitive field library to determine whether the text has security risks.

[0063] For example, the sensitive action library includes keywords such as "transferring money" and "providing verification codes," while the sensitive field library includes format rules such as bank card numbers and ID card numbers.

[0064] Low-risk text contains no sensitive content and aligns with the target persona and strategy, such as text asking for a job number. Medium-risk text carries a slight risk but does not cross security boundaries, such as "I think I remember the last four digits of my bank card number, should I give them to you?". High-risk text directly violates rigid security boundaries, such as "My bank card number is XXX, could you please check it for me?". By using preset sensitivity rules to stratify and filter candidate texts, different levels of security risk can be accurately identified.

[0065] If the risk level is low, the candidate response text is directly determined as the target response text without additional processing, ensuring efficient response generation. If the risk level is medium, the text review and security inspection unit optimizes the candidate response text according to a preset strategy. For example, "I think I remember the last four digits of my bank card number, should I tell you?" is rewritten as "I don't quite remember my bank card number, and I don't dare to just give it to you." The optimized text is the target response text. If the risk level is high, a security termination mechanism is triggered, directly ending the current round of the call to prevent the leakage of sensitive information.

[0066] Differentiated approaches are adopted for different risk levels. While ensuring safety, responses that meet the needs of the scenario are retained to the greatest extent possible, and losses are stopped in a timely manner when there is high risk.

[0067] As an optional embodiment, generating a response speech matching the target object based on the target response text includes: obtaining a reference timbre resource package of the target object, wherein the reference timbre resource includes reference timbre audio and reference text corresponding to the reference timbre audio, and the reference text is in the target format; inputting the target response text, the reference timbre audio, and the reference text into a preset synthesis model, and obtaining the response speech output by the synthesis model, wherein the preset synthesis model is used to obtain speech prosody based on the semantic association between the target response text and the reference text, and to extract speech timbre based on the reference timbre audio, so as to generate the response speech based on the speech prosody and speech timbre.

[0068] The system retrieves a reference timbre resource package corresponding to the target object from its pre-set resource library. This package includes reference timbre audio and reference text. The reference timbre audio is an audio segment recording the unique timbre, intonation, and tone of the target object's acoustic characteristics. The reference text is text generated by an automatic speech recognition unit, transcribed from the reference timbre audio, and its format matches the target response text format. By pre-storing the target object's timbre and corresponding text resources, a dual-source reference can be provided for subsequently generating speech that matches its characteristics, avoiding significant differences between the synthesized speech and the target object's timbre.

[0069] The target response text, reference audio timbre, and reference text are input together into a preset text-to-speech synthesis model.

[0070] The pre-defined synthesis model first analyzes the semantic relationship between the target response text and the reference text. By matching the sentence structure and semantic logic of both, it determines the prosodic features of the response speech, including pauses, speech rate, and intonation variations. Simultaneously, the pre-defined synthesis model extracts the timbre features of the target object from the reference audio, such as pitch, timbre distinctiveness, and pronunciation habits. Finally, the pre-defined synthesis model fuses the prosodic features with the timbre to generate a response speech that highly matches the timbre and tone of the target object and conforms to the semantics of the target response text.

[0071] This application leverages the complementary effect of dual-source input (text and audio) to ensure that the speech prosody matches the semantics through reference text, and to ensure that the timbre matches the target object through reference audio timbre.

[0072] Figure 5The flowchart of the single-round interaction for the fraudulent call provided in this application is shown in the figure. Each round starts with "Start / Next Round". After the call is connected, ASR (Automatic Speech Recognition) is performed first. If ASR recognition fails, the branch process of "playing transition words and re-recognizing" is triggered. If recognition is successful, the prompt word construction stage is entered. After the prompt word is input into LLM (Large Language Model) to generate a response, a security review is performed. If the review triggers sensitive words, the "reject or rewrite" operation is performed and the response is regenerated. If no sensitive words are triggered, the audio is synthesized by TTS (Text-to-Speech) and played back, and then the next round of interaction begins.

[0073] As an optional embodiment, the method further includes structurally reconstructing the multi-turn dialogue of a case in the following manner: obtaining case descriptions of multiple cases, wherein the case descriptions are obtained from call logs of fraudulent calls; traversing each case description; during the traversal, generating input instructions for the current case, wherein the input instructions include multiple scenario constraints; inputting the input instructions into a preset dialogue generation model to obtain candidate multi-turn dialogue texts output by the preset dialogue generation model; sequentially parsing and processing the candidate multi-turn dialogue texts to obtain target multi-turn dialogue texts; and storing the target multi-turn dialogue texts of the current case to obtain structured dialogue data.

[0074] From the audit data of fraudulent calls stored in the system, we extracted case descriptions of multiple processed cases. These case descriptions were all generated based on the call logs of the fraudulent calls. The call logs contain data such as ASR transcription text, sensitive event markers, and core call content recorded during the real-time call. The case descriptions are a concise summary of the fraud scenario in a single case. For example, a user receives a call impersonating a bank customer service representative and is tricked into providing a bank card verification code under the pretext of "account abnormality requiring verification".

[0075] Case materials are generated based on real call data from real-time call sessions to ensure that the final reconstructed dialogue closely matches actual fraud scenarios and avoids distortion of training data caused by fictional scenarios.

[0076] The case reconstruction unit initiates the traversal process, reading each case description in a preset order (such as case timestamp, case number). The currently read case description is treated as a "pending object," pausing the processing of subsequent cases until the entire dialogue reconstruction process of the current case is completed before proceeding to the processing of the next case description.

[0077] Based on the description of the current pending case, the case reconstruction unit constructs exclusive input instructions, which explicitly include multiple constraints related to the fraud scenario: First, role constraints (the two parties in the dialogue are "fraudster" and "victim"); second, style constraints (the overall language should be colloquial and natural, reflecting a realistic conversational tone); third, content constraints (it must cover the core fraud terms in the case description, such as "account anomaly" and "verification code," and the victim's response must not contain any statements related to reporting to the police); fourth, format constraints (the output is a fixed JSON structure, with "dialogue" as the key, and array elements alternately labeled with the roles "Left (fraudster)" and "Right (victim)"); and fifth, round constraints (multiple rounds of dialogue, no less than 10 rounds, to ensure the integrity of the dialogue.

[0078] By using multi-dimensional constraints, the case requirements are transformed into rules that can be recognized by the pre-defined dialogue generation model, providing a clear basis for generating structured dialogues that meet the requirements.

[0079] Input instructions containing scenario constraints are fed into a preset dialogue generation model (LLM). Based on the constraints in the instructions, the LLM simulates the inducement logic of fraudsters and the response logic of victims in fraud scenarios, generating candidate multi-turn dialogue text containing complete multi-turn interactions, and the text format initially conforms to the preset JSON requirements.

[0080] The case reconstruction unit performs JSON parsing on the candidate multi-turn dialogue text to verify whether the text conforms to the preset format requirements. If the parsing fails (e.g., missing "dialogue" key, incorrect role labeling), feedback is sent to the LLM with a format correction prompt, triggering the model to regenerate; if the parsing succeeds, regular expression matching is used to remove redundant content such as explanatory comments and Markdown tags from the text, retaining purely colloquial dialogue sentences, and finally obtaining the target multi-turn dialogue text with a standardized format and clean content.

[0081] Through dual processing of parsing and verification and text cleansing, the output dialogue text is ensured to meet both structured storage requirements and can be directly used for training or drill scenarios.

[0082] The case reconstruction unit stores the target multi-turn dialogue text of the current case, using "case number-case keywords" as identifiers (e.g., "case_001-impersonating bank verification code"), associating them with the text content. After processing all cases, it integrates this data to form a total structured dialogue data containing structured dialogues from multiple cases, and saves it to a designated database according to the naming rule "original log file name-dialogue reconstruction.json". Through standardized storage and integration, it forms a structured data resource that can be accessed in batches, meeting the data source requirements of anti-fraud training, drills, and quality inspection systems.

[0083] This application provides a method for processing fraudulent calls, including: when a target receives a fraudulent call, collecting the incoming call audio and dividing it into multiple audio segments; generating a text to be arranged for the current round based on the audio segments; injecting scenario-specific prompts into the text to be arranged to obtain prompt words, wherein the scenario-specific prompts include a persona template for the fraudulent scenario, rigid security boundary rules, and response strategies; generating a target response text corresponding to the prompt words based on a preset dialogue generation model; generating a response voice matching the target target based on the target response text, and sending the response voice back to the caller in the fraudulent call. By segmenting the incoming call audio and generating a text to be arranged, injecting prompt words containing a victim persona, security boundaries, and response strategies, then generating a target response text using a preset model, and finally generating a voice response matching the victim based on the target response text, a safe and controllable voice broadcast that conforms to the victim's characteristics is achieved, solving the problem of how to broadcast safe and controllable voice that conforms to the victim's characteristics in fraudulent calls.

[0084] According to another aspect of the embodiments of this application, this application provides a processing apparatus for fraudulent calls, such as... Figure 6 As shown, it includes: The acquisition module 601 is used to acquire the audio of the incoming call when the target receives a fraudulent call, and to divide the audio into multiple audio segments. The first generation module 602 is used to generate the text to be arranged for the current round based on the audio segment; The injection module 603 is used to inject scenario-specific prompts into the text to be edited to obtain prompt words. The scenario-specific prompts include persona templates for fraud scenarios, rigid security boundary rules, and response strategies. The second generation module 604 is used to generate target response text corresponding to prompt words according to a preset dialogue generation model; The third generation module 605 is used to generate a reply voice that matches the target object based on the target reply text, and send the reply voice back to the caller in the fraudulent call.

[0085] It should be noted that the acquisition module 601 in this embodiment can be used to execute step 201 in this application embodiment, the first generation module 602 in this embodiment can be used to execute step 202 in this application embodiment, the injection module 603 in this embodiment can be used to execute step 203 in this application embodiment, the second generation module 604 in this embodiment can be used to execute step 204 in this application embodiment, and the third generation module 605 in this embodiment can be used to execute step 205 in this application embodiment.

[0086] Optionally, the first generation module 602 is further configured to convert the audio segment into a text segment in the target format through speech recognition; to perform validity verification on the text segment; and, if the text segment passes the validity verification, to concatenate the text segment with the historical conversation text to obtain the text to be arranged for the current round.

[0087] Optionally, the first generation module 602 is further configured to, after validating the validity of the text segment, obtain a secure transition text if the text segment fails the validity check; generate a secure transition voice matching the target object based on the secure transition text; and send the secure transition voice back to the caller in the fraudulent call.

[0088] Optionally, the injection module 603 is also used to determine the current session stage based on the historical session text in the text to be arranged; call the fraud scenario strategy library and extract the response strategy from the fraud scenario strategy library according to the current session stage; retrieve the persona template and rigid security boundary rules, and determine the response strategy, persona template and rigid security boundary rules as scenario-specific prompts; inject the scenario-specific prompts into the text to be arranged to obtain prompt words.

[0089] Optionally, the second generation module 604 is further configured to input the prompt words into a preset dialogue generation model, obtain candidate response texts output by the preset dialogue generation model; perform sensitivity rule verification on the candidate response texts to obtain risk levels, wherein the risk levels include low risk, medium risk and high risk; if the risk level is medium risk, optimize the candidate response texts to obtain target response texts; if the risk level is low risk, determine the candidate response texts as target response texts; if the risk level is high risk, end the current round of the call.

[0090] Optionally, the third generation module 605 is further configured to obtain a reference timbre resource package of the target object, wherein the reference timbre resource includes a reference timbre audio and a reference text corresponding to the reference timbre audio, and the reference text is in the target format; input the target response text, the reference timbre audio, and the reference text into a preset synthesis model to obtain the response speech output by the synthesis model, wherein the preset synthesis model is configured to obtain the speech prosody based on the semantic association between the target response text and the reference text, and to extract the speech timbre based on the reference timbre audio, so as to generate the response speech based on the speech prosody and the speech timbre.

[0091] Optionally, the device further includes a reconstruction module for structurally reconstructing multi-turn conversations of a case in the following manner: acquiring case descriptions of multiple cases, wherein the case descriptions are obtained from call logs of fraudulent calls; traversing each case description; during the traversal, generating input instructions for the current case, wherein the input instructions include multiple scenario constraints; inputting the input instructions into a preset dialogue generation model to obtain candidate multi-turn dialogue texts output by the preset dialogue generation model; sequentially parsing and processing the candidate multi-turn dialogue texts to obtain the target multi-turn dialogue text; and storing the target multi-turn dialogue text of the current case to obtain structured dialogue data.

[0092] It should be noted that the examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should also be noted that the above modules, as part of a device, can operate in environments such as... Figure 1 The hardware environment shown.

[0093] According to another aspect of the embodiments of this application, this application provides an electronic device, such as... Figure 7 As shown, the device includes a memory 701, a processor 703, a communication interface 705, and a communication bus 707. The memory 701 stores a computer program that can run on the processor 703. The memory 701 and the processor 703 communicate through the communication interface 705 and the communication bus 707. When the processor 703 executes the computer program, it implements the steps of the above method.

[0094] The memory and processor in the aforementioned electronic devices communicate with each other via a communication bus and a communication interface. The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0095] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0096] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0097] According to another aspect of the embodiments of this application, a computer-readable medium having processor-executable non-volatile program code is also provided.

[0098] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments, and will not be repeated here.

[0099] In specific implementation, the embodiments of this application can be referred to the above embodiments and have corresponding technical effects.

[0100] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0101] For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or external to the processor.

[0102] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0103] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0104] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0105] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0107] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0108] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for processing fraudulent calls, characterized in that, include: When a target receives a fraudulent call, the incoming call audio is collected and divided into multiple audio segments. Generate the text to be arranged for the current round based on the audio segment; Inject scenario-specific prompts into the text to be arranged to obtain prompt words. The scenario-specific prompts include persona templates for fraud scenarios, rigid security boundary rules, and response strategies. Generate the target response text corresponding to the prompt words according to the preset dialogue generation model; Generate a response voice matching the target object based on the target response text, and send the response voice back to the caller in the fraudulent call.

2. The method according to claim 1, characterized in that, The step of generating the text to be arranged for the current round based on the audio segment includes: The audio segment is converted into a text segment in the target format using speech recognition. Perform a validity check on the text segment; If the text segment passes the validity check, the text segment is concatenated with the historical session text to obtain the text to be arranged for the current round.

3. The method according to claim 2, characterized in that, After validating the validity of the text segment, the method further includes: If the text segment fails the validity check, obtain the secure transition text; Generate a secure transition speech that matches the target object based on the secure transition text; The secure transition voice message is then sent back to the caller in the fraudulent call.

4. The method according to claim 2, characterized in that, The process of injecting scene-specific prompts into the text to be formatted, resulting in prompt words, includes: The current session stage is determined based on the historical session text in the text to be arranged; Invoke the fraud scenario strategy library and extract the response strategy from the fraud scenario strategy library according to the current session stage; Retrieve the persona template and the rigid safety boundary rule, and determine the response strategy, the persona template, and the rigid safety boundary rule as the scene-specific prompt; The scene-specific prompts are injected into the text to be arranged to obtain the prompt words.

5. The method according to claim 1, characterized in that, The step of generating the target response text corresponding to the prompt word according to the preset dialogue generation model includes: Input the prompt words into the preset dialogue generation model to obtain the candidate response text output by the preset dialogue generation model; Sensitive rule validation is performed on the candidate response texts to obtain risk levels, which include low risk, medium risk, and high risk. If the risk level is medium risk, the candidate response text is optimized to obtain the target response text. If the risk level is low risk, the candidate response text is determined as the target response text. If the risk level is high risk, the current round of the call ends.

6. The method according to claim 1, characterized in that, The step of generating a response voice matching the target object based on the target response text includes: Obtain the reference timbre resource package of the target object, wherein the reference timbre resource includes reference timbre audio and reference text corresponding to the reference timbre audio, and the reference text is in the target format; The target response text, the reference timbre audio, and the reference text are input into a preset synthesis model to obtain the response speech output by the synthesis model. The preset synthesis model is used to obtain speech prosody based on the semantic association between the target response text and the reference text, and to extract speech timbre based on the reference timbre audio, so as to generate the response speech based on the speech prosody and the speech timbre.

7. The method according to claim 1, characterized in that, The method also includes structurally reconstructing the multi-turn conversations of the case in the following manner: Obtain case descriptions for multiple cases, wherein the case descriptions are obtained based on the call logs of the fraudulent calls; Iterate through each of the case descriptions; During the traversal, input instructions for the current case are generated, wherein the input instructions include multiple scenario constraints; The input command is input into the preset dialogue generation model to obtain candidate multi-turn dialogue text output by the preset dialogue generation model; The candidate multi-turn dialogue texts are parsed and processed sequentially to obtain the target multi-turn dialogue text. The target multi-turn dialogue text of the current case is stored to obtain structured dialogue data.

8. A processing device for fraudulent calls, characterized in that, include: The acquisition module is used to acquire the audio of incoming calls when the target receives a fraudulent call, and to divide the audio into multiple audio segments; The first generation module is used to generate the text to be arranged for the current round based on the audio segment; The injection module is used to inject scenario-specific prompts into the text to be arranged to obtain prompt words. The scenario-specific prompts include persona templates for fraud scenarios, rigid security boundary rules, and response strategies. The second generation module is used to generate target response text corresponding to the prompt words according to a preset dialogue generation model; The third generation module is used to generate a response voice that matches the target object based on the target response text, and to send the response voice back to the caller in the fraudulent call.

9. An electronic device comprising a memory, a processor, a communication interface, and a communication bus, wherein the memory stores a computer program executable on the processor, and the memory and the processor communicate via the communication bus and the communication interface, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable medium having processor-executable non-volatile program code, characterized in that, The program code causes the processor to execute the method of any one of claims 1 to 7.

Images