Voice detection method and device, equipment and storage medium

Abstract

The invention provides a voice detection method and device, equipment and a storage medium, and relates to the field of big data processing, and is used for improving the timeliness and accuracy of voice detection and ensuring the normal development of businesses. The specific scheme is as follows: acquiring and analyzing voice data in a call process, converting the voice data into a text and matching a preset lexicon; automatic compliance detection from voice to text is realized; and when the preset word is detected, differentiated treatment is adopted according to the response level of the preset word, so that timely blocking and non-compliance control are realized, and non-compliance behavior diffusion is prevented. In the whole process, the timeliness of voice detection is improved, objective and consistent compliance judgment can be realized based on the unified word bank, the detection accuracy is remarkably improved, and standard development of call services is guaranteed.

Description

Technical Field

[0001] This application relates to the field of big data processing, and more particularly to voice detection methods, apparatus, devices, and storage media. Background Technology

[0002] In business scenarios involving remote voice interaction, compliance testing of the call process is a common requirement. Currently, the main testing method relies on manual quality inspection, where quality inspectors listen to call recordings or perform real-time testing to determine whether the call content complies with regulations.

[0003] However, this manual detection method has significant drawbacks: First, it lacks timeliness, typically only processing can be done after a call ends, making timely feedback difficult. Furthermore, manual processing is subjective, making it difficult to achieve accurate detection. This, in turn, affects the normal operation of the business. Summary of the Invention

[0004] This application provides a voice detection method, apparatus, device, and storage medium to improve the timeliness and accuracy of voice detection and ensure normal business operations.

[0005] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, a speech detection method is provided, which includes: Acquire the voice data of the first and second parties in the communication link during the call, and parse the first voice data of the first party to obtain the first voice data. The first speech data is converted into target text data, and it is detected whether the target text data contains preset words from a preset dictionary; wherein, the preset dictionary contains multiple preset words, each of which corresponds to a different response level; If a preset word from a preset dictionary is detected in the target text data, then the preset word detected in the target text data will be used as the target preset word. If the response level corresponding to the target preset word is the preset first response level, a prompt message is sent to the terminal device used by the first caller; the prompt message is used to remind the first caller to pay attention to the standard of language; if the response level corresponding to the target preset word is the preset second response level, the communication link is cut off.

[0006] This embodiment of the application acquires and parses voice data during a call, converts it into text, and matches it against a preset dictionary to achieve automated compliance detection from voice to text. When a preset word is detected, differentiated actions are taken based on its response level: if it is at the first response level, a prompt message is sent to the first party in the call to provide an immediate reminder and prevent the continued use of non-compliant language; if it is at the second response level, the communication link is directly cut off to achieve timely blocking and control and prevent the spread of non-compliant behavior. The entire process not only improves the timeliness of voice detection but also enables objective and consistent compliance judgments based on a unified dictionary, significantly improving detection accuracy and ensuring the standardized operation of business.

[0007] In one possible implementation of the first aspect, the method further includes: After detecting a communication link disconnection, complete voice recording data of the first and second parties in the call is acquired; the complete voice recording data includes at least one first voice segment and at least one second voice segment; the first voice segment corresponds to the first party in the call, and the second voice segment corresponds to the second party in the call. Speech recognition is performed on the complete voice recording data to obtain a complete text sequence; wherein, the complete text sequence includes at least one first text segment and at least one second text segment, the first text segment corresponds to the first caller, the second text segment corresponds to the second caller, and each text segment is timestamped; For each first text segment, the start time, end time, and target time interval of the first text segment are determined based on the timestamp corresponding to the first text segment, and at least one second text segment within the target time interval is taken as the associated text segment of the first text segment; the left value of the target time interval is the difference between the start time and the preset time value, and the right value of the target time interval is the sum of the end time and the preset time value; Extract each first text segment corresponding to the first speaker in the complete text sequence, as well as the associated text segments corresponding to the first text segments, to obtain at least one text sequence to be analyzed; Each text sequence to be analyzed is input into a preset natural language processing model to obtain the compliance judgment result output by the natural language processing model; the compliance judgment result is used to characterize whether the language used by the first party in the call is compliant.

[0008] Based on the aforementioned technical content, this embodiment acquires the complete voice recording after the communication link is disconnected and converts it into a timestamped text sequence. Time interval matching is used to determine the associated second text segment for each first text segment, forming a text sequence to be analyzed. This sequence is then input into a natural language processing model for compliance determination. This method enables traceability analysis of the entire call process, providing a systematic and context-aware assessment of the overall language compliance of the first party after the call ends. This avoids missed or incorrect judgments due to limitations in detection, further improving the comprehensiveness and accuracy of voice detection.

[0009] In one possible implementation of the first aspect, the natural language processing model is used to perform the following steps: The first text segment and the associated text segment in the text sequence to be analyzed are converted into corresponding semantic vectors, where the semantic vectors are used to represent the text semantic features of the corresponding text segments; Calculate the correlation weight between the semantic vector corresponding to the first text segment and the semantic vector of each second text segment in the associated text segment, where the correlation weight characterizes the degree of influence of the semantic features of each second text segment on the compliance of the first text segment; The semantic vectors of each second text segment in the associated text segment are weighted and summed according to the relevance weight to obtain the context semantic vector, where the context semantic vector represents the overall influence of the semantic features of the associated text segment on the compliance of the first text segment. The semantic vector of the first text segment is concatenated with the semantic vector of the context to obtain a comprehensive semantic vector. Based on the comprehensive semantic vector, the probability value of the first text segment belonging to each preset non-compliant term category is determined. If the probability value of the first text segment under any non-compliant term category is greater than the preset threshold, the first text segment is determined to be a non-compliant term.

[0010] Here, the semantic impact on the second text segment is quantitatively assessed by converting the first and related text segments into semantic vectors and calculating their relevance weights. Then, a weighted summation is used to obtain a contextual semantic vector, which is concatenated with the semantic vector of the first text segment to form a comprehensive semantic vector for compliance classification. This enables the model to understand the impact of the dialogue context on the compliance of the first speaker's language, avoiding misjudgments caused by isolated assessments and improving the contextual adaptability of compliance determination. Simultaneously, by using probability thresholds to determine non-compliant language, refined and configurable non-compliance identification is achieved, enhancing the flexibility and accuracy of detection.

[0011] In one possible implementation of the first aspect, the method further includes: Based on each first text segment, extract the text sentiment features corresponding to the first caller, and based on each first speech segment, extract the acoustic sentiment features corresponding to the first caller. Each text sentiment feature is fused with each acoustic sentiment feature to obtain each multimodal sentiment feature corresponding to the first caller. Based on each multimodal emotional feature, the emotional tendency of the first caller during this call is determined; emotional tendency includes positive tendency, neutral tendency, and negative tendency. If the emotional tendency of the first party in the call is negative, the complete voice recording data is marked according to the negative tendency, and the first review task is generated based on the marked complete voice recording data. Send the first review task to the review system and receive the review results from the review system for the first review task.

[0012] This method extracts and fuses textual and acoustic sentiment features to obtain multimodal sentiment features, which are then used to determine the emotional inclination of the first caller. If the inclination is negative, it is automatically flagged and a review task is generated. This approach enables compliance assessment of the first caller's emotional dimension, helping to identify potential emotionally charged non-compliant behaviors. Simultaneously, the automatic generation of review tasks streamlines the handling of abnormal emotional events during calls, improving review efficiency and targeting.

[0013] In one possible implementation of the first aspect, the method further includes: Based on the response level corresponding to the target preset word matched by the target text data, add corresponding tags to the first speech data. Based on the first speech data after adding tags and the target preset word corresponding to the first speech data, generate a second review task, send the second review task to the review system, and receive the review result of the review system for the second review task. And / or, For the first text segment that is judged to be non-compliant, the corresponding target speech data is extracted from the complete speech recording data according to the timestamp of the first text segment; the speech data, the first text segment and its corresponding non-compliant speech category are used to generate a third review task, and the third review task is sent to the review system to receive the review results of the review system for the third review task.

[0014] Based on the aforementioned technical content, this embodiment generates review tasks according to various scenarios such as voice detection results during a call and non-compliance detection results after the call ends, and pushes them to the review system. This mechanism achieves seamless integration of detection and review, ensuring that all kinds of suspicious or non-compliant events can be recorded and enter the review process, improving review efficiency and accuracy, forming a closed-loop management of "detection-review-feedback", and enhancing the reliability of the voice detection system.

[0015] In one possible implementation of the first aspect, the speech detection method is implemented using a trained intelligent model; the intelligent model includes a speech recognition model and a natural language processing model; the speech recognition model is used to convert the speech data of the first speaker into target text data; The method also includes: Before acquiring the voice data of the first and second parties in the communication link during the call, the initial intelligent model is trained using the training data in the training database to obtain the trained intelligent model; wherein, the training database includes multiple historical call voice data and the corresponding text data, compliance judgment results, sentiment tendencies and review results for each historical call voice data; After obtaining the compliance assessment result, sentiment tendency, and at least one review result of this call, the complete voice recording data, complete text sequence, compliance assessment result, sentiment tendency, and at least one review result of this call are stored as new training data in the training database for use in optimizing and training the intelligent model.

[0016] Here, an intelligent model incorporating speech recognition and natural language processing is constructed and trained using historical data to achieve end-to-end speech compliance detection. After each call, the call data and detection results are fed back into the database as new training data for model optimization. This enables continuous learning and evolution of the model. As data accumulates, the model's ability to recognize compliant language continuously improves, adapting to changes in business terminology and enhancing the accuracy of speech detection.

[0017] In one possible implementation of the first aspect, converting the voice data of the first party in the call into target text data includes: The speech data is segmented into frames to obtain the corresponding audio frame sequence, and acoustic features are extracted from the audio frame sequence to obtain the acoustic feature sequence. The acoustic feature sequence is input into a preset speech recognition model for decoding to obtain the corresponding text unit sequence; the text unit sequence is then combined into target text data.

[0018] In this embodiment, speech data is framed, acoustic features are extracted, and then input into a speech recognition model for decoding into text units, which are then combined into target text data. This method achieves efficient and accurate conversion from speech to text, providing reliable input for subsequent text analysis. Framed segmentation and acoustic feature extraction effectively capture the temporal and spectral characteristics of the speech signal, improving recognition robustness. The structured decoding of the speech recognition model ensures the continuity and semantic integrity of the text output, laying a high-quality data foundation for subsequent compliance testing.

[0019] Secondly, a speech detection device is provided, the device comprising: The acquisition unit is used to acquire the voice data of the first party and the second party in the communication link during the call, and to parse the first party's voice data. The first processing unit is used to convert the first speech data into target text data and detect whether the target text data contains preset words from a preset dictionary; wherein, the preset dictionary contains multiple preset words, and each preset word corresponds to a different response level; The determining unit is used to determine the target preset word if a preset word in the preset word library is detected in the target text data. The second processing unit is configured to send a prompt message to the terminal device used by the first caller if the response level corresponding to the target preset word is a preset first response level; the prompt message is used to remind the first caller to pay attention to the standard of language use; if the response level corresponding to the target preset word is a preset second response level, the communication link is cut off.

[0020] Thirdly, an electronic device is provided, the method comprising: a memory and at least one processor. The memory is communicatively connected to the processor. The memory is used to store computer program code, the computer program code including computer instructions. When the processor executes the computer instructions, it causes the electronic device to perform the method of the first aspect and any possible implementation thereof.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the method as described in the first aspect and any possible implementation thereof.

[0022] Fifthly, embodiments of this application provide a computer program product that, when running on a computer / executed by the computer's processor, implements the method described in the first aspect and any possible design thereof. The computer may be an electronic device as described in the third aspect and any possible implementation thereof.

[0023] Understandably, the beneficial effects achieved by the voice detection device of the second aspect, the electronic device of the third aspect, the computer-readable storage medium of the fourth aspect, and the computer program product of the fifth aspect provided above can be referred to as the beneficial effects of the first aspect and any possible implementation thereof, which will not be repeated here. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of a speech detection method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating another speech detection method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a voice detection device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this embodiment, unless otherwise stated, "a plurality of" means two or more.

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0027] The technical solutions provided in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data, comply with relevant laws and regulations and do not violate public order and good morals.

[0028] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0029] In business scenarios involving remote voice interaction, compliance testing of the call process is a common requirement. Currently, the main testing method relies on manual quality inspection, where quality inspectors listen to call recordings or perform tests to determine whether the call content complies with regulations. However, this manual testing method has significant drawbacks: firstly, it lacks timeliness, typically only processing after the call has ended, making timely feedback difficult; secondly, manual processing is subjective, making it difficult to achieve accurate testing. This ultimately impacts the normal operation of the business.

[0030] For example, banks typically outsource the collection of overdue payments to external collection agencies, resulting in a massive volume of daily collection calls. Human quality inspectors can only sample and review a small number of recordings, leading to a significant number of non-compliant calls going undetected. This not only allows non-compliant behavior to escape regulation but also risks damaging the bank's reputation due to complaints, and the cost of remediation afterward far outweighs the cost of prevention.

[0031] For example, when conducting marketing or promotion, customer service personnel should strictly adhere to the standards of information disclosure and compliant sales scripts. The lag in manual quality inspection means that misleading language such as exaggerated claims cannot be corrected in a timely manner during the call, which may lead to customers completing transactions without full knowledge, and subsequently, disputes are highly likely.

[0032] To improve the timeliness and accuracy of voice detection and ensure normal business operations, the implementation of this application acquires and parses voice data during calls, converts it into text, and matches it against a preset dictionary to achieve automated compliance detection from voice to text. When a preset word is detected, differentiated handling is applied based on its response level to achieve timely blocking and control, preventing the spread of non-compliant behavior. The entire process not only improves the timeliness of voice detection but also enables objective and consistent compliance judgments based on a unified dictionary, significantly improving detection accuracy and ensuring the standardized operation of business.

[0033] This application provides a voice detection method that can be applied to electronic devices. The electronic device can be a single server or a server cluster consisting of multiple servers, or a cloud computing platform with data processing capabilities, an edge computing device, a chip, or a device with computing capabilities. This application does not limit the specific form of the electronic device.

[0034] Figure 1 This is a schematic flowchart of a speech detection method provided in an embodiment of this application. Figure 1 As shown, the method in the embodiments of this application may include: S101. Acquire the voice data of the first and second parties in the communication link during the call, and parse the first voice data of the first party.

[0035] There can be multiple communication links. In this embodiment, voice detection can be performed on multiple communication links in parallel. The following explanation uses one communication link as an example.

[0036] For example, this embodiment can capture mixed voice data of the first and second parties during a call by using a voice acquisition module integrated in the communication link or by calling the application programming interface provided by the communication platform. Subsequently, this embodiment analyzes and separates the voice data according to communication signaling or voice features (such as voiceprint features, voice channel separation technology) to extract the first voice data containing the voice components of the first party.

[0037] For example, in a customer service scenario, the first party to the call can be the customer service representative. In this embodiment, the customer service representative's voice data is extracted from the recording of their call with the customer.

[0038] This step provides a voice data source for the first caller for subsequent analysis, ensuring the clarity of the detection target and avoiding analysis errors caused by mixed voices.

[0039] S102. Convert the first speech data into target text data and detect whether the target text data contains preset words from the preset word library; wherein, the preset word library has multiple preset words, and each preset word corresponds to a different response level.

[0040] For example, in this embodiment, an automatic speech recognition (ASR) engine can be used to convert the first speech data to obtain the corresponding target text data.

[0041] In one example, the conversion process may include framing the speech data, feature extraction, decoding with acoustic and language models, and finally outputting a coherent text sequence.

[0042] In this embodiment, a preset word library is pre-built. The preset word library contains multiple preset words that need to be detected (such as non-compliant terms, sensitive topic words, specific business prohibitions, etc.), and each preset word is assigned a specific response level.

[0043] This embodiment performs matching detection between the target text data and a preset dictionary. For example, string matching and regular expression matching are used to detect whether the target text data contains preset words from the preset dictionary.

[0044] This embodiment converts voice data into analyzable text and performs rapid matching using a preset thesaurus. This transforms subjective issues of language compliance into objective keyword matching problems, laying the foundation for automated judgment.

[0045] S103. If a preset word from a preset word library is detected in the target text data, the preset word detected in the target text data shall be used as the target preset word.

[0046] For example, if the target text contains a word or phrase that matches a preset word in the preset word library, this embodiment determines the specific preset word that is matched as the target preset word.

[0047] S104. If the response level corresponding to the target preset word is the preset first response level, then send a prompt message to the terminal device used by the first caller.

[0048] This notification is intended to remind the first party in the call to use appropriate language.

[0049] For example, this embodiment determines the response level associated with the target preset word. If the level is the preset first response level (e.g., "warning level"), it indicates that the current language is non-compliant but has not yet reached the point where the call must be terminated immediately. This embodiment will send a prompt message to the terminal device currently used by the first caller (such as the softphone interface of the agent's computer, headset, or screen pop-up system) through the communication link. This reminds the first caller that their current language may not be compliant, prompting them to adjust their communication strategy before the call ends, thereby achieving the effect of timely correction and preventing the situation from escalating.

[0050] This embodiment provides early warning intervention for non-compliance at the low response level, guiding the first party in the call to self-correct. This ensures compliance while maximizing call continuity.

[0051] S105. If the response level corresponding to the target preset word is the preset second response level, then disconnect the communication link.

[0052] For example, if the response level corresponding to the target preset word is a preset second response level (e.g., "blocking level"), it indicates that the detected terminology constitutes serious non-compliance and must be stopped immediately. In this case, this embodiment will trigger a control command to actively disconnect the communication link between the first and second parties. This operation can be achieved by calling the control interface of the communication server to forcibly hang up the current call connection.

[0053] This embodiment enforces a forced block for high-response-level non-compliance calls, immediately terminating calls with serious issues. This effectively prevents the further spread of non-compliant information.

[0054] In summary, this embodiment provides a tiered intervention-based automated voice detection mechanism. By acquiring and parsing voice data during a call, converting it into text, and performing preset word matching, differentiated automated processing is executed based on the preset response level associated with the matching results: for lower-level non-compliant words, a "warning-self-correction" mechanism is implemented by sending a prompt message; for higher-level non-compliant words, "timely blocking" is achieved by forcibly severing the link. The entire method runs concurrently with the call, transforming the traditional manual, delayed, and sampling-based quality inspection model into an automated, full-call coverage model that can complete analysis and intervention within the call cycle. This significantly enhances the timeliness and accuracy of voice detection and control while improving detection coverage and processing efficiency, ensuring the compliant and secure operation of voice interaction services.

[0055] In one feasible implementation, S102 may include the following steps: The speech data is segmented into frames to obtain the corresponding audio frame sequence, and acoustic features are extracted from the audio frame sequence to obtain the acoustic feature sequence.

[0056] The acoustic feature sequence is input into a preset speech recognition model for decoding to obtain the corresponding text unit sequence; the text unit sequence is then combined into target text data.

[0057] For example, in this embodiment, the first speech data is first processed by frame segmentation. For instance, a window function of fixed duration (e.g., 20-30 milliseconds) is used to slide at a certain step size to divide the continuous speech signal into a series of short-term stationary audio frames, forming an audio frame sequence. Then, acoustic features, such as Mel-frequency cepstral coefficients and filter bank energy, are extracted from each audio frame to form an acoustic feature sequence characterizing the speech signal. This feature sequence is then input into a pre-trained speech recognition model. The model decodes the feature sequence and maps it into a series of text units (e.g., phonemes, characters, or words), forming a text unit sequence. Finally, in this embodiment, the text unit sequence is combined, corrected, and normalized to generate the final coherent and readable target text data.

[0058] This implementation segments speech data into frames, extracts acoustic features, and inputs them into a speech recognition model for decoding into text units, ultimately combining them into target text data. This method achieves efficient and accurate conversion from speech to text, providing reliable input for subsequent text analysis. Frame segmentation and acoustic feature extraction effectively capture the temporal and spectral characteristics of the speech signal, improving recognition robustness. The structured decoding of the speech recognition model ensures the continuity and semantic integrity of the text output, laying a high-quality data foundation for subsequent compliance testing.

[0059] In one feasible implementation, this embodiment may further include: After detecting a communication link disconnection, complete voice recording data of the first and second parties in the call is acquired; the complete voice recording data includes at least one first voice segment and at least one second voice segment; the first voice segment corresponds to the first party in the call, and the second voice segment corresponds to the second party in the call.

[0060] Speech recognition is performed on the complete voice recording data to obtain a complete text sequence; wherein the complete text sequence includes at least one first text segment and at least one second text segment, the first text segment corresponds to the first caller, the second text segment corresponds to the second caller, and each text segment is timestamped.

[0061] For each first text segment, the start and end times of the first text segment are determined based on the timestamp corresponding to the first text segment. Based on the start and end times, a target time interval is determined, and at least one second text segment within the target time interval is taken as the associated text segment of the first text segment. The difference between the start time and the preset time value is the left value of the target time interval, and the sum of the end time and the preset time value is the right value of the target time interval. The left value and the right value constitute the target time interval.

[0062] Extract each first text segment corresponding to the first speaker in the complete text sequence, as well as the associated text segments corresponding to the first text segments, to obtain at least one text sequence to be analyzed.

[0063] Each text sequence to be analyzed is input into a preset natural language processing model to obtain the compliance judgment result output by the natural language processing model; the compliance judgment result is used to characterize whether the language used by the first party in the call is compliant.

[0064] For example, in this embodiment, after each call link ends, the complete voice recording data of each call link is obtained from the recording server.

[0065] Taking the processing of complete voice recording data from a call link as an example, the complete voice recording data is separated into multiple first voice segments for the first caller and multiple second voice segments for the second caller, based on the speaker. All voice segments are identified to generate a complete text sequence with timestamps, where each text segment is marked with its owner (first / second caller) and its start and end times.

[0066] For each first text segment spoken by the first caller, this embodiment uses the time of the first text segment as a benchmark and constructs a target time interval (△T1 and △T2 are preset values, such as 2 seconds) through forward expansion (left value of the interval = start time - △T1) and backward expansion (right value of the interval = end time + △T2). All second text segments appearing within this time interval are identified as associated text segments of the first text segment. Subsequently, this embodiment combines each "first text segment" with its corresponding "associated text segment" to form a text sequence to be analyzed. Finally, all text sequences to be analyzed are input into a natural language processing model, which outputs the compliance determination result for each segment spoken by the first caller.

[0067] The compliance assessment results can indicate whether the first party in the call used non-compliant language, etc.

[0068] This embodiment provides a mechanism for conducting a full and in-depth compliance review after a call ends. It can uncover complex, context-dependent non-compliance situations that may be missed during keyword detection (such as leading questions, out-of-context answers, etc.), greatly improving the comprehensiveness, accuracy, and depth of the analysis, and providing a reliable basis for refined management and source tracing.

[0069] The natural language processing model is used to perform the following steps: The first text segment and the associated text segments in the text sequence to be analyzed are converted into corresponding semantic vectors, whereby the semantic vectors are used to represent the text semantic features of the corresponding text segments.

[0070] Calculate the correlation weight between the semantic vector corresponding to the first text segment and the semantic vector of each second text segment in the associated text segment. The correlation weight represents the degree of influence of the semantic features of each second text segment on the compliance of the first text segment.

[0071] The semantic vectors of each second text segment in the associated text segment are weighted and summed according to the relevance weight to obtain the context semantic vector, where the context semantic vector represents the overall influence of the semantic features of the associated text segment on the compliance of the first text segment.

[0072] The semantic vector of the first text segment is concatenated with the semantic vector of the context to obtain a comprehensive semantic vector. Based on the comprehensive semantic vector, the probability value of the first text segment belonging to each preset non-compliant term category is determined. If the probability value of the first text segment under any non-compliant term category is greater than the preset threshold, the first text segment is determined to be a non-compliant term.

[0073] For example, the natural language processing model in this embodiment (such as a Transformer-based model) first converts the first text segment and each associated second text segment into high-dimensional semantic vectors through an encoding layer. Then, the model can use an attention mechanism to calculate the correlation weight between the semantic vector of the first text segment and the semantic vector of each second text segment, where the weight value reflects the importance of the second-party utterance to judging the compliance of the first-party current utterance.

[0074] Next, the model uses these weights to perform a weighted summation of the semantic vectors of all second text segments, generating a context semantic vector that condenses the context information.

[0075] Finally, the model concatenates the semantic vector of the first text segment itself with the semantic vector of the context to form a comprehensive semantic vector, which is then input into the classification layer to calculate the probability value of the first text segment belonging to each preset non-compliant category (such as "insult", "threat", "fraud"). If any probability value exceeds a preset threshold (such as 0.8), the segment is determined to be non-compliant language.

[0076] This embodiment utilizes a natural language processing model to simulate how humans understand dialogue, quantifying the impact of context on the compliance of the first speaker's utterance. By calculating relevance weights and generating contextual semantic vectors, the model can focus on the most relevant dialogue context. This significantly improves the model's ability to identify complex, obscure, or context-dependent non-compliant language, reducing the false positive rate. This makes compliance determination more intelligent and closer to human judgment, enhancing the reliability of the entire system in handling complex scenarios.

[0077] In one feasible implementation, this embodiment may further include: Based on the response level corresponding to the target preset word matched by the target text data, add corresponding tags to the first speech data. Based on the first speech data with added tags and the target preset word corresponding to the first speech data, generate a second review task, send the second review task to the review system, and receive the review results of the review system for the second review task.

[0078] And / or, for the first text segment that is determined to be non-compliant, extract the corresponding target speech data from the complete speech recording data according to the timestamp corresponding to the first text segment; generate a third review task from the speech data, the first text segment and its corresponding non-compliant speech category, and send the third review task to the review system to receive the review results of the review system for the third review task.

[0079] For example, for a target preset word that triggers real-time handling (prompt or cut-off) during a call, this embodiment adds a tag to the corresponding first voice data segment and records the preset word it matches and the response level.

[0080] Based on this tagged voice data and specific preset word information, a review task is generated and pushed to the pending queue of the manual review system so that the quality inspector can make a final confirmation of this real-time alarm.

[0081] For the first text segment deemed non-compliant during the full analysis after the call ends, this embodiment will locate and extract the corresponding original target speech data from the complete recording based on the timestamp of the first text segment. Then, combining this speech segment, the non-compliant text content, and the category of non-compliant language it was determined to be, a review task will be generated and pushed to the review system, whereby a human will review the machine's judgment result.

[0082] This embodiment automatically transforms detected suspicious or non-compliant events into structured review tasks, attaching raw data (voice and text) and judgment criteria. It achieves a collaborative workflow of "machine inspection and human review," leveraging the full and efficient inspection advantages of machines while ensuring the authority and accuracy of the final results through human review. Simultaneously, it provides valuable labeled feedback data for model optimization, forming a closed-loop business process.

[0083] In one feasible implementation, this embodiment further includes: Based on each first text segment, extract the text sentiment features corresponding to the first caller, and based on each first speech segment, extract the acoustic sentiment features corresponding to the first caller.

[0084] By fusing each text sentiment feature with each acoustic sentiment feature, we obtain each multimodal sentiment feature corresponding to the first caller.

[0085] Based on each multimodal emotional feature, the emotional tendency of the first caller during this call is determined; emotional tendency includes positive tendency, neutral tendency, and negative tendency.

[0086] If the emotional tendency of the first party in the call is negative, the complete voice recording data is marked according to the negative tendency, and the first review task is generated based on the marked complete voice recording data.

[0087] Send the first review task to the review system and receive the review results from the review system for the first review task.

[0088] For example, in this embodiment, for each first text segment after the call ends, its text sentiment features (such as sentiment polarity values ​​calculated by word vectors and sentiment dictionaries) are extracted using a sentiment analysis model. At the same time, for each corresponding first speech segment, acoustic sentiment features (such as acoustic parameters reflecting emotion, such as fundamental frequency, energy, and speech rate) are extracted from its acoustic features.

[0089] Next, this embodiment fuses textual sentiment features and acoustic sentiment features within the same time period, such as feature splicing or fusion based on attention mechanisms, to obtain multimodal sentiment features that more comprehensively reflect the emotions of the first caller at that time.

[0090] Based on multimodal emotional characteristics across all time periods, this embodiment can comprehensively determine the overall emotional tendency (negative, neutral, or positive) of the first caller throughout the entire call. If the overall or key emotional tendency is determined to be negative (such as anger or irritability), this embodiment will mark the entire complete voice recording data as "emotionally negative" and generate a review task to request human intervention to address potential service attitude or emotional control issues in this call.

[0091] This embodiment introduces sentiment analysis into the voice compliance inspection system. It comprehensively analyzes the emotional state of the first caller from both text content and voice signal modalities. Multimodal fusion can more accurately capture genuine emotions. It can identify calls that, although not triggering keywords, are accompanied by strong negative emotions. By generating specific review tasks, it reminds managers to pay attention to the first caller's emotional management and service quality, expanding the dimensions and depth of quality inspection.

[0092] In one example, the speech detection method utilizes a trained intelligent model; this model includes a speech recognition model and a natural language processing model; the speech recognition model is used to convert the speech data of the first speaker into target text data. The aforementioned speech detection method also includes: Before acquiring the voice data of the first and second parties in the communication link during the call, the initial intelligent model is trained using the training data in the training database to obtain the trained intelligent model. The training database includes multiple historical call voice data and the corresponding text data, compliance judgment results, sentiment tendencies and review results for each historical call voice data.

[0093] After obtaining the compliance assessment result, sentiment tendency, and at least one review result of this call, the complete voice recording data, complete text sequence, compliance assessment result, sentiment tendency, and at least one review result of this call are stored as new training data in the training database for use in optimizing and training the intelligent model.

[0094] For example, this embodiment can utilize a pre-trained artificial intelligence model to perform the above-described speech detection method.

[0095] Artificial intelligence models include speech recognition models and natural language processing models.

[0096] Before deploying the artificial intelligence model, this embodiment uses a training database to train the initial artificial intelligence model. This database contains a large amount of historical call voice data and its manually annotated text data, the final confirmed compliance judgment results, sentiment labels, and review results.

[0097] Furthermore, the AI ​​model is not static but possesses continuous learning capabilities. The complete voice recordings, transcribed text sequences, automatically generated compliance assessments, sentiment analysis results, and manually verified results from each call are all used as valuable new training data. After preprocessing, these are stored in the training database. This new data is periodically used for incremental or optimization training of the existing intelligent model, enabling it to continuously adapt to new language habits, non-compliant patterns, and business scenarios, thus achieving continuous performance evolution.

[0098] This embodiment constructs a self-iteratory, continuously optimizing intelligent model. It continuously optimizes the model by utilizing new data generated during business operations and feedback from manual review. Over time and with business development, the accuracy of speech detection is continuously improved.

[0099] In summary, this embodiment not only enables rapid keyword matching and tiered intervention during calls, but also allows for in-depth contextual compliance analysis and sentiment assessment after the call ends. Furthermore, through a review system and data feedback mechanism, it ensures the accuracy and reliability of the detection results and the continuous optimization of the model. This method systematically addresses the limitations of traditional manual quality inspection in terms of efficiency, coverage, objectivity, and depth, providing assurance for various remote voice interaction services with high requirements for call compliance and service quality.

[0100] For example, banks typically outsource the collection of overdue payments to external collection agencies, resulting in a massive volume of daily collection calls. Human quality inspectors can only sample and review a small number of recordings, leading to a significant number of non-compliant calls going undetected. This not only allows non-compliant behavior to escape regulation but also risks damaging the bank's reputation due to complaints, and the cost of remediation afterward far outweighs the cost of prevention.

[0101] Taking voice detection in the process of bank telephone collection as an example, the first party in the call is the collection officer, and the second party is the customer being collected from. The preset word library can be a pre-set sensitive word library, in which the preset words can be pre-set sensitive words. The response level corresponding to the sensitive words can be divided into high, medium and low levels. The sensitive word library is maintained and updated by business personnel according to business needs.

[0102] In one example, voice detection during a telephone debt collection process can be performed using an artificial intelligence model, which may include: a voice acquisition module, a voice translation module, a quality inspection and analysis module, and a self-learning module.

[0103] The system includes a voice acquisition module for recording collection call voice data; a voice escaping model for automatically escaping voice data into text using ASR speech recognition technology; a quality inspection and analysis module for conducting in-depth analysis of dialogue content using natural language processing (NLP) and sentiment analysis technologies, including non-compliant word detection and monitoring, and sentiment analysis of collection voices; and a self-learning module for optimizing model performance and improving the ability to recognize bank-specific terminology and non-compliant language using machine learning methods.

[0104] Figure 2 This is a flowchart illustrating another speech detection method provided in an embodiment of this application, such as... Figure 2 As shown, during a collection agent's telephone collection process, the call backend sends the call audio to an AI model. The model performs voice detection, and if the collector's speech matches the model's sensitive word library, it is marked as non-compliant. If the collector's language includes a sensitive word from the library, and the response level for that sensitive word is medium or low, this embodiment can display a pop-up notification to remind the collector to be mindful of their language. If the response level for that sensitive word is high, this embodiment can directly disconnect the call between the collector and the customer.

[0105] After each collection call ends, the call backend archives the entire call recording and sends it to an AI model. The model automatically judges whether there is any non-compliant behavior in the entire call recording every day. Through intelligent voice escaping technology, potentially non-compliant voices are pushed to the manual quality inspection system, minimizing the workload of manual quality inspection and improving operational efficiency.

[0106] Figure 3 This is a schematic diagram of the structure of a voice detection device provided in an embodiment of this application. Figure 3 As shown, the voice detection device includes: an acquisition unit 301, a first processing unit 302, a determination unit 303, and a second processing unit 304.

[0107] The acquisition unit 301 is used to acquire the voice data of the first party and the second party in the communication link during the call, and parse the first voice data of the first party.

[0108] The first processing unit 302 is used to convert the first speech data into target text data and detect whether the target text data contains preset words from a preset word library; wherein, the preset word library has multiple preset words, and each preset word corresponds to a different response level.

[0109] The determining unit 303 is used to determine the target preset word as the target preset word if a preset word in the preset word library is detected in the target text data.

[0110] The second processing unit 304 is configured to send a prompt message to the terminal device used by the first caller if the response level corresponding to the target preset word is a preset first response level; the prompt message is used to remind the first caller to pay attention to the standard of language use; if the response level corresponding to the target preset word is a preset second response level, the communication link is cut off.

[0111] In other embodiments, the apparatus further includes a third processing unit for: After detecting a communication link disconnection, complete voice recording data of the first and second parties in the call is acquired; the complete voice recording data includes at least one first voice segment and at least one second voice segment; the first voice segment corresponds to the first party in the call, and the second voice segment corresponds to the second party in the call.

[0112] Speech recognition is performed on the complete voice recording data to obtain a complete text sequence; wherein the complete text sequence includes at least one first text segment and at least one second text segment, the first text segment corresponds to the first caller, the second text segment corresponds to the second caller, and each text segment is timestamped.

[0113] For each first text segment, the start and end times of the first text segment are determined based on the timestamp corresponding to the first text segment. Based on the start and end times, a target time interval is determined, and at least one second text segment within the target time interval is taken as the associated text segment of the first text segment. The difference between the start time and the preset time value is the left value of the target time interval, and the sum of the end time and the preset time value is the right value of the target time interval. The left value and the right value constitute the target time interval.

[0114] Extract each first text segment corresponding to the first speaker in the complete text sequence, as well as the associated text segments corresponding to the first text segments, to obtain at least one text sequence to be analyzed.

[0115] Each text sequence to be analyzed is input into a preset natural language processing model to obtain the compliance judgment result output by the natural language processing model; the compliance judgment result is used to characterize whether the language used by the first party in the call is compliant.

[0116] In other embodiments, the natural language processing model is used to perform the following steps: The first text segment and the associated text segments in the text sequence to be analyzed are converted into corresponding semantic vectors, whereby the semantic vectors are used to represent the text semantic features of the corresponding text segments.

[0117] Calculate the correlation weight between the semantic vector corresponding to the first text segment and the semantic vector of each second text segment in the associated text segment. The correlation weight represents the degree of influence of the semantic features of each second text segment on the compliance of the first text segment.

[0118] The semantic vectors of each second text segment in the associated text segment are weighted and summed according to the relevance weight to obtain the context semantic vector, where the context semantic vector represents the overall influence of the semantic features of the associated text segment on the compliance of the first text segment.

[0119] The semantic vector of the first text segment is concatenated with the semantic vector of the context to obtain a comprehensive semantic vector. Based on the comprehensive semantic vector, the probability value of the first text segment belonging to each preset non-compliant term category is determined. If the probability value of the first text segment under any non-compliant term category is greater than the preset threshold, the first text segment is determined to be a non-compliant term.

[0120] In other embodiments, the apparatus further includes a fourth processing unit for: Based on each first text segment, extract the text sentiment features corresponding to the first caller, and based on each first speech segment, extract the acoustic sentiment features corresponding to the first caller.

[0121] By fusing each text sentiment feature with each acoustic sentiment feature, we obtain each multimodal sentiment feature corresponding to the first caller.

[0122] Based on each multimodal emotional feature, the emotional tendency of the first caller during this call is determined; emotional tendency includes positive tendency, neutral tendency, and negative tendency.

[0123] If the emotional tendency of the first party in the call is negative, the complete voice recording data is marked according to the negative tendency, and the first review task is generated based on the marked complete voice recording data.

[0124] Send the first review task to the review system and receive the review results from the review system for the first review task.

[0125] In other embodiments, the apparatus further includes a verification unit for: Based on the response level corresponding to the target preset word matched by the target text data, add corresponding tags to the first speech data. Based on the first speech data with added tags and the target preset word corresponding to the first speech data, generate a second review task, send the second review task to the review system, and receive the review results of the review system for the second review task.

[0126] And / or, for the first text segment that is determined to be non-compliant, extract the corresponding target speech data from the complete speech recording data according to the timestamp corresponding to the first text segment; generate a third review task from the speech data, the first text segment and its corresponding non-compliant speech category, and send the third review task to the review system to receive the review results of the review system for the third review task.

[0127] In other embodiments, the speech detection method is implemented using a trained intelligent model; the intelligent model includes a speech recognition model and a natural language processing model; the speech recognition model is used to convert the speech data of the first caller into target text data; the apparatus further includes: The training unit is used to: train the initial intelligent model using training data in the training database before acquiring the voice data of the first and second parties in the communication link during the call, so as to obtain the trained intelligent model; wherein, the training database includes multiple historical call voice data and the corresponding text data, compliance judgment results, sentiment tendency and review results for each historical call voice data.

[0128] The storage unit is used to: after obtaining the compliance judgment result, sentiment tendency and at least one review result of this call, store the complete voice recording data, complete text sequence, compliance judgment result, sentiment tendency and at least one review result of this call as new training data in the training database for optimizing training of the intelligent model.

[0129] In other embodiments, the first processing unit 302 is specifically used for: The speech data is segmented into frames to obtain the corresponding audio frame sequence, and acoustic features are extracted from the audio frame sequence to obtain the acoustic feature sequence.

[0130] The acoustic feature sequence is input into a preset speech recognition model for decoding to obtain the corresponding text unit sequence; the text unit sequence is then combined into target text data.

[0131] The speech detection device provided in this application embodiment can execute the method shown in the above method embodiment. Its implementation principle and beneficial effects can be referred to the relevant description in the method embodiment, and will not be repeated here.

[0132] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device includes a memory 401 and at least one processor 402.

[0133] The memory 401 is used to store computer program code, which includes computer instructions. These computer instructions run in the aforementioned electronic device to implement the method shown in the above-described method embodiments. For example, the memory may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk, or optical disc, etc.

[0134] Processor 402 can be a general-purpose processor, including a Central Processing Unit (CPU), a network processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Processor 402 can also be other general-purpose processors. The general-purpose processor can be a microprocessor or any conventional processor.

[0135] The memory 401 and processor 402 are communicatively connected. For example, the memory 401 can be connected to the processor 402 via a system bus and communicate with it. The system bus can be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, an industry standard architecture (ISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the figure, but this does not mean that there is only one bus or one type of bus.

[0136] Optionally, the memory 401 can be either standalone or integrated with the processor 402. When the memory 401 is set up independently, it is connected to the processor 402 via a system bus.

[0137] This application also provides a chip for executing instructions, which is used to execute the technical solution of the speech detection method in the above embodiments.

[0138] This application also provides a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the technical solution of the speech detection method described in the above embodiments. Specifically, when the computer instructions are executed by a processor, the electronic device can perform the technical solution of the speech detection method described in the above embodiments.

[0139] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solution of the speech detection method in the above embodiments.

[0140] The aforementioned computer-readable storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage media can be any available medium accessible to a general-purpose or special-purpose computer.

[0141] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic control unit or main control device; this application does not limit this.

[0142] 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 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 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 indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0143] The modules described as separate components may or may not be physically separate. The components shown as modules 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 modules can be selected to implement the solution of this embodiment according to actual needs.

[0144] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0145] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0146] It should be understood that the steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0147] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A voice detection method characterized by, The method comprises: acquiring voice data of a first call party and a second call party in a communication link during a call, and parsing to obtain first voice data of the first call party; converting the first voice data into target text data, and detecting whether the target text data contains a preset word in a preset word library; wherein the preset word library is preset with a plurality of preset words, and each preset word corresponds to a different response level; if it is detected that the target text data contains a preset word in the preset word library, the detected preset word in the target text data is taken as a target preset word; if the response level corresponding to the target preset word is a preset first response level, a prompt information is sent to a terminal device used by the first call party; the prompt information is used to remind the first call party to pay attention to language norms; if the response level corresponding to the target preset word is a preset second response level, the communication link is cut off.

2. The voice detection method of claim 1, wherein, The method further comprises: after detecting that the communication link is disconnected, acquiring complete voice recording data of the first call party and the second call party; the complete voice recording data comprises at least one first voice segment and at least one second voice segment; the first voice segment corresponds to the first call party, and the second voice segment corresponds to the second call party; performing voice recognition on the complete voice recording data to obtain a complete text sequence; wherein the complete text sequence comprises at least one first text segment and at least one second text segment, the first text segment corresponds to the first call party, the second text segment corresponds to the second call party, and each text segment corresponds to a time stamp; for each first text segment, according to the time stamp corresponding to the first text segment, determining a start time and an end time corresponding to the first text segment and a target time interval, and taking at least one second text segment in the target time interval as an associated text segment of the first text segment; the interval left value of the target time interval is the difference between the start time and a preset time value, and the interval right value of the target time interval is the sum of the end time and the preset time value; extracting each first text segment corresponding to the first call party in the complete text sequence and the associated text segment corresponding to the first text segment to obtain at least one text sequence to be analyzed; inputting each text sequence to be analyzed into a preset natural language processing model to obtain a compliance determination result output by the natural language processing model; wherein the compliance determination result is used to represent whether the language used by the first call party in the call is compliant.

3. The voice detection method of claim 2, wherein, The natural language processing model is used to perform the following steps: convert the first text segment and the associated text segment in the text sequence to be analyzed into corresponding semantic vectors, wherein the semantic vector is used to represent the text semantic features of the corresponding text segment; calculate a correlation weight between the semantic vector corresponding to the first text segment and a semantic vector of each second text segment in the associated text segment, where the correlation weight represents an influence degree of the semantic feature of each second text segment on the compliance of the first text segment; perform weighted summation on the semantic vectors of each second text segment in the associated text segment according to the correlation weight, to obtain a context semantic vector, where the context semantic vector represents a total influence degree of the semantic features of the associated text segment on the compliance of the first text segment; concatenate the semantic vector of the first text segment with the context semantic vector to obtain a comprehensive semantic vector, and determine a probability value of the first text segment belonging to each preset non-compliant language category based on the comprehensive semantic vector; if the probability value of the first text segment under any non-compliant language category is greater than a preset threshold, it is determined that the first text segment is non-compliant language.

4. The voice detection method of claim 2, wherein, The method further comprises: extracting a text sentiment feature corresponding to the first call party based on each first text segment, and extracting an acoustic sentiment feature corresponding to the first call party based on each first voice segment; fusing each text sentiment feature with each acoustic sentiment feature to obtain each multi-modal sentiment feature corresponding to the first call party; determining a sentiment tendency of the first call party in the current call process based on each multi-modal sentiment feature; the sentiment tendency includes positive tendency, neutral tendency and negative tendency; if the sentiment tendency of the first call party in the current call process is negative, marking the complete voice recording data according to the negative tendency, and generating a first review task according to the marked complete voice recording data; sending the first review task to a review system and receiving a review result of the review system for the first review task.

5. The voice detection method of claim 4, wherein, The method further comprises: adding a corresponding label to the first voice data according to the response level corresponding to the target preset word matched by the target text data, generating a second review task according to the first voice data after adding the label and the target preset word corresponding to the first voice data, and sending the second review task to a review system and receiving a review result of the review system for the second review task; and / or, for the first text segment determined as non-compliant language, extracting corresponding target voice data from the complete voice recording data according to the time stamp corresponding to the first text segment; generating a third review task by taking the voice data, the first text segment and the corresponding non-compliant language category thereof, and sending the third review task to a review system to receive a review result of the review system for the third review task.

6. The voice detection method of claim 5, wherein, The voice detection method is realized by using a trained intelligent model; the intelligent model includes a voice recognition model and the natural language processing model; the voice recognition model is used to convert the voice data of the first call party into the target text data; the method further comprises: In the obtaining conversation process, the initial intelligent model is trained by using the training data in the training database before the voice data of the first conversation party and the second conversation party in the communication link, to obtain a trained intelligent model; wherein the training database includes a plurality of historical conversation voice data, and text data, compliance determination results, emotional tendencies and review results corresponding to each historical conversation voice data; After obtaining the compliance determination result, the emotional tendency and at least one review result of the current conversation, the complete voice recording data, the complete text sequence, the compliance determination result, the emotional tendency and at least one review result of the current conversation are stored in the training database as new training data, for optimizing training of the intelligent model.

7. The voice detection method of any of claims 1-6, wherein, The conversion of the voice data of the first conversation party into target text data includes: frame processing is performed on the voice data to obtain a corresponding audio frame sequence, and acoustic features are extracted from the audio frame sequence to obtain an acoustic feature sequence; the acoustic feature sequence is input into a preset speech recognition model for decoding to obtain a corresponding text unit sequence; and the text unit sequence is combined into target text data.

8. A voice detection apparatus characterized by comprising: The device includes: An obtaining unit is configured to obtain voice data of a first conversation party and a second conversation party in a communication link in a conversation process, and analyze the voice data to obtain first voice data of the first conversation party. A first processing unit is configured to convert the first voice data into target text data, and detect whether the target text data contains preset words in a preset word library; wherein the preset word library is preset with a plurality of preset words, and each preset word corresponds to a different response level. A determining unit is configured to, if it is detected that the target text data contains a preset word in the preset word library, take the detected preset word in the target text data as a target preset word. A second processing unit is configured to, if the target preset word corresponds to a preset first response level, send a prompt information to a terminal device used by the first conversation party; the prompt information is used to remind the first conversation party to pay attention to the language; and if the target preset word corresponds to a preset second response level, cut off the communication link.

9. An electronic device, comprising: It includes: a memory and at least one processor; the memory is in communication connection with the processor; the memory is used to store computer program code, the computer program code includes computer instructions; when the processor executes the computer instructions, the electronic device executes the voice detection method in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are executed by the processor to implement the voice detection method in any one of claims 1-7.

11. A computer program product, characterised in that, When the computer program product runs on the computer / is executed by the processor of the computer, the voice detection method in any one of claims 1-7 is implemented.

Images