Method and device for locating tampered audio

Abstract

The present invention provides a method and device for locating tampered audio, which can be applied to the fields of computer technology and audio signal processing technology. The method for locating tampered audio includes: performing feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected; determining a fusion feature based on the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected; and inputting the fusion feature into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

Description

Technical Field

[0001] The present invention relates to the fields of computer technology and audio signal processing technology, and more specifically, to a method and device for locating tampered audio. Background Art

[0002] With the development of artificial intelligence technology, high-quality audio synthesis methods are becoming increasingly difficult to identify, especially for partially tampered audio. Since only a small number of fragments of this type of audio have been tampered with, it is highly concealed and difficult to locate.

[0003] In one example, a feature extractor may be used to extract features from the audio to be tested, and a classifier may be trained using labels in the data, thereby locating the locally tampered audio in the audio to be tested.

[0004] In the process of realizing the concept of the present invention, the inventors found that there are at least the following problems in the related art: the efficiency and accuracy of local tampering audio positioning cannot be effectively guaranteed. Summary of the invention

[0005] In view of this, the present invention provides a method and device for locating tampered audio.

[0006] According to one aspect of the present invention, there is provided a method for locating tampered audio, comprising: performing feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected; determining a fusion feature based on the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected; and inputting the fusion feature into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

[0007] According to an embodiment of the present invention, the trained reinforcement learning network is trained in the following manner: obtaining a training sample set, wherein the training sample set includes multiple sample audios and a label sequence for each of the sample audios, the sample audios include multiple sample frames, the label sequence includes a label for each of the sample frames, and the label represents whether the sample frame has been tampered with; performing feature extraction processing on the sample audios to obtain a first sample feature and a second sample feature, wherein the first sample feature represents a speech feature of the sample audio, and the second sample feature represents a temporal self-consistency feature of the sample audio; determining a sample fusion feature based on the first sample feature and the second sample feature, wherein the dimension of the sample fusion feature is the same as the number of frames of the sample audio; inputting the sample fusion feature into the reinforcement learning network to be trained to obtain a sample positioning result, wherein the sample positioning result represents the position of the sample tampered audio in the sample audio; and, training the reinforcement learning network based on the label sequence and the sample positioning result to obtain the trained reinforcement learning network.

[0008] According to an embodiment of the present invention, the above method also includes: defining the state space and strategy space of the above reinforcement learning network to be trained; wherein the dimension of the above state space is the same as the number of frames of the above sample audio; wherein the candidate strategies in the above strategy space include at least one of the following: keeping the position unchanged, extending a first number of positions in a first direction, and extending a second number of positions in a second direction, and the above first direction is opposite to the above second direction.

[0009] According to an embodiment of the present invention, the above-mentioned sample fusion features are input into the reinforcement learning network to be trained to obtain the sample positioning result, which includes: based on the target decision function, selecting the target strategy for the current round from the multiple candidate strategies mentioned above; and, according to the target strategy for the current round, processing the above-mentioned state space to obtain the positioning result of the current round.

[0010] According to an embodiment of the present invention, the above-mentioned selection of the target strategy for the current round from the multiple candidate strategies based on the target decision function includes: determining the decision value of each of the above-mentioned candidate strategies based on the above-mentioned target decision function; and determining the candidate strategy corresponding to the maximum decision value among the multiple decision values ​​as the above-mentioned target strategy for the current round.

[0011] According to an embodiment of the present invention, the above-mentioned target decision function is determined in the following manner: according to the state loss value, determining the reward loss value for the current round, wherein the above-mentioned reward loss value for the current round is inversely proportional to the above-mentioned state loss value; according to the above-mentioned reward loss value for the current round, determining the reward loss value for each round; and, according to the above-mentioned reward loss value for the current round and the above-mentioned reward loss value for each round, determining the above-mentioned target decision function.

[0012] According to an embodiment of the present invention, the above-mentioned state space is processed according to the above-mentioned target strategy to obtain the positioning result of the current round, including: when the above-mentioned target strategy is to keep the position unchanged, the above-mentioned state space is determined as the positioning result of the above-mentioned current round; when the above-mentioned target strategy is to extend a first number of positions in a first direction, the label of at least one position corresponding to the above-mentioned first direction and the first number in the above-mentioned state space is changed to a preset value, and the positioning result of the above-mentioned current round is obtained; and when the above-mentioned target strategy is to extend a second number of positions in a second direction, the label of at least one position corresponding to the above-mentioned second direction and the second number in the above-mentioned state space is changed to the preset value, and the positioning result of the above-mentioned current round is obtained.

[0013] According to an embodiment of the present invention, the reinforcement learning network is trained according to the label sequence and the sample positioning result to obtain the trained reinforcement learning network, including: determining a state loss value according to the label sequence and the state space based on a state loss function; and, in response to the state loss value and a preset loss threshold satisfying a preset condition, determining the state space corresponding to the state loss value as the sample positioning result, and determining the reinforcement learning network corresponding to the state loss value as the trained reinforcement learning network.

[0014] According to an embodiment of the present invention, the above-mentioned feature extraction processing of the audio to be detected to obtain the first feature and the second feature includes: inputting the above-mentioned audio to be detected into a first feature extraction branch to obtain the above-mentioned first feature, wherein the above-mentioned first feature extraction branch includes a pre-trained speech model; and inputting the above-mentioned audio to be detected into a second feature extraction branch to obtain the above-mentioned second feature, wherein the above-mentioned second feature extraction branch includes a feature extraction network and a fully connected network, and the above-mentioned feature extraction network is constructed based on at least one of the following methods: wavelet transform, empirical wavelet transform, Fourier transform, and empirical mode decomposition.

[0015] According to another aspect of the present invention, there is provided a device for locating tampered audio, comprising: a feature extraction processing module, used to perform feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected; a determination module, used to determine a fusion feature based on the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected; and a positioning module, used to input the fusion feature into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

[0016] According to another aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more instructions, wherein when the one or more instructions are executed by the one or more processors, the one or more processors implement the method described in the present invention.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, on which executable instructions are stored. When the executable instructions are executed by a processor, the processor implements the method according to the present invention.

[0018] According to another aspect of the present invention, a computer program product is provided. The computer program product comprises computer executable instructions. When the computer executable instructions are executed, they are used to implement the method according to the present invention.

[0019] According to an embodiment of the present invention, by respectively extracting the first feature and the second feature of the audio to be detected, the key information of the audio to be detected can be captured in multiple dimensions. By fusing the first feature representing the speech feature and the second feature representing the temporal self-consistency feature, a fused feature with the same dimension as the number of frames of the audio to be detected can be obtained, thereby ensuring the comprehensiveness and consistency of the feature, and being able to more accurately retain the time series information of the audio to be detected during subsequent processing. On this basis, by introducing a reinforcement learning framework, that is, inputting the fused feature into a trained reinforcement learning network, the accuracy and continuity of the positioning results can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings.

[0021] Figure 1 The system architecture to which the method for locating tampered audio can be applied according to an embodiment of the present invention is schematically shown.

[0022] Figure 2The flowchart of the method for locating tampered audio according to an embodiment of the present invention is schematically shown.

[0023] Figure 3 An example schematic diagram of a process for obtaining the first feature and the second feature according to an embodiment of the present invention is schematically shown.

[0024] Figure 4 An example schematic diagram of a training process of a reinforcement learning network according to an embodiment of the present invention is schematically shown.

[0025] Figure 5 An example schematic diagram of a training process of a reinforcement learning network according to another embodiment of the present invention is schematically shown.

[0026] Figure 6 An example schematic diagram of a state space update process according to an embodiment of the present invention is schematically shown.

[0027] Figure 7 A block diagram of a device for locating tampered audio according to an embodiment of the present invention is schematically shown.

[0028] Figure 8 A block diagram of an electronic device suitable for implementing a method for locating tampered audio according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0029] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0030] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0031] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0032] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0033] In the technical solution of the present invention, the user information (including but not limited to user personal information, user image information, user device information, such as location information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, invention and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0034] In one example, the problem of locating partially tampered audio can be transformed into a frame-level audio authenticity detection task or an audio splicing boundary detection task.

[0035] The task of authenticity detection of frame-level audio is only applicable to the case where the tampered clips are fake audio generated by the model. It achieves positioning through the feature differences between real and fake audio. However, it cannot cope with the case where the tampered clips are also real audio, especially when they come from the same source as the audio in the non-tampered area, resulting in poor model robustness and low generalization ability.

[0036] For the task of detecting audio splicing boundaries, training data bias is often prone to occur. For example, in a 7-second audio clip with one tampering, there are 700 frame-level labels (i.e., 10ms for one frame and a frame shift of 10ms), of which only 2 to 4 labels may represent the splicing boundaries. This will greatly affect the model training efficiency and recognition ability.

[0037] In addition, since the tampered areas of locally tampered audio in real scenarios often have large distribution differences and there is a large amount of data of unknown classification, the core task of the locally tampered audio localization problem is to judge the consistency between the tampered area and the non-tampered area, not its absolute authenticity, the model cannot stably exert its recognition ability in practical applications.

[0038] In addition, since the detection process of the tampered audio localization method in the related art is completed frame by frame, it is easy for discontinuous intervals to appear in the detection results of the same tampered area.

[0039] In summary, relevant technologies are unable to effectively guarantee the accuracy of localizing audio tampering.

[0040] In order to at least partially solve the technical problems existing in the related art, the present invention provides a method and device for locating tampered audio, which can be applied to the fields of computer technology and audio signal processing technology. The method for locating tampered audio includes: performing feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected; determining a fusion feature based on the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected; and inputting the fusion feature into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

[0041] Figure 1 The system architecture of the method for locating tampered audio according to an embodiment of the present invention is schematically shown. It should be noted that: Figure 1 The examples shown are merely examples of system architectures to which the embodiments of the present invention may be applied, to help those skilled in the art understand the technical content of the present invention, but do not mean that the embodiments of the present invention may not be used in other devices, systems, environments or scenarios.

[0042] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0043] The user may use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only for example).

[0044] The first terminal device 101, the second terminal device 102, and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0045] The server 105 may be a server that provides various services, such as a background management server (only as an example) that provides support for websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process the received data such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0046] It should be noted that the method for locating tampered audio provided in the embodiment of the present invention can generally be executed by the server 105. Accordingly, the device for locating tampered audio provided in the embodiment of the present invention can generally be set in the server 105. The method for locating tampered audio provided in the embodiment of the present invention can also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Accordingly, the device for locating tampered audio provided in the embodiment of the present invention can also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105.

[0047] Alternatively, the method for locating tampered audio provided in the embodiment of the present invention may also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103, or may also be executed by other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the apparatus for locating tampered audio provided in the embodiment of the present invention may also be provided in the first terminal device 101, the second terminal device 102, or the third terminal device 103, or may be provided in other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0048] It should be understood that Figure 1 The number of the first terminal device, the second terminal device, the third terminal device, the network and the server in the embodiment is only for illustration. According to the implementation requirements, there may be any number of the first terminal device, the second terminal device, the third terminal device, the network and the server.

[0049] It should be noted that the sequence numbers of the operations in the following method are only used as representations of the operations for the purpose of description, and should not be regarded as representing the execution order of the operations. Unless explicitly stated, the method does not need to be executed completely in the order shown.

[0050] Figure 2 The flowchart of the method for locating tampered audio according to an embodiment of the present invention is schematically shown.

[0051] like Figure 2 As shown, the method 200 for locating tampered audio includes operations S210 to S230.

[0052] In operation S210, feature extraction processing is performed on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents a speech feature of the audio to be detected, and the second feature represents a temporal self-consistency feature of the audio to be detected.

[0053] In operation S220, a fusion feature is determined according to the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected.

[0054] In operation S230, the fused features are input into the trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

[0055] The audio to be detected refers to audio including tampered frames of unknown positions. The audio to be detected may include multiple frames arranged in chronological order. When some of the multiple frames are tampered, by processing the audio to be detected using the tampered audio positioning method provided by the present invention, a positioning result representing the position of the tampered audio in the audio to be detected can be obtained.

[0056] In one example, based on the first method, feature extraction processing can be performed on the audio to be detected to obtain a first feature that characterizes the speech feature of the audio to be detected. The first method can be configured according to actual business needs and is not limited here. For example, the first method may include at least one of the following: a feature extraction method based on Mel Frequency Cepstrum Coefficient (MFCC), a feature extraction method based on Linear Predictive Coding (LPC), a feature extraction method based on spectrogram, etc.

[0057] In another example, based on the second method, the target frequency can be extracted from the audio to be detected to obtain a second feature that characterizes the temporal self-consistency feature of the audio to be detected. The target frequency can be a frequency in the audio to be detected that can reflect the difference between natural speech and tampered speech. The temporal self-consistency feature can be understood as a similarity feature between multiple frames, which can highlight the traces of splicing to achieve the detection of tampered frames. The second method can be configured according to actual business needs and is not limited here. For example, the second method may include at least one of the following: a feature extraction method based on dynamic time warping (DTW), a feature extraction method based on cosine similarity, a feature extraction method based on correlation coefficients, and a feature extraction method based on spectral similarity.

[0058] After obtaining the first feature and the second feature, the first feature and the second feature can be fused to obtain a fused feature that combines different types of features. The fusion processing method can be configured according to actual business needs and is not limited here. For example, the fusion method can include at least one of the following: a feature fusion method based on weighted fusion, a feature fusion method based on splicing, a feature fusion method based on principal component analysis (PCA), etc.

[0059] Before executing the tampered audio localization method of the present invention, a trained reinforcement learning network can be obtained in advance by training with a training sample set, so that after obtaining the fusion feature, the fusion feature can be directly input into the trained reinforcement learning network, thereby eliminating the need for strategy selection and making decisions directly to obtain the positioning result. The trained reinforcement learning network can be obtained by using the feedback mechanism of reinforcement learning to continuously improve the model performance, and by performing multiple rounds of training using the label sequence and sample localization results. The positioning result can be a time series sequence, which includes a tampering flag for each frame, and the tampering flag is used to characterize whether the frame has been tampered with. For example, when the tampering flag is 0, it indicates that the frame has not been tampered with. Alternatively, when the tampering flag is 1, it indicates that the frame has been tampered with.

[0060] According to an embodiment of the present invention, by respectively extracting the first feature and the second feature of the audio to be detected, the key information of the audio to be detected can be captured in multiple dimensions. By fusing the first feature representing the speech feature and the second feature representing the temporal self-consistency feature, a fused feature with the same dimension as the number of frames of the audio to be detected can be obtained, thereby ensuring the comprehensiveness and consistency of the feature, and being able to more accurately retain the time series information of the audio to be detected during subsequent processing. On this basis, by introducing a reinforcement learning framework, that is, inputting the fused feature into a trained reinforcement learning network, the accuracy and continuity of the positioning results can be improved.

[0061] Reference below Figure 3~Figure 6 , the method 200 for locating tampered audio according to an embodiment of the present invention is further described.

[0062] Figure 3 An example schematic diagram of a process for obtaining the first feature and the second feature according to an embodiment of the present invention is schematically shown.

[0063] like Figure 3 As shown, in 300, the manner of performing feature extraction processing on the audio to be detected 301 may include feature extraction processing based on a first feature extraction branch and feature extraction processing based on a second feature extraction branch.

[0064] For example, the first feature extraction branch may include a pre-trained speech model 302. In this case, the audio to be detected 301 may be input into the pre-trained speech model 302 to obtain a first feature 304. The dimension of the first feature 304 is the same as the number of frames of the audio to be detected 301.

[0065] In one example, the pre-trained speech module 302 may include at least one of the following: wav2vec, HuBert, and AASIST, etc.

[0066] Alternatively, the second feature extraction branch may include a feature extraction network 303_1 and a fully connected network 303_2. In this case, the audio to be detected 301 may be input into the feature extraction network 303_1 to obtain an initial second feature 305; the initial second feature 305 may be input into the fully connected network 303_2 to obtain a second feature 306. The fully connected network 303_2 is used to perform a dimension transformation on the initial second feature 305 so that the dimension of the second feature 306 is the same as the first feature 304, that is, the dimension of the second feature 306 is the same as the number of frames of the audio to be detected 301.

[0067] In one example, the feature extraction network 301_1 may be constructed based on at least one of the following methods: wavelet transform, empirical wavelet transform, Fourier transform, empirical mode decomposition, etc. Thus, the audio to be detected 301 may be decomposed by the feature extraction network 301_1 constructed by the above method, so as to obtain the second feature 306 that can reflect the difference between natural speech and disguised speech.

[0068] According to an embodiment of the present invention, since the first feature is obtained by processing the audio to be detected using a pre-trained speech model, the speech features in the audio to be detected can be effectively extracted. Since the second feature is obtained by processing the audio to be detected using a feature extraction network and a fully connected network, the temporal self-consistency feature of the audio to be detected can be extracted. This feature focuses more on the core discriminant feature of local tampering audio positioning, and can cope with the situation where the tampered area is real audio in real scenes, which helps to improve the continuity of the recognition results and avoid the problem of discontinuities in the recognition results of the same tampered area. Therefore, by using two feature extraction branches respectively, features of different dimensions of the audio to be detected can be obtained, which is beneficial to improving the robustness and generalization ability of the model in real scene applications, and further helps to improve the positioning accuracy of the model.

[0069] Figure 4 An example schematic diagram of a training process of a reinforcement learning network according to an embodiment of the present invention is schematically shown.

[0070] like Figure 4As shown, in 400, in the process of training the reinforcement learning network, a training sample set can be obtained. The training sample set may include multiple sample audios 401 and a label sequence 402 for each sample audio 401. The sample audio 401 includes multiple sample frames. The label sequence 402 includes a label for each sample frame, and the label indicates whether the sample frame has been tampered with. For example, when the label is 0, it indicates that the sample frame has not been tampered with. Alternatively, when the label is 1, it indicates that the sample frame has been tampered with.

[0071] For each sample audio 401, feature extraction processing can be performed on the sample audio 401 to obtain a first sample feature 403 representing the speech feature of the sample audio 401 and a second sample feature 404 representing the temporal self-consistency feature of the sample audio 401. According to the first sample feature 403 and the second sample feature 404, a sample fusion feature 405 is determined. The dimension of the sample fusion feature 405 is the same as the number of frames of the sample audio 401.

[0072] The sample fusion feature 405 is input to the reinforcement learning network 406 to be trained, and a sample positioning result 407 representing the position of the sample tampered audio in the sample audio is obtained. On this basis, the reinforcement learning network 406 can be trained according to the label sequence 402 and the sample positioning result 407 to obtain a trained reinforcement learning network.

[0073] According to an embodiment of the present invention, by extracting features from the sample audio, a first sample feature and a second sample feature are obtained, so that the model can simultaneously capture the basic acoustic information of the sample audio and the similarities between different frames. Since the sample fusion feature is determined based on the first sample feature and the second sample feature, the dimension of the sample fusion feature is the same as the number of frames of the sample audio, ensuring the integrity of the feature and the retention of time series information, so that the model can accurately map and locate the tampering position. On this basis, by inputting the sample fusion feature into the reinforcement learning network to be trained, the model performance can be continuously improved with the help of the feedback mechanism of reinforcement learning, and by using the label sequence and sample positioning results to train the reinforcement learning network, the network can continuously optimize its own parameters, thereby improving the accuracy of recognition and positioning.

[0074] In one example, the state space of the reinforcement learning network 406 to be trained may be predefined. The dimension of the state space is the same as the number of frames of the sample audio 401. For example, if the sample audio 401 includes N sample frames, the dimension of the state space may be (1, N).

[0075] In another example, the strategy space of the reinforcement learning network 406 to be trained can also be predefined. The candidate strategies in the strategy space include at least one of the following: the position remains unchanged, the first number of positions are extended in the first direction, and the second number of positions are extended in the second direction, and the first direction and the second direction are opposite. The first direction, the second direction, the first number and the second number can be configured according to actual business needs and are not limited here. For example, the first direction is the left side and the second direction is the right side; the first number and the second number can be set to 1, 2, 5 and 10. In this case, all natural numbers can be obtained by combining the first number or the second number.

[0076] According to an embodiment of the present invention, by setting the state space dimension of the reinforcement learning network to be trained to be the same as the number of frames of the sample audio, it is ensured that each state can clearly correspond to a specific frame of the sample audio. In addition, by defining multiple candidate strategies in the strategy space, a variety of decision-making options can be provided for the reinforcement learning network, so that the model can select the best action according to the specific audio features and situations, thereby more flexibly responding to different tampering situations.

[0077] Figure 5 An example schematic diagram of a training process of a reinforcement learning network according to another embodiment of the present invention is schematically shown.

[0078] like Figure 5 As shown, in 500, the sample fusion feature 503 of the sample audio 501 can be input into the reinforcement learning network 504, so that the reinforcement learning network 504 performs multiple rounds of reinforcement learning on the state space to obtain the sample positioning result.

[0079] For each round, a state loss value 507 can be determined based on a state loss function 506 according to the label sequence 502 and the state space 505 of the current round. The state loss function 506 refers to a function used to evaluate or quantify the gap between the performance of the model in a specific state and the target. The specific form of the state loss function 506 can be configured according to actual business needs and is not limited here. For example, the state loss function 506 may include at least one of the following: value function loss, strategy loss, imitation learning loss, and entropy loss. In an example, the state loss function 506 may be shown as follows (1).

[0080]

[0081] in, represents the i-th position, Characterizes the tag at the i-th position of the tag sequence 502, Represents the character at the i-th position of the state space of the current round, Characterize the state loss function. When the label at the i-th position of the label sequence 502 is not equal to the character at the i-th position of the state space of the current round, it means that the position contributes to the Hamming distance; when the label at the i-th position of the label sequence 502 is not equal to the character at the i-th position of the state space of the current round, it means that the position does not contribute to the Hamming distance. Based on this, the state loss value 507 can be determined according to the sum of each position.

[0082] After obtaining the state loss value 507 of each round, it can be determined whether it is necessary to continue to perform the next round of iterative reinforcement according to whether the state loss value 507 and the preset loss threshold meet the preset condition. The preset condition can be configured according to actual business needs and is not limited here. For example, the preset condition can be that the state loss value 507 is equal to the preset loss threshold.

[0083] In one example, when the state loss value 507 and the preset loss threshold do not meet the preset conditions, it means that it is necessary to continue the next round of iterative reinforcement. Then, based on the reward loss function 508 and the state loss value 507, the reward loss value 509 for the current round can be determined, and the reinforcement learning network 504 can be guided to make the next round of decision-making based on the reward loss value 509 of the current round.

[0084] The state loss function 506 and the reward loss function 508 may be inversely proportional, that is, the larger the state loss value 507 is, the smaller the reward loss value 509 is. The reward loss function 508 refers to a function used to evaluate the gap between the reward obtained by the action taken by the agent in a specific state and the expected reward. The specific form of the reward loss function 508 can be configured according to actual business needs and is not limited here. For example, the reward loss function 508 may include at least one of the following: expected return loss, advantage function, strategy loss, and entropy loss. In one example, the reward loss function 508 may be shown as follows (2).

[0085]

[0086] in, Characterizing the reward loss function 508, Characterizes the number of frames of the sample audio 501 or the dimension of the sample fusion feature 503.

[0087] In another example, when the state loss value 507 and the preset loss threshold meet the preset conditions, it means that there is no need to continue the next round of iterative reinforcement. In this case, the state space corresponding to the state loss value 507 can be determined as the sample positioning result, and the reinforcement learning network corresponding to the state loss value 507 can be determined as the trained reinforcement learning network.

[0088] According to an embodiment of the present invention, the state loss value is calculated based on the state loss function, which can effectively quantify the gap between the performance of the network in a specific state and the expectation. Through the real-time feedback mechanism that compares the state loss value with the preset loss threshold, the reinforcement learning network can dynamically adjust its behavior. On this basis, by determining the state space when the state loss value meets the preset conditions as the sample positioning result, the state related to the tampering position can be effectively screened out in the high-dimensional state space, which can improve the accuracy of the model audio tampering detection.

[0089] Figure 6 An example schematic diagram of a state space update process according to an embodiment of the present invention is schematically shown.

[0090] like Figure 6 As shown, in 600, the sample fusion feature 603 of the sample audio 601 can be input into the reinforcement learning network 604, so that the reinforcement learning network 604 performs multiple rounds of reinforcement learning on the strategy space 602 to obtain the sample positioning result. The following takes the sample audio 601 including 10 sample frames as an example to illustrate the reinforcement learning process.

[0091] In the first round, based on the sample fusion feature 603, one or several initial decision positions can be selected in the state space 602 of dimension (1, 10) and the initial decision positions are assigned 1, and the remaining positions are assigned 0. The initial decision position can be configured according to actual business needs and is not limited here. For example, the selection method of the initial decision position may include at least one of the following: a method based on random selection and a method of inferring based on the first sample feature and the second sample feature. For example, the initial decision position is the 4th and 6th positions.

[0092] In the nth round, any target strategy for the current round can be selected from multiple candidate strategies in the strategy space, and the state space 602 is processed according to the target strategy for the current round to obtain the positioning result of the current round. The selection method of the target strategy for the current round can be configured according to actual business needs and is not limited here. For example, the selection method of the target strategy can include at least one of the following: a selection method based on maximizing value, a selection method based on a model, a selection method based on a policy gradient method, a selection method based on a distance metric, and a selection method based on an integrated strategy.

[0093] Selection methods based on maximizing value refer to selecting the action with the highest expected reward. Selection methods based on models refer to building a model of the environment and selecting actions based on the results predicted by the model. Selection methods based on policy gradient methods refer to learning the target policy by maximizing the expected value of the cumulative reward. Selection methods based on distance metrics refer to selecting the policy closest to the target by comparing the distance between the current policy and the best known policy. Selection methods based on integrated strategies refer to integrating multiple strategies and selecting the best action.

[0094] In one example, the decision value of each candidate strategy can be determined based on the target decision function, and the candidate strategy corresponding to the maximum decision value among the multiple decision values ​​can be determined as the target strategy for the current round. The target decision function refers to a function used to select a strategy, and the target decision function can maximize the sum of future decisions, that is, the target strategy selected in each round is better than the reinforcement learning of subsequent rounds.

[0095] The specific form of the target decision function can be configured according to actual business needs and is not limited here. In one example, the reward loss value for the current round can be determined based on the state loss value, and the reward loss value for each round can be determined based on the reward loss value for the current round. On this basis, the target decision function can be determined based on the reward loss value for the current round and the reward loss value for each round. For example, the target decision function can be shown in the following equations (3) and (4).

[0096]

[0097]

[0098] in, Represents the reward loss value for the tth round, Represents the reward loss value for the t+1th round, Represents the reward loss value for the t+2th round, and so on. Represents the sum of the reward loss values ​​of all rounds, Characterize the target decision function.

[0099] After obtaining the target strategy for the current round, the state space can be processed based on the target strategy for the current round to obtain the state space of the current round. Processing the state space 602 according to the target strategy for the current round means executing the target strategy for the current round on the state space 602 so as to change the labels of the extended positions in the state space 602 to 1, and the labels of the other positions remain unchanged, thereby obtaining the state space of the current round.

[0100] According to an embodiment of the present invention, by calculating the decision value of the candidate strategy based on the target decision function, a quantitative evaluation of each candidate strategy is achieved. On this basis, by determining the candidate strategy corresponding to the maximum decision value as the target strategy, it is ensured that the most promising strategy is selected in each decision-making process, thereby optimizing the performance of the model in the audio tampering detection task.

[0101] For example, when the target strategy for the current round is to keep the position unchanged, the state space 602 can be determined as the state space 605 of the current round, and the state space 605 of the current round can be determined as the positioning result of the current round.

[0102] Alternatively, when the target strategy for the current round is to extend a first number of positions in a first direction, the first direction is upward, and the first number is 2, the label of at least one position corresponding to the first direction and the first number in the state space 602 can be changed to a preset value, that is, the label of the two upward positions is changed to 1, to obtain the state space 606 of the current round, and the state space 606 of the current round is determined as the positioning result of the current round.

[0103] Alternatively, when the target strategy for the current round is to extend a second number of positions in the second direction, the second direction is downward, and the second number is 2, the label of at least one position corresponding to the second direction and the second number in the state space 602 can be changed to a preset value, that is, the label of the two downward positions is changed to 1, to obtain the state space 607 of the current round, and the state space 607 of the current round is determined as the positioning result of the current round.

[0104] According to an embodiment of the present invention, when the target strategy for the current round is to keep the position unchanged, by determining the state space as the positioning result of the current round, it is ensured that the positioning result under specific conditions is stable and repeatable, which helps to improve the reliability and accuracy of the detection or positioning task. When the target strategy for the current round is to extend a specific position along the first direction or the second direction, by adjusting the relevant labels in the label sequence to preset values, the consistency between the label and the state space is ensured. By dynamically adjusting the label sequence, the label data can reflect the real-time state space changes, enhance data consistency, help reduce the model's erroneous judgment during training and reasoning, and improve the accuracy of the positioning results.

[0105] The above are only exemplary embodiments, but are not limited thereto, and may also include other tampered audio positioning methods known in the art, as long as they can improve the accuracy and continuity of the positioning results.

[0106] Figure 7A block diagram of a device for locating tampered audio according to an embodiment of the present invention is schematically shown.

[0107] like Figure 7 As shown, the device 700 for locating tampered audio may include a feature extraction processing module 710 , a determination module 720 and a positioning module 730 .

[0108] The feature extraction processing module 710 is used to perform feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected.

[0109] The determination module 720 is used to determine a fusion feature according to the first feature and the second feature, wherein the dimension of the fusion feature is the same as the number of frames of the audio to be detected.

[0110] The positioning module 730 is used to input the fused features into the trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected.

[0111] According to the embodiments of the present invention, any multiple of the modules, or at least part of the functions of any multiple of them, can be implemented in one module. According to the embodiments of the present invention, any one or more of the modules can be split into multiple modules for implementation. According to the embodiments of the present invention, any one or more of the modules can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware in any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in any appropriate combination of any of them. Alternatively, according to the embodiments of the present invention, one or more of the modules can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0112] For example, any multiple of the feature extraction processing module 710, the determination module 720 and the positioning module 730 can be combined in one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module / unit / subunit. According to an embodiment of the present invention, at least one of the feature extraction processing module 710, the determination module 720 and the positioning module 730 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in any appropriate combination of any of them. Alternatively, at least one of the feature extraction processing module 710, the determination module 720 and the positioning module 730 can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0113] It should be noted that the part of the positioning device for tampering with audio in the embodiment of the present invention corresponds to the part of the positioning method for tampering with audio in the embodiment of the present invention. The description of the part of the positioning device for tampering with audio specifically refers to the part of the positioning method for tampering with audio, which will not be repeated here.

[0114] Figure 8 A block diagram of an electronic device suitable for implementing a method for locating tampered audio according to an embodiment of the present invention is schematically shown. Figure 8 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0115] like Figure 8 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 to a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An I / O (input / output) interface 805 is also connected to the bus 804.

[0116] Multiple components in the electronic device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the electronic device 800 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0117] The computing unit 801 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above, such as the positioning method of tampering with audio. For example, in some embodiments, the positioning method based on tampering with audio may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the positioning method of tampering with audio described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to execute the method for locating tampered audio in any other appropriate manner (eg, by means of firmware).

[0118] According to an embodiment of the present invention, the method flow according to an embodiment of the present invention can be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product, which includes a computer program carried on a computer-readable storage medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication unit 809. When the computer program is executed by the computing unit 801, the above-mentioned functions defined in the system of the embodiment of the present invention are executed. According to an embodiment of the present invention, the system, equipment, device, module, unit, etc. described above can be implemented by a computer program module.

[0119] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiment; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present invention is implemented.

[0120] According to an embodiment of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium. For example, it may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, an apparatus or a device.

[0121] For example, according to an embodiment of the present invention, the computer-readable storage medium may include the ROM 802 and / or the RAM 803 described above and / or one or more memories other than the ROM 802 and the RAM 803 .

[0122] An embodiment of the present invention also includes a computer program product, which includes a computer program, which contains program code for executing the method provided by the embodiment of the present invention. When the computer program product runs on an electronic device, the program code is used to enable the electronic device to implement the method for locating tampered audio provided by the embodiment of the present invention.

[0123] When the computer program is executed by the computing unit 801, the above functions defined in the system / device of the embodiment of the present invention are executed. According to the embodiment of the present invention, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0124] In one embodiment, the computer program may rely on tangible storage media such as optical storage devices, magnetic storage devices, etc. In another embodiment, the computer program may also be transmitted and distributed in the form of signals on a network medium, and downloaded and installed through the communication unit 809. The program code included in the computer program may be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0125] According to an embodiment of the present invention, the program code for executing the computer program provided by the embodiment of the present invention can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0126] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions. It can be understood by those skilled in the art that the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways, even if such a combination or combination is not explicitly recorded in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways. All these combinations and / or combinations fall within the scope of the present invention.

[0127] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A method for locating tampered audio, characterized in that: include: Performing feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents a speech feature of the audio to be detected, and the second feature represents a temporal self-consistency feature of the audio to be detected; Determine a fusion feature according to the first feature and the second feature, wherein a dimension of the fusion feature is the same as the number of frames of the audio to be detected; and Inputting the fused features into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected; The trained reinforcement learning network is trained in the following manner: Acquire a training sample set, wherein the training sample set includes a plurality of sample audios and a label sequence of each of the sample audios, the sample audios include a plurality of sample frames, the label sequence includes a label of each of the sample frames, and the label indicates whether the sample frame has been tampered with; Performing feature extraction processing on the sample audio to obtain a first sample feature and a second sample feature, wherein the first sample feature represents a speech feature of the sample audio, and the second sample feature represents a temporal self-consistency feature of the sample audio; Determine a sample fusion feature according to the first sample feature and the second sample feature, wherein the dimension of the sample fusion feature is the same as the number of frames of the sample audio; Inputting the sample fusion feature into the reinforcement learning network to be trained to obtain a sample positioning result, wherein the sample positioning result represents the position of the sample tampered audio in the sample audio; and The reinforcement learning network is trained according to the label sequence and the sample positioning result to obtain the trained reinforcement learning network.

2. The method according to claim 1, characterized in that The method further comprises: Defining the state space and policy space of the reinforcement learning network to be trained; Wherein, the dimension of the state space is the same as the number of frames of the sample audio; The candidate strategies in the strategy space include at least one of the following: keeping the position unchanged, extending a first number of positions in a first direction, and extending a second number of positions in a second direction, and the first direction and the second direction are opposite.

3. The method according to claim 2, characterized in that The step of inputting the sample fusion features into the reinforcement learning network to be trained to obtain the sample positioning result includes: Based on the target decision function, selecting a target strategy for the current round from the plurality of candidate strategies; and The state space is processed according to the target strategy for the current round to obtain the positioning result of the current round.

4. The method according to claim 3, characterized in that The selecting a target strategy for the current round from a plurality of candidate strategies based on the target decision function includes: Based on the target decision function, determining a decision value of each candidate strategy; and The candidate strategy corresponding to the maximum decision value among the multiple decision values ​​is determined as the target strategy for the current round.

5. The method according to claim 4, characterized in that The target decision function is determined as follows: Determine a reward loss value for the current round according to the state loss value, wherein the reward loss value for the current round is inversely proportional to the state loss value; Determining a reward loss value for each round based on the reward loss value for the current round; and The target decision function is determined according to the reward loss value for the current round and the reward loss value for each round.

6. The method according to claim 3, characterized in that The processing of the state space according to the target strategy to obtain the positioning result of the current round includes: When the target strategy is to keep the position unchanged, determining the state space as the positioning result of the current round; When the target strategy is to extend a first number of positions in a first direction, changing the label of at least one position corresponding to the first direction and the first number in the state space to a preset value, to obtain the positioning result of the current round; and When the target strategy is to extend a second number of positions in a second direction, the label of at least one position corresponding to the second direction and the second number in the state space is changed to the preset value to obtain the positioning result of the current round.

7. The method according to any one of claims 2 to 6, characterized in that The step of training the reinforcement learning network according to the tag sequence and the sample positioning result to obtain the trained reinforcement learning network includes: Based on the state loss function, determining a state loss value according to the label sequence and the state space; and In response to the state loss value and the preset loss threshold satisfying a preset condition, the state space corresponding to the state loss value is determined as the sample positioning result, and the reinforcement learning network corresponding to the state loss value is determined as the trained reinforcement learning network.

8. The method according to claim 1, characterized in that The feature extraction process of the audio to be detected to obtain the first feature and the second feature includes: Inputting the audio to be detected into a first feature extraction branch to obtain the first feature, wherein the first feature extraction branch includes a pre-trained speech model; and The audio to be detected is input into a second feature extraction branch to obtain the second feature, wherein the second feature extraction branch includes a feature extraction network and a fully connected network, and the feature extraction network is constructed based on at least one of the following methods: wavelet transform, empirical wavelet transform, Fourier transform, and empirical mode decomposition.

9. A device for locating tampered audio, characterized in that: include: A feature extraction processing module, used to perform feature extraction processing on the audio to be detected to obtain a first feature and a second feature, wherein the first feature represents the speech feature of the audio to be detected, and the second feature represents the temporal self-consistency feature of the audio to be detected; a determining module, configured to determine a fusion feature according to the first feature and the second feature, wherein a dimension of the fusion feature is the same as the number of frames of the audio to be detected; and A positioning module, used for inputting the fusion feature into a trained reinforcement learning network to obtain a positioning result, wherein the positioning result represents the position of the tampered audio in the audio to be detected; The trained reinforcement learning network is trained in the following manner: Acquire a training sample set, wherein the training sample set includes a plurality of sample audios and a label sequence of each of the sample audios, the sample audios include a plurality of sample frames, the label sequence includes a label of each of the sample frames, and the label indicates whether the sample frame has been tampered with; Performing feature extraction processing on the sample audio to obtain a first sample feature and a second sample feature, wherein the first sample feature represents a speech feature of the sample audio, and the second sample feature represents a temporal self-consistency feature of the sample audio; Determine a sample fusion feature according to the first sample feature and the second sample feature, wherein the dimension of the sample fusion feature is the same as the number of frames of the sample audio; Inputting the sample fusion feature into the reinforcement learning network to be trained to obtain a sample positioning result, wherein the sample positioning result represents the position of the sample tampered audio in the sample audio; and The reinforcement learning network is trained according to the label sequence and the sample positioning result to obtain the trained reinforcement learning network.

Images