Live broadcast audio auditing method and device, computer equipment and storage medium

Abstract

The invention relates to a live audio auditing method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring live audio entering a client of a live room; inputting the live audio into a pre-trained audio auditing model to obtain confidence coefficients of various normal sounds and confidence coefficients of various abnormal sounds for the live audio; under the condition that the confidence coefficient of the abnormal sound is greater than a preset confidence coefficient threshold value, inputting the confidence coefficient of each normal sound and the confidence coefficient of each abnormal sound into a sound decision model to obtain a normal sound result and / or an abnormal sound result of the live audio of the live room; wherein the sound abnormal result comprises the sound type of the abnormal sound; wherein the sound decision model is generated through the sound abnormity prediction result and the sound abnormity actual result of the audio auditing model. By adopting the method, the auditing accuracy of the live audio can be improved.

Description

Technical Field

[0001] The present application relates to the field of multimedia technology, and in particular to a live audio review method, device, computer equipment and storage medium. Background Art

[0002] With the development of multimedia technology, live streaming has become an important application in daily interactive entertainment and social interaction. Monitoring of live streaming is a necessary regulatory requirement to maintain healthy network security, especially for online live streaming, which often contains negative live streaming information.

[0003] Among the related technologies, the auditing technologies for live content mainly include image recognition technology (such as image classification, image detection, image comparison, etc.), audio recognition technology (such as ASR voice content recognition, sound event detection, voiceprint recognition, audio fingerprint retrieval, etc.), and NLP text recognition (such as classification, detection, blacklist and whitelist, etc.). In addition, the auditing subjects for live content mainly apply online audit or on-end audit.

[0004] However, during the implementation process, the applicant discovered that the relevant technology had at least the problem of low accuracy in live audio review. Summary of the invention

[0005] Based on this, the purpose of this application is to solve at least one of the above-mentioned technical defects, especially the technical defect of low accuracy of live audio review in the prior art. This application provides a live audio review method, device, computer equipment and storage medium.

[0006] In a first aspect, the present application provides a live audio review method, which is applied to a client, and the method includes:

[0007] Get the live audio from the client that enters the live room;

[0008] The live audio is input into a pre-trained audio review model, and the confidence of various normal sounds and various abnormal sounds of the live audio are obtained through the audio review model; wherein the audio review model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds;

[0009] When the confidence of the abnormal sound is greater than a preset confidence threshold, the confidence of each normal sound and the confidence of each abnormal sound are input into the sound decision model to obtain a normal sound result and / or an abnormal sound result for the live audio of the live broadcast room; wherein the abnormal sound result includes the sound type of the abnormal sound;

[0010] Among them, the sound decision model is generated by the sound anomaly prediction results and the actual results of the sound anomaly of the audio review model.

[0011] In one embodiment, after obtaining the normal sound result and / or abnormal sound result of the live audio of the live broadcast room, the method further includes:

[0012] Determine the sound type of the abnormal sound;

[0013] In the case where the sound type of the abnormal sound is an uncivilized sound, based on a pre-set voting mechanism for uncivilized sounds, a penalty result for the user corresponding to the uncivilized sound is determined.

[0014] In one embodiment, after obtaining the normal sound result and / or abnormal sound result of the live audio of the live broadcast room, the method further includes:

[0015] Determine the sound type of the abnormal sound;

[0016] If the sound type of the abnormal sound is suspected of being a suspicious sound of an abnormal live broadcast that is on the edge of the preset rules, the suspicious sound is sent to the background monitoring end;

[0017] The background monitoring terminal is used to determine the monitoring status of the suspected sound, and the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of the rule;

[0018] Receive the monitoring status sent from the background monitoring terminal.

[0019] In one embodiment, the method further comprises:

[0020] If the number of times that the suspected sound is determined to be an abnormal sound is greater than a preset number threshold, a penalty result for the user who made the abnormal sound is output based on the number of abnormal sounds.

[0021] In one embodiment, the method further comprises:

[0022] receiving the audio review model and the sound decision model sent by the server;

[0023] The server is used to obtain an audio review model by fine-tuning the initial audio model after self-supervised training using normal sound samples and abnormal sound samples; and to generate a sound decision model using the sound abnormality prediction results and actual sound abnormality results of the audio review model.

[0024] In a second aspect, the present application also provides a live audio review method, which is applied to a server and includes:

[0025] Obtain sample live audio; and input the sample live audio into a pre-trained audio review model to obtain confidences of various candidate normal sounds and confidences of various candidate abnormal sounds;

[0026] Acquire actual sound annotation data for each candidate abnormal sound; wherein the actual sound annotation data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound;

[0027] Generate a sound decision model using actual sound annotation data, the confidence of each candidate normal sound, and the confidence of each candidate abnormal sound; and send the sound decision model to the client;

[0028] The client is used to obtain normal sound results and / or abnormal sound results for the live audio of the live room based on the sound decision model.

[0029] In one embodiment, the training method of the audio review model includes:

[0030] Obtain training audio obtained from the audio of the live broadcast room, and obtain sample normal sounds and sample abnormal sounds;

[0031] The training audio is input into the audio review model to be trained, and an initial audio model is obtained through self-supervised training of the audio review model;

[0032] The sample normal sound and the sample abnormal sound are input into the initial audio model to train the initial audio model to obtain the audio review model.

[0033] In one embodiment, the method further comprises:

[0034] Obtaining suspicious sounds sent by the client, where the suspicious sounds are abnormal live broadcast sounds that are suspected to be on the edge of preset rules;

[0035] The suspected sound is displayed through the background monitoring terminal, and the monitoring status of the suspected sound on the background monitoring terminal is obtained; the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of the rule;

[0036] Send monitoring status to the client.

[0037] In a third aspect, the present application provides a live audio review device, which is applied to a client, and the device includes:

[0038] The audio receiving module is used to receive the live audio of the live room to be reviewed in the client;

[0039] The confidence module is used to input the live audio into a pre-trained audio review model, and obtain the confidence of various normal sounds and various abnormal sounds of the live audio through the audio review model; wherein the audio review model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds;

[0040] A sound result module, for inputting the confidence of each normal sound and the confidence of each abnormal sound into the sound decision model when the confidence of the abnormal sound is greater than a preset confidence threshold, to obtain a normal sound result and / or an abnormal sound result for the live audio of the live broadcast room; wherein the abnormal sound result includes the sound type of the abnormal sound;

[0041] Among them, the sound decision model is generated by the sound anomaly prediction results and the actual results of the sound anomaly of the audio review model.

[0042] In a fourth aspect, the present application further provides a live audio review device, which is applied to a server, and the device includes:

[0043] A sample audio module is used to obtain sample live audio; and input the sample live audio into a pre-trained audio review model to obtain the confidence of various candidate normal sounds and the confidence of various candidate abnormal sounds;

[0044] A labeling data module is used to obtain actual sound labeling data for each candidate abnormal sound; wherein the actual sound labeling data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound;

[0045] A decision model module is used to generate a sound decision model using actual sound annotation data, the confidence of each candidate normal sound and the confidence of each candidate abnormal sound; and send the sound decision model to the client;

[0046] The client is used to obtain normal sound results and / or abnormal sound results for the live audio of the live room based on the sound decision model.

[0047] In a fifth aspect, the present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0048] In a sixth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when the computer program is executed by a processor.

[0049] In a seventh aspect, the present application provides a computer program product, comprising a computer program, which implements the steps of the above method when executed by a processor.

[0050] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0051] The live audio review method, device, computer equipment and storage medium provided by the present application can enable the client to directly apply the audio review model and sound decision model sent by the server after receiving them by pre-training the audio review model and establishing the sound decision model on the server side, that is, the live audio review can be implemented on the client side, thereby reducing the cost of online review. Specifically, on the client side, the present application can predict the confidence of various abnormal sounds by inputting the live audio into the pre-trained audio review model. When the confidence of the abnormal sound is greater than the preset confidence threshold, the confidence of the normal sound and the confidence of the abnormal sound can be further input into the sound decision model. Through the sound decision model, the normal sound result and / or abnormal sound result of the live audio can be obtained. In this way, the present application further decides the audio review result by utilizing the confidence of the normal sound and the confidence of the abnormal sound, thereby improving the review accuracy of the live audio. That is, compared with traditional technology, this application not only predicts the confidence of abnormal sounds through the model, but also predicts the confidence of normal sounds, realizes a multi-label audio review method, and improves the accuracy of live audio review on the terminal. Therefore, this application can reduce the cost of live audio review while improving the review accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 A schematic diagram of the application environment of the live audio review method provided in the embodiment of the present application;

[0054] Figure 2 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 1 ;

[0055] Figure 3 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 2 ;

[0056] Figure 4 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 3 ;

[0057] Figure 5 A flowchart of a training method for an audio review model provided in an embodiment of the present application;

[0058] Figure 6 A schematic diagram of the structure of a live audio review device provided in an embodiment of the present application;

[0059] Figure 7 A schematic diagram of the structure of another live audio review device provided in an embodiment of the present application;

[0060] Figure 8 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0062] With the development of multimedia technology, live streaming has become an important application in daily interactive entertainment and social interaction. Monitoring of live streaming is a necessary regulatory requirement for maintaining healthy network security. Especially for online live streaming, there is usually bad live streaming information. Audio and video live streaming companies need to monitor the content generated by their users and impose penalties such as banning users for any illegal information.

[0063] Among the related technologies, the auditing technologies for live content mainly include image recognition technology (such as image classification, image detection, image comparison, etc.), audio recognition technology (such as ASR voice content recognition, sound event detection, voiceprint recognition, audio fingerprint retrieval, etc.), and NLP text recognition (such as classification, detection, blacklist and whitelist, etc.). In addition, the auditing subjects for live content mainly apply online audit or on-end audit.

[0064] The online audit solution records audio through the APP, uploads image / audio data to the cloud for storage, and requests the audit algorithm model service. After obtaining the audit results, the auditor will review the suspected illegal content and finally take corresponding actions on the user.

[0065] The on-device audit solution deploys the algorithm model on the mobile phone / client to detect illegal audio in real time and impose penalties.

[0066] However, the applicant discovered during the implementation process that the above technical solution has at least the following technical problems:

[0067] Online audit solution: The penalty timeliness is poor and it is impossible to prevent users from violating the rules in real time; the cost is high, and the cost of traffic, storage, algorithm model services, and manual audit increases with the scale of business.

[0068] On-end review solution: The on-end algorithm model is subject to computing power and size limitations, which leads to model capacity limitations, or may lead to problems such as missed recalls of illegal content and incorrect penalties for normal content; and it is difficult to flexibly adjust the model.

[0069] In other words, the relevant technology at least has the problem of low accuracy of live audio review.

[0070] Based on the above technical problems, the present application provides a live audio review method, device, computer equipment and storage medium. By inputting the live audio into a pre-trained audio review model on the client, the present application can predict the confidence of various abnormal sounds. When the confidence of the abnormal sound is greater than the preset confidence threshold, the confidence of the normal sound and the confidence of the abnormal sound can be further input into the sound decision model, and the normal sound result and / or abnormal sound result of the live audio can be obtained. In this way, the present application further decides the audio review result by utilizing the confidence of the normal sound and the confidence of the abnormal sound, thereby improving the review accuracy of the live audio.

[0071] Figure 1 A schematic diagram of the application environment of the live audio review method provided in the embodiment of the present application; Figure 1 As shown, the live audio review method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown in FIG. 1 , the terminal 101 communicates with the server 102 via a network. The terminal 101 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The server 102 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0072] In an exemplary embodiment, Figure 2 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 1 ,like Figure 2 As shown, a live audio review method is provided, which is applied to Figure 1 Taking the client in as an example, the method includes the following S201 to S203. Among them:

[0073] S201. Obtain the live audio from the client entering the live broadcast room.

[0074] The client refers to a client that can be used for live audio and video broadcasting. For example, the client can be a program for providing local live broadcasting services to users. The live broadcast room can refer to a room where live broadcasting is being conducted. Users in the live broadcast room can include the anchor and users who watch the live broadcast, and can also include anchors or users in other live broadcast rooms connected to the anchor. Live broadcast audio can refer to the audio generated in real time during live broadcasting in the live broadcast room.

[0075] Exemplarily, in a terminal (e.g., a mobile phone), the live audio generated by any live room in the client can be received. For example, any live room can be subject to live supervision and review, that is, any live room can be used as a live room to be reviewed.

[0076] This application can directly review the live audio of the live broadcast room in the terminal to reduce the cost of online review, improve the timeliness of the review, and prevent users from violating regulations in the live broadcast room in real time.

[0077] S202. Input the live audio into a pre-trained audio review model, and obtain the confidence of various normal sounds and the confidence of various abnormal sounds of the live audio through the audio review model; wherein the audio review model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds.

[0078] The audio review model may refer to a model used to review whether live audio has any violations, and the confidence of abnormal sounds may be used to determine whether there are any violations in live broadcasts. The confidence may refer to the degree of certainty about the sound type predicted by the model, and the confidence may be represented by a score value. The initial audio model may refer to the initial model obtained after self-supervised training of the audio review model to be trained.

[0079] Exemplarily, the audio review model can be pre-trained in the server, and the terminal can receive and save the audio review model sent by the server. The server can use the audio review model to be trained for self-supervised training to obtain a trained initial audio model. Further, the initial audio model can be fine-tuned using sample normal sounds and sample abnormal sounds. For example, the branch parameters of the initial audio model can be adjusted to obtain a trained audio review model. The server can send the pre-trained audio review model to the client of the terminal to enable live audio review on the terminal.

[0080] The client can input the live audio into the above-mentioned audio review model. Through the audio review model, the confidence of various normal sounds and the confidence of various abnormal sounds that may exist in the above-mentioned live audio that needs to be reviewed can be predicted. In this way, the present application not only predicts the confidence of abnormal sounds, but also predicts the confidence of normal sounds. The confidence of normal sounds can be used to further judge whether abnormal sounds are truly abnormal sounds, so that the abnormal sounds predicted by the audio review model can be further judged for accuracy, thereby improving the accuracy of live audio review. The present application reviews abnormal sounds of live audio through multi-label results (confidence of normal sounds and confidence of abnormal sounds), which is conducive to improving the accuracy of live audio review.

[0081] S203. When the confidence of the abnormal sound is greater than a preset confidence threshold, the confidence of each normal sound and the confidence of each abnormal sound are input into the sound decision model to obtain a normal sound result and / or an abnormal sound result for the live audio of the live room; wherein the abnormal sound result includes the sound type of the abnormal sound.

[0082] Among them, the sound decision model is generated by the sound anomaly prediction results and the actual results of the sound anomaly of the audio review model.

[0083] The sound decision model may be a model pre-established in the server. The sound abnormality prediction result may refer to the abnormal sound predicted by the sound decision model; the sound abnormality actual result may refer to the actual result of the abnormal sound obtained by the above sound abnormality actual result. For example, the abnormal sound predicted by the sound decision model may actually be a normal sound, or may actually be an abnormal sound. That is, the abnormal sound predicted by the sound decision model may not be accurate, and it can be determined whether the abnormal sound is a real abnormal sound.

[0084] As an example, the sound decision model may be a model established based on a decision tree structure.

[0085] The confidence threshold may refer to a threshold used to determine whether an abnormal sound can be confirmed as a real abnormal sound. The confidence threshold may be pre-set according to actual conditions. For example, different confidence thresholds may be set for different types of live broadcast rooms to determine abnormal sounds.

[0086] Exemplarily, the server can use the audio audit model to predict a certain segment of live audio, obtain the sound anomaly prediction result of the live audio, and establish a sound decision model through the actual sound anomaly result corresponding to the sound anomaly prediction result. For example, the normal sound label and abnormal sound label contained in the live audio segment pre-marked by the user can be used to determine whether the sound anomaly prediction result output by the audio audit model is correct. If it is not correct, the sound anomaly prediction result can be annotated and corrected. Further, the prediction result and annotation data output by the audio audit model can be used to establish a sound decision model. The server can send the pre-established sound decision model to the client and the sound decision model can take effect directly in the client.

[0087] When the confidence level of the abnormal sound is greater than a preset confidence level threshold, that is, when the abnormal sound is truly an abnormal sound, the terminal can make a further accuracy judgment on the abnormal sound, and can input the confidence levels of various normal sounds and various abnormal sounds into the sound decision model, and make further judgments through the sound decision model to determine the audio review results of the live audio, that is, the normal sound results and / or abnormal sound results contained in the live audio can be obtained. When there is an abnormal sound in the live audio of the live room, the penalty result for the user who made the abnormal sound in the live room can be further output to ensure the normal and standardized live broadcast of the live room.

[0088] For example, if the confidence of the abnormal sound is greater than the preset confidence threshold, the prediction results output by all audio review models (including the confidence of normal sounds and the confidence of abnormal sounds) can be input into the sound decision model to obtain the result of whether there are any illegal sounds in the live audio, and if there are any illegal sounds, the result of what type of violation it is can also be obtained.

[0089] Optionally, the types of sound violations may include uncivilized sounds, abnormal sounds on the edge of the rules, etc. For example, in UGC (user-generated content) voice live broadcasts, there may be abnormal sounds on the edge of the live broadcast rules. In this case, the live broadcast platform or live broadcast software needs to be regulated for abnormal sounds. Uncivilized sounds include quarrels, room bombing, etc. Abnormal sounds on the edge of the rules may refer to sounds close to the boundaries or edges of the rules, such as some vulgar behavior sounds.

[0090] In this way, through the above-mentioned step S203, the confidence level of normal sounds and the confidence level of each abnormal sound can be utilized to further judge the accuracy of the predicted abnormal sound through the sound decision model, thereby improving the accuracy of live audio review.

[0091] In this embodiment, by pre-training the audio review model and establishing the sound decision model on the server side, the client can directly apply the audio review model and the sound decision model after receiving them from the server, that is, the live audio review can be realized on the client, thereby reducing the cost of online review. Specifically, on the client side, the present application can predict the confidence of various abnormal sounds by inputting the live audio into the pre-trained audio review model. When the confidence of the abnormal sound is greater than the preset confidence threshold, the confidence of the normal sound and the confidence of the abnormal sound can be further input into the sound decision model. Through the sound decision model, the normal sound result and / or abnormal sound result of the live audio can be obtained. In this way, the present application further determines the audio review result by utilizing the confidence of the normal sound and the confidence of the abnormal sound, thereby improving the review accuracy of the live audio. That is, compared with traditional technology, this application not only predicts the confidence of abnormal sounds through the model, but also predicts the confidence of normal sounds, realizes a multi-label audio review method, and improves the accuracy of live audio review on the terminal. Therefore, this application can reduce the cost of live audio review while improving the review accuracy.

[0092] In an exemplary embodiment, after step S202, the method further includes:

[0093] When the confidence level of the abnormal sound is less than or equal to a preset confidence threshold, the normal sound result of the live audio is output.

[0094] Exemplarily, if the confidence level of the abnormal sound is less than or equal to a preset confidence threshold, that is, the result predicted by the audio review model shows that the abnormal sound is not a true abnormal sound, then the terminal can directly output that there is no abnormal sound in the live audio, that is, it can directly output the normal sound result of the live audio review.

[0095] In this embodiment, when the confidence level of the abnormal sound is less than or equal to the preset confidence threshold, the normal sound result of the live audio review is directly output without the need to perform the judgment step of the decision sound model, thereby improving the efficiency of the live audio review.

[0096] In an exemplary embodiment, before step S201, the method may further include:

[0097] receiving the audio review model and the sound decision model sent by the server;

[0098] The server is used to obtain an audio review model by fine-tuning the initial audio model after self-supervised training using normal sound samples and abnormal sound samples; and to generate a sound decision model using the sound abnormality prediction results and actual sound abnormality results of the audio review model.

[0099] Exemplarily, on the server side, the audio review model and the sound decision model can be pre-trained, and the server can send the trained audio review model and the sound decision model to the terminal. The terminal can install the audio review model and the sound decision model in the client so that the two models can take effect and be used, thereby realizing real-time review of live audio on the terminal and improving the review timeliness.

[0100] Optionally, the server can compress the lightweight deep neural network that can be used on the client (for example, proportionally compress the number of channels of the mobilnetV2 model), use the compressed model as the backbone model, and then use the actual audio data of the APP for large-scale self-supervised training (without manual labeling) to complete the training of the backbone and obtain the initial audio model.

[0101] After completing backbone training, the corresponding illegal audio (such as abnormal live broadcast sound, uncivilized sound, etc. that are on the edge of the rules) can be collected as sound labels for abnormal sounds, and normal audio data (such as speaking, singing, background music, laughter, panting, TV sound, etc.) can be used as labels for normal sounds; and the labels of abnormal sounds and normal sounds can be used to train the branch model of the above model to identify various labels, thus obtaining a trained audio review model.

[0102] In this embodiment, by receiving the audio review model and the sound decision model sent by the server, it is helpful to realize the real-time review of the live audio on the terminal, improve the timeliness of the live audio review, and compared with realizing the live audio review on the server, it is helpful to reduce the cost.

[0103] In an exemplary embodiment, Figure 3 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 2 ,like Figure 3 As shown, in Figure 2 Based on the above, the steps included in the live audio review method are exemplarily described. In step S203, after obtaining the normal sound result and / or abnormal sound result for the live audio of the live room, the following steps may also be used:

[0104] S301: Determine the sound type of the abnormal sound.

[0105] Exemplarily, the sound type of the abnormal sound may include uncivilized sound and abnormal sound on the edge of the rule. The client can determine the specific sound type of the abnormal sound, and further output different abnormal sound processing results for different abnormal sound types.

[0106] S302: When the sound type of the abnormal sound is an uncivilized sound, based on a pre-set voting mechanism for uncivilized sounds, determine a penalty result for the user corresponding to the uncivilized sound.

[0107] Among them, uncivilized sounds may refer to uncivilized sounds, for example, insults, room bombings, harsh sounds, etc.

[0108] For example, if uncivilized sounds appear in the live broadcast room, a voting mechanism for uncivilized sounds can be pre-set to initiate a vote by other users to decide whether to punish the suspected violating user. For example, users can vote to determine whether the suspected violating user needs to be punished for uncivilized sounds such as room bombing and insults, thereby further improving the accuracy of live broadcast audio review. When punishment is required, the punishment result for the user who made the uncivilized sound is output. The punishment result can be determined based on the pre-set punishment mechanism. For example, for insults and other sounds, the user who made the sound can be blacklisted to ensure that the user can no longer make uncivilized sounds, so as to ensure the normal and standardized live broadcast of the live broadcast room.

[0109] In this embodiment, since the audio review model and the sound decision model output on the end are not manually reviewed, even if the review accuracy is very high, there is a possibility of erroneous recall. Through step S301, this application can conduct user voting on suspected uncivilized sounds to determine whether they are truly uncivilized sounds, thereby further improving the accuracy of live audio review.

[0110] Alternatively, if Figure 3 As shown, in step S203, after obtaining the normal sound result and / or abnormal sound result of the live audio of the live room, and after step S301, the following may be specifically included:

[0111] S303. If the sound type of the abnormal sound is a suspected sound of an abnormal live broadcast that is suspected to be on the edge of a preset rule, the suspected sound is sent to a background monitoring terminal.

[0112] The background monitoring terminal is used to determine the monitoring status of the suspected sound, and the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of the rule;

[0113] S304: Receive the monitoring status sent from the background monitoring terminal.

[0114] For example, if the sound type of the abnormal sound is an abnormal live broadcast sound that is suspected to be on the edge of the rules, that is, if the abnormal sound is a sound close to the boundary or edge of the rules, at this time, the users in the live broadcast room may be tolerant of the sound, and the voting mechanism may not be able to accurately determine whether the sound is an illegal sound. At this time, the suspected sound can be sent to the background monitoring end for monitoring. The background monitoring end can manually determine whether the suspected sound is an illegal sound (abnormal sound), obtain the monitoring status of the suspected sound, and can send the monitoring status of the suspected sound to the client. If the monitoring status of the suspected sound indicates that the suspected sound is an abnormal sound, the user who makes the abnormal sound in the live broadcast room can be punished.

[0115] In this embodiment, since the audio review model and the sound decision model output on the end are not manually reviewed, through the above steps S302 to S303, when the abnormal sound is suspected to be an abnormal live sound on the edge of the preset rules, the background monitoring end can determine whether the suspected sound is a real illegal sound, thereby improving the accuracy of the review of the live audio.

[0116] Alternatively, if Figure 3 As shown, after step S304, the method may further include:

[0117] S305: If the number of times the suspected sound is determined to be an abnormal sound is greater than a preset number threshold, a penalty result for the user who made the abnormal sound is output based on the number of abnormal sounds.

[0118] For example, if it is determined that the suspicious sound emitted by the user in the live broadcast room is an abnormal live broadcast sound at the edge of the rules, and the number of times the user emits the abnormal live broadcast sound at the edge of the rules is greater than the preset number threshold, the user can be punished, for example, the user can be punished by suspending the live broadcast, speaking during the live broadcast, etc. This is conducive to improving the supervision of live broadcasts.

[0119] For example, if the anchor user who hosts the live broadcast room makes abnormal sounds more than a preset threshold number of times, the anchor user may be punished by being banned from live broadcast.

[0120] As an example, if the number of abnormal sounds is greater than the preset first threshold, the anchor user will be banned for the first preset time; if the number of abnormal sounds is greater than the preset second threshold, the anchor user will be banned for the second preset time; if the number of abnormal sounds is greater than the preset third threshold, the anchor user will be blacklisted and the live broadcasting behavior of the anchor user will be prohibited.

[0121] For another example, if the number of times an audience user in the live broadcast room makes abnormal sounds is greater than a preset threshold, such as when the audience user is conducting a live broadcast with the host user, the number of times the audience user makes abnormal sounds is greater than the preset threshold, then the live broadcast function of the audience user can be prohibited.

[0122] For another example, if the number of times that the anchor user in the live broadcast room and the anchor users in other live broadcast rooms make abnormal sounds exceeds a preset number threshold, the two anchor users can be banned from speaking at the same time.

[0123] In some examples, the type of live broadcast room and the number of users are obtained in advance, and the threshold of the number of abnormal sounds in the live broadcast room can be dynamically determined based on the type of live broadcast room and the number of users.

[0124] In some other examples, the type of live broadcast room and the number of users are further obtained in advance, and the penalty method for exceeding the threshold of the number of abnormal sounds in the live broadcast room is dynamically determined based on the selected type of live broadcast room and the number of users. When the types of live broadcast rooms or the number of users reach a certain number, the penalty intensity is increased. For example, when the live broadcast room is a relatively active type and there are a large number of participants, the penalty result of mute or banning the user's live broadcast is output, so that the bad abnormal sounds can be prevented from spreading widely; when the live broadcast room is a relatively serious type, the penalty result of mute or banning the user's live broadcast is output, so that the bad abnormal sounds can be prevented from spreading in serious scenes. In some further examples, the penalty intensity can be adjusted based on the classification of the user population, for example, based on user characteristics, including the interaction characteristics of live broadcast resources, the level of live broadcast users, the number of user violations, etc.

[0125] If the number of times the anchor user in the live broadcast room makes abnormal sounds exceeds the preset number threshold, based on the type of the live broadcast room of the anchor user, the penalty result corresponding to the type of the live broadcast room is output; for example, if the type of the live broadcast room is a type with a higher tolerance, a lighter penalty result can be output. Specifically, the type of live broadcast with a higher tolerance and the lighter penalty result can be determined based on the actual situation, and no specific restrictions are made here. For another example, when the number of users in the live broadcast room is small, the penalty result for the anchor user can be reduced.

[0126] In this embodiment, through step S304, the live broadcast is reviewed from the user's perspective, thereby ensuring the accuracy of the review of the user's live broadcast, thereby improving the effectiveness of the review of the users in the live broadcast room during the live broadcast process.

[0127] In an exemplary embodiment, Figure 4 Schematic diagram of the live audio review method provided in the embodiment of the present application Figure 3 ,like Figure 4 As shown, a live audio review method is provided, which is applied to Figure 1 Taking the server in as an example, the method includes S401 to S403, wherein:

[0128] S401, obtaining sample live audio; and inputting the sample live audio into a pre-trained audio review model to obtain confidences of various candidate normal sounds and confidences of various candidate abnormal sounds;

[0129] S402, obtaining actual sound annotation data for each candidate abnormal sound; wherein the actual sound annotation data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound;

[0130] S403, generating a sound decision model using the actual sound annotation data, the confidence of each candidate normal sound, and the confidence of each candidate abnormal sound; and sending the sound decision model to the client;

[0131] The client is used to obtain normal sound results and / or abnormal sound results for the live audio of the live room based on the sound decision model.

[0132] The sample live audio may refer to the audio training data of the audio review model. The candidate normal sound may refer to the normal sound result to be determined by the model prediction output. The candidate abnormal sound may refer to the abnormal sound result to be determined by the model prediction output. The actual sound annotation data may refer to the data annotated for the candidate abnormal sound, and the actual sound annotation data is used to mark whether the candidate abnormal sound is really an abnormal sound (illegal sound).

[0133] Exemplarily, the server can obtain sample live audio, and through the audio review model, review and predict the sample live audio, and output the confidence of various candidate normal sounds (such as the score of each normal sound label) and the confidence of various candidate abnormal sounds (such as the score of each abnormal sound label). The server can obtain manual annotation data for various candidate abnormal sounds, that is, it can determine whether various candidate abnormal sounds are truly abnormal sounds. In this way, the server can use the confidence of each candidate normal sound and the confidence of each candidate abnormal sound, and use a pre-set algorithm framework in the actual sound annotation data to generate a sound decision model. For example, the pre-set algorithm framework can be a decision model preset by the user, a decision tree model, etc.

[0134] The server can send the sound decision model to the client in the form of a file, and the client can load the trained sound decision model file and use the sound decision model after loading the file. Figure 1The step implementation example can improve the accuracy of live audio review and reduce the cost of online review.

[0135] In this embodiment, through steps S401 to S403, a sound decision model basis can be provided to the client, which is beneficial to the review and use of live audio, thereby helping to improve the review accuracy of the live audio.

[0136] In an exemplary embodiment, Figure 5 A flowchart of a training method for an audio audit model provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, in Figure 4 Based on the above, the steps of the live audio review method are exemplarily described, and the steps include S501 to S503, wherein:

[0137] S501, obtaining training audio obtained by using the audio of the live broadcast room, and obtaining sample normal sound and sample abnormal sound;

[0138] S502, inputting the training audio into the audio audit model to be trained, and obtaining an initial audio model through self-supervised training of the audio audit model;

[0139] S503. Input the sample normal sound and the sample abnormal sound into the initial audio model to train the initial audio model to obtain an audio review model.

[0140] The training audio may be the actual audio data in the live broadcast APP, and the actual audio data is used as the training data for the training model backbone. The sample normal sound may refer to the normal sound label data for the training model branch. The sample abnormal sound may refer to the abnormal sound label data for the training model branch.

[0141] Exemplarily, the server can obtain the training audio for training the model from the actual audio data in the live broadcast room, and can obtain the normal sound sample label data and the abnormal sound sample label data. The server can input the training audio into the audio audit model to be trained, and obtain the trained backbone model through the self-supervised training of the audio audit model, that is, obtain the initial audio model. The service can input the sample normal sound and the sample abnormal sound into the initial audio model, and train the branch model of the initial audio model to achieve fine-tuning of the initial audio model, and obtain the trained audio audit model.

[0142] In this embodiment, through the above S501 to S503, the basis of the audio review model can be provided to the client, which is beneficial to the review and use of the live audio, thereby facilitating the improvement of the review accuracy of the live audio.

[0143] In an exemplary embodiment, the method applied to the server further includes:

[0144] Obtaining suspicious sounds sent by the client, where the suspicious sounds are abnormal live broadcast sounds that are suspected to be on the edge of preset rules;

[0145] The suspected sound is displayed through the background monitoring end, and the monitoring status of the suspected sound on the background monitoring end is obtained; the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of the rule.

[0146] Send monitoring status to the client.

[0147] For example, if the abnormal sound type is an abnormal live broadcast sound suspected of being on the edge of the rules, that is, if the abnormal sound is a sound close to the boundary or edge of the rules, at this time, the users in the live broadcast room may be tolerant of the sound, and the voting mechanism may not be able to accurately determine whether the sound is a violation sound. At this time, the client can send the suspected sound to the backend monitoring end for monitoring. The server can obtain the suspected sound sent by the client, and can send the suspected sound sent by the client to the backend monitoring end.

[0148] The background monitoring end can manually determine whether the suspected sound is an illegal sound (abnormal sound), obtain the monitoring status of the suspected sound, and send the monitoring status of the suspected sound to the client through the server. If the monitoring status of the suspected sound indicates that the suspected sound is an abnormal sound, the user who makes the abnormal sound in the live broadcast room can be punished.

[0149] In this embodiment, the server can manually judge whether the suspected sound is a real abnormal sound, which is conducive to improving the accuracy of judging abnormal sounds, thereby improving the accuracy of live audio review.

[0150] In some specific embodiments, the present application can improve the review accuracy of the small model on the terminal through a multi-label (including abnormal sound labels and normal sound labels) model and decision-making strategy; the present application can configure different sound decision models according to different live broadcast room / voice room types, improve the flexibility of live audio review, and can also adjust the decision-making strategy in real time; some violation labels of the present application can be judged by other users whether they are violations, saving online data interaction costs, ensuring the timeliness of penalties, and improving the experience of ordinary users.

[0151] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0152] The live audio review device provided in the embodiment of the present application is described below. The live audio review device has the same inventive concept as the above-mentioned live audio review method. The implementation solution for solving the problem provided by the device is similar to the implementation solution recorded in the above-mentioned method. Therefore, the specific limitations in one or more live audio review device embodiments provided below can refer to the limitations on the live audio review method in the above text. The live audio review device described below and the live audio review method described above can be referenced to each other, and will not be repeated here.

[0153] In an exemplary embodiment, Figure 6 A schematic diagram of the structure of a live audio review device provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, applied to the client, the live audio review device 60 includes: an audio receiving module 610, a confidence module 620 and a sound result module 630, wherein:

[0154] The audio receiving module 610 is used to receive the live audio of the live room to be reviewed in the client;

[0155] The confidence module 620 is used to input the live audio into a pre-trained audio audit model, and obtain the confidence of various normal sounds and various abnormal sounds of the live audio through the audio audit model; wherein the audio audit model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds;

[0156] The sound result module 630 is used to input the confidence of each normal sound and the confidence of each abnormal sound into the sound decision model when the confidence of the abnormal sound is greater than a preset confidence threshold, so as to obtain a normal sound result and / or an abnormal sound result for the live audio of the live broadcast room; wherein the abnormal sound result includes the sound type of the abnormal sound;

[0157] Among them, the sound decision model is generated by the sound anomaly prediction results and the actual results of the sound anomaly of the audio review model.

[0158] In an exemplary embodiment, the sound result module is also used to determine the punishment result for the user corresponding to the uncivilized sound based on a pre-set voting mechanism for the uncivilized sound when the sound type of the abnormal sound is an uncivilized sound.

[0159] In an exemplary embodiment, the sound result module is also used to send the suspected sound to the background monitoring end if the sound type of the abnormal sound is a suspected sound of abnormal live broadcast suspected to be on the edge of preset rules; the background monitoring end is used to determine the monitoring status of the suspected sound, and the monitoring status is used to indicate that the suspected sound is a normal sound or an abnormal sound on the edge of the rules; and receive the monitoring status sent from the background monitoring end.

[0160] In an exemplary embodiment, the sound result module is further configured to punish the user corresponding to the abnormal sound if the number of times the suspected sound is determined to be an abnormal sound is greater than a preset number threshold.

[0161] In an exemplary embodiment, the sound result module is also used to output a normal sound result of the live audio when the confidence of the abnormal sound is less than or equal to a preset confidence threshold.

[0162] In an exemplary embodiment, the apparatus further includes a model receiving module.

[0163] The model receiving module is used to receive the audio review model and the sound decision model sent by the server; the server is used to obtain the audio review model by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds; and to generate the sound decision model with the sound abnormality prediction results and actual sound abnormality results of the audio review model.

[0164] In an exemplary embodiment, Figure 7 A schematic diagram of the structure of another live audio review device provided in an embodiment of the present application, such as Figure 7 As shown, applied to the server, the live audio review device 70 includes: a sample audio module 710, a sample audio module 711 and a decision model module 730, wherein:

[0165] The sample audio module 710 is used to obtain sample live audio; and input the sample live audio into the pre-trained audio audit model to obtain the confidence of various candidate normal sounds and the confidence of various candidate abnormal sounds.

[0166] The annotation data module 720 is used to obtain actual sound annotation data for each candidate abnormal sound; wherein the actual sound annotation data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound.

[0167] The decision model module 730 is used to generate a sound decision model using actual sound annotation data, the confidence of each candidate normal sound and the confidence of each candidate abnormal sound; and send the sound decision model to the client.

[0168] The client is used to obtain normal sound results and / or abnormal sound results for the live audio of the live room based on the sound decision model.

[0169] In an exemplary embodiment, the device includes a model training module. The model training module is used to obtain training audio obtained by using the audio of the live broadcast room, and obtain sample normal sounds and sample abnormal sounds; and input the training audio into the audio audit model to be trained, and obtain an initial audio model through self-supervised training of the audio audit model; and input the sample normal sounds and sample abnormal sounds into the initial audio model to train the initial audio model and obtain the audio audit model.

[0170] In an exemplary embodiment, a suspicious sound determination module. The suspicious sound determination module is used to obtain a suspicious sound sent by a client, where the suspicious sound is an abnormal live broadcast sound suspected to be on the edge of a preset rule; display the suspicious sound through a background monitoring terminal, and obtain the monitoring status of the suspicious sound on the background monitoring terminal; the monitoring status is used to indicate whether the suspicious sound is a normal sound or an abnormal sound on the edge of a rule. The monitoring status is sent to the client.

[0171] In an exemplary embodiment, the present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by one or more processors, the one or more processors execute the steps of the live audio review method as described in any one of the above embodiments.

[0172] In an exemplary embodiment, the present application also provides a computer device having a computer program stored therein, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the live audio review method as described in any one of the above embodiments.

[0173] In an exemplary embodiment, the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the live audio review method as described in any one of the above embodiments.

[0174] Indicatively, if Figure 8 As shown, Figure 8This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 800 may be provided as a server. Figure 8 The computer device 800 includes a processing component 802, which further includes one or more processors, and a memory resource represented by a memory 801 for storing instructions executable by the processing component 802, such as an application. The application stored in the memory 801 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 802 is configured to execute the instructions to perform the text recognition method of any of the above embodiments.

[0175] The computer device 800 may further include a power supply component 803 configured to perform power management of the computer device 800, a wired or wireless network interface 804 configured to connect the computer device 800 to a network, and an input / output (I / O) interface 805. The computer device 800 may operate based on an operating system stored in the memory 801, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM, or the like.

[0176] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0177] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0178] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0179] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0180] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A live audio audit method, characterized in that: Applied to a client, the method comprises: Get the live audio from the client that enters the live room; The live audio is input into a pre-trained audio audit model, and the confidence of various normal sounds and various abnormal sounds of the live audio are obtained through the audio audit model; wherein the audio audit model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds; In the case where the confidence of the abnormal sound is greater than a preset confidence threshold, the confidence of each normal sound and the confidence of each abnormal sound are input into the sound decision model to obtain a normal sound result and / or an abnormal sound result for the live audio of the live room; wherein the abnormal sound result includes the sound type of the abnormal sound; Among them, the sound decision model is generated by the sound abnormality prediction results and the sound abnormality actual results of the audio review model.

2. The method according to claim 1, characterized in that After obtaining the normal sound result and / or abnormal sound result for the live audio of the live room, the method further includes: determining a sound type of the abnormal sound; In the case where the sound type of the abnormal sound is an uncivilized sound, based on a pre-set voting mechanism for the uncivilized sound, a punishment result for the user corresponding to the uncivilized sound is determined.

3. The method according to claim 1, characterized in that After obtaining the normal sound result and / or abnormal sound result for the live audio of the live room, the method further includes: determining a sound type of the abnormal sound; If the sound type of the abnormal sound is a suspected sound of an abnormal live broadcast that is suspected to be on the edge of a preset rule, the suspected sound is sent to a background monitoring terminal; The background monitoring terminal is used to determine the monitoring status of the suspected sound, and the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of a rule; Receive the monitoring status sent from the background monitoring terminal.

4. The method according to claim 3, characterized in that The method further comprises: If the number of times the suspected sound is determined to be an abnormal sound is greater than a preset number threshold, a penalty result for the user who made the abnormal sound is output based on the number of abnormal sounds.

5. The method according to claim 1, characterized in that: The method further comprises: receiving the audio review model and the sound decision model sent by the server; The server is used to obtain the audio audit model by fine-tuning the initial audio model after self-supervised training using sample normal sounds and sample abnormal sounds; and to generate the sound decision model using the sound abnormality prediction results and actual sound abnormality results of the audio audit model.

6. A live audio audit method, characterized in that: Applied to a server, the method comprises: Obtain sample live audio; and input the sample live audio into a pre-trained audio audit model to obtain confidences of various candidate normal sounds and confidences of various candidate abnormal sounds; Acquiring actual sound annotation data for each of the candidate abnormal sounds; wherein the actual sound annotation data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound; Generate a sound decision model using the actual sound annotation data, the confidence of each candidate normal sound and the confidence of each candidate abnormal sound; and send the sound decision model to the client; The client is used to obtain a normal sound result and / or an abnormal sound result of the live audio of the live broadcast room based on the sound decision model.

7. The method according to claim 6, characterized in that The training method of the audio review model includes: Obtain training audio obtained from the audio of the live broadcast room, and obtain sample normal sounds and sample abnormal sounds; The training audio is input into the audio audit model to be trained, and an initial audio model is obtained through self-supervised training of the audio audit model; The sample normal sound and the sample abnormal sound are input into the initial audio model to train the initial audio model to obtain the audio audit model.

8. The method according to claim 6, characterized in that The method further comprises: Acquiring a suspicious sound sent by the client, wherein the suspicious sound is an abnormal live broadcast sound suspected to be on the edge of a preset rule; The suspected sound is displayed through a background monitoring terminal, and a monitoring status of the suspected sound at the background monitoring terminal is obtained; the monitoring status is used to indicate whether the suspected sound is a normal sound or an abnormal sound on the edge of a rule; The monitoring status is sent to the client.

9. A live audio review device, characterized in that: Applied to a client, the device comprises: An audio receiving module, used to receive the live audio of the live room to be reviewed in the client; A confidence module is used to input the live audio into a pre-trained audio audit model, and obtain the confidence of various normal sounds and various abnormal sounds of the live audio through the audio audit model; wherein the audio audit model is obtained by fine-tuning the initial audio model after self-supervised training with sample normal sounds and sample abnormal sounds; A sound result module, for inputting the confidence of each normal sound and the confidence of each abnormal sound into a sound decision model when the confidence of the abnormal sound is greater than a preset confidence threshold, to obtain a normal sound result and / or an abnormal sound result for the live audio of the live room; wherein the abnormal sound result includes a sound type of the abnormal sound; Among them, the sound decision model is generated by the sound abnormality prediction results and the sound abnormality actual results of the audio review model.

10. A live audio review device, characterized in that: Applied to a server, the device comprises: A sample audio module is used to obtain sample live audio; and input the sample live audio into a pre-trained audio review model to obtain the confidence of various candidate normal sounds and the confidence of various candidate abnormal sounds; A labeling data module, used to obtain actual sound labeling data for each candidate abnormal sound; wherein the actual sound labeling data is used to indicate whether the candidate abnormal sound is an abnormal sound or a normal sound; A decision model module, for generating a sound decision model by using the actual sound annotation data, the confidence of each candidate normal sound and the confidence of each candidate abnormal sound; and sending the sound decision model to the client; The client is used to obtain a normal sound result and / or an abnormal sound result for the live audio of the live broadcast room based on the sound decision model.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 or 7 to 9 are implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 or 6 to 8 are implemented.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 or 6 to 8 are implemented.

Images