Call voice quality inspection method, device, equipment and storage medium

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a call voice quality inspection method, device, equipment and storage medium, wherein the method comprises: obtaining call voice and converting the call voice into call text, the call text including customer call text and agent reply text; inputting the call text into a pre-trained text detection model to obtain a target text segment including sensitive content; based on a preset violation detection method, performing a detail check on the target text segment; if the target text segment violates the rule, confirming the violation type corresponding to the target text segment as the target violation type; and generating quality inspection result information based on the target text segment and the target violation type. The present invention preliminarily screens text segments that are likely to violate the rule in calls between agents and customers through a model, and then performs a detail check on the text segments based on a preset detection method to improve the accuracy of the quality inspection results.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a call voice quality inspection method, apparatus, device and storage medium. Background Art

[0002] At present, call centers have become an important way for enterprises to provide comprehensive business information online. The call center agents carry out outbound or inbound telephone services according to the needs of the enterprise to accept customer feedback, consulting suggestions and other services, or conduct market research, telephone sales, after-sales tracking and other services for the company's products. In the process of the call center carrying out the above-mentioned telephone services, it is necessary to ensure the standardization and professionalism of the content of the calls between the agents and the customers in order to ensure the quality and efficiency of the telephone services provided by the agents. Therefore, it is particularly important to inspect and control the content of the calls. As for the quality inspection, the existing technology usually adopts the method of manually listening to the call recordings to conduct quality inspection and evaluation, which requires the call center to be equipped with a corresponding number of quality inspectors. This method has high manpower requirements and a large workload for quality inspection, low efficiency and low coverage, and it is difficult to effectively evaluate the overall service quality. Summary of the Invention

[0003] The present application provides a call voice quality inspection method, apparatus, device and storage medium to solve the problems of high labor cost and low efficiency in existing call voice quality inspection.

[0004] In order to solve the above technical problems, a technical solution adopted in this application is: to provide a call voice quality inspection method, including: obtaining call voice and converting the call voice into call text, the call text includes customer call text and agent reply text; inputting the call text into a pre-trained text detection model to obtain a target text segment including sensitive content; based on a preset violation detection method, performing detail verification on the target text segment; if the target text segment violates the rules, confirming the violation type corresponding to the target text segment as the target violation type; generating quality inspection result information based on the target text segment and the target violation type.

[0005] As a further improvement of the present application, the target text segment is verified in detail based on the preset violation detection method, including: obtaining the preset violation detection method corresponding to a plurality of pre-set violation types; and performing detailed verification on the target text segment using the preset violation detection method corresponding to each violation type.

[0006] As a further improvement of the present application, the text detection model includes a first encoding layer, a second encoding layer and a softmax layer; the call text is input into a pre-trained text detection model to obtain a target text segment including sensitive content, including: randomly initializing the call text based on the normal distribution to obtain a position vector; inputting the call text into the first encoding layer for encoding to obtain a sentence vector; splicing the sentence vector with the position vector to obtain a splicing vector; using the second encoding layer to encode the splicing vector to obtain a comprehensive vector; inputting the comprehensive vector into the softmax layer for prediction to obtain a target text segment.

[0007] As a further improvement of the present application, call voice is obtained and converted into call text, and the call text includes customer call text and agent reply text, including: obtaining call voice; identifying the voiceprint features of each voice segment in the call voice to obtain customer voice segments and agent voice segments; converting the customer voice segments into customer call text, and converting the agent voice segments into agent reply text.

[0008] As a further improvement of the present application, it is characterized in that the violation type includes information not being informed type violation; based on a preset violation detection method, the target text segment is verified in detail, including: detecting whether the target text segment meets the preset characteristics; if it meets the preset characteristics, detecting whether there is text content related to the preset characteristics in the target text segment; if there is no relevant text content, the target text segment triggers the information not being informed type violation.

[0009] As a further improvement of the present application, the violation types include incomplete information notification or error type violations; based on a preset violation detection method, the target text segment is verified in detail, including: obtaining the previous text segment and the next text segment of the target text segment, and forming the previous text segment and the target text segment into a first text segment, forming the target text segment and the next text segment into a second text segment, and forming the previous text segment, the target text segment, and the next text segment into a third text segment; vectorizing the target text segment, the first text segment, the second text segment, and the third text segment respectively to obtain a target vector, a first vector, a second vector, and a third vector; inputting the target vector, the first vector, the second vector, and the third vector into the BERT model for encoding, respectively, and matching them one by one with the preset violation vectors in the preset violation vector library to obtain the matching degrees corresponding to the target vector, the first vector, the second vector, and the third vector respectively; selecting the final text content corresponding to the vector with the highest matching degree, and outputting the final text content triggering incomplete information notification or error type violations.

[0010] As a further improvement of the present application, the violation types include violations of the type of notification of out-of-scope information; based on a preset violation detection method, the target text fragment is verified in detail, including: using preset rules to match the target text fragments respectively; when the match is successful, the target text fragment triggers a violation of the type of notification of out-of-scope information.

[0011] In order to solve the above technical problems, another technical solution adopted in this application is: to provide a call voice quality inspection device, including: an acquisition module, used to acquire call voice and convert the call voice into call text, the call text includes customer call text and agent reply text; a screening module, used to input the call text into a pre-trained text detection model to obtain a target text segment including sensitive content; a verification module, used to perform detail verification on the target text segment based on a preset violation detection method; a confirmation module, used to confirm the violation type corresponding to the target text segment as the target violation type when the target text segment violates the law; an output module, used to generate quality inspection result information based on the target text segment and the target violation type.

[0012] In order to solve the above technical problems, another technical solution adopted in this application is: to provide a computer device, which includes a processor and a memory coupled to the processor, wherein program instructions are stored in the memory, and when the program instructions are executed by the processor, the processor executes the steps of the call voice quality inspection method as described in any one of the above items.

[0013] In order to solve the above technical problems, another technical solution adopted in this application is: providing a storage medium storing program instructions that can implement the above call voice quality inspection method.

[0014] The beneficial effect of the present application is that the call voice quality inspection method of the present application converts the voice of the conversation between the agent and the customer into call text, and then inputs the call text into a pre-trained text detection model, thereby screening out target text segments including sensitive content from the call text. It is not limited to the content of the agent's reply, and combines the customer's call content for comprehensive screening, so that the screened target text segments including sensitive content are more accurate, and then the target text segments are detected for violations based on a preset detection method to further confirm whether there is actually any illegal content in the target text segment, and it can also identify the violation type of the target text segment with illegal content, so that the final output call voice quality inspection result is more intuitive, helping quality inspectors to quickly confirm the quality and effect of the agent's call voice reply. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 1 is a flow chart of a method for quality inspection of call speech according to an embodiment of the present invention;

[0016] Figure 2 Schematic diagram of the structure of a text detection model according to an embodiment of the present invention;

[0017] Figure 3 yes Figure 1 A specific flow chart of step S102;

[0018] Figure 4 yes Figure 1 A specific flow chart of step S103;

[0019] Figure 5 yes Figure 1 Another specific flow chart of step S103

[0020] Figure 6 yes Figure 1 Another specific flow chart of step S103

[0021] Figure 7 1 is a schematic diagram of the functional modules of a call voice quality inspection device according to an embodiment of the present invention;

[0022] Figure 8 is a schematic structural diagram of a computer device according to an embodiment of the present invention;

[0023] Figure 9 It is a schematic structural diagram of a storage medium according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] The terms "first," "second," and "third" in this application are used only for descriptive purposes and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of such features. In the description of this application, "multiple" means at least two, for example, two, three, etc., unless otherwise specifically defined. All directional indications in the embodiments of this application (such as up, down, left, right, front, back...) are only used to explain the relative positional relationship, movement, etc. between the components under a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications also change accordingly. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products, or devices.

[0026] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0027] Figure 1 It is a flow chart of the method for quality inspection of call voice according to an embodiment of the present invention. It should be noted that if there is substantially the same result, the method of the present invention is not limited to the method of Figure 1 The process sequence shown is limited. Figure 1 As shown, the method includes the steps of:

[0028] Step S101: Acquire call audio and convert the call audio into call text, where the call text includes the customer call text and the agent reply text.

[0029] Specifically, when a call center agent connects to or makes an outgoing call, once a call connection is established between the agent and the customer, the two typically engage in voice communication around a specific topic, typically including product sales, after-sales support, product research, product consultation, user troubleshooting, complaints, and suggestions. Step S101 begins recording the conversation between the agent and the customer once the call connection is successfully established, thereby obtaining the conversation audio between the agent and the customer, and then converting the conversation audio into text. This speech-to-text conversion is achieved using an Automatic Speech Recognition (ASR) model.

[0030] Furthermore, after acquiring the call voice, it is necessary to identify the customer voice and the agent's reply voice in the call voice. Therefore, step S101 specifically includes:

[0031] 1. Get the call audio.

[0032] 2. Identify the voiceprint features of each voice segment in the call to obtain customer voice segments and agent voice segments.

[0033] 3. Convert customer voice clips into customer call text, and convert agent voice clips into agent reply text.

[0034] When using a speech recognition model for speech recognition, the voiceprint features in the call speech must first be identified. Typically, a voice call is between two people and contains two voiceprint features. By dividing the speech segments with different voiceprint features, the customer's speech segments and the agent's speech segments are separated. Each speech segment can also be labeled by its position in the speech. The speech segments are divided according to the alternating time points when the customer and the agent communicate. The customer speech segments and the agent speech segments with position labels are then input into the speech recognition model to obtain customer text segments and agent text segments. The customer text segments and the agent text segments are then concatenated in sequence according to the position labels to obtain the complete call text. The customer text segments in the call text are labeled with the customer label, and the agent text segments are labeled with the agent label.

[0035] It should be noted that the collected call voice may include sounds other than human voices made by agents and customers, such as sounds in the environment where the terminal device is located. At this time, in order to avoid noise interference from sounds other than human voices in the call voice, the call voice can be first subjected to voice enhancement processing to eliminate the noise in the call voice. Among them, when performing voice enhancement on the call voice, different voice enhancement algorithms can be selected according to needs to eliminate noise and retain a clearer human voice. It should be noted that since human voices are usually translated during the translation process, all sounds other than human voices can be defined as noise. When performing voice enhancement, all noise in the call voice can be eliminated.

[0036] Step S102: Input the call text into a pre-trained text detection model to obtain a target text segment including sensitive content.

[0037] It should be noted that the text detection model is trained based on pre-acquired samples, and the samples include texts with sensitive content and texts without sensitive content, and the sensitive content is pre-set.

[0038] Specifically, after obtaining the call text, the call text is input into the text detection model to detect the content of each text segment in the text, thereby filtering out target text segments containing sensitive content from all text segments. Sensitive content refers to content that is likely to lead to violations. For example, when an agent informs a customer about the return of a financial product in a text segment, this text segment contains sensitive content. If the return information is in accordance with the expected return of the product, the information is considered correct. If the return information significantly exceeds the expected return of the product, the agent is considered to have exaggerated the product return and committed a violation.

[0039] Further, in this embodiment, please refer to Figure 2 , the text detection model includes a first encoding layer, a second encoding layer and a softmax layer. Among them, the first encoding layer and the second encoding layer are both implemented based on the Encoder layer of the Transformer model. The first encoding layer is used to perform sentence encoding on the call text to obtain a sentence vector, and the second encoding layer is used to fully encode the splicing vector composed of the sentence vector and the position vector obtained by initializing the call text. It should be understood that the purpose of including two encoding layers in this embodiment is to obtain the sentence vector of the call text through the first encoding layer, splice each sentence vector with the position vector generated according to the call text, and then fully encode the spliced ​​vector, so that the text detection model combines the sentence vector and the position vector for comprehensive and comprehensive judgment, thereby improving the text detection model's ability to recognize target text segments containing sensitive content.

[0040] Specifically, see Figure 3 In step S102, the call text is input into a pre-trained text detection model to obtain a target text segment containing sensitive content, specifically including:

[0041] Step S201: Randomly initialize the call text based on normal distribution to obtain a position vector.

[0042] Step S202: Input the call text into the first encoding layer for encoding to obtain a sentence vector.

[0043] Step S203: concatenate the sentence vector and the position vector to obtain a concatenated vector.

[0044] Step S204: Encode the concatenated vector using the second coding layer to obtain a full vector.

[0045] Step S205: Input the comprehensive vector into the softmax layer for prediction to obtain the target text segment.

[0046] Furthermore, the training process of the text detection model includes:

[0047] 1. Obtain training samples and randomly initialize them based on normal distribution to obtain sample position vectors. Texts with sensitive content in the training samples are labeled with sensitive labels, and texts without sensitive content are labeled with normal labels.

[0048] 2. Input the training sample into the first encoding layer of the text detection model for encoding to obtain a sample sentence vector;

[0049] 3. Concatenate the sample sentence vector and the sample position vector to obtain the sample concatenation vector.

[0050] 4. Use the second encoding layer of the text detection model to encode the sample splicing vector to obtain the sample comprehensive vector.

[0051] 5. Input the sample full vector into the softmax layer of the text detection model for prediction to obtain the sample prediction result.

[0052] 6. Reversely update the parameters of the text detection model based on the sample prediction results, the labels of the training samples, and the preset loss function until the accuracy of the text detection model reaches the preset requirements or the number of training times reaches the preset requirements.

[0053] Step S103: Based on a preset violation detection method, perform detail verification on the target text segment.

[0054] In step S103, after obtaining the target text segment, the target text segment is checked for violations based on a pre-defined violation detection method to determine whether the target text segment contains any violations. This violation detection method is designed based on the types of violations found in historical call records between customers and agents, and performs fine-grained detection on the text segment to determine whether the text segment contains any violations.

[0055] It should be understood that in this embodiment, the violation type can be obtained based on the violation type appearing in the historical call records between the customer and the agent. Therefore, there is more than one violation type. In order to reasonably detect each violation type, the detection method for each violation type is separately designed according to the appearance method of the violation type, thereby obtaining a preset violation detection method corresponding to each violation type.

[0056] Therefore, in some embodiments, step S103 specifically includes:

[0057] 1. Obtain preset violation detection methods corresponding to multiple pre-set violation types.

[0058] 2. Perform detailed verification on the target text segment using the preset violation detection method corresponding to each violation type.

[0059] It should be understood that the target text segment may contain content belonging to different violation types. Therefore, after obtaining the target text segment, each target text segment is verified according to each preset violation detection method to confirm whether the target text segment violates the violation. Step S104: If the target text segment violates the violation, the violation type corresponding to the target text segment is confirmed as the target violation type.

[0060] Specifically, these violation types include failure to disclose information, incomplete or incorrect information disclosure, and inappropriate information disclosure. A failure to disclose information violation occurs when a call involves content in a specific section, for which the agent is required to provide the customer with relevant information. If the agent fails to provide this information promptly, the agent's response is considered to have triggered a failure to disclose information violation. For example, when discussing an insurance product with a customer, the agent is required to provide the customer with information such as the product's name, content, and premium. If this information is not detected in the relevant text segment of the call, the relevant text segment is considered to have committed a failure to disclose information violation. Incomplete or incorrect information disclosure violations refer to when a call involves the content of a certain section, the agent must inform the customer of the relevant information. When the relevant information provided by the agent is incomplete or incorrect, the agent's response is considered to have triggered an incomplete or incorrect information disclosure violation. For example, when communicating with a customer about a certain insurance product, if the agent only informs the customer of the name and content of the insurance product but does not inform the customer of the premium information, the agent's response is considered to have incomplete information disclosure. For example, when the agent replies that the income of the insurance product exceeds the expected income, the agent's response is considered to have incorrect information disclosure. Violations of information disclosure beyond the scope refer to when the agent uses prohibited words or sentences when communicating with the customer, or violates the rules and provides incorrect advice to the customer. The agent's response is considered to have triggered an incomplete information disclosure violation. For example, when applying for insurance, the agent prompts the customer to falsify their own information.

[0061] For further information, see Figure 4 For the violation of not informing the user, in step S103, the target text segment is verified in detail based on the preset violation detection method, specifically including:

[0062] Step S301: Detect whether the target text segment meets the preset characteristics. If it meets the preset characteristics, execute step S302; if it does not meet the preset characteristics, perform other violation types detection.

[0063] Specifically, the preset feature is pre-set according to actual conditions. For example, when conducting product sales over the phone, when a customer inquires about the product's usage restrictions or contraindications, the agent needs to inform the customer of the product's applicable population, precautions, usage scenarios, etc., and the "product's usage restrictions or contraindications" are the preset features.

[0064] Step S302: Detect whether there is text content related to the preset feature in the target text segment. If there is no relevant text content, execute step S303.

[0065] Specifically, when it is confirmed that the target text segment meets the preset feature, the preset text content corresponding to the preset feature is obtained, and the preset text content is matched with the text content in the target text segment to confirm whether there is text content related to the preset feature in the target text segment.

[0066] Step S303: The target text segment trigger information does not inform the class violation.

[0067] Specifically, when the target text segment does not contain text content related to the preset feature, it is considered that the target text segment trigger information does not notify the class violation.

[0068] For further information, see Figure 5 For incomplete information disclosure or error violations, in step S103, the target text segment is verified in detail based on a preset violation detection method, specifically including:

[0069] Step S401: Obtain the previous text segment and the next text segment of the target text segment, and combine the previous text segment and the target text segment to obtain a first text segment, combine the target text segment and the next text segment to obtain a second text segment, and combine the previous text segment, the target text segment, and the next text segment to obtain a third text segment.

[0070] Specifically, in this embodiment, a sliding window of size 3 is used to perform context association on the target text segment, thereby obtaining a previous text segment and a next text segment associated with the target text.

[0071] Step S402: vectorize the target text segment, the first text segment, the second text segment, and the third text segment respectively to obtain a target vector, a first vector, a second vector, and a third vector.

[0072] Specifically, the target text segment, the first text segment, the second text segment, and the third text segment are respectively vectorized to obtain a target vector corresponding to the target text segment, a first vector corresponding to the first text segment, a second vector corresponding to the second text segment, and a third vector corresponding to the third text segment. Text vectorization is the process of representing text as a series of vectors that can express the semantics of the text. In this embodiment, text vectorization is implemented based on the word2vec model.

[0073] Step S403: After the target vector, the first vector, the second vector, and the third vector are respectively input into the BERT model for encoding, they are matched one by one with the preset violation vectors in the preset violation vector library to obtain the matching degrees corresponding to the target vector, the first vector, the second vector, and the third vector respectively.

[0074] Specifically, the target vector, the first vector, the second vector, and the third vector are respectively input into the BERT model for encoding. The encoded results are then matched one by one with the preset violation vectors in the preset violation vector library to obtain the target matching degree corresponding to the target vector, the first matching degree corresponding to the first vector, the second matching degree corresponding to the second vector, and the third matching degree corresponding to the third vector. The preset violation vector library is pre-established.

[0075] Step S404: selecting the final text content corresponding to the vector that exceeds the preset matching degree threshold and has the highest matching degree, and outputting the final text content trigger information to notify the incomplete or error type violation.

[0076] Specifically, determine whether the target matching degree, the first matching degree, the second matching degree, and the third matching degree exceed the preset matching degree threshold. If there is a matching degree that exceeds the preset matching degree threshold, the vector corresponding to the matching degree is confirmed, and then the final text content corresponding to the vector is confirmed, and finally the final text content trigger information notification is output to notify incomplete or error-type violations. If there are two or more matching degrees that exceed the preset matching degree threshold, the final text content corresponding to the vector with the highest matching degree is selected, and the final text content trigger information notification is output to notify incomplete or error-type violations. If there is no matching degree that exceeds the preset matching degree threshold, then the target text segment, the first text segment, the second text segment, and the third text segment do not trigger information notification of incomplete or error-type violations.

[0077] In this embodiment, a sliding window strategy is adopted to integrate contextual feature information for violation point detection, thereby fully utilizing the contextual information between the agent and the customer, thereby improving the accuracy of the violation detection result.

[0078] For further information, see Figure 6 For violations of the "out-of-scope information notification" category, in step S103, based on a preset violation detection method, the target text segment is subjected to detailed verification, specifically including:

[0079] Step S501: Matching with target text segments respectively using preset rules.

[0080] Specifically, the preset rules include at least one of regular expression matching and keyword matching. When using regular expression matching, a regular template is pre-built for illegal content that violates the rubric and provides information. After obtaining the target text segment, the target text segment is converted into a regular expression and then matched against the regular template. When using keyword matching, keywords that represent illegal content are pre-specified, such as "price fraud" and "false identity," and then the words in the target text segment are matched against the keywords.

[0081] Step S502: When the match is successful, the target text segment triggers out-of-scope information to notify the class violation.

[0082] Step S105: Generate quality inspection result information based on the target text segment and the target violation type.

[0083] Specifically, after obtaining the target text segment with violations and the target violation type, quality inspection result information is generated based on the target text segment and the target violation type. This quality inspection result information includes the content of the target text segment and the target violation type. To facilitate viewing of the quality inspection result, this quality inspection result information can be highlighted in the call text. For example, the target text segment with violations is displayed in a color that is different from other text segments, and an annotation is added to the target text segment, indicating the target violation type.

[0084] Furthermore, in order to further evaluate the call quality of the agent, in this embodiment, after obtaining the target text segment of the violation and the violation type, the following steps are further included:

[0085] Get the deduction value corresponding to each violation type;

[0086] The number of occurrences of each violation type in the target text segment where the violations were identified;

[0087] The total deduction value is calculated based on the number of occurrences and deduction value of each violation type, and the total deduction value is subtracted from the preset full score to obtain the score record of the current seat for this call;

[0088] Count all the rating records of the current seat within a period of time and take the average to obtain the average rating;

[0089] The call quality of the current agent is determined based on the average score. Specifically, the score range is divided into different levels, and each level corresponds to a specific level. The level to which the average score belongs can be used to determine the call quality of the current agent.

[0090] The call voice quality inspection method of the embodiment of the present invention converts the voice of the conversation between the agent and the customer into call text, and then inputs the call text into a pre-trained text detection model, thereby screening out target text segments including sensitive content from the call text. It is not limited to the content of the agent's reply, and combines the customer's call content for comprehensive screening, so that the screened target text segments including sensitive content are more accurate. Then, based on a preset detection method, the target text segment is detected for violations to further confirm whether there is actually any illegal content in the target text segment, and it can also identify the violation type of the target text segment with illegal content, so that the final output call voice quality inspection result is more intuitive, helping quality inspectors to quickly confirm the quality and effect of the agent's call voice reply.

[0091] Figure 7Schematic diagram of the functional modules of the call voice quality inspection device according to an embodiment of the present invention. Figure 7 As shown, the call voice quality inspection device 60 includes an acquisition module 61, a screening module 62, a verification module 63, a confirmation module 64 and an output module 65.

[0092] An acquisition module 61 is used to acquire the call voice and convert the call voice into call text, which includes the customer call text and the agent reply text;

[0093] A screening module 62 is configured to input the call text into a pre-trained text detection model to obtain a target text segment containing sensitive content;

[0094] Verification module 63, used to perform detail verification on the target text segment based on a preset violation detection method;

[0095] A confirmation module 64 is configured to confirm the violation type corresponding to the target text segment as a target violation type when the target text segment violates the rule;

[0096] The output module 65 is used to generate quality inspection result information according to the target text segment and the target violation type.

[0097] Optionally, the inspection module 63 performs detailed verification of the target text segment based on the preset violation detection method, specifically including: obtaining the preset violation detection method corresponding to multiple pre-set violation types; and performing detailed verification of the target text segment using the preset violation detection method corresponding to each violation type.

[0098] Optionally, the text detection model includes a first encoding layer, a second encoding layer and a softmax layer; the screening module 62 executes the operation of inputting the call text into a pre-trained text detection model to obtain a target text segment including sensitive content, specifically including: randomly initializing the call text based on normal distribution to obtain a position vector; inputting the call text into the first encoding layer for encoding to obtain a sentence vector; splicing the sentence vector with the position vector to obtain a splicing vector; using the second encoding layer to encode the splicing vector to obtain a comprehensive vector; inputting the comprehensive vector into the softmax layer for prediction to obtain a target text segment.

[0099] Optionally, the acquisition module 61 executes the operations of acquiring call voice and converting the call voice into call text, where the call text includes customer call text and agent reply text. Specifically, the operations include: acquiring call voice; performing voiceprint feature recognition on each voice segment in the call voice to obtain customer voice segments and agent voice segments; converting the customer voice segments into customer call text, and converting the agent voice segments into agent reply text.

[0100] Optionally, the violation type includes an information not-informed violation; the inspection module 63 performs detailed verification of the target text segment based on a preset violation detection method, specifically including: detecting whether the target text segment meets the preset characteristics; if it meets the preset characteristics, detecting whether there is text content related to the preset characteristics in the target text segment; if there is no relevant text content, the target text segment triggers an information not-informed violation.

[0101] Optionally, the violation type includes incomplete information notification or error type violation; the verification module 63 performs detailed verification of the target text segment based on the preset violation detection method, specifically including: obtaining the previous text segment and the next text segment of the target text segment, and combining the previous text segment and the target text segment into a first text segment, combining the target text segment and the next text segment into a second text segment, and combining the previous text segment, the target text segment, and the next text segment into a third text segment; vectorizing the target text segment, the first text segment, the second text segment, and the third text segment respectively to obtain a target vector, a first vector, a second vector, and a third vector; inputting the target vector, the first vector, the second vector, and the third vector into the BERT model for encoding, and matching them one by one with the preset violation vectors in the preset violation vector library to obtain the matching degrees corresponding to the target vector, the first vector, the second vector, and the third vector respectively; selecting the final text content corresponding to the vector with the highest matching degree, and outputting the final text content triggering incomplete information notification or error type violation.

[0102] Optionally, the violation type includes an out-of-scope information notification violation; the inspection module 63 performs detailed verification of the target text segment based on a preset violation detection method, specifically including: matching the target text segment separately using preset rules; when the match is successful, the target text segment triggers an out-of-scope information notification violation.

[0103] For other details about the technical solutions for implementing each module in the call voice quality inspection device in the above embodiment, please refer to the description of the call voice quality inspection method in the above embodiment, which will not be repeated here.

[0104] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similarities between the various embodiments can be referred to in conjunction with each other. For device embodiments, since they are generally similar to method embodiments, their description is relatively simple, and for relevant details, reference can be made to the description of the method embodiments.

[0105] See also Figure 8 , Figure 8 FIG. 1 is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Figure 8As shown, the computer device 70 includes a processor 71 and a memory 72 coupled to the processor 71. The memory 72 stores program instructions. When the program instructions are executed by the processor 71, the processor 71 executes the steps of the call voice quality inspection method described in any of the above embodiments.

[0106] The processor 71 may also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip having signal processing capabilities. The processor 71 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The general-purpose processor may be a microprocessor or any conventional processor.

[0107] See Figure 9 , Figure 9 Schematic diagram of the structure of the storage medium of an embodiment of the present invention. The storage medium of the embodiment of the present invention stores program instructions 81 that can implement all the above methods, wherein the program instructions 81 can be stored in the above storage medium in the form of a software product, including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a computer device such as a computer, a server, a mobile phone, or a tablet.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed computer devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0109] In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software functional units. The above is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for checking call voice quality, characterized in that: include: Acquire call audio and convert the call audio into call text, wherein the call text includes the customer call text and the agent reply text; Input the call text into a pre-trained text detection model to obtain the target text segment containing sensitive content; Performing detailed verification on the target text segment based on a preset violation detection method; If the target text segment violates the rule, the violation type corresponding to the target text segment is confirmed as the target violation type; Generating quality inspection result information according to the target text segment and the target violation type; The types of violations include incomplete or erroneous information disclosure; The detailed verification of the target text segment based on the preset violation detection method includes: Acquire a preceding text segment and a succeeding text segment of the target text segment, and combine the preceding text segment and the target text segment into a first text segment, combine the target text segment and the succeeding text segment into a second text segment, and combine the preceding text segment, the target text segment, and the succeeding text segment into a third text segment; Vectorizing the target text segment, the first text segment, the second text segment, and the third text segment respectively to obtain a target vector, a first vector, a second vector, and a third vector; After inputting the target vector, the first vector, the second vector, and the third vector into the BERT model for encoding, they are matched one by one with the preset violation vectors in the preset violation vector library to obtain the matching degrees corresponding to the target vector, the first vector, the second vector, and the third vector respectively; The final text content corresponding to the vector with the highest matching degree is selected, and the final text content is output to trigger the information notification of incomplete or error-type violation.

2. The call voice quality inspection method according to claim 1, characterized in that: The detailed verification of the target text segment based on the preset violation detection method includes: Obtain preset violation detection methods corresponding to multiple pre-set violation types; The target text segment is verified in detail using a preset violation detection method corresponding to each violation type.

3. The call voice quality inspection method according to claim 1, characterized in that: The text detection model includes a first encoding layer, a second encoding layer, and a softmax layer; the call text is input into a pre-trained text detection model to obtain a target text segment containing sensitive content, including: Randomly initialize the call text based on normal distribution to obtain a position vector; Inputting the call text into the first encoding layer for encoding to obtain a sentence vector; Concatenating the sentence vector with the position vector to obtain a concatenated vector; encoding the concatenated vector using the second coding layer to obtain a full vector; The comprehensive vector is input into the softmax layer for prediction to obtain the target text segment.

4. The call voice quality inspection method according to claim 1, characterized in that: The acquisition of call voice and conversion of the call voice into call text, wherein the call text includes the customer call text and the agent reply text, includes: Obtaining the call voice; Identify the voiceprint features of each voice segment in the call to obtain customer voice segments and agent voice segments; The customer voice segment is converted into the customer call text, and the agent voice segment is converted into the agent reply text.

5. The call voice quality inspection method according to claim 1, characterized in that: The violation type includes a violation of information not being disclosed; and the detailed verification of the target text segment based on the preset violation detection method includes: Detecting whether the target text segment meets the preset characteristics; If the preset features are met, detecting whether there is text content related to the preset features in the target text segment; If there is no relevant text content, the target text segment triggers the information not informed violation.

6. The call voice quality inspection method according to claim 1, characterized in that: The types of violations include violations of informing of information beyond the scope of the regulations; The detailed verification of the target text segment based on the preset violation detection method includes: Matching the target text segments respectively with the preset rules; When the match is successful, the target text segment triggers the out-of-scope information notification class violation.

7. A call voice quality inspection device, the device being used to implement the steps of the call voice quality inspection method according to any one of claims 1 to 6, characterized in that: include: An acquisition module is used to acquire call voice and convert the call voice into call text, wherein the call text includes the customer call text and the agent reply text; A screening module, configured to input the call text into a pre-trained text detection model to obtain a target text segment containing sensitive content; A verification module, configured to perform detailed verification on the target text segment based on a preset violation detection method; a confirmation module, configured to confirm the violation type corresponding to the target text segment as a target violation type when the target text segment violates the rule; The output module is used to generate quality inspection result information according to the target text segment and the target violation type.

8. A computer device, characterized in that: The computer device includes a processor and a memory coupled to the processor, wherein program instructions are stored in the memory. When the program instructions are executed by the processor, the processor executes the steps of the call voice quality inspection method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that: The device stores program instructions capable of implementing the call voice quality inspection method according to any one of claims 1 to 6.

Images