Cross-domain playback voice detection method, device, equipment and medium

Abstract

The embodiments of the present application provide a cross-domain playback speech detection method, apparatus, device and medium, the method comprising: obtaining playback speech samples of multiple speech domains to obtain a first playback speech and a second playback speech; determining a first acoustic feature and a second acoustic feature; determining a first general feature corresponding to the first acoustic feature and a second general feature corresponding to the second acoustic feature based on a general feature extraction network; obtaining a first reconstructed signal corresponding to the first general feature and a second reconstructed signal corresponding to the second general feature based on a reconstruction network, and determining a first loss and a second loss to obtain a target loss; jointly training the general feature extraction network and the reconstruction network according to the target loss; determining a playback detection result of the speech to be tested based on a cross-domain playback speech detection model; according to the embodiments of the present application, the generalization of the cross-domain playback speech detection model can be improved, and the accuracy of cross-domain playback speech detection can be improved.

Description

Technical Field

[0001] The present application relates to, but is not limited to, the field of speech recognition technology, and in particular to a cross-domain playback speech detection method, device, equipment, and medium. Background Art

[0002] Speaker verification systems are usually deployed with recording playback detection models, which can detect whether the speech input into the speaker verification system is non-playback speech generated in real time from a real sound source, or a pre-recorded playback recording. Speaker verification systems are widely used in the field of social biometrics and play a vital role in improving security.

[0003] At present, in the playback speech detection scenario, there is a problem of speech domain offset in recording playback detection, for example, speech domain offset between different recording devices, different playback devices or different recording environments; the recording playback detection model is trained based on sample speech in one of the speech domains, and when the trained recording playback detection model is used to detect the test speech in the same speech domain as the sample speech, good results can be achieved; however, when the trained recording playback detection model is used to detect the test speech in a different speech domain from the sample speech, its detection performance will be significantly reduced; it can be seen that the generalization performance of the existing recording playback detection model is poor, and the accuracy of cross-domain recording playback detection is low. Summary of the invention

[0004] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0005] The embodiments of the present application provide a cross-domain playback voice detection method, apparatus, device and medium, which can improve the generalization of the cross-domain playback voice detection model, thereby improving the accuracy of cross-domain recording playback detection.

[0006] To achieve the above-mentioned purpose, the first aspect of an embodiment of the present application proposes a cross-domain playback speech detection method, including: obtaining playback speech samples of multiple speech domains, taking any one of the speech domains as the target domain, taking the playback speech sample corresponding to the target domain as the first playback speech, and taking the other playback speech samples except the first playback speech as the second playback speech, wherein each of the speech domains corresponds to its own reconstruction network; determining a first acoustic feature according to the first playback speech, and determining a second acoustic feature according to the second playback speech; determining a first general feature corresponding to the first acoustic feature based on a preset general feature extraction network, and determining a second general feature corresponding to the second acoustic feature based on the general feature extraction network; respectively extracting the first general feature and the second general feature based on the reconstruction network corresponding to the target domain. The method comprises the steps of: reconstructing a first universal feature and a second universal feature to obtain a first reconstructed signal corresponding to the first universal feature, and obtaining a second reconstructed signal corresponding to the second universal feature; determining a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and determining a second loss according to the second reconstructed signal and the corresponding second acoustic feature; performing weighted summation on the first loss and the second loss to obtain a target loss, and jointly training the universal feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss; obtaining a speech to be tested, inputting the speech to be tested into a trained cross-domain playback speech detection model, and determining a playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes the universal feature extraction network.

[0007] In some embodiments, after the general feature extraction network and the reconstruction network corresponding to the target domain are jointly trained according to the target loss, the cross-domain playback speech detection method further includes: updating the target domain to a speech domain corresponding to any untrained reconstruction network; and jointly training the general feature extraction network and the reconstruction network corresponding to the target domain again until each of the reconstruction networks is trained.

[0008] In some embodiments, determining a first acoustic feature based on the first playback speech and determining a second acoustic feature based on the second playback speech include: performing frequency domain transform processing on the first playback speech to obtain a first frequency domain signal, calculating the power spectrum of the first frequency domain signal to obtain a first power spectrum, and calculating the logarithm of the first power spectrum to obtain the first acoustic feature; performing frequency domain transform processing on the second playback speech to obtain a second frequency domain signal, calculating the power spectrum of the second frequency domain signal to obtain a second power spectrum, and calculating the logarithm of the second power spectrum to obtain the second acoustic feature.

[0009] In some embodiments, the speech domain corresponding to the second playback speech is used as the auxiliary domain, and the target loss is calculated as follows:

[0010]

[0011] Wherein, L is the target loss, M is the total number of the speech domains, l0 is the first loss, ω0 is the weight corresponding to the first loss, l m is the second loss corresponding to the mth auxiliary domain, ω m for l m The corresponding weights;

[0012] The calculation formula of l0 is as follows:

[0013]

[0014] in, is the i-th first acoustic feature, for corresponding to the first reconstructed signal, n0 is the total number of the first acoustic features;

[0015] The l m The calculation formula is as follows:

[0016]

[0017] in, is the kth second acoustic feature corresponding to the mth auxiliary domain, for The corresponding second reconstructed signal, n m is the total number of the second acoustic features corresponding to the mth auxiliary domain.

[0018] In some embodiments, the speech domain corresponding to the second playback speech is used as an auxiliary domain, and the number of the auxiliary domains is multiple. The first loss and the second loss are weightedly summed to obtain a target loss, including: obtaining the domain similarity between each of the auxiliary domains and the target domain; determining the first weight corresponding to the target domain and the second weight corresponding to each of the auxiliary domains according to each of the domain similarities, wherein the sum of the first weight and all of the second weights is equal to one, and the second weight is positively correlated with the corresponding domain similarity; using the first weight as the weight of the first loss, using each of the second weights as the weight of the second loss corresponding to each of the second losses, weighted summing the first loss and the second loss to obtain a target loss.

[0019] In some embodiments, the speech domain corresponding to the second playback speech is used as an auxiliary domain, and the first loss and the second loss are weightedly summed to obtain a target loss, including: obtaining the first recording scene information of the target domain and the second recording scene information of each of the auxiliary domains; performing feature extraction on the first recording scene information to obtain a first scene feature; performing feature extraction on each of the second recording scene information to obtain a second scene feature corresponding to each; inputting the first scene feature and each of the second scene features into a prediction network for prediction to obtain a second weight of each of the auxiliary domains; determining the first weight of the target domain according to the second weight of each of the auxiliary domains, wherein the sum of the first weight and the second weight is equal to one; using the first weight as the weight of the first loss, and using each of the second weights as the weight of the second loss corresponding to each of the second losses, weighted summing the first loss and the second loss to obtain a target loss.

[0020] In some embodiments, the cross-domain playback speech detection model also includes a classifier, and determining the playback detection result of the speech to be tested based on the cross-domain playback speech detection model includes: determining the target acoustic feature according to the speech to be tested, and determining the target universal feature corresponding to the target acoustic feature based on the universal feature extraction network; inputting the target universal feature into the classifier to obtain the playback detection result of the speech to be tested, wherein the playback detection result is used to indicate whether the speech type of the speech to be tested is a playback type or a non-playback type.

[0021] To achieve the above-mentioned purpose, the second aspect of an embodiment of the present application proposes a cross-domain playback speech detection device, including: an acquisition module, used to acquire playback speech samples of multiple speech domains, taking any one of the speech domains as a target domain, taking the playback speech sample corresponding to the target domain as a first playback speech, and taking the other playback speech samples except the first playback speech as a second playback speech, wherein each of the speech domains corresponds to its own reconstruction network; an acoustic feature determination module, used to determine a first acoustic feature according to the first playback speech, and to determine a second acoustic feature according to the second playback speech; a general feature extraction module, used to determine a first general feature corresponding to the first acoustic feature based on a preset general feature extraction network, and to determine a second general feature corresponding to the second acoustic feature based on the general feature extraction network; a reconstruction module, used to respectively reconstruct the first general feature and the second general feature according to the reconstruction network corresponding to the target domain. The second universal feature is reconstructed to obtain a first reconstructed signal corresponding to the first universal feature, and a second reconstructed signal corresponding to the second universal feature; a loss calculation module is used to determine a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and to determine a second loss according to the second reconstructed signal and the corresponding second acoustic feature; a training module is used to perform weighted summation of the first loss and the second loss to obtain a target loss, and to jointly train the universal feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss; a detection module is used to obtain a speech to be tested, input the speech to be tested into a trained cross-domain playback speech detection model, and determine a playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes the universal feature extraction network.

[0022] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the cross-domain playback speech detection method described in the first aspect when executing the computer program.

[0023] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium, and the storage medium stores a computer program. When the computer program is executed by a processor, it implements the cross-domain playback speech detection method described in the first aspect above.

[0024] The embodiments of the present application include at least the following beneficial effects: by acquiring playback speech samples of multiple speech domains, based on the target domain in the multiple speech domains, dividing the playback speech samples into a first playback speech and a second playback speech, determining a first acoustic feature according to the first playback speech, determining a second acoustic feature according to the second playback speech, extracting a first general feature corresponding to the first acoustic feature through a general feature extraction network of a cross-domain playback speech detection model, and extracting a second general feature corresponding to the second acoustic feature, then reconstructing a first reconstructed signal corresponding to the first general feature through a reconstruction network, and calculating a first loss, and reconstructing a second reconstructed signal corresponding to the second general feature through a reconstruction network, and calculating a second loss, then performing weighted summation on the first loss and the second loss, and performing weighted summation on the weighted summation The obtained target loss is used to jointly train the universal feature extraction network and the reconstruction network. Since the weight of the second loss is smaller than the weight of the first loss, the reconstruction effect of the reconstruction network on the first universal feature of the target domain can be greatly improved during the training process, and the reconstruction effect of the reconstruction network on the second universal feature of other speech domains can be improved to a certain extent, so that the universal feature extraction network can more accurately extract the common features of different speech domains from all the feature information of the playback speech sample, which can improve the generalization performance of the cross-domain playback speech detection model. In the subsequent cross-domain playback speech detection scenario, the playback detection result of the speech to be tested can be determined based on the cross-domain playback speech detection model including the trained universal feature extraction network, which can improve the accuracy of cross-domain recording playback detection.

[0025] On this basis, in a cross-domain playback speech detection scenario, a speech to be tested is obtained, where the speech to be tested refers to speech in any speech domain, and then the playback detection result of the speech to be tested is determined based on a cross-domain playback speech detection model including a trained general feature extraction network. Since the general feature extraction network can more accurately extract common features of different speech domains from all feature information of the first acoustic feature and the second acoustic feature, the generalization performance of the cross-domain playback speech detection model is relatively high, so the accuracy of cross-domain recording playback detection can be improved, and thus the playback detection results of the speech to be tested in each speech domain can be accurately detected. The cross-domain playback speech detection method provided in the embodiment of the present application can be better applied to cross-domain playback speech detection scenarios.

[0026] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings are used to provide further understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.

[0028] Figure 1 An optional flow chart of a cross-domain playback speech detection method provided in an embodiment of the present application;

[0029] Figure 2 A schematic diagram of an optional process of updating a target domain provided in an embodiment of the present application;

[0030] Figure 3 An optional specific flow chart of determining the first acoustic feature and the second acoustic feature provided in an embodiment of the present application;

[0031] Figure 4 An optional flow chart of determining target loss provided in an embodiment of the present application;

[0032] Figure 5 Another optional flow chart of determining target loss provided in an embodiment of the present application;

[0033] Figure 6 An optional flow chart of a cross-domain playback speech detection model provided in an embodiment of the present application;

[0034] Figure 7 A schematic diagram of an optional architecture for playing back voice samples provided in an embodiment of the present application;

[0035] Figure 8 A schematic diagram of an optional architecture for calculating the first loss in the target domain provided in an embodiment of the present application;

[0036] Fig. 9 A schematic diagram of an optional architecture for calculating the second loss in the auxiliary domain provided in an embodiment of the present application;

[0037] Fig.10 A schematic diagram of an optional architecture for jointly training a general feature extraction network and a reconstruction network with target loss provided in an embodiment of the present application;

[0038] Fig.11 A schematic diagram of an optional architecture for calculating classification loss provided in an embodiment of the present application;

[0039] Fig.12 A schematic diagram of an optional architecture for jointly training a general feature extraction network and a classifier for classification loss provided in an embodiment of the present application;

[0040] Fig.13An optional architecture diagram of a cross-domain playback speech detection method provided in an embodiment of the present application;

[0041] Fig.14 An optional structural diagram of a cross-domain playback speech detection device provided in this embodiment;

[0042] Fig.15 A schematic diagram of an optional hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0044] In the description of this application, “several” means one or more, “more” means more than two, “greater than”, “less than”, “exceed”, etc. are understood to exclude the number, and “above”, “below”, “within”, etc. are understood to include the number.

[0045] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims or the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0046] It should be noted that in each specific embodiment of the present application, when it comes to the need to perform relevant processing based on data related to the characteristics of the target object such as target object attribute information or attribute information set, the permission or consent of the target object will be obtained first, and the collection, use and processing of these data will comply with relevant laws, regulations and standards. Among them, the target object can be a user. In addition, when the embodiment of the present application needs to obtain the target object attribute information, the target object's separate permission or separate consent will be obtained by means of a pop-up window or jumping to a confirmation page, etc., and after the separate permission or separate consent of the target object is clearly obtained, the necessary target object-related data for enabling the normal operation of the embodiment of the present application will be obtained.

[0047] At present, cross-domain recording playback detection has the problem of speech domain offset, for example, speech domain offset between different recording devices, different playback devices or different recording environments; the cross-domain playback speech detection model currently in use has poor generalization performance and can only detect playback recordings in one or more speech domains, but cannot recognize playback recordings in other speech domains.

[0048] Based on this, the embodiments of the present application provide a cross-domain playback voice detection method, apparatus, device and medium, which can improve the generalization of the cross-domain playback voice detection model, thereby improving the accuracy of cross-domain playback voice detection.

[0049] The cross-domain playback voice detection method, device, equipment and medium provided in the embodiments of the present application are specifically illustrated through the following embodiments. First, the cross-domain playback voice detection method in the embodiments of the present application is described.

[0050] The cross-domain playback voice detection method provided in the embodiment of the present application relates to the field of computer technology. The cross-domain playback voice detection method provided in the embodiment of the present application can be applied to a terminal, can also be applied to a server side, and can also be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements a cross-domain playback voice detection method, etc., but is not limited to the above forms.

[0051] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0052] It should be noted that in each specific implementation of the present application, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, and the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or consent, the necessary user-related data for the normal operation of the embodiment of the present application will be obtained.

[0053] The embodiments of the present application are further described below in conjunction with the accompanying drawings.

[0054] like Figure 1 As shown, Figure 1 An optional flow chart of a cross-domain playback voice detection method provided in an embodiment of the present application, the cross-domain playback voice detection method can be executed by a server, or can also be executed by a terminal, or can also be executed by a server in cooperation with a terminal, and the cross-domain playback voice detection method includes but is not limited to the following steps S110 to S170:

[0055] Step S110: obtaining playback speech samples of multiple speech domains, taking any speech domain as a target domain, taking the playback speech sample corresponding to the target domain as a first playback speech, and taking other playback speech samples except the first playback speech as a second playback speech, wherein each speech domain corresponds to its own reconstruction network;

[0056] Step S120: determining a first acoustic feature according to the first playback voice, and determining a second acoustic feature according to the second playback voice;

[0057] Step S130: determining a first universal feature corresponding to the first acoustic feature based on a preset universal feature extraction network, and determining a second universal feature corresponding to the second acoustic feature based on the universal feature extraction network;

[0058] Step S140: reconstructing the first general feature and the second general feature respectively based on the reconstruction network corresponding to the target domain to obtain a first reconstructed signal corresponding to the first general feature and a second reconstructed signal corresponding to the second general feature;

[0059] Step S150: determining a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and determining a second loss according to the second reconstructed signal and the corresponding second acoustic feature;

[0060] Step S160: performing a weighted summation of the first loss and the second loss to obtain a target loss, and jointly training the general feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss;

[0061] Step S170: obtaining the speech to be tested, inputting the speech to be tested into the trained cross-domain playback speech detection model, and determining the playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes a general feature extraction network.

[0062] Among them, the playback voice sample is a pre-recorded voice input into the speaker verification system, and the voice domain is used to indicate the recording scene of the playback voice sample. For example, the recording scene includes a recording device, a playback device, and a recording environment. Different voice domains are used to indicate different recording scenes. After the playback voice sample is feature extracted by a general feature extraction network, since the feature information contained in the playback voice samples of different recording scenes is usually different, it is necessary to configure a corresponding reconstruction network for each voice domain in order to effectively reconstruct the feature extraction results of different voice domains.

[0063] Based on this, by obtaining playback speech samples of multiple speech domains, based on the target domain in the multiple speech domains, the playback speech samples are divided into a first playback speech and a second playback speech, the first acoustic feature is determined according to the first playback speech, and the second acoustic feature is determined according to the second playback speech, and the first general feature corresponding to the first acoustic feature is extracted through the general feature extraction network of the cross-domain playback speech detection model, and the second general feature corresponding to the second acoustic feature is extracted, and then the first reconstructed signal corresponding to the first general feature is reconstructed through the reconstruction network, and the first loss is calculated, and the second reconstructed signal corresponding to the second general feature is reconstructed through the reconstruction network, and the second loss is calculated, and then the first loss and the second loss are weighted summed, and the target loss obtained by the weighted sum is applied to the general The feature extraction network and the reconstruction network are jointly trained. Since the weight of the second loss is smaller than the weight of the first loss, the reconstruction effect of the reconstruction network on the first universal feature of the target domain can be greatly improved during the training process, and the reconstruction effect of the reconstruction network on the second universal feature of other speech domains can be improved to a certain extent, so that the universal feature extraction network can more accurately extract the common features of different speech domains from all the feature information of the first acoustic feature and the second acoustic feature, which can improve the generalization performance of the cross-domain playback speech detection model. Subsequently, in the cross-domain playback speech detection scenario, the playback detection result of the speech to be tested can be determined based on the cross-domain playback speech detection model including the trained universal feature extraction network, which can improve the accuracy of cross-domain recording playback detection.

[0064] On this basis, in a cross-domain playback speech detection scenario, a speech to be tested is obtained, where the speech to be tested refers to speech in any speech domain, and then the playback detection result of the speech to be tested is determined based on a cross-domain playback speech detection model including a trained general feature extraction network. Since the general feature extraction network can more accurately extract common features of different speech domains from all feature information of the playback speech sample, the cross-domain playback speech detection model has a high generalization performance, so it can improve the accuracy of cross-domain playback speech detection, and thus can accurately detect and obtain the playback detection results of the speech to be tested in each speech domain. The playback speech detection method provided in the embodiment of the present application can be better applied to cross-domain playback speech detection scenarios.

[0065] In addition, refer to Figure 2 , in one embodiment, Figure 1 After step S160 in the illustrated embodiment, the cross-domain playback speech detection method further includes but is not limited to the following steps:

[0066] Step S210: updating the target domain to a speech domain corresponding to any untrained reconstruction network;

[0067] Step S220: jointly train the general feature extraction network and the reconstruction network corresponding to the target domain again until all reconstruction networks are trained.

[0068] Based on this, by continuously updating the target domain and the reconstruction network corresponding to the target domain, different speech domains are used as target domains to jointly train the universal feature extraction network and the reconstruction network corresponding to the target domain. The universal feature extraction network uses all the first acoustic features and the second acoustic features as training data sets. The first acoustic features have a greater training effect on the corresponding reconstruction network. During the training process, the reconstruction effect of the reconstruction network on the first universal feature of the target domain can be greatly improved, so that the universal feature extraction network can more accurately extract the common features of different speech domains from all the feature information of the playback speech samples, avoiding the universal feature extraction network from overfitting one of the speech domains as the target domain and underfitting other speech domains, thereby improving the specificity of the reconstruction network and the robustness and generalization of the universal feature extraction network.

[0069] In addition, refer to Figure 3 , in one embodiment, Figure 1 Step S120 in the illustrated embodiment includes but is not limited to the following steps:

[0070] Step S310: performing frequency domain transformation processing on the first playback speech to obtain a first frequency domain signal, calculating the power spectrum of the first frequency domain signal to obtain a first power spectrum, and calculating the logarithm of the first power spectrum to obtain a first acoustic feature;

[0071] Step S320: performing frequency domain transformation processing on the second playback speech to obtain a second frequency domain signal, calculating the power spectrum of the second frequency domain signal to obtain a second power spectrum, and calculating the logarithm of the second power spectrum to obtain a second acoustic feature.

[0072] In some possible implementations, the calculation method of the first acoustic feature is as follows:

[0073] For the first playback voice x(n), the first playback voice x(n) is a time domain signal. The first playback voice x(n) is transformed into a frequency domain to obtain a first frequency domain signal Y(k, n). The formula of the frequency domain transformation is as follows:

[0074] Among them, k is the frequency domain index, n is the time domain index, and N k is the variable window length, j represents the index value of the sampling point in the first playback speech, Represents α k The complex conjugate of k is a complex time-frequency atom, specifically defined as:

[0075]

[0076] Among them, f k is the center frequency of k frequency bands, f s is the sampling rate, v is the window function, φ k is the phase shift, C is the scaling factor, and the scaling factor C is calculated as:

[0077]

[0078] In addition, the center frequency of the kth frequency band obeys the following rule: Among them, f1 is the center frequency of the lowest frequency band, B is the number of spectral lines in each octave, and the bandwidth of the kth frequency band is δ f The calculation formula is:

[0079]

[0080] The first power spectrum PS can be expressed as:

[0081] PS=|Y(k,n)| 2

[0082] Taking the logarithm of the first power spectrum PS, we get the first acoustic feature LPS:

[0083] LPS=logPS=log|Y(k,n)| 2

[0084] Based on this, by transforming the power spectrum of the first playback speech from a time domain signal to a frequency domain signal, the distribution of power energy of each frequency component in the first playback speech can be reflected, the dynamic characteristics in the first playback speech can be revealed, and the form and composition of noise and interference in the first playback speech can be observed, which can largely reflect the characteristics of the speech domain in which the first playback speech is located. Taking the logarithm of the first power spectrum to obtain the first logarithmic power spectrum can increase the components with lower amplitudes in the first power spectrum, so as to observe the periodic signals concealed in the low-amplitude noise, and facilitate the general feature extraction network to extract the first general feature from the first logarithmic power spectrum.

[0085] For the second playback speech, the second logarithmic power spectrum can be obtained by adopting the above calculation method, and its beneficial effect is similar to the beneficial effect of obtaining the first logarithmic power spectrum by calculating from the first playback speech.

[0086] In one embodiment, the speech domain corresponding to the second playback speech is used as the auxiliary domain, and the target loss is calculated as follows:

[0087]

[0088] Among them, L is the target loss, M is the total number of speech domains, l0 is the first loss, ω0 is the weight corresponding to the first loss, l m is the second loss corresponding to the mth auxiliary domain, ω m for l m The corresponding weights;

[0089] The calculation formula of l0 is as follows:

[0090]

[0091] in, is the i-th first acoustic feature, for The corresponding first reconstructed signal, n0 is the total number of the first acoustic features;

[0092] l m The calculation formula is as follows:

[0093]

[0094] in, is the kth second acoustic feature corresponding to the mth auxiliary domain, for The corresponding second reconstructed signal, n m is the total number of second acoustic features corresponding to the mth auxiliary domain.

[0095] In addition, refer to Figure 4In one embodiment, the speech domain corresponding to the second playback speech is used as the auxiliary domain. Figure 1 Step S160 in the illustrated embodiment includes but is not limited to the following steps:

[0096] Step S410: obtaining domain similarities between each auxiliary domain and the target domain;

[0097] Step S420: determining a first weight corresponding to the target domain and a second weight corresponding to each auxiliary domain according to each domain similarity, wherein the sum of the first weight and all the second weights is equal to one, and the second weight is positively correlated with the corresponding domain similarity;

[0098] Step S430: Using the first weight as the weight of the first loss, using each second weight as the weight of the corresponding second loss, performing weighted summation on the first loss and the second loss to obtain the target loss.

[0099] Based on this, by obtaining the domain similarity between different auxiliary domains and the target domain, and determining the second weight that is positively correlated with the domain similarity based on the domain similarity, the second loss is calculated based on the second weight, and the second playback speech in the auxiliary domain with small domain similarity adopts a second weight with a small value, thereby reducing the training effect of the auxiliary domain on the reconstruction network corresponding to the target domain, improving the reconstruction effect of the reconstruction network on the first general feature, and ensuring the specificity of the reconstruction network to its corresponding target domain.

[0100] In some possible implementations, an optional way to determine the first weight and the second weight is: obtain recording scene information of each speech domain, the recording scene information includes multiple scene factors, take one of the auxiliary domains as an example, calculate the similarity between the scene factors of the auxiliary domain and the scene factors corresponding to the target domain, obtain the factor similarity, and then add the similarities of each factor corresponding to the auxiliary domain to obtain the domain similarity; after determining the domain similarity of each auxiliary domain, the second weight of each auxiliary domain can be determined based on the domain similarity, and the second weight is positively correlated with the domain similarity.

[0101] Specifically, the domain similarity between the mth auxiliary domain and the target domain is calculated as follows:

[0102]

[0103] Among them, Sim m represents the domain similarity between the mth auxiliary domain and the target domain, N represents the number of scene factors, and sc m,i represents the similarity of the i-th factor in the m-th auxiliary domain;

[0104] The second weight is calculated as follows:

[0105]

[0106] Among them, ω0 is the preset weight of the target domain, and the value range of ω0 is [0,1]. m represents the second weight of the mth auxiliary domain, and M-1 is the number of auxiliary domains.

[0107] In some possible implementations, there are three scene factors of the recording scene, including a recording environment, a recording device, and a playback device.

[0108] In addition, refer to Figure 5 , in one embodiment, Figure 1 Step S160 in the embodiment includes but is not limited to the following steps:

[0109] Step S510: obtaining first recording scene information of the target domain and second recording scene information of each auxiliary domain;

[0110] Step S520: extracting features from the first recording scene information to obtain first scene features;

[0111] Step S530: extracting features from each second recording scene information to obtain respective corresponding second scene features;

[0112] Step S540: inputting the first scene feature and each second scene feature into the prediction network for prediction, and obtaining a second weight of each auxiliary domain;

[0113] Step S550: determining a first weight of the target domain according to the second weights of each auxiliary domain, wherein the sum of the first weight and the second weight is equal to one;

[0114] Step S560: Use the first weight as the weight of the first loss, use each second weight as the weight of the corresponding second loss, perform weighted summation on the first loss and the second loss, and obtain the target loss.

[0115] Based on this, the second weight is predicted by the prediction network, so that the second weight can accurately characterize the domain similarity between the auxiliary domain and the target domain, thereby improving the training effect of the general feature extraction network and the reconstruction network, improving the reconstruction effect of the reconstruction network on the first general feature, and improving the generalization of the general feature extraction network.

[0116] In addition, refer to Figure 6 In one embodiment, the cross-domain playback speech detection model further includes a classifier, Figure 1 Step S160 in includes but is not limited to the following steps S610 to S630.

[0117] Step S610: determining a target acoustic feature according to the speech to be tested, and determining a target universal feature corresponding to the target acoustic feature based on a universal feature extraction network;

[0118] Step S620: input the target universal feature into the classifier to obtain a playback detection result of the speech to be tested, wherein the playback detection result is used to indicate whether the speech type of the speech to be tested is a playback type or a non-playback type.

[0119] Based on this, the target universal features of the test speech are extracted through the universal feature extraction network. If the test speech belongs to one of the speech domains, the target universal features will contain common features. By using the classifier to identify whether the target universal features contain common features, it can be determined whether the test speech is a recorded playback.

[0120] In some possible implementations, the classifier is trained through the following steps.

[0121] Obtain a training sample set, wherein the training sample set includes a plurality of training speech samples and sample labels corresponding to each training speech sample, the training speech samples include playback speech samples and real speech samples, and the sample labels are used to characterize the probability that the corresponding training speech sample is playback speech, and the probability that it is non-playback speech. For example, the sample label may be [0,1], which represents that the probability that the corresponding training speech sample is playback speech is 0, and the probability that it is non-playback speech is 1; for another example, the sample label may be [1,0], which represents that the probability that the corresponding training speech sample is playback speech is 1, and the probability that it is non-playback speech is 0, that is, the sample label is used to indicate the probability distribution of the target speech;

[0122] determining training acoustic features according to the training speech samples;

[0123] Determine a training universal feature corresponding to the training acoustic feature based on a universal feature extraction network;

[0124] Input the training universal features into the classifier, and determine the classification results corresponding to the training universal features based on the classifier, wherein the classification results are used to indicate the predicted speech probability distribution. For example, the classification results may be [0.7, 0.3], which indicates that the probability that the corresponding training speech sample is a playback speech is 0.7, and the probability that it is a non-playback speech is 0.3;

[0125] Determine the classification loss according to the classification result and the corresponding sample label, for example, by calculating the cross entropy loss between the classification result and the corresponding sample label, and use the calculated cross entropy loss as the classification loss;

[0126] According to the classification loss, the classifier and the general feature extraction network are jointly trained until the cross-domain playback speech detection model meets the convergence conditions.

[0127] The whole process of the cross-domain playback speech detection method is described in detail below.

[0128] refer to Figures 7 to 13 , Figure 7 A schematic diagram of an optional architecture for playing back voice samples provided in an embodiment of the present application, Figure 8 A schematic diagram of an optional architecture for calculating the first loss in the target domain provided in an embodiment of the present application, Fig. 9 A schematic diagram of an optional architecture for calculating the second loss in the auxiliary domain provided in an embodiment of the present application, Fig.10 A schematic diagram of an optional architecture for jointly training a general feature extraction network and a reconstruction network with target loss provided in an embodiment of the present application, Fig.11 A schematic diagram of an optional architecture for calculating classification loss provided in an embodiment of the present application, Fig.12 A schematic diagram of an optional architecture for jointly training a general feature extraction network and a classifier for classification loss provided in an embodiment of the present application, Fig.13 An optional architecture diagram of a cross-domain playback speech detection method provided in an embodiment of the present application;

[0129] The following is a detailed description of the training process of the general feature extraction network in the cross-domain playback speech detection model.

[0130] The method for training the general feature extraction network can be divided into multiple training rounds. The number of training rounds is consistent with the number of speech domains. The steps of each training round are similar. One of the training rounds is described in detail below.

[0131] First, a plurality of playback speech samples are obtained, each of which belongs to one of the speech domains;

[0132] Then, one of the speech domains is defined as a target domain, the playback speech samples in the target domain are defined as first playback speech, the speech domain outside the target domain is defined as an auxiliary domain, the playback speech samples in the auxiliary domain are defined as second playback speech, a reconstruction network corresponding to the target domain is determined, and the number of reconstruction networks is equal to the number of speech domains;

[0133] Then, a first weight is determined for the target domain and a second weight is determined for each auxiliary domain;

[0134] Then, determining a first universal feature corresponding to the first acoustic feature based on a preset universal feature extraction network, and determining a second universal feature corresponding to the second acoustic feature based on the universal feature extraction network;

[0135] Then, based on the reconstruction network corresponding to the target domain, the first general feature and the second general feature are reconstructed respectively to obtain a first reconstructed signal corresponding to the first general feature and a second reconstructed signal corresponding to the second general feature;

[0136] Then, determining a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and determining a second loss according to the second reconstructed signal and the corresponding second acoustic feature;

[0137] Then, the first loss is weighted by the first weight, and the second loss is weighted by the second weight. The target loss is obtained by summing the weighted first weight and the second weight. The general feature extraction network and the reconstruction network corresponding to the target domain are jointly trained according to the target loss.

[0138] Then, the target domain is updated to the speech domain corresponding to any untrained reconstruction network, the playback speech sample in the target domain is defined as the first playback speech, the speech domain outside the target domain is defined as the auxiliary domain, the playback speech sample in the auxiliary domain is defined as the second playback speech, a reconstruction network corresponding to the target domain is determined, and the above training rounds are repeated until all reconstruction networks are trained;

[0139] The following is a detailed description of the training process of the general feature extraction network and classifier in the cross-domain playback speech detection model.

[0140] First, a training sample set is obtained, wherein the training sample set includes a plurality of training speech samples and sample labels corresponding to each training speech sample, the training speech samples include playback speech samples and real speech samples, and the sample labels are used to characterize the probability that the corresponding training speech sample is playback speech, and the probability that it is non-playback speech. For example, the sample label may be [0,1], which represents that the probability that the corresponding training speech sample is playback speech is 0, and the probability that it is non-playback speech is 1; for another example, the sample label may be [1,0], which represents that the probability that the corresponding training speech sample is playback speech is 1, and the probability that it is non-playback speech is 0, that is, the sample label is used to indicate the probability distribution of the target speech;

[0141] Then, determining the training acoustic features according to the training speech samples;

[0142] Then, determining the training universal features corresponding to the training acoustic features based on the universal feature extraction network;

[0143] Then, the training universal features are input into the classifier, and the classification results corresponding to the training universal features are determined based on the classifier, wherein the classification results are used to indicate the predicted speech probability distribution. For example, the classification result may be [0.7, 0.3], which represents that the probability that the corresponding training speech sample is a playback speech is 0.7, and the probability that it is a non-playback speech is 0.3;

[0144] Then, according to the classification result and the corresponding sample label, the classification loss is determined. For example, by calculating the cross entropy loss between the classification result and the corresponding sample label, the calculated cross entropy loss is used as the classification loss.

[0145] Then, according to the classification loss, the classifier and the general feature extraction network are jointly trained until the cross-domain playback speech detection model meets the convergence conditions;

[0146] The application process of the cross-domain playback speech detection model is described in detail below.

[0147] First, obtain the voice to be tested;

[0148] Then, the target acoustic features are determined according to the speech to be tested;

[0149] Then, the target universal features corresponding to the target acoustic features are obtained based on the trained universal feature extraction network;

[0150] Then, a playback detection result of the speech to be tested is obtained based on the trained classifier.

[0151] Based on this, by obtaining playback speech samples of multiple speech domains, based on the target domain in the multiple speech domains, the playback speech samples are divided into a first playback speech and a second playback speech, the first acoustic feature is determined according to the first playback speech, and the second acoustic feature is determined according to the second playback speech, and the first general feature corresponding to the first acoustic feature is extracted through the general feature extraction network of the cross-domain playback speech detection model, and the second general feature corresponding to the second acoustic feature is extracted, and then the first reconstructed signal corresponding to the first general feature is reconstructed through the reconstruction network, and the first loss is calculated, and the second reconstructed signal corresponding to the second general feature is reconstructed through the reconstruction network, and the second loss is calculated, and then the first loss and the second loss are weighted summed, and the target loss obtained by the weighted sum is applied to the general The feature extraction network and the reconstruction network are jointly trained. Since the weight of the second loss is smaller than the weight of the first loss, the reconstruction effect of the reconstruction network on the first universal feature of the target domain can be greatly improved during the training process, and the reconstruction effect of the reconstruction network on the second universal feature of other speech domains can be improved to a certain extent, so that the universal feature extraction network can more accurately extract the common features of different speech domains from all the feature information of the first acoustic feature and the second acoustic feature, which can improve the generalization performance of the cross-domain playback speech detection model. Subsequently, in the cross-domain playback speech detection scenario, the playback detection result of the speech to be tested can be determined based on the cross-domain playback speech detection model including the trained universal feature extraction network, which can improve the accuracy of cross-domain recording playback detection.

[0152] On this basis, in a cross-domain playback speech detection scenario, a speech to be tested is obtained, where the speech to be tested refers to speech in any speech domain, and then the playback detection result of the speech to be tested is determined based on a cross-domain playback speech detection model including a trained general feature extraction network. Since the general feature extraction network can more accurately extract common features of different speech domains from all feature information of the playback speech samples, the cross-domain playback speech detection model has a high generalization performance, so it can improve the accuracy of cross-domain recording playback detection, and thus can accurately detect and obtain the playback detection results of the speech to be tested in each speech domain. The playback speech detection method provided in the embodiment of the present application can be better applied to cross-domain playback speech detection scenarios.

[0153] In addition, refer to Fig.14 The present application also provides a cross-domain playback speech detection device 1400, including:

[0154] The acquisition module 1401 is used to acquire playback speech samples of multiple speech domains, take any speech domain as a target domain, take the playback speech sample corresponding to the target domain as a first playback speech, and take other playback speech samples except the first playback speech as a second playback speech, wherein each speech domain corresponds to its own reconstruction network;

[0155] An acoustic feature determination module 1402, configured to determine a first acoustic feature according to the first playback speech, and to determine a second acoustic feature according to the second playback speech;

[0156] A universal feature extraction module 1403, configured to determine a first universal feature corresponding to the first acoustic feature based on a preset universal feature extraction network, and to determine a second universal feature corresponding to the second acoustic feature based on the universal feature extraction network;

[0157] A reconstruction module 1404 is used to reconstruct the first general feature and the second general feature respectively based on the reconstruction network corresponding to the target domain to obtain a first reconstructed signal corresponding to the first general feature and a second reconstructed signal corresponding to the second general feature;

[0158] A loss calculation module 1405, configured to determine a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and to determine a second loss according to the second reconstructed signal and the corresponding second acoustic feature;

[0159] A training module 1406 is used to perform a weighted summation of the first loss and the second loss to obtain a target loss, and to jointly train the general feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss;

[0160] The detection module 1407 is used to obtain the speech to be tested, input the speech to be tested into the trained cross-domain playback speech detection model, and determine the playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes a general feature extraction network.

[0161] The above-mentioned cross-domain playback speech detection device 1400 and the cross-domain playback speech detection method are based on the same inventive concept, which will not be described in detail here.

[0162] In addition, refer to Fig.15 , Fig.15 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0163] The processor 1501 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0164] The memory 1502 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1502 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1502, and the processor 1501 calls and executes the cross-domain playback speech detection method of the embodiment of the present application, for example, executing the above-described Figure 1 Steps S110 to S170 of the method, Figure 2 Steps S210 to S220 of the method, Figure 3 Steps S310 to S320 of the method, Figure 4 Steps S410 to S430 of the method, Figure 5 Steps S510 to S560 of the method and Figure 6 Steps S610 to S620 of the method;

[0165] Input / output interface 1503, used to implement information input and output;

[0166] Communication interface 1504, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0167] A bus 1505 that transmits information between the various components of the device (e.g., the processor 1501, the memory 1502, the input / output interface 1503, and the communication interface 1504);

[0168] The processor 1501 , the memory 1502 , the input / output interface 1503 and the communication interface 1504 are connected to each other in communication within the device via the bus 1505 .

[0169] The embodiment of the present application also provides a storage medium, which is a computer-readable storage medium for computer-readable storage, and the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned cross-domain playback speech detection method, for example, to execute the above-described Figure 1 Steps S110 to S170 of the method, Figure 2 Steps S210 to S220 of the method, Figure 3 Steps S310 to S320 of the method, Figure 4 Steps S410 to S430 of the method and Figure 5 Steps S510 to S560 of the method and Figure 6 Method steps S610 to S620.

[0170] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0171] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0172] It can be understood by those skilled in the art that Figures 1 to 6The technical solutions shown in the figure do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than those shown in the figure, or a combination of certain steps, or different steps.

[0173] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0174] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0175] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0176] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0177] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0178] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0179] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0180] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0181] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A cross-domain playback speech detection method, characterized in that: include: Acquire playback speech samples of multiple speech domains, take one of the speech domains as a target domain, take the playback speech sample corresponding to the target domain as a first playback speech, and take the other playback speech samples except the first playback speech as a second playback speech, wherein the playback speech samples recorded based on the same recording device, playback device and recording environment belong to the same speech domain, and the playback speech samples recorded based on different recording devices, playback devices or recording environments belong to different speech domains, and each speech domain corresponds to its own reconstruction network; Determining a first acoustic feature according to the first playback speech, and determining a second acoustic feature according to the second playback speech; Determine a first universal feature corresponding to the first acoustic feature based on a preset universal feature extraction network, and determine a second universal feature corresponding to the second acoustic feature based on the universal feature extraction network; Reconstructing the first general feature and the second general feature based on the reconstruction network corresponding to the target domain, respectively, to obtain a first reconstructed signal corresponding to the first general feature, and to obtain a second reconstructed signal corresponding to the second general feature; Determine a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and determine a second loss according to the second reconstructed signal and the corresponding second acoustic feature; Performing a weighted summation on the first loss and the second loss to obtain a target loss, and jointly training the general feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss; Acquire a speech to be tested, input the speech to be tested into a trained cross-domain playback speech detection model, and determine a playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes the universal feature extraction network.

2. The cross-domain playback speech detection method according to claim 1, characterized in that: After jointly training the general feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, the cross-domain playback speech detection method further includes: Updating the target domain to a speech domain corresponding to any untrained reconstruction network; The general feature extraction network and the reconstruction network corresponding to the target domain are jointly trained again until all the reconstruction networks are trained.

3. The cross-domain playback speech detection method according to claim 1, characterized in that: The determining a first acoustic feature according to the first playback voice and determining a second acoustic feature according to the second playback voice includes: Performing frequency domain transformation processing on the first playback speech to obtain a first frequency domain signal, calculating a power spectrum of the first frequency domain signal to obtain a first power spectrum, and calculating a logarithm of the first power spectrum to obtain a first acoustic feature; The second playback speech is subjected to frequency domain transformation processing to obtain a second frequency domain signal, a power spectrum of the second frequency domain signal is calculated to obtain a second power spectrum, and a logarithm of the second power spectrum is calculated to obtain a second acoustic feature.

4. The cross-domain playback speech detection method according to claim 1, characterized in that: The speech domain corresponding to the second playback speech is used as the auxiliary domain, and the target loss is calculated as follows: Wherein, L is the target loss, M is the total number of the speech domains, l0 is the first loss, ω0 is the weight corresponding to the first loss, l m is the second loss corresponding to the mth auxiliary domain, ω m for l m The corresponding weights; The calculation formula of l0 is as follows: in, is the i-th first acoustic feature, corresponding to the first reconstructed signal, n0 is the total number of the first acoustic features; The l m The calculation formula is as follows: in, is the kth second acoustic feature corresponding to the mth auxiliary domain, for The corresponding second reconstructed signal, n m is the total number of the second acoustic features corresponding to the mth auxiliary domain.

5. The cross-domain playback speech detection method according to claim 1, characterized in that: The method further comprises: taking the speech domain corresponding to the second playback speech as an auxiliary domain, wherein the number of the auxiliary domains is multiple, and performing weighted summation on the first loss and the second loss to obtain a target loss, including: Obtaining domain similarities between each of the auxiliary domains and the target domain; Determine, according to each of the domain similarities, a first weight corresponding to the target domain and a second weight corresponding to each of the auxiliary domains, wherein the sum of the first weight and all of the second weights is equal to one, and the second weight is positively correlated with the corresponding domain similarity; The first weight is used as the weight of the first loss, and each of the second weights is used as the weight of the second loss corresponding to each other, and the first loss and the second loss are weightedly summed to obtain a target loss.

6. The cross-domain playback speech detection method according to claim 1, characterized in that: The method further comprises: taking the speech domain corresponding to the second playback speech as an auxiliary domain, and performing weighted summation on the first loss and the second loss to obtain a target loss, comprising: Acquire first recording scene information of the target domain and second recording scene information of each of the auxiliary domains; Extracting features from the first recording scene information to obtain first scene features; Extracting features from each of the second recording scene information to obtain second scene features corresponding to each of the second recording scene information; Inputting the first scene feature and each of the second scene features into a prediction network for prediction, and obtaining a second weight of each of the auxiliary domains; determining a first weight of the target domain according to the second weights of the auxiliary domains, wherein a sum of the first weight and the second weight is equal to one; The first weight is used as the weight of the first loss, and each of the second weights is used as the weight of the second loss corresponding to each other, and the first loss and the second loss are weightedly summed to obtain a target loss.

7. The cross-domain playback speech detection method according to claim 1, characterized in that: The cross-domain playback speech detection model further includes a classifier, and determining the playback detection result of the speech to be tested based on the cross-domain playback speech detection model includes: Determine a target acoustic feature according to the speech to be tested, and determine a target universal feature corresponding to the target acoustic feature based on the universal feature extraction network; The target universal feature is input into the classifier to obtain a playback detection result of the speech to be tested, wherein the playback detection result is used to indicate whether the speech type of the speech to be tested is a playback type or a non-playback type.

8. A cross-domain playback speech detection device, characterized in that: include: An acquisition module, used to acquire playback speech samples of multiple speech domains, take any one of the speech domains as a target domain, take the playback speech sample corresponding to the target domain as a first playback speech, and take the other playback speech samples except the first playback speech as a second playback speech, wherein each of the speech domains corresponds to its own reconstruction network; an acoustic feature determination module, configured to determine a first acoustic feature according to the first playback speech, and determine a second acoustic feature according to the second playback speech; A universal feature extraction module, configured to determine a first universal feature corresponding to the first acoustic feature based on a preset universal feature extraction network, and to determine a second universal feature corresponding to the second acoustic feature based on the universal feature extraction network; A reconstruction module, configured to reconstruct the first general feature and the second general feature respectively based on the reconstruction network corresponding to the target domain, to obtain a first reconstructed signal corresponding to the first general feature, and to obtain a second reconstructed signal corresponding to the second general feature; a loss calculation module, configured to determine a first loss according to the first reconstructed signal and the corresponding first acoustic feature, and to determine a second loss according to the second reconstructed signal and the corresponding second acoustic feature; A training module, configured to perform a weighted summation of the first loss and the second loss to obtain a target loss, and jointly train the general feature extraction network and the reconstruction network corresponding to the target domain according to the target loss, wherein the weight of the first loss is greater than the weight of the second loss; A detection module is used to obtain a speech to be tested, input the speech to be tested into a trained cross-domain playback speech detection model, and determine a playback detection result of the speech to be tested based on the cross-domain playback speech detection model, wherein the cross-domain playback speech detection model includes the general feature extraction network.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the cross-domain playback speech detection method according to any one of claims 1 to 7 when executing the computer program.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the cross-domain playback speech detection method according to any one of claims 1 to 7 is implemented.

Images