Illegal voice detection method, device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to the fields of finance, medical treatment and the like, and provides a violation voice detection method, device, equipment and medium, and the method comprises the following steps: carrying out suspicious fragment detection on an input voice, and outputting a low-definition voice fragment; performing voice enhancement processing on the low-definition voice segment to obtain an optimized audio; converting the optimized audio into text data through voice recognition; and based on the text data, judging whether violation guide content exists or not and generating a violation analysis report. Compared with the prior art, real-time monitoring and accurate recognition of the violation guiding behaviors of the salesman are achieved, a whole-process closed loop covering detection, enhancement, conversion and judgment is formed, and it is ensured that the business process is compliant.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology and can be applied to fields such as finance and medical care. More specifically, it relates to a method, device, equipment and medium for detecting illegal speech. Background Art

[0002] In key business scenarios involving voice interaction, existing quality inspection technologies suffer from systemic flaws. High-risk processes like identity verification often involve deliberately lowered voice volume, leading to high rates of missed detection of low-definition speech. Environmental noise interference exacerbates the problem, making it difficult for conventional speech enhancement solutions to isolate effective human voices, resulting in ineffective semantic analysis. Especially for guidance sentences with gentle tones, the reduced signal-to-noise ratio leads to distorted speech recognition output, eliminating the basis for determining violations. Existing technologies also lack a comprehensive chain of evidence. The system outputs judgment labels in isolation, unable to correlate the temporal and spatial characteristics of the original speech with the processing process, rendering subsequent tracing ineffective. These three flaws create a chain of failures: hidden blind spots in speech detection, gaps in noise environment analysis, and an unfounded judgment result. Summary of the Invention

[0003] The purpose of the present invention is to overcome the defects of the existing technology and provide a method, device, equipment and medium for detecting illegal speech, which aims to solve the technical problem that the existing speech detection system cannot effectively locate, analyze and trace the evidence chain of low-definition illegal speech in key business scenarios.

[0004] In a first aspect, an embodiment of the present invention provides a method for detecting illegal speech, comprising:

[0005] Detect suspicious segments of input speech and output low-definition speech segments;

[0006] Performing speech enhancement processing on the low-definition speech segment to obtain optimized audio;

[0007] Utilizing the optimized audio, converting it into text data through speech recognition;

[0008] Based on the text data, determine whether there is any illegal guidance content and generate a violation analysis report.

[0009] In a second aspect, an embodiment of the present invention provides a device for detecting illegal speech, comprising:

[0010] A suspicious segment detection unit, configured to detect suspicious segments of input speech and output low-definition speech segments;

[0011] a speech enhancement processing unit, configured to perform speech enhancement processing on the low-definition speech segment to obtain optimized audio;

[0012] A speech recognition conversion unit, configured to utilize the optimized audio to convert the optimized audio into text data through speech recognition;

[0013] The violation analysis and reporting unit is used to determine whether there is any illegal guidance content based on the text data and generate a violation analysis report.

[0014] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned illegal speech detection method when executing the computer program.

[0015] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for detecting illegal speech as described above is implemented.

[0016] The embodiments of the present invention provide a method, apparatus, device, and medium for detecting illegal speech. The method detects suspicious segments of input speech and outputs low-definition speech segments. The method then performs speech enhancement processing on the low-definition speech segments to obtain optimized audio. The optimized audio is then converted into text data through speech recognition. Finally, based on the text data, it is determined whether there is illegal content and a violation mark status set is generated. The suspicious segment detection step accurately locates low-definition segments in the input speech, significantly improving the detection rate of covert speech. Speech enhancement processing is performed based on the segments, effectively improving the audio signal-to-noise ratio and parsability, providing a qualified input source for subsequent processing. The optimized audio is used for speech recognition conversion to ensure that the low-quality speech content is completely converted into analyzable text data. Finally, the illegal content is determined based on the text and a violation mark status set is generated, establishing a complete chain of evidence from the original speech signal to the violation conclusion. This solution systematically solves the three core defects of the existing technology: missed detection of covert speech, loss of semantics in low signal-to-noise ratio environments, and lack of traceable evidence of illegal behavior. It forms a closed loop covering the entire process of detection, enhancement, conversion, and determination, providing technical support for business compliance.

[0017] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following preferred embodiments are specifically cited and described in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of an application environment of a method for detecting illegal speech in one embodiment of the present invention;

[0019] Figure 2A flowchart of a method for detecting illegal speech in accordance with an embodiment of the present invention is provided;

[0020] Figure 3 for Figure 2 A schematic flow chart of a specific implementation of step S100;

[0021] Figure 4 for Figure 2 A schematic flow chart of a specific implementation of step S200;

[0022] Figure 5 for Figure 4 A schematic flow chart of a specific implementation of step S210;

[0023] Figure 6 for Figure 2 A schematic flow chart of a specific implementation of step S300;

[0024] Figure 7 for Figure 2 A schematic flow chart of a specific implementation of step S400;

[0025] Figure 8 A schematic structural diagram of an illegal speech detection device according to an embodiment of the present invention;

[0026] Figure 9 It is a schematic block diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0030] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0031] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0032] The embodiment of the present invention provides a method for detecting illegal speech, which can be applied in the following situations: Figure 1 In an application environment, wherein the client communicates with the server through a network, the server can detect suspicious segments of the input voice and output low-definition voice segments; perform voice enhancement processing on the low-definition voice segments to obtain optimized audio; utilize the optimized audio to convert it into text data through voice recognition; based on the text data, determine whether there is illegal guidance content and generate a violation mark status set. In the present invention, for the financial and medical fields, the low-definition segments in the input voice are accurately located through the suspicious segment detection step, which significantly improves the detection rate of concealed voice; based on the segment, voice enhancement processing is performed to effectively improve the audio signal-to-noise ratio and resolvability, providing a qualified input source for subsequent processing; utilize the optimized audio to perform voice recognition conversion to ensure that the low-quality voice content is completely converted into analyzable text data; finally, based on the text, illegal content determination is achieved and a violation mark status set is generated, establishing a complete chain of evidence from the original voice signal to the violation conclusion. This solution systematically solves the three core defects of the existing technology: missed detection of concealed voice, loss of semantics in a low signal-to-noise ratio environment, and no traceable evidence of illegal behavior, forming a closed loop covering the entire process of detection, enhancement, conversion, and determination, providing technical support for business compliance. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablet computers, and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0033] It is understandable that the illegal speech detection method in the embodiment of the present invention is explained in the application scenario in the financial field, taking the face-to-face interview process of a bank credit card as an example. In the audio recording quality inspection of the face-to-face interview process of a bank credit card, the system detects a voice clip with a sudden drop in volume from the salesperson, and obtains parseable audio after voice enhancement processing; the voice recognition conversion outputs the text "check no other loan options"; it is determined that the content contains illegal guidance that induces customers to conceal their debts, generates a violation mark state set, and records the illegal conclusion of "instructing customers to make false declarations." This process effectively prevents traditional systems from missing low-volume critical voices, ensures the integrity of the chain of evidence from physical signals to violation conclusions, and avoids the risk of regulatory accountability due to lack of evidence.

[0034] For a specific application scenario in healthcare, we'll use the example of remote surgery consent in an internet hospital. During the informed consent process, the system identifies audio clips where the doctor's speech speed suddenly increases, enhances them, and removes background noise. The system then translates the speech recognition into text: "Signature required today for scheduling." Based on medical standards, the system determines that the consent is reasonable, generates a violation status set, and notes "No violation for time-sensitive medical consent." This process precisely prevents misjudgments caused by audio quality distortion, ensures that the conclusions for high-risk medical consents are consistent with business practices, and avoids improper accountability for legitimate medical treatments.

[0035] See also Figure 2 The embodiment of the present invention discloses a method for detecting illegal speech, including S100-S400:

[0036] S100, detecting suspicious segments of input speech and outputting low-definition speech segments;

[0037] Specifically, by extracting speech feature parameters and inputting them into a specific neural network model, it is possible to accurately identify low-volume, low-definition segments from the input speech, such as the whispering and illegal guidance voices that may be heard by salespeople in financial services. Furthermore, this step also marks low-definition speech segments and their timestamps, which not only accurately locates the problematic speech but also provides a time-based reference for subsequent operations such as speech enhancement and semantic analysis. This facilitates full-process tracing and analysis, ensuring that no potential illegal speech segments are missed in scenarios such as finance, thereby improving the accuracy and reliability of detection.

[0038] In one embodiment, if Figure 3 As shown, step S100 includes:

[0039] S110, extracting feature parameters for characterizing characteristics of the speech signal;

[0040] Specifically, by extracting feature parameters used to characterize the characteristics of speech signals, the original speech signal can be converted into quantitative features that can be understood by computers. These features can capture key attributes of speech, such as spectral distribution, energy changes, and time-domain dynamics, providing a data basis for subsequent detection of low-definition speech segments. Whether it is high-frequency resonance peaks, low-frequency energy distribution, or time-domain features such as short-term zero-crossing rate in speech, they can all be effectively captured through the feature parameter extraction process, allowing the system to identify low-volume, poor-intelligibility segments from complex speech environments. For example, in financial business scenarios, the whispering and illegal guidance of salespeople often differs from normal speech in energy distribution and spectral structure. Feature parameter extraction can highlight these differences, providing accurate input for subsequent model analysis, ensuring the accuracy of low-definition speech segment detection, and avoiding missed detection problems caused by incomplete feature extraction.

[0041] It is understood that the "feature parameters" in this embodiment refer to quantitative indicators extracted from speech signals that can characterize their physical or perceptual characteristics. These include time domain features (such as short-time energy and short-time zero-crossing rate), frequency domain features (such as spectral entropy and frequency band energy ratio), and cepstral features (such as Mel-frequency cepstral coefficients and linear prediction cepstral coefficients). These parameters are extracted from the original speech through mathematical transformations or signal processing algorithms, effectively reducing data dimensionality while retaining key information. They are the fundamental input for speech processing tasks.

[0042] S120, inputting a self-attention mechanism neural network model based on the feature parameters;

[0043] Specifically, the extracted speech feature parameters are fed into a self-attention neural network model. Leveraging its unique self-attention mechanism, the model dynamically captures long-range dependencies between speech features at different time points and frequency dimensions, enabling accurate identification of characteristic patterns in low-definition speech. The self-attention neural network model automatically learns the characteristic distribution patterns of low-definition speech in both the time and frequency domains, making it more adaptable to complex scenarios such as background noise and overlapping voices compared to traditional models. For example, in the financial lending business, a salesperson's whispers may be accompanied by the ambient sounds of keyboards or colleagues chatting. By calculating attention weights for each feature parameter, the self-attention model focuses on frequency components that are critical for low-definition speech detection, suppressing noise interference and improving low-volume speech detection accuracy. Furthermore, the model's multi-layered neural network architecture abstracts speech features layer by layer, extracting more discriminative high-level representations from the raw feature parameters, making the recognition of low-definition speech segments more accurate and robust.

[0044] It is understood that the self-attention neural network model in this embodiment is a deep learning model based on the self-attention mechanism. Its core principle is to dynamically adjust the importance of different features by calculating the attention weights between each element in the input sequence, thereby capturing long-range dependencies in the sequence. In speech processing, this model can effectively analyze the temporal and frequency domain correlations of speech signals and is suitable for feature analysis in complex scenarios such as low-definition speech. Compared with traditional models such as recurrent neural networks (RNNs), it has stronger parallel computing capabilities and advantages in long-sequence processing.

[0045] S130: Mark the low-definition speech segment and timestamp according to the model output probability.

[0046] Specifically, the probability value output by the self-attention mechanism neural network model reflects the possibility that the input speech segment is low-definition speech. By setting a reasonable probability threshold, segments above the threshold can be marked as low-definition speech and accurate timestamp information can be added to them. This marking operation realizes the spatiotemporal positioning of low-definition speech. The timestamp can clearly record the specific time period when the illegal speech occurs (such as 00:02:15-00:02:18), providing a reference coordinate in the time dimension for subsequent operations such as speech enhancement and semantic analysis. For example, in a financial customer service call, when the model outputs a low-definition probability of 0.85 for a certain speech segment (higher than the threshold of 0.7), the system automatically marks the segment and records the timestamp. Subsequently, the speech in this period can be enhanced to ensure that illegal content is not missed due to insufficient clarity. In addition, the structured list formed by the marking results facilitates batch processing and retrospective review by the system. Financial institutions or medical institutions can quickly retrieve the original speech segments for manual review based on the timestamp, improving the traceability and accuracy of the detection process.

[0047] It should be understood that the timestamp in this embodiment identifies the specific time position of a speech segment within the original audio, typically expressed in the format of "hour:minute:second.millisecond." In speech processing, timestamps are used to map speech content to the temporal dimension. They are key information for accurately locating, editing, and analyzing speech segments, supporting subsequent operations such as multimodal evidence association and backtracking of offending segments.

[0048] S200, performing correlation analysis on the alarm set, the network topology, and the change records, constructing a fault propagation graph, and outputting a list of root cause nodes sorted by probability;

[0049] Specifically, by performing a series of processing on low-definition voice clips, including noise suppression, spectrum compensation, and dynamic range control, it can effectively remove background noise interference, repair spectral distortion, and balance volume, thereby improving the quality of the voice signal and making the optimized audio clearer and more intelligible. This provides a high-quality input foundation for subsequent speech recognition and semantic analysis, ensuring more accurate speech-to-text conversion, thereby ensuring the accuracy of illegal content detection and avoiding misjudgments or missed detections due to voice quality issues.

[0050] In one embodiment, if Figure 4 As shown, step S200 includes:

[0051] S210, performing noise suppression processing on the low-definition speech segment to obtain preliminary enhanced audio;

[0052] Specifically, low-definition voice clips are often difficult to discern due to interference from background noise (such as environmental noise and equipment current noise). Noise suppression processing analyzes the noise characteristics in the voice clips and uses adaptive filtering, spectral subtraction and other technologies to remove or attenuate the noise components, thereby improving the signal-to-noise ratio of the voice. This step can effectively reduce the masking of voice content by noise, making the voice signal in the initial enhanced audio clearer and more recognizable, laying the foundation for subsequent spectrum compensation and dynamic range control. For example, in a financial customer service call, the background keyboard sound or the sound of colleagues talking will interfere with the salesperson's voice. Noise suppression can highlight the salesperson's speech content, avoid misjudgment of voice features due to noise, and ensure the accuracy of subsequent semantic analysis.

[0053] It should be understood that noise suppression in this embodiment refers to the process of estimating and removing background noise from noisy speech using signal processing techniques to restore the pure speech signal. Its core is to utilize the statistical properties or time-frequency domain characteristics of noise to design filters or algorithms to suppress noise. Common methods include spectral subtraction, Wiener filtering, and adaptive filtering.

[0054] S220: Perform spectrum compensation processing based on the preliminary enhanced audio;

[0055] Specifically, after noise suppression, the initial enhanced audio may still have spectral distortion (such as attenuation of high-frequency components and fuzzy low-frequency resonance peaks) caused by transmission loss, equipment frequency response limitations, etc. Spectral compensation processing analyzes the spectral characteristics of the speech, compensates for the missing or attenuated frequency components, and restores the complete spectral structure of the speech, thereby enhancing the high-frequency details and resonance peak clarity of the speech and improving the intelligibility of the speech. For example, the high-frequency details of low-volume speech are often lost due to insufficient microphone sensitivity. Spectral compensation can specifically enhance the energy of the high-frequency band, further improving the clarity and intelligibility of the speech. In addition, this processing can balance the energy distribution of different frequency components, making the speech closer to the original pronunciation characteristics, providing better quality audio input for subsequent speech recognition and semantic analysis, and reducing text conversion errors caused by spectral distortion.

[0056] S230: Perform dynamic range control on the compensated audio and output the optimized audio.

[0057] Specifically, the compensated audio may have the problem of excessive volume fluctuations (such as low-definition voice clips with too low volume, and background sudden noise with too high volume). Dynamic range control adjusts the dynamic range of the audio through tools such as compressors and limiters to control the volume within a reasonable range. Specifically, dynamic range control will compress the parts with too high volume and increase the gain of the voice parts with too low volume, so that the overall volume of the optimized audio is uniform and consistent, which is convenient for the subsequent stable processing of the speech recognition model. For example, in a financial customer service call, dynamic range control can increase the salesperson's whispered guidance words (originally too low volume) to an appropriate volume, while suppressing sudden noises in the background, ensuring that there will be no text conversion errors due to volume fluctuations during speech recognition, and thus accurately identifying illegal guidance content.

[0058] In one embodiment, if Figure 5 As shown, step S210 includes:

[0059] S211, extracting background noise features of the low-definition speech segment;

[0060] Specifically, by analyzing the background noise characteristics in low-definition voice clips, the characteristic parameters of the noise in the time domain, frequency domain or other dimensions are obtained, providing a key basis for subsequent noise suppression. These characteristic parameters can reflect the frequency distribution, energy changes and periodicity of the background noise, enabling the system to accurately identify the difference between noise and voice signals, thereby more specifically removing interference components during noise reduction processing to avoid excessive damage to the original voice signal. For example, in customer service calls in financial services, after extracting background noise characteristics such as keyboard tapping and hall ambient sound, a more accurate noise reduction solution can be designed based on this, ensuring that the salesperson's voice content can be clearly identified in subsequent processing, avoiding the omission of illegal guidance content due to noise interference.

[0061] S212, constructing an adaptive filter based on the background noise characteristics;

[0062] Specifically, based on the extracted background noise features, an adaptive filtering model is constructed, which can dynamically adjust the filtering parameters according to the real-time changes in the noise, thereby effectively suppressing different types of background noise. This adaptive filter can automatically optimize the filtering effect for the complex and changeable noise environments in financial or medical scenarios (such as the noise of the crowd in the financial business hall and the operating noise of the hospital equipment), while removing the noise while retaining the key features of the speech signal to the maximum extent, improving the intelligibility of low-definition speech, and laying the foundation for subsequent speech enhancement and semantic analysis. For example, in medical remote consultations, the adaptive filter constructed based on the background noise characteristics of the ward can suppress interference such as the alarm sound of the monitor in real time, ensuring that the doctor's diagnostic voice is clear and discernible.

[0063] S213: Apply the filter to perform noise reduction processing on the speech segment to generate the preliminary enhanced audio.

[0064] Specifically, by applying the constructed adaptive filter to low-definition speech clips, targeted suppression of background noise is achieved, thereby generating a preliminary enhanced audio signal. The audio after noise reduction processing can effectively reduce interference from environmental noise, equipment noise, and other sources, thereby improving the signal-to-noise ratio of the speech signal and significantly improving the clarity and intelligibility of the speech content. This process provides high-quality audio input for subsequent operations such as spectrum compensation and dynamic range control, ensuring that the whispered words of salespeople can be accurately captured in the detection of financial irregularities, or that the professional expressions of doctors can be clearly recorded in medical scenarios, avoiding speech recognition errors and semantic misunderstandings caused by noise.

[0065] S300, generating a device parameterized repair script according to the root cause node list;

[0066] Specifically, by acquiring optimized audio and converting it into text data, the voice signal can be converted into editable and analyzable text, providing a foundation for the generation of parameterized device repair scripts. Based on correlation analysis between text data and root cause nodes, the generated repair scripts can accurately repair device parameter anomalies, automating the entire process from voice detection to device repair. This improves the efficiency and accuracy of network troubleshooting and provides efficient solutions for device operation and maintenance in fields such as finance and healthcare.

[0067] In one embodiment, if Figure 6 As shown, step S300 includes:

[0068] S310: Obtain the optimized audio and divide it into audio frames of fixed length;

[0069] Specifically, the optimized audio remains a continuous speech signal after processing, and is divided into audio frames of fixed duration (such as 20ms or 30ms), which can make speech processing more refined. The fixed division of audio frames facilitates the extraction of speech features (such as short-time energy and zero-crossing rate) frame by frame, providing a stable input unit for the speech recognition model. In addition, inter-frame overlapping technology (such as 50% overlap) can ensure the temporal continuity of the speech signal and avoid feature breaks caused by segmentation. This processing method can adapt to the short-term stationary characteristics of the speech signal (that is, the speech features are relatively stable in a short time), so that the speech recognition model can more accurately capture the acoustic features of each time segment, laying the foundation for subsequent text generation, and is especially suitable for the precise positioning of keywords in low-definition speech, such as the recognition of short words such as "concealment" and "only need" in financial violation guidance.

[0070] S320: Generate text for the audio frame using a pre-trained speech recognition model.

[0071] Specifically, the pre-trained speech recognition model has learned the mapping relationship between acoustic features and text through a large amount of speech data. For the segmented audio frames, the model can convert the acoustic features of each frame (such as Mel-frequency cepstral coefficients, spectrograms) into corresponding text sequences. For low-definition speech, the model adaptively learns speech features in a noisy environment through its deep neural network structure (such as convolutional neural network + recurrent neural network), and combines language models (such as n-gram or Transformer language models) to predict the most likely text content, thereby realizing the conversion from speech to text. The generated text data can be directly used for subsequent semantic analysis, which enables the detection of illegal guidance from the audio level to the text semantic level. For example, after the financial clerk's whispered guidance voice "This loan interest is very low" is converted into text, it is convenient to match it with the preset violation template library to determine whether there is any misleading statement.

[0072] It can be understood that the pre-trained speech recognition model in this embodiment refers to a model that has been pre-trained using a large-scale speech data set (such as speech data containing multiple languages, accents, and scenarios), and has the ability to convert speech signals into text. In actual applications, it can be fine-tuned to adapt to the speech recognition needs of specific fields.

[0073] S400: Based on the text data, determine whether there is any illegal guidance content and generate a violation mark status set.

[0074] Specifically, by extracting and matching semantic features from text data, we can deeply explore the implicit intent within the text and accurately identify content that violates regulations, such as "inducing the concealment of information" in the financial sector or "exaggerating efficacy" in the medical sector. Generating a status set containing violation type identifiers and confidence scores provides a quantitative basis for subsequent violation handling, allowing relevant departments to take targeted measures such as early warnings, verification, and accountability based on different violation situations, ensuring the compliance of business processes in the financial and medical sectors and safeguarding user rights and industry standards.

[0075] In one embodiment, if Figure 7 As shown, step S400 includes:

[0076] S410, extracting a semantic feature vector based on the text data;

[0077] Specifically, semantic feature vector extraction of text data involves converting natural language text into a numerical representation that can be understood by computers. Deep learning models (such as BERT and Transformer) are then used to analyze the context of the text, lexical semantic associations, and other aspects, to generate a vector that can represent the overall semantics of the text. For text in the financial or medical fields, this vector can capture the sentiment of keywords (such as guiding words like "must" and "only need"), semantic relationships (such as causal relationships and conditional relationships), and domain-specific semantics (such as the financial term "liquidated damages" and the medical term "surgical risk"). The extracted semantic feature vector can remove the surface grammatical structure of the text while retaining the deep semantic connotation, providing a unified semantic representation space for subsequent matching analysis with a preset violation template library. This allows the system to accurately identify the implicit violation-guiding intent in the text, such as extracting the semantic feature of "concealing information" from the sentence "This loan application is very simple and there is no need to provide real bank statements."

[0078] S420, matching and analyzing the semantic feature vector with a preset violation template library;

[0079] Specifically, the preset violation template library stores semantic feature vectors or template patterns of known violation-inducing texts in the financial or medical fields. By calculating the similarity (such as cosine similarity or Euclidean distance) between the semantic feature vector of the text to be analyzed and the vectors in the template library, it can be determined whether the text contains any violation content. The matching analysis process combines domain knowledge and historical violation cases. For example, the template library in the financial field contains violation semantic templates such as "inducing customers to conceal income" and "misleading promises of benefits", while the template library in the medical field contains templates such as "exaggerating efficacy" and "concealing risks". This matching mechanism can quickly locate the violation semantics in the text and is more capable of capturing semantic variations (such as synonymous expressions and implicit intentions) than simple keyword matching. For example, matching the template "You just need to meet the income target" with the template "inducing concealment of income" improves the comprehensiveness and accuracy of violation detection.

[0080] S430: Generate a violation mark status set according to the matching result, wherein the status set includes a violation type identifier and a confidence score.

[0081] Specifically, based on the matching results of semantic feature vectors with pre-set templates, the system generates a structured violation flagging status set. The violation type identifier (e.g., "Induced to Withhold Information" or "Misleading Promises") identifies the nature of the violation, and the confidence score (e.g., 0.85) reflects the reliability of the match. This status set provides a quantitative basis for subsequent violation handling. For example, a confidence score above 0.8 triggers a high-priority alert, while a score below 0.8 requires manual review. Furthermore, the structured design of the status set facilitates statistical analysis, such as generating distribution reports by violation type. This helps financial or medical institutions identify high-frequency violation scenarios and optimize business processes accordingly. For example, in loan operations, if the violation type of "Inducing Customers to Simplify Documents" frequently occurs, sales staff training on document review can be strengthened. In the medical field, the status set can identify the type and confidence level of any unauthorized exaggerations in pharmacist recommendations, such as "This drug is absolutely effective for diabetes," allowing drug regulatory authorities to prioritize high-confidence violations.

[0082] It can be understood that the violation marking status set in this embodiment refers to a structured data set containing information such as the violation type and confidence score, which is used to record and identify the violation content detected in the text and is the core data carrier for subsequent violation processing and analysis.

[0083] Furthermore, the steps after step S430 further include:

[0084] S440: Locate a timestamp interval corresponding to the original speech according to the violation mark state set;

[0085] Specifically, the violation marking state set is associated with the timestamp interval of the original speech, which can accurately map the text-level violation judgment to the specific time period of the speech, and realize the retrospective mechanism of "text violation-speech positioning". The positioning of the timestamp interval is based on the frame-level time information recorded during speech recognition (such as the starting milliseconds corresponding to the audio frame), combined with the context length of the violation text, to determine the time range of the violation content in the original speech. This step facilitates the rapid positioning of the violation voice segment during manual review. For example, financial compliance personnel can directly play the voice of the marked time period, and further confirm the violation by combining acoustic features such as tone and background sound. In addition, the timestamp interval provides a time anchor for the subsequent extraction of multimodal feature evidence (such as video images, operation logs), ensuring the spatiotemporal consistency of multi-dimensional evidence and improving the credibility of violation judgments.

[0086] S450, extracting associated multimodal feature evidence based on the timestamp interval;

[0087] Specifically, the extraction of multimodal feature evidence is based on the timestamp interval, integrating multi-dimensional data such as voice, video, and text to form a comprehensive chain of evidence of violations. The voice modality can extract acoustic features (such as sudden changes in volume, accelerated speaking speed, and spectral anomalies), the video modality can extract visual features (such as salesperson's gesture occlusion, facial expression avoidance, and abnormal operating interface), and the text modality can extract semantic features (such as the context of illegal keywords). For example, in a financial violation scenario, in addition to the illegal statements in the text, the whispered tone of the voice (acoustic features) and the gesture of the salesperson covering the contract terms in the video (visual features) can jointly prove the intention of illegal guidance. The fusion analysis of multimodal features can avoid the misjudgment of a single modality and improve the sufficiency of evidence. For example, the word "concealment" appearing only in the text may be a neutral statement, but combined with the whispered voice and the covering action, it can be confirmed as illegal guidance.

[0088] It can be understood that the multimodal feature evidence in this embodiment refers to feature information from multiple modalities such as voice, video, text, sensor data, etc. After being associated through timestamps, this information can be used together as evidence to determine violations, and is multidimensional and complementary.

[0089] S460: Integrate the multimodal feature evidence to generate a violation analysis report.

[0090] Specifically, integrating multimodal evidence and generating a violation analysis report is a process of systematizing and structuring scattered evidentiary elements. Reports typically include an overview of the violation (time, setting, and parties involved), detailed multimodal evidence (voice feature maps, video screenshots, and text highlights), the basis for determining the violation type, a confidence analysis, and recommended actions. For example, a report in the financial sector might read: "On June 15, 2025, at 2:30 PM, a loan officer whispered (average volume 20dB) during a video interview that 'no bank statement is required' (a text violation), while simultaneously covering the bank statement entry screen with his hand. This was comprehensively determined to be a 'induced concealment of financial information' violation, with a confidence level of 0.92. A compliance investigation is recommended." This report format provides an intuitive and comprehensive reference for compliance decision-making in financial or healthcare institutions. It can also serve as training material, helping practitioners identify the multimodal characteristics of violations and mitigating the risk of violations at the source.

[0091] It can be seen that the illegal speech detection method of the embodiment of the present invention detects suspicious fragments of the input speech, accurately identifies low-definition speech fragments and marks the timestamps, obtains optimized audio through speech enhancement processing such as noise suppression, spectrum compensation and dynamic range control, and then converts it into text data. Combined with semantic feature extraction and matching analysis of the preset violation template library, it generates a marked state set containing the violation type and confidence level, and finally realizes real-time monitoring and accurate identification of the salesperson's illegal guidance behavior, effectively solving the problem of insufficient detection ability of traditional technology for low-volume and hidden illegal speech, improving the model's generalization ability and robustness, and forming a dual detection mechanism of speech and text to ensure compliance of business processes and protect customer rights.

[0092] The embodiment of the present invention also provides a device for detecting illegal speech, which is used to execute any embodiment of the above-mentioned illegal speech detection method. Figure 8 , Figure 8 1 is a schematic block diagram of a device for detecting illegal speech provided by an embodiment of the present invention. The device 500 for detecting illegal speech includes:

[0093] The suspicious segment detection unit 510 is used to detect suspicious segments of the input speech and output low-definition speech segments;

[0094] a speech enhancement processing unit 520, configured to perform speech enhancement processing on the low-definition speech segment to obtain optimized audio;

[0095] A speech recognition conversion unit 530 is configured to convert the optimized audio into text data through speech recognition;

[0096] The violation analysis reporting unit 540 is configured to determine whether there is any illegal guidance content based on the text data and generate a violation analysis report.

[0097] In one embodiment, the suspicious segment detection unit 510 is specifically configured to:

[0098] Extracting feature parameters for characterizing speech signal characteristics;

[0099] Based on the feature parameters, input a self-attention mechanism neural network model;

[0100] Based on the model output probability, low-definition speech segments and timestamps are labeled.

[0101] In one embodiment, the speech enhancement processing unit 520 is specifically configured to:

[0102] Performing noise suppression processing on the low-definition speech segment to obtain preliminary enhanced audio;

[0103] performing spectrum compensation processing based on the preliminary enhanced audio;

[0104] Dynamic range control is performed on the compensated audio, and the optimized audio is output.

[0105] In one embodiment, more specifically, the speech enhancement processing unit 520 further includes:

[0106] A background noise feature extraction unit, configured to extract background noise features of the low-definition speech segment;

[0107] An adaptive filter construction unit, configured to construct an adaptive filter based on the background noise characteristics;

[0108] The noise reduction processing unit is used to apply the filter to perform noise reduction processing on the speech segment to generate the preliminary enhanced audio.

[0109] In one embodiment, the speech recognition conversion unit 530 is specifically configured to:

[0110] Obtaining the optimized audio and dividing it into audio frames of fixed duration;

[0111] For the audio frame, text is generated by using a pre-trained speech recognition model.

[0112] In one embodiment, the violation analysis and reporting unit 540 is specifically configured to:

[0113] Extracting semantic feature vectors based on the text data;

[0114] Matching and analyzing the semantic feature vector with a preset violation template library;

[0115] A violation mark state set is generated according to the matching result, wherein the state set includes a violation type identifier and a confidence score.

[0116] Furthermore, the violation analysis and reporting unit 540 is further specifically configured to:

[0117] Locating a timestamp interval corresponding to the original speech according to the violation mark state set;

[0118] Extracting associated multimodal feature evidence based on the timestamp interval;

[0119] Integrate the multimodal feature evidence to generate a violation analysis report.

[0120] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned illegal voice detection device and each unit can refer to the corresponding description in the aforementioned method embodiment. For the convenience and brevity of description, it will not be repeated here.

[0121] The above-mentioned illegal speech detection device 500 can be implemented in the form of a computer program. The computer program can be used in Figure 9 Runs on the computer equipment shown.

[0122] See also Figure 9 , Figure 9 This is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 600 can be a terminal or a server. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, personal digital assistant, wearable device, or other electronic device with communication capabilities. The server can be a standalone server or a server cluster consisting of multiple servers.

[0123] The computer device 600 includes a processor 620 , a memory, and a network interface 650 connected via a system bus 610 , wherein the memory may include a non-volatile storage medium 630 and an internal memory 640 .

[0124] The non-volatile storage medium 630 can store an operating system 631 and a computer program 632. The computer program 632 includes program instructions, which, when executed, can enable the processor 620 to perform an illegal speech detection method.

[0125] The processor 620 is used to provide computing and control capabilities to support the operation of the entire computer device 600.

[0126] The internal memory 640 provides an environment for the operation of the computer program 632 in the non-volatile storage medium 630. When the computer program 632 is executed by the processor 620, the processor 620 can execute a method for detecting illegal speech.

[0127] The network interface 650 is used to communicate with other devices through the network. Figure 9 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 600 to which the solution of the present application is applied. The specific computer device 600 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0128] It should be understood that in the embodiment of the present application, the processor 620 may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0129] Those skilled in the art will appreciate that all or part of the steps in the method of the above-described embodiment can be implemented by instructing the relevant hardware through a computer program. The computer program includes program instructions, which can be stored in a storage medium that is computer-readable. The program instructions are executed by at least one processor in the computer system to implement the steps in the method of the above-described embodiment.

[0130] Therefore, the present application also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When the program instructions are executed by a processor, the processor performs the following steps:

[0131] S100, detecting suspicious segments of input speech and outputting low-definition speech segments;

[0132] S200, performing speech enhancement processing on the low-definition speech segment to obtain optimized audio;

[0133] S300, using the optimized audio to convert it into text data through speech recognition;

[0134] S400: Based on the text data, determine whether there is any illegal guidance content and generate a violation mark status set.

[0135] The storage medium may be any computer-readable storage medium that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk.

[0136] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0137] The non-Company software tools or components appearing in the embodiments of this application are merely examples and do not represent actual use.

[0138] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and other division methods may be used in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not implemented.

[0139] The steps in the method of the embodiment of the present application can be adjusted in order, combined, and deleted according to actual needs. The units in the device of the embodiment of the present application can be combined, divided, and deleted according to actual needs. In addition, the functional units in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit.

[0140] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, terminal, or network device, etc.) to execute all or part of the steps of the various embodiments of the present application.

[0141] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for detecting illegal speech, characterized in that: include: Detect suspicious segments of input speech and output low-definition speech segments; Performing speech enhancement processing on the low-definition speech segment to obtain optimized audio; Utilizing the optimized audio, converting it into text data through speech recognition; Based on the text data, it is determined whether there is any illegal guidance content and a violation mark status set is generated.

2. The illegal speech detection method according to claim 1, characterized in that: The step of detecting suspicious segments of input speech and outputting low-definition speech segments comprises: Extracting feature parameters for characterizing speech signal characteristics; Based on the feature parameters, input a self-attention mechanism neural network model; Based on the model output probability, low-definition speech segments and timestamps are labeled.

3. The illegal speech detection method according to claim 1, characterized in that: The step of performing speech enhancement processing on the low-definition speech segment to obtain optimized audio includes: Performing noise suppression processing on the low-definition speech segment to obtain preliminary enhanced audio; performing spectrum compensation processing based on the preliminary enhanced audio; Dynamic range control is performed on the compensated audio, and the optimized audio is output.

4. The illegal speech detection method according to claim 3, characterized in that: The steps of the noise suppression process include: Extracting background noise features of the low-definition speech segment; constructing an adaptive filter based on the background noise characteristics; The filter is applied to perform noise reduction processing on the speech segment to generate the preliminary enhanced audio.

5. The illegal speech detection method according to claim 1, characterized in that: The step of converting the optimized audio into text data through speech recognition includes: Obtaining the optimized audio and dividing it into audio frames of fixed duration; For the audio frame, text is generated by using a pre-trained speech recognition model.

6. The illegal speech detection method according to claim 1, characterized in that: The step of determining whether there is illegal guidance content based on the text data and generating an illegal marking state set includes: Extracting semantic feature vectors based on the text data; Matching and analyzing the semantic feature vector with a preset violation template library; A violation mark state set is generated according to the matching result, wherein the state set includes a violation type identifier and a confidence score.

7. The illegal speech detection method according to claim 6, characterized in that: After determining whether there is illegal content based on the text data and generating a violation mark status set, the following steps may also be performed: Locating a timestamp interval corresponding to the original speech according to the violation mark state set; Extracting associated multimodal feature evidence based on the timestamp interval; Integrate the multimodal feature evidence to generate a violation analysis report.

8. A device for detecting illegal speech, for implementing the method for detecting illegal speech according to any one of claims 1 to 7, characterized in that: include: A suspicious segment detection unit, configured to detect suspicious segments of input speech and output low-definition speech segments; a speech enhancement processing unit, configured to perform speech enhancement processing on the low-definition speech segment to obtain optimized audio; A speech recognition conversion unit, configured to utilize the optimized audio to convert the optimized audio into text data through speech recognition; The violation analysis and reporting unit is used to determine whether there is any illegal guidance content based on the text data and generate a violation analysis report.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the illegal speech detection method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the illegal speech detection method according to any one of claims 1 to 7 are implemented.

Images