Audio identification method and device, equipment and medium

Abstract

The invention relates to an audio identification method and device in the technical field of social contact, equipment and a medium, and the method comprises the steps: responding to a message interception event triggered by a social contact chat window, and obtaining an original audio in a message; performing segment detection processing on the original audio, determining each mute segment in the original audio, removing all the mute segments in the audio and splicing all the remaining voice segments to obtain pure voice frequency; performing framing identification processing on the pure voice frequency, and determining each voice frame in the pure voice frequency and a violation score thereof; and determining a violation mean value according to the plurality of violation scores meeting a preset condition, and when the violation mean value does not exceed a violation threshold value, sending the message to the social chat window. According to the invention, the information in the social chat can be efficiently and accurately managed and controlled, and illegal communication is strictly prohibited.

Description

Technical Field

[0001] The present application relates to the field of social technology, and in particular to an audio identification method and its corresponding apparatus, computer equipment, and computer-readable storage medium. Background Art

[0002] In recent years, the audio and video social industry has developed rapidly around the world, especially driven by the mobile Internet, the number of applications and users has surged, forming a huge user scale. Due to the increasing supervision of online audio and video content, a series of regulations have been introduced, requiring platforms to strengthen content review, which has promoted the rapid development of audio violation identification technology, which is an important guarantee for maintaining a clear cyberspace and promoting the healthy development of the social industry containing voice.

[0003] In traditional technology, the neural network model for identifying whether audio and video are illegal has a high accuracy rate, but the model structure is complex, the calculation delay is large, and the resource consumption is huge. With the surge in the volume of audio and video social networking, the use of such identification technology will bring huge hardware costs and delays, which is not conducive to social scenarios with high real-time requirements.

[0004] In view of the shortcomings of traditional technology, the applicant has made corresponding explorations. Summary of the invention

[0005] The primary purpose of the present application is to solve at least one of the above problems and provide an audio identification method and its corresponding device, computer equipment, and computer-readable storage medium.

[0006] In order to meet the various objectives of this application, this application adopts the following technical solutions:

[0007] An audio identification method provided for one of the purposes of this application comprises the following steps:

[0008] In response to a message interception event triggered by the social chat window, obtaining original audio in the message;

[0009] Performing segment detection processing on the original audio to determine each silent segment in the original audio, removing all silent segments in the audio and splicing all remaining speech segments to obtain pure speech audio;

[0010] Performing frame identification processing on the pure speech audio to determine each speech frame in the pure speech audio and its violation score;

[0011] An average violation value is determined according to a plurality of violation scores that meet a preset condition, and when the average violation value does not exceed a violation threshold value, the message is sent to the social chat window.

[0012] On the other hand, an audio identification device provided to meet one of the purposes of the present application includes an event response module, a first processing module, a second processing module, and a message sending module. Among them, the event response module is used to respond to a message interception event triggered by a social chat window and obtain the original audio in the message; the first processing module is used to perform segment detection processing on the original audio, determine each silent segment in the original audio, remove all silent segments in the audio, and splice the remaining all speech segments to obtain a pure speech audio; the second processing module is used to perform frame-by-frame identification processing on the pure speech audio to determine each speech frame and its violation score in the pure speech audio; the message sending module is used to determine a violation average value according to multiple violation scores that meet a preset condition, and when the violation average value does not exceed a violation threshold, send the message to the social chat window.

[0013] On the other hand, a computer device provided to meet one of the purposes of the present application includes a central processing unit and a memory. The central processing unit is used to call and run a computer program stored in the memory to execute the steps of the audio identification method described in the present application.

[0014] On the other hand, a computer-readable storage medium provided to meet another purpose of the present application stores a computer program implemented according to the audio identification method in the form of computer-readable instructions. When the computer program is called and run by a computer, it executes the steps included in the method.

[0015] The technical solution of the present application has multiple advantages, including but not limited to the following aspects:

[0016] After the message interception is triggered, the present application first identifies and removes silent segments from the original audio, effectively eliminating the interference of invalid audio data on subsequent identification and reducing the consumption of computing resources. On this basis, a frame-by-frame identification technology is used to extract fine-grained violation features from a continuous speech stream. Combining the comparison mechanism of dynamic violation average value and preset threshold can adaptively balance the contradiction between the false positive rate and the missed detection rate: the accidental error of single-frame detection is effectively suppressed through the average value calculation of multiple-frame violation scores, and the dynamic adjustment mechanism of threshold determination takes into account the actual application scenario requirements. While ensuring the timeliness of real-time message transmission, it realizes the efficient and accurate identification of whether there is illegal content in the audio in a social environment, serving the strict supervision of audio communication on social platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present application will become obvious and easy to understand from the following description of the embodiments in conjunction with the drawings, where:

[0018] Figure 1 is a schematic flowchart of a typical embodiment of the audio identification method of the present application;

[0019] Figure 2 Schematic diagram of the model structure and inference process of an audio authentication model as an exemplary illustration of this application;

[0020] Figure 3 Principle block diagram of the audio authentication device of this application;

[0021] Figure 4 Schematic diagram of the structure of a computer device adopted by this application. Detailed implementation manners

[0022] The embodiments of the present application will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application and should not be construed as a limitation to the present application.

[0023] Those skilled in the art of the present technology can understand that unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present application means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any unit and all combinations of one or more related listed items.

[0024] Those skilled in the art of the present technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art and will not be interpreted with an idealized or overly formal meaning unless specifically defined as here.

[0025] Those skilled in the art can understand that the "client", "terminal", and "terminal device" used herein include both devices with a wireless signal receiver that only has the ability to receive and no ability to transmit, and devices with receiving and transmitting hardware that can perform two-way communication on a two-way communication link. Such devices can include: cellular or other communication devices such as personal computers and tablet computers, which have a single-line display or a multi-line display or a cellular or other communication device without a multi-line display; PCS (Personal Communications Service), which can combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant), which can include a radio frequency receiver, a pager, Internet / intranet access, a web browser, a notepad, a calendar, and / or a GPS (Global Positioning System) receiver; conventional laptop and / or palm-held computers or other devices, which are conventional laptop and / or palm-held computers or other devices with and / or including a radio frequency receiver. The "client", "terminal", and "terminal device" used herein can be portable, transportable, installed in a vehicle (air, sea, and / or land), or suitable for and / or configured to run locally, and / or run in a distributed form at any other location on the earth and / or in space. The "client", "terminal", and "terminal device" used herein can also be a communication terminal, an Internet access terminal, a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback functions, or can also be devices such as a smart TV and a set-top box.

[0026] The hardware referred to by names such as "server", "client", and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer, and is a hardware device with the necessary components disclosed by the von Neumann principle, including a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device. The computer program is stored in its memory, and the central processing unit loads the program stored in the external memory into the internal memory for execution, executes the instructions in the program, and interacts with the input / output devices to complete specific functions.

[0027] It should be noted that the concept of "server" in this application can similarly be extended to apply to server clusters. According to the network deployment principles understood by those skilled in the art, the various servers should be logically divided. Physically, these servers can either be independent of each other but can be invoked through interfaces, or be integrated into a single physical computer or a set of computer clusters. Those skilled in the art should understand this flexibility and should not be restricted by this when implementing the network deployment method of this application.

[0028] One or several technical features of this application, unless expressly specified, can either be deployed on the server and accessed by the client remotely invoking the online service interface provided by the server, or be directly deployed and run on the client for access.

[0029] The neural network models cited or possibly cited in this application, unless expressly specified, can either be deployed on a remote server and remotely invoked by the client, or be deployed on a client capable of handling the device and directly invoked. In some embodiments, when it runs on the client, its corresponding intelligence can be obtained through transfer learning to reduce the requirements for the client's hardware operating resources and avoid over-occupying the client's hardware operating resources.

[0030] All kinds of data involved in this application, unless expressly specified, can either be remotely stored on the server or stored on the local terminal device, as long as it is suitable for being invoked by the technical solution of this application.

[0031] Those skilled in the art should be aware that although the various methods of this application are described based on the same concept and thus show commonality with each other, unless otherwise specified, these methods can be executed independently. Similarly, for the various embodiments disclosed in this application, they are all proposed based on the same inventive concept. Therefore, concepts with the same expression, as well as concepts that are only appropriately transformed for convenience although the concept expressions are different, should be equivalently understood.

[0032] For the various embodiments to be disclosed in this application, unless expressly stated that there is a mutually exclusive relationship between them, the relevant technical features involved in each embodiment can be cross-combined to flexibly construct new embodiments, as long as this combination does not deviate from the creative spirit of this application and can meet the requirements in the prior art or solve certain deficiencies in the prior art. Those skilled in the art should be aware of this flexibility.

[0033] An audio authentication method of the present application can be programmed as a computer program product and implemented by running on a client or a server. For example, in an exemplary application scenario of the present application, it can be implemented by being deployed on the server of an e-commerce platform. Thereby, by accessing the interface opened after the computer program product runs, human-computer interaction can be performed with the process of the computer program product through a graphical user interface to execute this method.

[0034] Please refer to Figure 1 , in a typical embodiment of the audio authentication method of the present application, it includes the following steps:

[0035] Step S1100: Respond to the message interception event triggered by the social chat window, and obtain the original audio in the message;

[0036] The social chat window is an interface in social software, where users can communicate using information such as text, pictures, audio, and video. When a user sends a packet of audio and video messages to the social chat window, the message will first be submitted to the server. After receiving the message, the server will not directly send the message to the recipient. Instead, it will first intercept the message, authenticate the original audio in the message, and ensure that the original audio does not involve any violations before sending the message to the social chat window.

[0037] Users can send messages by recording audio or audio-video in real time, or by selecting and uploading audio or audio-video in their terminal devices, or by making a voice call or a video call, and correspondingly sending the audio or audio-video during the call as a message.

[0038] It can be understood that for the audio in the message, it is regarded as the original audio; for the audio-video in the message, the audio corresponding to its audio track can be separated as the original audio.

[0039] Step S1200: Perform segmented detection processing on the original audio, determine each silent segment in the original audio, remove all silent segments in the audio, and splice the remaining all voice segments to obtain a pure voice frequency;

[0040] According to a preset segment length and a preset segment interval, and sliding the window in chronological order to split the original audio, each audio segment therein is obtained. The window size of this sliding window is the preset segment length, and the window moving step size is the preset segment interval. It is recommended that the preset segment interval accounts for half of the preset segment length (the proportion here can also be set by those skilled in the art as needed, such as 25%), so that there is partial overlapping content between two adjacent audio segments split out. The specific preset segment length can be set by those skilled in the art as needed, such as 15 seconds.

[0041] In one embodiment, a silence detection algorithm based on short-time energy and / or short-time zero-crossing rate can be used to quickly determine whether an audio segment belongs to a silent segment. Thus, each silent segment in the original audio can be determined.

[0042] In other embodiments, a VAD algorithm (Voice Activity Detection) or a silence detection algorithm implemented by combining any one or more of a signal-to-noise ratio detection algorithm, a deep neural network, and a decoder on the basis of this algorithm can be used to determine whether an audio segment belongs to a silent segment. Thus, each silent segment in the original audio can be determined.

[0043] Step S1300: Perform frame identification processing on the pure speech audio to determine each speech frame and its violation score in the pure speech audio.

[0044] According to a preset frame length and a preset frame interval, the pure speech audio is split by sliding a window in chronological order to obtain each original frame therein. The window size of this sliding window is the preset frame length, and the window moving step size is the preset frame interval. It is recommended that the preset frame interval account for half of the preset frame length (alternatively, the ratio here can be set as required by those skilled in the art, such as 25%), so that there is some overlapping content between two adjacent audio segments split out. The specific preset frame length can be set as required by those skilled in the art, such as 40 ms. Each original frame is windowed to obtain each processed frame as a speech frame. The window recommended for processing is the Hamming window or the Hanning window.

[0045] An audio discrimination model pre-trained to a convergent state is used. The ability learned through training is to identify the violation features of the input speech frame and quantify the probability score corresponding to the fact that the speech frame contains violation content. The selection of the audio discrimination model can be any one of an audio classification model based on a convolutional neural network (such as VGGish, a CNN-Transformer hybrid architecture), a sequence modeling model based on a recurrent neural network (such as a bidirectional LSTM), a self-attention mechanism model based on Transformer (such as the AST model), and a transfer learning model based on pre-trained speech representations (such as a fine-tuned version of Wav2Vec 2.0), etc. Those skilled in the art can flexibly implement it.

[0046] In one embodiment, an audio classification model based on a convolutional neural network pre-trained to a convergent state is used as the audio discrimination model. Its inference process and pre-training method are as follows:

[0047] In a specific embodiment, the inference process of the audio discrimination model is divided into two stages: feature extraction and violation determination. First, the time-domain waveform of the speech frame is preprocessed and converted into a Mel spectrogram. This conversion process includes: performing a fast Fourier transform on the speech frame to convert the time-domain signal into a frequency-domain energy distribution, i.e., a spectrum; taking the modulus square of the complex spectrum after the fast Fourier transform to obtain a power spectrum; passing the power spectrum through a set of Mel-scale triangular filter banks to map the linear frequency scale to the Mel scale and obtain the output energy of the Mel filter bank; taking the logarithm of the output energy for dynamic range compression, and finally generating a time-frequency two-dimensional Mel spectrogram. Subsequently, this spectrogram is input into a convolutional neural network composed of multiple convolutional layers, batch normalization layers, and activation function layers. Each convolutional layer extracts local frequency-domain and time-domain features through a sliding convolutional kernel, and finally the network extracts deep semantic features. For example, during the inference process of this network, shallow convolutional kernels can capture short-time spectral patterns at the phoneme level, while deep convolutional kernels identify the acoustic features of violation keywords. Furthermore, the fully connected layer connected to the end of the convolutional neural network in this model maps the deep semantic features to a binary violation distribution, and this violation distribution is normalized and used as the violation score of the input speech frame.

[0048] The pre-training of this model adopts a supervised learning method, which specifically includes the following steps: constructing a large-scale audio training set containing violation and non-violation annotations, where the training samples of violations include speech content that violates morality and / or laws, such as abuse and inappropriate language, and the training samples of non-violations do not contain such speech content. In addition, each training sample is correspondingly labeled with a supervision label according to whether it is a violation; furthermore, perform the same preprocessing operations as the aforementioned frame splitting process on the training samples to ensure the consistency of the input features with the inference stage. Use the audio training set to train the audio discrimination model, and the model infers the predicted violation score corresponding to the training sample. Then, use the cross-entropy loss function to calculate the loss value corresponding to this predicted violation score according to the supervision label of the training sample. When this loss value exceeds the preset threshold, adjust the network parameters accordingly through backpropagation. Then, call other training samples and their supervision labels for the next round of training; after multiple rounds of training, when the loss value is reduced below the preset threshold, confirm that the model is trained to the convergence state and the training ends. Thus, during the process of training the model to the convergence state, the model gradually establishes a non-linear mapping relationship from Mel spectrogram features to violation classification by iteratively learning multiple labeled speech frames. The preset threshold can be set by those skilled in the art as needed.

[0049] Step S1400: Determine the violation mean based on multiple violation scores that meet the preset conditions. When the violation mean does not exceed the violation threshold, send the message to the social chat window.

[0050] The preset condition can be defined as selecting the top K violation scores from the results of sorting the violation scores from high to low, where K is a parameter obtained by multiplying the total number of speech frames contained in the pure speech audio by a preset ratio, and the preset ratio can be set by a person skilled in the art as needed. Then, the average of the selected violation scores is calculated to obtain the violation average.

[0051] When the average violation value does not exceed the violation threshold, the original audio in the message is judged to be compliant, the blocking status will be lifted and the message will be sent to the social chat window normally; if the average violation value exceeds the violation threshold, the content blocking mechanism will be triggered and a violation event log will be generated to record the judgment time, the user being judged, and the message judged to be in violation for traceability and user appeal.

[0052] The violation threshold can be configured as a threshold that is dynamically adjusted based on the time period, and / or the user reputation of the message sender, and / or the user attributes of the message recipient. For example, the threshold can be automatically lowered during the night period to strengthen the control of sensitive content, and / or stricter threshold standards can be applied to underage accounts.

[0053] The content blocking mechanism can be configured as needed by those skilled in the art, for example, not sending the illegal message and informing the user who submitted the message that the message he wanted to send to the social chat window contains illegal voice content.

[0054] In addition, it can also support explainable feedback on the interception decision results by constructing a visual violation score distribution map and / or a key violation frame (a speech frame selected as one that meets preset conditions) positioning function to assist manual or large language models in giving the aforementioned feedback.

[0055] It is not difficult to understand from the above embodiments that, compared with the prior art, the present application has many advantages, including at least:

[0056] This application first identifies and removes silent segments of the original audio after the message interception is triggered, effectively eliminating the interference of invalid audio data on subsequent identification, and also reducing the consumption of computing resources. On this basis, the frame identification technology is used to extract fine-grained violation features from continuous voice streams, and the dynamic violation mean and preset threshold comparison mechanism are combined to adaptively balance the contradiction between the false positive rate and the missed detection rate: the average calculation of multi-frame violation scores effectively suppresses the accidental errors of single-frame detection, while the dynamic adjustment mechanism of threshold judgment takes into account the needs of actual application scenarios. While ensuring the timeliness of real-time message transmission, it realizes the efficient and accurate identification of whether there is illegal content in the audio in a social environment, serving the strict supervision of audio communication on social platforms.

[0057] In a further embodiment, step S1300, performing frame identification processing on the pure speech audio to determine each speech frame in the pure speech audio and its violation score, comprises the following steps:

[0058] Step S1310: Split the pure speech audio according to a preset frame length and a preset frame interval, and perform windowing processing on each obtained original frame to obtain corresponding speech frames;

[0059] According to the preset frame length and the preset frame interval, and sliding the window in chronological order to split the pure speech audio accordingly, each original frame therein is obtained. The window size of this sliding window is the preset frame length, and the window moving step size is the preset frame interval. It is recommended that the preset frame interval account for half of the preset frame length (the ratio here can also be set as required by those skilled in the art, such as 25%), so that there is some overlapping content between two adjacent audio segments split out. The specific preset frame length can be set as required by those skilled in the art, such as 40 ms.

[0060] Perform windowing processing on each original frame, and use each processed frame as a speech frame respectively. The window recommended for processing is the Hamming window or the Hanning window. Windowing the original frame can make the amplitude of the signal therein gradually become zero at both ends. This gradual change is beneficial to the subsequent Fourier transform, which can make each peak on the spectrum thinner and not easily blend together, that is, reduce spectral leakage.

[0061] Step S1320: For each of the speech frames, perform convolutional pooling processing on the speech frame by using a preset audio discrimination model to extract corresponding deep semantic features;

[0062] The preset audio discrimination model is pre-trained to a converged state and has the ability to learn and determine the violation score of the input speech frame.

[0063] Please refer to Figure 2, the model structure of an audio discrimination model includes, in sequential connection order, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a second pooling layer, and a fully connected layer. The first convolutional layer includes: 16 convolutional kernels with the same parameters but different internal learning weights, the parameters of the convolutional kernels: [height: 1, width: 7, channel: 1], the stride of the convolutional kernels: [stride: 1]; the first pooling layer includes: the pooling window size: [height: 1, width: 4], the stride of the window: [stride: (1, 4)], the pooling type: max pooling; the second convolutional layer includes: 19 convolutional kernels with the same parameters but different internal learning weights, the parameters of the convolutional kernels: [height: 1, width: 10, channel: 16], the stride of the convolutional kernels: [stride: 1]; the third convolutional layer includes: 38 convolutional kernels with the same parameters but different internal learning weights, the parameters of the convolutional kernels: [height: 7, width: 1, channel: 19, padding: 3], the stride of the convolutional kernels: [stride: 1]; the fourth convolutional layer includes: 19 convolutional kernels with the same parameters but different internal learning weights, the parameters of the convolutional kernels: [height: 7, width: 1, channel: 38, padding: 3], the stride of the convolutional kernels: [stride: 1]; the second pooling layer includes: the pooling window size: [height: 49, width: 1], the pooling type: global max pooling;

[0064] For each of the said speech frames, first preprocess the speech frame to convert it into the form of a Mel spectrogram, and then reshape it into a four-dimensional tensor (batch_size, time_steps, frequency_bins, channels), where batch_size is 256, time_step is 49, frequency_bins is 40, and channels is 1.

[0065] Input the tensor into the audio discrimination model. The first convolutional layer of the model performs a convolution operation on the audio signal in the time dimension and uses 16 convolutional kernels with different internal learning weights to perform a convolution operation in the frequency dimension, capable of extracting local features of the audio in the frequency domain, such as energy changes and frequency components within a specific frequency range. Since the height of the convolutional kernel in the time dimension is 1, the information in the time dimension remains unchanged, and the focus is mainly on the patterns in the frequency domain;

[0066] Subsequently, through the first pooling layer of the model, based on the information of 16 output channels in the previous layer, downsampling is performed in the frequency dimension to retain the most important features (maximum values) in the frequency domain, reducing the amount of data while maintaining the robustness of the features. The time dimension remains unchanged, and the frequency dimension is compressed, making the model less sensitive to minor changes in frequency and highlighting the main frequency components;

[0067] Immediately afterwards, through the second convolutional layer of the model, more complex features are further extracted in the frequency dimension. Based on the information of 16 output channels in the previous layer, it can capture more advanced frequency patterns, such as frequency combinations and correlations between frequencies. Since the width of the convolutional kernel is relatively large (10), it can cover a wider frequency range, extract more representative frequency features, and use 19 convolutional kernels with different internal learning weights to extract advanced features in the time dimension;

[0068] Immediately afterwards, through the third convolutional layer of the model, further convolutional operations are performed in the time dimension. Based on the information of 19 output channels in the previous layer, 38 convolutional kernels with different internal learning weights are used, which can further extract more comprehensive advanced features of the audio in time, such as the frequency change trend and rhythm within a specific time window. Since the height of the convolutional kernel is 7 and there is padding, it can capture patterns within a longer time range in the time dimension while maintaining the time resolution.

[0069] Immediately afterwards, through the fourth convolutional layer of the model, convolutional operations continue to be performed in the time dimension. Based on the information of 38 output channels in the previous layer, 19 convolutional kernels with different internal learning weights are used, reducing the number of output channels to 19. It can compress and integrate the features, abstract and highlight the most important time features, and optimize the time features.

[0070] Immediately afterwards, through the second pooling layer of the model, based on the information of 19 output channels in the previous layer, global max pooling is performed in the time dimension, compressing the time features at each frequency position into a single value and retaining the most prominent time features. This can highlight the key time point features of the audio at each frequency position, such as the time when the peak appears, further reducing the amount of data and enhancing the expressive power of the features.

[0071] Step S1330: Determine the violation score based on the deep semantic feature classification mapping.

[0072] Finally, the fully connected layer in the audio authentication model integrates the information of the 19 output channels of the previous layer into a single feature representation, which is used as the deep semantic feature obtained by integrating the features extracted from the previous layers. Then, a non-linear activation function, such as the sigmoid function, is applied to map the deep semantic feature to the binary classification category space, and the violation score of the input speech frame is output, which quantifies the probability corresponding to the fact that the speech frame contains violation content.

[0073] Those skilled in the art can flexibly implement the pre-training of the model to the convergence state according to the inference process of the audio authentication model disclosed above. In one embodiment, during the training process of the model, a strategy for dynamically adjusting the learning rate and dropout probability can be designed to prevent overfitting. In addition, in order to ensure the diversity of the training samples corresponding to the training set prepared for training the model, based on the corresponding violation audios covering different scenarios, languages, and tones, and the corresponding non-violation audios, the corresponding audio frames are selected and constructed for training the model to improve its robustness; and for each training sample, whether it is a violation is marked with a corresponding supervision label.

[0074] In this embodiment, the inference process of the audio authentication model is disclosed, which can efficiently and accurately determine each speech frame and its violation score in the pure speech audio.

[0075] In a further embodiment, in step S1320, a preset audio authentication model is used to perform convolutional pooling on the speech frame to extract the corresponding deep semantic features, including:

[0076] In step S1321, after convolving the speech frame, a preset activation function is used to activate the convolution result, and the corresponding activation result is used for subsequent processing. The preset activation function is used to take the maximum value of two values, and the two values include the original parameter value and the value obtained by weighting the difference between the sine value and the cosine value corresponding to the original parameter value.

[0077] The preset activation function, an exemplary formula is as follows:

[0078] Ac(x) = max(λ * (sin(x) - cos(x)), x)

[0079] Where: x represents a single parameter in the feature map output after convolution, Ac(x) represents the result after non-linear activation of the parameter, max(λ * (sin(x) - cos(x)), x) represents taking the maximum value of the two parameters separated by a comma in the input parameters of the maximum value function, and λ represents an adjustment weight coefficient, which can be preset by those skilled in the art according to prior knowledge, for example: 0.0025, or set as a learnable parameter.

[0080] It can be understood that the core of the preset activation function lies in the dual-path screening of the convolution result: the original parameter retention path: directly passing the input value x, similar to the linear characteristic of ReLU, to retain important features; the trigonometric function modulation path: introducing a non-linear transformation through λ*(sin(x) - cos(x)), where λ controls the weight. Since The output of this path shows periodic oscillation and is sensitive to the time-frequency characteristics (such as pitch, harmonics) of the audio signal. When λ*(sin(x) - cos(x)) > x, the activation value is dominated by the trigonometric function modulation path; otherwise, it degrades to linear transmission, that is, the original parameter retention path is dominant, similar to the sparsity of ReLU.

[0081] In this embodiment, it is revealed that after convolving the speech frames, introducing non-linear changes can map the convolution result to a non-linear space, enabling the model to have the ability to fit complex non-linear relationships, so that it can learn and represent more complex patterns and features. Moreover, by implementing dual-path screening, dynamic non-linear enhancement can be achieved through the trigonometric function modulation path, that is, different non-linear responses can be given to inputs in different intervals, which can improve the model's representation ability for complex speech features (such as formants, fundamental frequencies). Compared with the zero-gradient region of the traditional ReLU, the gradient of this function contains λ*(sin(x) - cos(x)) when λ*(sin(x) - cos(x)) > x, which can alleviate the problem of gradient disappearance.

[0082] In a further embodiment, after step S1400, determining the violation mean according to multiple violation scores that meet the preset conditions, the following steps are included:

[0083] Step S1500, when the violation mean exceeds the violation threshold, determine each speech frame corresponding to the difference between the violation score and the violation mean exceeding the preset threshold;

[0084] At this time, it means that the original audio is in violation. For this, further locate the specific speech frames that cause the violation to support subsequent interpretability feedback. Subtract the violation mean from the violation score of each speech frame to obtain the corresponding differences, and screen out each speech frame with a difference exceeding the preset threshold. Such speech frames have the significant violation characteristic that the violation score is much higher than the violation mean. The preset threshold can be dynamically adjusted by those skilled in the art according to actual needs. For example, a more stringent standard can be set for high-risk users or specific periods (such as at night), that is, the threshold is preset to a relatively lower value.

[0085] Step S1510, perform speech masking processing on each speech frame in the original audio, and send the processed message to the social chat window;

[0086] For each of the selected voice frames, perform voice masking processing. The audio signal in the voice frame can be set to zero to become a silent frame; or, the audio signal in the voice frame can be replaced with a preset compliant audio signal (such as a prompt ringtone: beep beep beep...), so as to cover the illegal content.

[0087] Re - splice the processed voice frame with other audio frames in the original audio to obtain compliant original audio. In this process, it is necessary to ensure time alignment and waveform continuity to avoid cliff - like jumps during playback. In this way, the message containing the original audio can be sent to the social chat window.

[0088] Step S1520: In response to the original audio playback event after the voice masking process, display an audio violation explanation pop - up window in the social chat window.

[0089] Apply a pre - built listener for monitoring front - end user operations. When the recipient of the message clicks to play the processed original audio, the listener monitors this user operation and triggers the original audio playback event. In response to this event, an audio violation explanation pop - up window is displayed in the social chat window. Among them, the position of the masked voice frame is highlighted on the audio playback progress bar, and a "Feedback on Misjudgment" entry is provided for the user to submit a review appeal. Furthermore, text is used to explain why the audio is identified as illegal and how to handle the identified illegal audio.

[0090] In this embodiment, by screening voice frames where the difference between the violation score and the violation average value exceeds a preset threshold, the system can accurately locate the key audio segments that lead to the overall violation determination, thereby achieving the masking process only for local violation frames. This not only blocks high - risk content but also maximally retains compliant voice information, avoiding a decline in user experience caused by misjudgment or over - processing. Furthermore, explaining the audio violation with a pop - up window can inform the user of the specific details and help the user understand the relevant masking process, ensuring the user's right to know.

[0091] In a further embodiment, before step S1400: Determine the violation average value based on multiple violation scores that meet the preset conditions, the following steps are included:

[0092] Step S2400: Obtain the prior violation value corresponding to the user who submitted the message;

[0093] Based on the pre-established user credit assessment rules, the total number of corresponding violations of the user in the recent period or accumulated from the beginning, and / or the average value of each violation as the severity, the historical violation behavior of the user can be quantified to obtain a numerical prior violation value. The specific duration corresponding to the recent period can be set by those skilled in the art as needed. For example, taking the total number of violations as the prior violation value; calculating the average value of the average values corresponding to each violation as the prior violation value; multiplying the average value of the average values corresponding to each violation by the total number of violations as the prior violation value.

[0094] Step S2410: When the prior violation value exceeds the tolerance upper limit value, according to the ratio between the tolerance upper limit value and the prior violation value, reduce the preset violation threshold, and use a time decay function to decay the prior violation value correspondingly as time increases.

[0095] The tolerance upper limit value is used to measure whether the prior violation value is large enough. If the prior violation value exceeds the tolerance upper limit value, it means that the prior violation value is large enough and exceeds the tolerance upper limit, and further measures need to be taken; if the prior violation value does not exceed the tolerance upper limit value, it means that the prior violation value is not large enough and has not exceeded the tolerance upper limit, and this measure is not needed temporarily. Those skilled in the art can set the tolerance upper limit value as needed according to the disclosure here.

[0096] The above measure is to divide the tolerance upper limit value by the ratio between the prior violation value exceeding this value, and multiply by the preset violation threshold. The obtained multiplication result is used as the violation threshold. Accordingly, it is easier to trigger interception during subsequent voice frame identification.

[0097] In addition, the time decay function includes any one of an exponential decay function, a Gaussian decay function, a linear decay function, a natural logarithm decay function, etc. Those skilled in the art can select one to implement according to the business and the function characteristics matching the business.

[0098] In one embodiment, taking the exponential decay function as the time decay function, an exemplary formula is as follows:

[0099] V(t) = V0 * e ―kt

[0100] Where: V(t) and V0 represent the prior violation values corresponding before and after decay; k represents the decay rate constant (controlling the weakening speed of the violation impact), and the larger the value, the faster the decay; t represents the duration between before and after decay.

[0101] In this embodiment, a closed-loop feedback system is constructed by quantitatively comparing the prior possibility of a user sending a violation voice with the tolerance upper limit in real time. When the tolerance upper limit is exceeded, it automatically enters the enhanced review mode, matches the result of reasonably evaluating the user's behavior, and correspondingly reduces the violation threshold required for identification, making it easier to intercept voice frames and improving the accuracy of violation investigation. Furthermore, based on the time decay function, a time dimension change is added to the user credit assessment to ensure the distinction between occasional violations and malicious behaviors. For example, dense violation behaviors will cause the decay curve to not be effectively smoothed, continuously triggering strict reviews; on the contrary, the reduction in violation behaviors due to the user's improvement will effectively smooth the decay function, correspondingly reducing the review intensity or even getting out of the strict review.

[0102] In a further embodiment, before step S1400 of determining the violation mean value according to multiple violation scores that meet the preset conditions, the following steps are included:

[0103] Step S3400: Screen out each voice frame corresponding to the violation score exceeding the first threshold. For every two adjacent voice frames in the original audio in terms of position, when the violation score of any one of the voice frames exceeds the first threshold, the exceeded increment is added to the violation score of the other semantic frame.

[0104] First, screen out the voice frames with violation scores exceeding the first threshold (for example, frames with scores > 0.8). Subsequently, perform incremental compensation for every two adjacent screened voice frames in the original audio: If the violation score of any one of the frames exceeds the first threshold, then the incremental part exceeding the threshold of this frame, that is, the difference obtained by subtracting the first threshold from the exceeded violation score, is added to the violation score of the frame adjacent to this frame. For example, when the violation score of adjacent frame A is 0.85, its exceeded increment of 0.05 will be added to the violation score of adjacent frame B. Since the violation characteristics often have a diffusion effect on the time axis, there is a positive correlation between the violation possibilities of adjacent frames. Through the aforementioned incremental compensation mechanism, the positive influence of the subsequent obtained violation mean value can be effectively enhanced, and the problem of feature breakage caused by frame segmentation can be reduced.

[0105] Step S3410: When the violation scores of two adjacent voice frames in terms of position both exceed the second threshold, add the minimum value of the respectively exceeded increments to the violation score of each voice frame.

[0106] Furthermore, optimize for high-confidence violation scenarios: when the violation scores of two adjacent screened speech frames both exceed a second threshold lower than the first threshold, the system will take the minimum value of the increments by which they exceed the second threshold as the two-way compensation value and append it to the violation scores of each speech frame. For example, if the violation scores of adjacent frames C and D are 0.6 and 0.58 respectively (the second threshold is 0.55), the excess increments are 0.05 and 0.02 respectively. Take the minimum value of 0.02 and append it to the scores of the two frames, so that the scores of C and D are increased to 0.62 and 0.6 respectively. In this way, three optimizations are achieved: First, strengthen the decision-making weight of consecutive high-violation frames through two-way compensation to avoid compensation bias in a single direction; second, use the minimum value as the compensation benchmark to prevent extreme values from overly interfering with the score distribution; third, by increasing the violation weight of high-score frames, provide a more significant violation feature representation for the subsequent mean calculation process, thereby enhancing the reliability and accuracy of the overall determination.

[0107] Those skilled in the art can set the first threshold and the second threshold as needed according to the above disclosure.

[0108] In this embodiment, by designing a phased violation score compensation mechanism, a double breakthrough in feature continuity and determination robustness is achieved in the field of audio violation detection. By introducing a one-way incremental compensation strategy, the technical pain point of broken violation features in traditional speech frame segmentation algorithms is effectively solved: by screening speech frames whose violation scores exceed the first threshold and appending the incremental value of the exceeded part to adjacent frames, this operation essentially simulates the diffusion effect of violation features on the time axis. Also, by introducing a two-way compensation strategy, in-depth processing is carried out for adjacent double-high violation confidence scenarios: by extracting the minimum value of the excess increments of the two and appending it in both directions, it not only avoids the excessive interference of the maximum value on the adjacent area but also strengthens the decision-making weight of consecutive violation segments.

[0109] In a further embodiment, after step S1400, determining the violation mean according to multiple violation scores that meet the preset conditions, the following steps are included:

[0110] Step S1600, obtaining the cumulative violation times corresponding to the user who submitted the message;

[0111] The cumulative violation times include the violation times accumulated recently or from the beginning, as well as the specific timestamps of each violation. The specific duration corresponding to recently can be set as needed by those skilled in the art.

[0112] Step S1610, when the cumulative violation times meet the preset conditions, restricting the corresponding social chat permissions of the user.

[0113] The social permissions of users can be dynamically adjusted by quantifying the matching degree between their history of violation behaviors and preset conditions to construct a gradient constraint mechanism. Specifically, when the cumulative number of violations of a user reaches the preset conditions, it indicates that the user has a pattern of continuously violating the platform rules, and it is necessary to increase the cost of violations by restricting the social function permissions, so as to inhibit the occurrence of repeated violation behaviors.

[0114] When specifically implemented, the preset conditions can be configured in any one or a combination of the following forms:

[0115] Threshold trigger type: Set a fixed number threshold (for example, the cumulative number of violations is 3 times). When the number of violations of a user exceeds this threshold, the permission restriction is automatically triggered;

[0116] Time window type: Define time-sensitive conditions (for example, the number of violations within the most recent 7 days ≥ 5 times). Combining with the timeliness characteristics of violation behaviors, stronger constraints are imposed on short-term intensive violation behaviors;

[0117] Weight superposition type: Different weight values are assigned according to the severity of each violation (for example, the amplitude by which the violation mean exceeds the threshold, that is, the absolute value of their subtraction). When the cumulative weight value exceeds the set upper limit, the restriction is triggered, so that a single event of a high-risk violation behavior can trigger the permission adjustment;

[0118] Dynamic attenuation type: Introduce a time decay function (such as exponential decay), so that the influence of the historical number of violations weakens over time, and only users with high-frequency violations recently are restricted in permissions, avoiding the decline in user experience caused by permanent punishment.

[0119] When the conditions are met, different permission restriction strategies will be executed according to the cumulative characteristics of violation behaviors, including but not limited to:

[0120] Message sending frequency limit: Reduce the upper limit of the number of messages that a user can send per unit time (for example, from 5 messages per minute to 1 message);

[0121] Function disabling: Prohibit the use of high-risk functions such as audio messages and video calls, and only retain the basic permission of text chat;

[0122] Temporary mute: Completely prohibit sending any messages within a set duration (such as 24 hours);

[0123] Social demotion: Restrict the visibility of the user's messages in the recipient's chat window (such as folding display or delayed push);

[0124] Account banning: For extremely malicious users (for example, the cumulative number of violations exceeds 10 times), permanently freeze the account and clear the social relationship chain.

[0125] The enforcement of permission restrictions is implemented through a server - side policy engine: when a user triggers a message - sending request, the server first queries the user's cumulative violation count and current permission status. If the preset conditions are met, the request is intercepted and an error code is returned to the user's client, and at the same time, a permission - restricted prompt is displayed on the front - end interface (such as a pop - up window explaining the reason for the restriction and the conditions for lifting the restriction). In addition, customizing the conditions for lifting the restriction is supported. For example:

[0126] Time - based lifting of restrictions: The corresponding restricted permissions are automatically restored after the expiration of the permission restriction period;

[0127] Appeal - based lifting of restrictions: The user submits a manual review request, and the restriction is lifted after being reviewed and confirmed as a misjudgment;

[0128] Learning - based lifting of restrictions: The user is forced to complete the learning of platform rules and / or pass a compliance test to restore some permissions.

[0129] In this embodiment, by constructing a dynamic coupling mechanism for historical violation behaviors and real - time permission control, the long - term behavior pattern management of social platform user governance is realized. The core of the technical solutions in steps S1600 to S1610 lies in deeply binding the user's cumulative violation count (including multi - dimensional data such as timestamps, frequencies, and severity levels) with a gradient - based permission constraint strategy, forming a closed - loop management system of "behavior record - risk assessment - dynamic response". Specifically, through the coordination of composite rules such as threshold triggering, time window, weight superposition, and dynamic attenuation in the preset conditions (for example, temporarily muting users who have violated the rules ≥ 5 times in the last 7 days, or permanently banning users whose cumulative violation weight value exceeds 10 points), two types of typical risk groups can be accurately identified: short - term high - frequency violation users and long - term latent violation users, and differential restrictions are imposed according to their behavior characteristics. From the technical implementation level, the linkage design of the server - side policy engine and the front - end prompt system (such as returning an error code and guidance for lifting the restriction when intercepting a request), through the human - machine interaction logic of "transparent violation reasons - clear path for lifting restrictions", transforms the traditional banning operation from a black - box control into an interpretable compliance - guiding process. Particularly crucial is the introduction of a dynamic attenuation model and learning - based lifting of restriction conditions (such as the violation impact value decaying according to an exponential function, or the user being forced to pass a compliance test to restore permissions), which breaks the technical limitations of traditional "permanent banning" or "one - size - fits - all" management. Finally, a full - link governance paradigm of "real - time detection - historical traceability - elastic constraint - behavior correction" is constructed, providing an innovative solution to the balance problem between user experience and compliance control for social platforms.

[0130] Please refer to Figure 3, An audio authentication device provided to meet one of the purposes of the present application is a functional embodiment of the audio authentication method of the present application. The device includes an event response module 1100, a first processing module 1200, a second processing module 1300, and a message sending module 1400. Among them, the event response module 1100 is used to respond to the message interception event triggered by the social chat window and obtain the original audio in the message; the first processing module 1200 is used to perform segment detection processing on the original audio, determine each silent segment in the original audio, remove all silent segments in the audio and splice the remaining all voice segments to obtain a pure voice frequency; the second processing module 1300 is used to perform frame-by-frame authentication processing on the pure voice frequency to determine each voice frame and its violation score in the pure voice frequency; the message sending module 1400 is used to determine the violation average value according to multiple violation scores that meet the preset conditions, and when the violation average value does not exceed the violation threshold, send the message to the social chat window.

[0131] In a further embodiment, the second processing module 1300 includes: a voice frame construction sub-module, which is used to split the pure voice frequency according to a preset frame length and a preset frame interval, and perform windowing processing on each obtained original frame to obtain the corresponding voice frame; a feature extraction sub-module, which is used for each voice frame, and uses a preset audio authentication model to perform convolutional pooling processing on the voice frame to extract the corresponding deep semantic features; a classification mapping sub-module, which is used for classification mapping based on the deep semantic features to determine the violation score.

[0132] In a further embodiment, the feature extraction sub-module includes: a convolutional activation sub-module, which is used to perform convolution on the voice frame, apply a preset activation function to activate the convolution result, and use the corresponding activation result for subsequent processing. The preset activation function is used to take the maximum value of two items, and the two items include the original parameter value and the value obtained by weighting the difference between the sine value and the cosine value corresponding to the original parameter value.

[0133] In a further embodiment, after the message sending module 1400, it includes: a voice frame determination sub-module, which is used to determine each voice frame corresponding to the difference between the violation score and the violation average value exceeding the preset threshold when the violation average value exceeds the violation threshold; a message sending sub-module, which is used to perform voice shielding processing on each voice frame in the original audio and send the processed message to the social chat window; an event response sub-module, which is used to respond to the original audio playback event after the voice shielding processing and display an audio violation explanation pop-up window in the social chat window.

[0134] In a further embodiment, before the message sending module 1400, it includes: a prior value obtaining sub-module, configured to obtain a prior violation value corresponding to the user submitting the message; an over-tolerance processing sub-module, configured to, when the prior violation value exceeds the upper tolerance limit value, reduce a preset violation threshold according to a ratio between the upper tolerance limit value and the prior violation value, and decay the prior violation value with time using a time decay function as time increases.

[0135] In a further embodiment, before the message sending module 1400, it includes: a first fine-tuning sub-module, configured to screen out each voice frame corresponding to the violation score exceeding a first threshold, and for every two adjacent voice frames in the original audio in terms of position, when the violation score of any one of the voice frames exceeds the first threshold, append the exceeded increment to the violation score of the other semantic frame; a second fine-tuning sub-module, configured to, when the violation scores of two adjacent voice frames in terms of position both exceed a second threshold, append the minimum value of the respectively exceeded increments to the violation score of each of the voice frames.

[0136] In a further embodiment, after the message sending module 1400, it includes: a times obtaining sub-module, configured to obtain the cumulative violation times corresponding to the user submitting the message; a permission restriction sub-module, configured to restrict the corresponding social chat permissions of the user when the cumulative violation times meet a preset condition.

[0137] To solve the above technical problems, an embodiment of the present application further provides a computer device. As Figure 4 shown, it is a schematic internal structure diagram of the computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected through a system bus. Among them, the computer-readable storage medium of the computer device stores an operating system, a database, and computer-readable instructions. Control information sequences can be stored in the database. When the computer-readable instructions are executed by the processor, the processor can implement an audio identification method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. Computer-readable instructions can be stored in the memory of the computer device. When the computer-readable instructions are executed by the processor, the processor can execute the audio identification method of the present application. The network interface of the computer device is used to connect and communicate with a terminal. Those skilled in the art can understand that Figure 4 the structure shown in

[0138] In this embodiment, the processor is used to execute Figure 3The specific functions of each module and its sub-modules in it, and the memory stores the program codes and various types of data required to execute the above-mentioned modules or sub-modules. The network interface is used for data transmission between user terminals or servers. The memory in this embodiment stores the program codes and data required to execute all modules / sub-modules in the audio authentication device of the present application, and the server can call the program codes and data of the server to execute the functions of all sub-modules.

[0139] The present application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to execute the steps of the audio authentication method according to any embodiment of the present application.

[0140] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments of the present application can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a magnetic disk, an optical disc, a read-only memory (ROM), or a random access memory (RAM), etc.

[0141] In summary, the present application can efficiently and accurately manage the information in social chats and strictly prohibit illegal exchanges.

[0142] Those skilled in the art of the present technology can understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can be alternated, changed, combined, or deleted. Further, the other steps, measures, and solutions in the various operations, methods, and processes discussed in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted. Further, the steps, measures, and solutions in the prior art that are the same as those disclosed in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0143] The above are only some embodiments of the present application. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

[0144] All actions of obtaining signals, information, or data in the present application are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where it is located and obtaining the authorization given by the owner of the corresponding device.

Claims

1. An audio identification method, characterized in that: The steps include: In response to a message interception event triggered by the social chat window, obtaining original audio in the message; Performing segment detection processing on the original audio to determine each silent segment in the original audio, removing all silent segments in the audio and splicing all remaining speech segments to obtain pure speech audio; Performing frame identification processing on the pure speech audio to determine each speech frame in the pure speech audio and its violation score; An average violation value is determined according to a plurality of violation scores that meet a preset condition, and when the average violation value does not exceed a violation threshold value, the message is sent to the social chat window.

2. The audio identification method according to claim 1, characterized in that: The pure speech audio is subjected to frame identification processing to determine each speech frame in the pure speech audio and its violation score, including the following steps: Splitting the pure speech audio according to a preset frame length and a preset frame interval, performing windowing processing on each of the obtained original frames to obtain corresponding speech frames; For each of the speech frames, a preset audio identification model is used to perform convolution pooling processing on the speech frame to extract corresponding deep semantic features; Based on the deep semantic feature classification mapping, a violation score is determined.

3. The audio identification method according to claim 2, characterized in that: The preset audio identification model is used to perform convolution pooling on the speech frame to extract the corresponding deep semantic features, including: After convolving the speech frame, a preset activation function is applied to activate the convolution result, and the corresponding activation result is used for subsequent processing. The preset activation function is used to take the maximum value of two numerical values, and the two numerical values ​​include the original parameter value and the value obtained by weighting the difference between the sine value and the cosine value corresponding to the original parameter value.

4. The audio identification method according to claim 1, characterized in that: After determining the average violation value according to multiple violation scores that meet the preset conditions, the following steps are included: When the violation mean exceeds the violation threshold, determining each speech frame corresponding to which the difference between the violation score and the violation mean exceeds a preset threshold; Performing voice masking processing on each voice frame in the original audio, and sending the processed message to the social chat window; In response to the original audio playback event after the voice masking processing, a pop-up window explaining the audio violation is displayed in the social chat window.

5. The audio identification method according to claim 1, characterized in that: Before determining the average violation value according to multiple violation scores that meet the preset conditions, the following steps are included: Obtaining a violation prior value corresponding to the user who submitted the message; When the violation prior value exceeds the upper tolerance limit, the preset violation threshold is lowered according to the ratio between the upper tolerance limit and the violation prior value, and the time decay function is used to decay the violation prior value accordingly as time increases.

6. The audio identification method according to claim 1, characterized in that: Before determining the average violation value according to multiple violation scores that meet the preset conditions, the following steps are included: Filter out the speech frames corresponding to the violation scores exceeding the first threshold, and for every two speech frames adjacent to each other in the original audio, when the violation score of any of the speech frames exceeds the first threshold, add the excess increment to the violation score of the other semantic frame; When the violation scores of the two adjacent speech frames both exceed the second threshold, the minimum value of the respectively exceeded increments is added to the violation score of each of the speech frames.

7. The audio identification method according to claim 1, characterized in that: After determining the average violation value according to multiple violation scores that meet the preset conditions, the following steps are included: Obtain the cumulative number of violations corresponding to the user who submitted the message; When the accumulated number of violations meets a preset condition, the corresponding social chat authority of the user is restricted.

8. An audio identification device, characterized in that: include: An event response module, used to respond to a message interception event triggered by the social chat window and obtain the original audio in the message; The first processing module is used to perform segment detection processing on the original audio, determine each silent segment in the original audio, remove all silent segments in the audio and splice all remaining voice segments to obtain pure speech audio; The second processing module is used to perform frame identification processing on the pure speech audio to determine each speech frame in the pure speech audio and its violation score; The message sending module is used to determine an average violation value according to multiple violation scores that meet preset conditions, and when the average violation value does not exceed a violation threshold, send the message to the social chat window.

9. A computer device comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images