Offline personalized voiceprint learning method and speaker separation method

Abstract

The invention provides an offline personalized voiceprint learning method and a speaker separation method, and relates to the technical field of voice processing, and the method comprises the steps: determining a built-in universal voiceprint recognition model and target universal voiceprint learning data at a device end, and obtaining personalized voiceprint learning data; performing meta-learning on the universal voiceprint recognition model by using the training data and the personalized voiceprint learning data to obtain an initial personalized voiceprint recognition model; and finally, testing the universal voiceprint recognition model and the initial personalized voiceprint recognition model by using the test data, and determining a target personalized voiceprint recognition model based on an obtained first test result. According to the method, off-line personalized training can be realized by performing meta-learning and testing on the universal voiceprint recognition model by utilizing personalized voiceprint learning data and built-in training data of the equipment end, and a model updating packet does not need to be transmitted to each equipment end, so that the model updating efficiency can be greatly improved, the cost is reduced, and the user experience is improved. And the security problem caused by update packet transmission is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of speech processing, and in particular to an offline personalized voiceprint learning method and a speaker separation method. Background Art

[0002] As an advanced biometric technology that authenticates identity based on voice features, the accuracy and robustness of voiceprint recognition technology are highly dependent on the accuracy and timeliness of the voiceprint recognition model. With the continuous advancement of technology and the continuous evolution of attack methods, voiceprint recognition models deployed on electronic devices with high confidentiality requirements may gradually expose defects such as being unable to effectively resist new attacks or reduced recognition accuracy.

[0003] Since these electronic devices cannot be directly connected to the network, the traditional online update mechanism of voiceprint recognition models becomes unfeasible. This means that every time the voiceprint recognition model needs to be updated to adapt to new sound characteristics, optimize the recognition algorithm, or fix known vulnerabilities, the update package must be transmitted to each target device and installed locally. The above process is not only inefficient and costly, but may also introduce new risks due to human error or security issues during the transmission of the update package.

[0004] In addition, the long-term lack of updates may also cause the voiceprint recognition model to gradually lag behind the pace of technological development and fail to fully utilize the latest algorithm optimization and performance improvement, thus affecting the overall performance of the system and user experience. Therefore, how to achieve efficient and secure updates of voiceprint recognition models and algorithms while ensuring confidentiality has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The present invention provides an offline personalized voiceprint learning method and a speaker separation method, which are used to solve the defects existing in the related technology.

[0006] The present invention provides an offline personalized voiceprint learning method, which is applied to a device end and includes: Determine a built-in universal voiceprint recognition model and target universal voiceprint learning data, and obtain personalized voiceprint learning data; the target universal voiceprint learning data includes training data and test data; Based on the training data and the personalized voiceprint learning data, meta-learning the general voiceprint recognition model is performed to obtain an initial personalized voiceprint recognition model; Based on the test data, the universal voiceprint recognition model and the initial personalized voiceprint recognition model are tested respectively to obtain a first test result, and based on the first test result, a target personalized voiceprint recognition model is determined.

[0007] According to an offline personalized voiceprint learning method provided by the present invention, the first test result includes a second test result corresponding to the universal voiceprint recognition model and a third test result corresponding to the initial personalized voiceprint recognition model; The step of determining a target personalized voiceprint recognition model based on the first test result includes: If the second test result is better than the third test result, determining the universal voiceprint recognition model as the target personalized voiceprint recognition model; If the third test result is better than the second test result, the initial personalized voiceprint recognition model is determined to be the target personalized voiceprint recognition model.

[0008] According to an offline personalized voiceprint learning method provided by the present invention, the obtaining of personalized voiceprint learning data includes: Get sample speech; Based on the target personalized voiceprint recognition model, voiceprint recognition is performed on the sample speech to determine a first voiceprint recognition result, and based on the first voiceprint recognition result, speaker separation is performed on the sample speech to obtain a speaker separation result; Receive feedback information from a user regarding the speaker separation result, and construct the personalized voiceprint learning data based on the feedback information and the sample speech.

[0009] According to an offline personalized voiceprint learning method provided by the present invention, the feedback information includes the interested speaker in the sample speech and the modification data of the user when an error occurs in the first voiceprint recognition result.

[0010] The present invention also provides an offline personalized voiceprint learning method, which is applied in the cloud, comprising: Based on the first initial universal voiceprint learning data, meta-learning is performed on the neural network model to obtain a pre-trained model, and based on multiple single-task universal voiceprint learning data, meta-learning is performed on the pre-trained models respectively to obtain multiple first training models; Calculating discrete indicators of structural parameters at the same position in the plurality of first training models, and selecting target parameters in the pre-training model for freezing based on the discrete indicators; Based on the first initial universal voiceprint learning data, meta-learning is performed on the frozen model to obtain the universal voiceprint recognition model; Based on the multiple single-task universal voiceprint learning data, target universal voiceprint learning data is determined, and the universal voiceprint recognition model and the target universal voiceprint learning data are configured to the device end.

[0011] According to an offline personalized voiceprint learning method provided by the present invention, the target parameter is a structural parameter whose discrete index is less than a preset threshold.

[0012] According to an offline personalized voiceprint learning method provided by the present invention, the determining target universal voiceprint learning data based on the plurality of single-task universal voiceprint learning data includes: Based on the multiple single-task universal voiceprint learning data, meta-learning is performed on the universal voiceprint recognition model to obtain a second training model, and based on the second initial universal voiceprint learning data, the universal voiceprint recognition model and the second training model are tested respectively to obtain a fourth test result; Based on the accuracy of the universal voiceprint recognition model and the second training model for the multiple single-task universal voiceprint learning data, and / or the fourth test result, the multiple single-task universal voiceprint learning data are screened to obtain the target universal voiceprint learning data.

[0013] The present invention also provides a speaker separation method, which is applied to a device end and includes: Determine user separation configuration information; the user separation configuration information includes at least one of a real-time separation mode and a non-real-time separation mode; In the real-time separation mode, a real-time voice stream is received, and based on the target personalized voiceprint recognition model obtained by the above-mentioned offline personalized voiceprint learning method, voiceprint recognition is performed on the real-time voice stream to obtain a second voiceprint recognition result, and based on the second voiceprint recognition result, or based on the speaker registration information and the second voiceprint recognition result, speaker separation is performed on the real-time voice stream; In the non-real-time separation mode, a global speech is obtained, and voiceprint recognition is performed on the global speech based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result, and speaker separation is performed on the global speech based on the third voiceprint recognition result, or based on the speaker registration information and the third voiceprint recognition result.

[0014] According to a speaker separation method provided by the present invention, the speaker registration information includes offline speaker voiceprint registration information and / or dynamic speaker marking information.

[0015] According to a speaker separation method provided by the present invention, the speaker separation is performed on the real-time voice stream based on the speaker registration information and the second voiceprint recognition result, and then further includes: storing the speaker separation result and the second voiceprint recognition result in a temporary voiceprint database in correspondence; If the real-time voice stream is interrupted, receiving the voice after the interruption, and performing voiceprint recognition on the voice after the interruption based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result; Based on the third voiceprint recognition result, the temporary voiceprint library is applied to perform speaker separation on the interrupted speech.

[0016] The present invention also provides a speaker separation integrated machine, comprising a memory, a processor, a target personalized voiceprint recognition model stored in the memory, and a computer program stored in and runnable on the processor; When the processor executes the computer program, the target personalized voiceprint recognition model is called to implement the above-mentioned speaker separation method.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described offline personalized voiceprint learning methods or speaker separation methods.

[0018] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned offline personalized voiceprint learning methods or speaker separation methods.

[0019] The offline personalized voiceprint learning method and speaker separation method provided by the present invention first determine the built-in universal voiceprint recognition model and target universal voiceprint learning data on the device side, and obtain personalized voiceprint learning data; use the training data and personalized voiceprint learning data to meta-learn the universal voiceprint recognition model to obtain an initial personalized voiceprint recognition model; finally, use the test data to test the universal voiceprint recognition model and the initial personalized voiceprint recognition model respectively to obtain a first test result, and determine the target personalized voiceprint recognition model based on the first test result. This method uses personalized voiceprint learning data and the training data built into the device side to meta-learn and test the universal voiceprint recognition model built into the device side to achieve offline personalized training, without the need to transmit the model update package to each device side, which can not only greatly improve the model update efficiency and reduce costs, but also avoid security issues introduced by the transmission of update packages. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is a flow chart of the offline personalized voiceprint learning method applied to the device end provided by the present invention.

[0022] Figure 2 It is a flow chart of the discrete training in the offline personalized voiceprint learning method applied to the device side provided by the present invention.

[0023] Figure 3 It is a schematic diagram of the process of obtaining personalized voiceprint learning data in the offline personalized voiceprint learning method applied to the device side provided by the present invention.

[0024] Figure 4 It is a schematic diagram of multiple candidate results corresponding to sample voices in the offline personalized voiceprint learning method applied to the device side provided by the present invention.

[0025] Figure 5 It is one of the schematic diagrams of feedback information and its application effect in the offline personalized voiceprint learning method applied to the device end provided by the present invention.

[0026] Figure 6 This is the second schematic diagram of feedback information and its application effect in the offline personalized voiceprint learning method applied to the device side provided by the present invention.

[0027] Figure 7 This is one of the flow charts of the offline personalized voiceprint learning method applied to the cloud provided by the present invention.

[0028] Figure 8 This is the second flow chart of the offline personalized voiceprint learning method applied to the cloud provided by the present invention.

[0029] Fig. 9 It is a schematic diagram of the overall process of the offline personalized voiceprint learning method provided by the present invention.

[0030] Fig.10 This is one of the flow charts of the speaker separation method provided by the present invention.

[0031] Fig.11 This is the second flow chart of the speaker separation method provided by the present invention.

[0032] Fig.12 It is a schematic diagram of the blind speaker separation effect in the speaker separation method provided by the present invention.

[0033] Fig.13 It is a schematic diagram of offline speaker voiceprint registration information and its application effect in the speaker separation method provided by the present invention.

[0034] Fig.14 It is a schematic diagram of dynamic speaker marking information and its application effect in the speaker separation method provided by the present invention.

[0035] Fig.15 It is a schematic diagram of existing speaker separation results before and after the real-time voice stream is interrupted in the speaker separation method provided by the present invention.

[0036] Fig.16 It is a schematic diagram of the speaker separation result desired by the user before and after the real-time voice stream is interrupted in the speaker separation method provided by the present invention.

[0037] Fig.17 It is a flow chart of the speaker separation method provided by the present invention after the real-time speech stream is interrupted.

[0038] Fig.18 This is one of the structural schematic diagrams of the offline personalized voiceprint learning device provided by the present invention.

[0039] Fig.19 This is the second structural schematic diagram of the offline personalized voiceprint learning device provided by the present invention.

[0040] Fig. 20 It is a structural schematic diagram of the speaker separation system provided by the present invention.

[0041] Fig.21 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] When the voiceprint recognition model on existing electronic devices is updated offline, it is usually necessary to transfer the update package to each target device through physical media such as USB drives, Secure Digital Cards (SD) or other removable storage media, and install and replace it locally on the target device; or, use intermediate devices such as dedicated update servers or administrators' mobile devices as transfer stations to transmit the update package to the target device through wireless methods such as Bluetooth and near-field communication.

[0044] In the above offline update solution, each target device needs to be updated separately, which is time-consuming and costly. In addition, the physical media may be lost, damaged or tampered with during transmission.

[0045] Based on this, an offline personalized voiceprint learning method is provided in an embodiment of the present invention.

[0046] Figure 1 FIG. 1 is a flow chart of an offline personalized voiceprint learning method provided in an embodiment of the present invention. Figure 1 As shown, the method is applied to the device side, including: S11, determining a built-in universal voiceprint recognition model and target universal voiceprint learning data, and obtaining personalized voiceprint learning data; the target universal voiceprint learning data includes training data and test data; S12, performing meta-learning on the general voiceprint recognition model based on the training data and the personalized voiceprint learning data to obtain an initial personalized voiceprint recognition model; S13, based on the test data, respectively testing the universal voiceprint recognition model and the initial personalized voiceprint recognition model to obtain a first test result, and based on the first test result, determining a target personalized voiceprint recognition model.

[0047] Specifically, the offline personalized voiceprint learning method provided in the embodiment of the present invention is executed by an offline personalized voiceprint learning device, which can be configured in a target device. The target device can be an electronic device that uses the target personalized voiceprint recognition model obtained by the offline personalized voiceprint learning method to perform downstream personalized voiceprint recognition tasks. No specific limitation is made here.

[0048] First, execute step S11 to determine the built-in universal voiceprint recognition model M and the target universal voiceprint learning data U. The universal voiceprint recognition model M and the target universal voiceprint learning data U can be determined by the cloud and configured to the target device when the target device leaves the factory.

[0049] The general voiceprint recognition model M may be a voiceprint recognition model suitable for different tasks. By inputting audio, it may output the voiceprint features in the audio, and further distinguish different speakers.

[0050] The target universal voiceprint learning data U may be universal voiceprint learning data used to assist the target device in performing meta-learning on a universal voiceprint recognition model, and may include universal speech data and speaker labels therein.

[0051] The target universal voiceprint learning data U can be divided into training data u1 and test data u2 according to their functions. The training data u1 is used to meta-learn the universal voiceprint recognition model, and the test data u2 is used to test the initial personalized voiceprint recognition model obtained by meta-learning. It can be understood that the training data u1 and the test data u2 do not overlap.

[0052] Then, personalized voiceprint learning data V can be obtained by interacting with the user. The personalized voiceprint learning data V can be the voiceprint learning data for the target task that the target device needs to perform, and can include personalized speech data and the speaker label therein.

[0053] After that, the meta-learning process and the testing process are carried out to realize the offline personalized training of the general voiceprint recognition model M on the device side, such as Figure 2 shown.

[0054] Execute step S12, use the training data u1 and the personalized voiceprint learning data V as training samples, perform meta-learning on the general voiceprint recognition model M, and obtain the initial personalized voiceprint recognition model P0. Through meta-learning, new tasks can be learned quickly, and then new data can be quickly adapted.

[0055] Step S13 is executed, and the universal voiceprint recognition model M and the initial personalized voiceprint recognition model P0 are tested respectively using the test data u2 to obtain a first test result. The first test result includes a second test result corresponding to the universal voiceprint recognition model M and a third test result corresponding to the initial personalized voiceprint recognition model P0. The second test result can represent the accuracy of the universal voiceprint recognition model M, and the third test result can represent the accuracy of the initial personalized voiceprint recognition model P0.

[0056] By using the second test result and the third test result, the one with better performance can be selected from the universal voiceprint recognition model M and the initial personalized voiceprint recognition model P0 as the target personalized voiceprint recognition model P. That is to say, if the performance of the initial personalized voiceprint recognition model P0 is improved compared with the universal voiceprint recognition model M on the test data u2, the initial personalized voiceprint recognition model P0 is used to replace the universal voiceprint recognition model M as the target personalized voiceprint recognition model P, otherwise the universal voiceprint recognition model M continues to be used as the target personalized voiceprint recognition model P to avoid degradation of the model effect.

[0057] The offline personalized voiceprint learning method provided in the embodiment of the present invention is applied to the device side, firstly determining the built-in universal voiceprint recognition model and the target universal voiceprint learning data, and obtaining the personalized voiceprint learning data; using the training data and the personalized voiceprint learning data, meta-learning the universal voiceprint recognition model is performed to obtain the initial personalized voiceprint recognition model; finally, using the test data, the universal voiceprint recognition model and the initial personalized voiceprint recognition model are tested respectively to obtain the first test result, and based on the first test result, the target personalized voiceprint recognition model is determined. This method uses the personalized voiceprint learning data and the training data built into the device side to perform meta-learning and testing on the universal voiceprint recognition model built into the device side to achieve offline personalized training, without the need to transmit the model update package to each device side, which can not only greatly improve the model update efficiency and reduce the cost, but also avoid the security issues introduced by the transmission of the update package.

[0058] Based on the above embodiment, the first test result includes the second test result corresponding to the universal voiceprint recognition model and the third test result corresponding to the initial personalized voiceprint recognition model; The step of determining a target personalized voiceprint recognition model based on the first test result includes: If the second test result is better than the third test result, determining the universal voiceprint recognition model as the target personalized voiceprint recognition model; If the third test result is better than the second test result, the initial personalized voiceprint recognition model is determined to be the target personalized voiceprint recognition model.

[0059] Specifically, when determining the target personalized voiceprint recognition model, the second test result can be compared with the third test result. If the second test result is better than the third test result, that is, the accuracy of the universal voiceprint recognition model on the test data is higher than the accuracy of the initial personalized voiceprint recognition model on the test data, then it is considered that the performance of the universal voiceprint recognition model is better than the initial personalized voiceprint recognition model, and then the universal voiceprint recognition model is determined as the target personalized voiceprint recognition model, that is, the universal voiceprint recognition model built into the device is retained for subsequent applications.

[0060] If the third test result is better than the second test result, that is, the accuracy of the initial personalized voiceprint recognition model on the test data is higher than the accuracy of the universal voiceprint recognition model on the test data, then it is considered that the performance of the initial personalized voiceprint recognition model is better than the universal voiceprint recognition model, and then the initial personalized voiceprint recognition model is determined as the target personalized voiceprint recognition model, that is, the universal voiceprint recognition model built into the device is replaced with the initial personalized voiceprint recognition model for subsequent applications.

[0061] In the embodiment of the present invention, the target personalized voiceprint recognition model can always be a voiceprint recognition model with excellent performance.

[0062] Based on the above embodiment, the step of obtaining personalized voiceprint learning data includes: Get sample speech; Based on the target personalized voiceprint recognition model, voiceprint recognition is performed on the sample speech to determine a first voiceprint recognition result, and based on the first voiceprint recognition result, speaker separation is performed on the sample speech to obtain a speaker separation result; Receive feedback information from a user regarding the speaker separation result, and construct the personalized voiceprint learning data based on the feedback information and the sample speech.

[0063] Specifically, Figure 3 As shown, when obtaining personalized voiceprint learning data, a sample voice can be obtained first, and then the sample voice can be input into a target personalized voiceprint recognition model, and the target personalized voiceprint recognition model can be used to perform voiceprint recognition on the sample voice to obtain a first voiceprint recognition result. The first voiceprint recognition result is the voiceprint feature contained in the sample voice.

[0064] Thereafter, the speaker separation of the sample speech is performed using the first voiceprint recognition result to obtain a speaker separation result. It is understandable that different speakers have different voiceprint features, so different speakers can be determined based on the different voiceprint features in the first voiceprint recognition result. Here, the speaker separation result may include the transcribed text of the speech segments of different speakers in the sample speech.

[0065] The speaker separation result may include multiple candidate results, and each candidate result may be sorted from high to low according to the score. Figure 4 For session 1 in , the speaker separation results obtained include result 1 and result 2, and the accuracy score of result 1 is higher than that of result 2.

[0066] Finally, the device may provide a user interface to display the speaker separation results, and may allow the user to input feedback information on the speaker separation results by marking and correcting them.

[0067] The feedback information may include multiple feedback types, including speaker recognition errors, voice confusion, incomplete separation, etc. Users can select a specific feedback type and submit detailed feedback information. The content of the feedback information may include the error part marked by the user, the correct result after correction, and the time and scene of the feedback.

[0068] The user can input feedback information on the display interface of the device, and the device can receive the feedback information and analyze the feedback information to extract key information from the feedback information, such as analyzing the error patterns and frequencies marked by the user.

[0069] Furthermore, the device can use the key information in the feedback information and the sample voice to build personalized voiceprint learning data.

[0070] In an embodiment of the present invention, according to the user's feedback information on the speaker separation results, and based on the feedback information and sample speech, the constructed personalized voiceprint learning data can be made more in line with user needs, thereby improving the execution effect and user satisfaction of downstream tasks of the target personalized voiceprint recognition model on the subsequent device end.

[0071] Based on the above embodiment, the feedback information includes the interested speaker in the sample speech and the modification data of the user when an error occurs in the first voiceprint recognition result.

[0072] Specifically, the feedback information may include the speakers of interest in the sample speech. Since different users pay attention to different speakers, the user's feedback information can only mark the speakers of interest, and the others are not separated, further improving the speaker separation effect. For example, the speaker separation result is as follows: Figure 5 As shown in Session 1 in , after receiving the speaker of interest in the feedback information, the speaker separation result can be changed to Figure 5 In conversation 2, "boss" is the speaker of interest.

[0073] The feedback information may also include the user's modification data when an error occurs in the first voiceprint recognition result. Figure 6 As shown in the conversation 1 in the figure, in the conversation 1, the speaker Li Si is mistakenly identified as Zhang San. The user actively changes Zhang San to Li Si. After the device receives the modified data in the feedback information, the speaker separation result can be changed to the new conversation 1. Moreover, in the subsequent conversation 2, the speakers Zhang San and Li Si will also be correctly separated.

[0074] Based on the above embodiments, Figure 7 As shown, an embodiment of the present invention further provides an offline personalized voiceprint learning method, which is applied to the cloud, including: S21, performing meta-learning on the neural network model based on the first initial universal voiceprint learning data to obtain a pre-trained model, and performing meta-learning on the pre-trained models based on multiple single-task universal voiceprint learning data to obtain multiple first training models; S22, calculating discrete indicators of structural parameters at the same position in the plurality of first training models, and selecting target parameters in the pre-training model for freezing based on the discrete indicators; S23, performing meta-learning on the frozen model based on the first initial universal voiceprint learning data to obtain the universal voiceprint recognition model; S24, based on the multiple single-task universal voiceprint learning data, determine the target universal voiceprint learning data, and configure the universal voiceprint recognition model and the target universal voiceprint learning data to the device end.

[0075] Specifically, the offline personalized voiceprint learning method provided in the embodiment of the present invention is executed by an offline personalized voiceprint learning device, which can be configured in the cloud.

[0076] First, execute step S21, such as Figure 8 As shown, a large amount of first initial universal voiceprint learning data D is used to perform meta-learning on the neural network model to obtain a pre-trained model A. The first initial universal voiceprint learning data D can be determined based on public voice data, and can include universal voice data and speaker labels therein. The specific structure of the neural network model can be selected as needed and is not specifically limited here.

[0077] Using multiple single-task universal voiceprint learning data, meta-learning is performed on the pre-trained models to obtain multiple first training models. The number of single-task universal voiceprint learning data can be set as needed, for example, it can be set to n, respectively represented as d1, d2, ..., dn. Each single-task universal voiceprint learning data can include a small amount of universal voiceprint learning data under the single task.

[0078] By meta-learning the pre-trained model through each single-task universal voiceprint learning data, a first training model can be obtained, that is, the single-task universal voiceprint learning data corresponds to the first training model one by one. Each first training model can be represented as B1, B2, ..., Bn.

[0079] Then, step S22 is performed to calculate the discrete index of the structural parameters at the same position in each first training model. The discrete index may include index parameters such as variance and standard deviation for indicating the degree of discretization of the structural parameters at the same position.

[0080] By using the discrete index, the target parameter in the pre-trained model A can be selected for freezing. Here, the target parameter can be a structural parameter whose discrete index is less than a preset threshold, and the preset threshold can be set as needed and is not specifically limited here.

[0081] Here, the structural parameters whose discrete index is greater than or equal to the preset threshold vary greatly in different tasks and are sensitive to personalized voiceprint learning data, while the target parameters whose discrete index is less than the preset threshold vary less in different tasks and are insensitive to personalized voiceprint learning data. Therefore, freezing the target parameters whose discrete index is less than the preset threshold in the pre-trained model A can reduce the amount of computation required for subsequent device-side training of the general voiceprint recognition model.

[0082] Thereafter, step S23 is executed to perform meta-learning on the frozen model using the first initial universal voiceprint learning data D to obtain a universal voiceprint recognition model M.

[0083] Finally, execute step S24, and use multiple single-task universal voiceprint learning data to determine the target universal voiceprint learning data. For example, multiple single-task universal voiceprint learning data can be directly used as the target universal voiceprint learning data, or the target universal voiceprint learning data can be obtained by screening from multiple single-task universal voiceprint learning data. No specific limitation is made here.

[0084] Furthermore, the universal voiceprint recognition model and the target universal voiceprint learning data can be configured to the device side.

[0085] The offline personalized voiceprint learning method provided in the embodiment of the present invention is applied to the cloud. A universal voiceprint recognition model is obtained by training in the cloud. During the training process, the target parameters in the pre-trained model are selected and frozen, and then the first initial universal voiceprint learning data is used to perform meta-learning on the frozen model to obtain a universal voiceprint recognition model. This can greatly reduce the amount of computation required for subsequent model training on the device side.

[0086] On the basis of the above embodiment, the step of determining target universal voiceprint learning data based on the plurality of single-task universal voiceprint learning data includes: Based on the multiple single-task universal voiceprint learning data, meta-learning is performed on the universal voiceprint recognition model to obtain a second training model, and based on the second initial universal voiceprint learning data, the universal voiceprint recognition model and the second training model are tested respectively to obtain a fourth test result; Based on the accuracy of the universal voiceprint recognition model and the second training model for the multiple single-task universal voiceprint learning data, and / or the fourth test result, the multiple single-task universal voiceprint learning data are screened to obtain the target universal voiceprint learning data.

[0087] Specifically, when determining the target universal voiceprint learning data, multiple single-task universal voiceprint learning data can be used simultaneously to perform meta-learning on the universal voiceprint recognition model M to obtain the second training model W. The universal voiceprint recognition model M and the second training model W are used to screen multiple single-task universal voiceprint learning data for their accuracy to obtain the target universal voiceprint learning data. For example, data that cannot be correctly recognized by the universal voiceprint recognition model M but can be correctly recognized by the second training model W can be selected to verify the effectiveness of meta-learning.

[0088] Then, according to the second initial universal voiceprint learning data, the universal voiceprint recognition model M and the second training model W are tested respectively to obtain a fourth test result. The second initial universal voiceprint learning data can be universal voiceprint learning data that is different from the first initial universal voiceprint learning data and each single-task universal voiceprint learning data.

[0089] The fourth test result may include the fifth test result corresponding to the universal voiceprint recognition model M and the sixth test result corresponding to the second training model W. The fifth test result and the sixth test result are used to screen multiple single-task universal voiceprint learning data. For example, the universal voiceprint recognition model M and the second training model W can both identify correct data to verify that the meta-learning does not deteriorate on the universal data. Finally, the target universal voiceprint learning data can be obtained after screening.

[0090] In summary, the offline personalized voiceprint learning method provided in the embodiment of the present invention obtains a universal voiceprint recognition model through meta-learning training in the cloud and configures the device side. On the device side, the universal voiceprint recognition model is fine-tuned. The overall process is as follows: Fig. 9 As shown, including: The cloud can obtain a universal voiceprint recognition model M using the first initial universal voiceprint learning data D and multiple single-task universal voiceprint learning data, which is preset on the device end before leaving the factory.

[0091] By using the universal voiceprint recognition model M to filter multiple single-task universal voiceprint learning data, the target universal voiceprint learning data can be obtained, which is preset to the local device end at the factory.

[0092] The personalized voiceprint learning data V with high credibility is obtained through the previous user interaction steps.

[0093] On the device side, the target universal voiceprint learning data U and the personalized voiceprint learning data V are used to train the universal voiceprint recognition model M to obtain the target personalized voiceprint recognition model P.

[0094] Existing speaker separation technology mainly adopts the "cloud + terminal" implementation solution, where the device initiates a session request, uploads the voice data to the cloud service for processing through the Internet, and then transmits the speaker separation result to the device. However, the existing cloud processing solution requires the device to be connected to the Internet, and the user needs to upload data such as voice, which poses a risk of privacy leakage. Moreover, considering the real-time and non-real-time separation effects, different algorithms are generally used to achieve them, so the cloud needs to deploy real-time and non-real-time services separately. With this independent deployment method, users need to call two sets of services to achieve the real-time and non-real-time separation functions, which has high deployment costs and maintenance costs.

[0095] Based on the above embodiments, Fig.10 As shown, an embodiment of the present invention further provides a speaker separation method, which is applied to a device end and includes: S31, determining user separation configuration information; the user separation configuration information includes at least one of a real-time separation mode and a non-real-time separation mode; S32, in the real-time separation mode, receiving a real-time voice stream, and performing voiceprint recognition on the real-time voice stream based on the target personalized voiceprint recognition model obtained by the offline personalized voiceprint learning method provided in the above embodiments to obtain a second voiceprint recognition result, and performing speaker separation on the real-time voice stream based on the second voiceprint recognition result or based on the speaker registration information and the second voiceprint recognition result; S33, in a non-real-time separation mode, obtaining global speech, and performing voiceprint recognition on the global speech based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result, and performing global speaker separation on the global speech based on the third voiceprint recognition result.

[0096] Specifically, the speaker separation method provided in the embodiment of the present invention is executed by a speaker separation system, which can be configured in a target device. The target device can be an electronic device that performs a speaker separation task using a target personalized voiceprint recognition model obtained by an offline personalized voiceprint learning method, which is not specifically limited here.

[0097] First, step S31 is performed to determine user separation configuration information. The user separation configuration information may be a separation mode pre-configured by the user, and may include at least one of a real-time separation mode and a non-real-time separation mode, that is, the user may configure only the real-time separation mode, only the non-real-time separation mode, or both the real-time separation mode and the non-real-time separation mode.

[0098] The real-time separation mode refers to a mode in which the speaker separation is performed directly after the voiceprint recognition is performed on the real-time voice stream, and the non-real-time separation mode refers to a mode in which the voiceprint recognition is performed on the global voice after the real-time voice stream transmission is completed, and the speaker separation is performed.

[0099] Execute step S32, such as Fig.11 As shown, in the real-time separation mode, a real-time voice stream is received. Thereafter, voice activity detection (VAD) can be performed on the real-time voice stream to detect whether a voice signal exists.

[0100] The target personalized voiceprint recognition model obtained by the offline personalized voiceprint learning method provided in the above embodiments is used to perform voiceprint recognition on the real-time voice stream to obtain a second voiceprint recognition result. The second voiceprint recognition result refers to the voiceprint feature in the real-time voice stream.

[0101] After that, the second voiceprint recognition result can be used to separate the speakers of the real-time voice stream and obtain the real-time speaker separation result. At this time, the speaker blind separation can be used by default, that is, if there are several speakers in the real-time voice stream, the speech fragments of several speakers are separated, and the speaker separation result is displayed in the form of speaker 1, speaker 2, speaker 3, ..., speaker N. Fig.12 The conversation shown includes speech segment texts of Speaker 1 and Speaker 2.

[0102] The speaker registration information and the second voiceprint recognition result can also be used to separate the speakers of the real-time voice stream. The speaker registration information refers to the corresponding relationship between the speaker and the voiceprint feature.

[0103] The speaker registration information may include offline speaker voiceprint registration information and / or dynamic speaker tag information. Offline speaker voiceprint registration information refers to the correspondence between the speaker and the voiceprint feature input by the user before the session. The voiceprint feature in the second voiceprint recognition result is matched with the voiceprint feature in the offline speaker voiceprint registration information, and when the match is successful, the matched speaker information in the offline speaker voiceprint registration information is displayed, such as Fig.13 If the match is recognized, the default speaker blind analysis result will continue to be displayed, such as Fig.13 Speaker 2 in .

[0104] Dynamic speaker tag information refers to the speaker tag information manually entered by the user during the conversation. After the user tags, the current tagged speaker and the speaker in the subsequent real-time voice stream are replaced with the speaker tag information. Fig.14 As shown in FIG. 1 , the user marks the speaker 2 on the left as Li Si's voiceprint, then the currently marked speaker 2 and the speaker 2 in the subsequent real-time voice stream are replaced by Li Si, as shown in FIG. Fig.14 Right side.

[0105] In real-time separation mode, it is possible to recognize and separate speech segments of multiple speakers in real time, with a response time of less than 800ms.

[0106] Execute step S33, in the non-real-time separation mode, the global voice can be obtained by caching the real-time voice stream, and the target personalized voiceprint recognition model is used to perform voiceprint recognition on the global voice to obtain a third voiceprint recognition result. The third voiceprint recognition result refers to the voiceprint feature in the global voice.

[0107] Finally, the third voiceprint recognition result can be used to perform speaker separation on the global voice by default blind speaker separation to obtain a non-real-time speaker separation result. The speaker registration information and the third voiceprint recognition result can also be used to perform speaker separation on the global voice, which is not specifically limited here.

[0108] It is understandable that in non-real-time separation mode, asynchronous thread processing can fully utilize multi-core performance and improve program execution efficiency. One hour of global speech can give speaker separation results within 5 seconds. Both real-time separation mode and real-time separation mode use the target personalized voiceprint recognition model when performing voiceprint recognition, which can greatly reduce the amount of calculation, improve efficiency, and reduce memory usage. In addition, both real-time separation mode and non-real-time separation mode can be configured to facilitate flexible use by users.

[0109] In the embodiment of the present invention, the real-time and non-real-time separation functions can be supported simultaneously, and there is no need to deploy two sets of engines separately, which effectively reduces the deployment cost. Moreover, the training of the target personalized voiceprint recognition model and the speaker separation operation are both performed on the device side, without the need for networking, and there is no risk of user privacy leakage.

[0110] On the basis of the above embodiment, the speaker separation is performed on the real-time voice stream based on the speaker registration information and the second voiceprint recognition result, and then further includes: storing the speaker separation result and the second voiceprint recognition result in a temporary voiceprint database in correspondence; If the real-time voice stream is interrupted, receiving the voice after the interruption, and performing voiceprint recognition on the voice after the interruption based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result; Based on the third voiceprint recognition result, the temporary voiceprint library is applied to perform speaker separation on the interrupted speech.

[0111] Specifically, in the embodiment of the present invention, the method can also realize automatic association of the same speaker in different sessions, that is, session continuation, which can greatly improve the user experience.

[0112] In actual conference scenarios, a conference may be interrupted due to active or passive reasons, and the same speaker may be divided into different roles before and after the interruption. Fig.15 As shown in the figure, assume that before the interruption, speakers A, B, and C are identified as speakers 1, 2, and 3. Because the session after the interruption is generally treated as a new session, if only speakers B and C speak in turn in the new session, then speaker B will be identified as speaker 1 and speaker C will be identified as speaker 2. From the user's perspective, it is definitely hoped that speaker B will continue to be identified as speaker 2 and speaker C will still be speaker 3. Fig.16 .

[0113] Therefore, after the real-time voice stream is speaker separated, the speaker separation result and the second voiceprint recognition result can be stored in a temporary voiceprint library correspondingly. The storage process will not be interrupted due to the interruption of the session. Therefore, the temporary voiceprint library contains the voiceprint features of each speaker in the speaker separation result before the session interruption and the voiceprint features of each speaker in the speaker separation result after the session interruption, which may include the voiceprint features of registered speakers and the voiceprint features of unregistered speakers. The temporary voiceprint library can be stored in a disk.

[0114] Therefore, when the real-time voice stream is interrupted, the voice after the interruption can be received, and then the voice after the interruption can be input into the target personalized voiceprint recognition model, and the voiceprint recognition of the voice after the interruption can be performed through the target personalized voiceprint recognition model to obtain a third voiceprint recognition result.

[0115] Afterwards, if Fig.17 As shown, the third voiceprint recognition result is matched with the voiceprint features in the temporary voiceprint library. If the match is successful, the speaker with the matching voiceprint features in the temporary voiceprint library can be used as the speaker in the voice after the interruption. If the match fails, it can be used as the new speaker. In this way, the speakers before and after the interruption can be associated with each other and the conversation can be resumed.

[0116] In summary, in order to solve the problem of difficulty in upgrading voiceprint models and algorithms for confidential and network-inaccessible devices, an offline personalized voiceprint learning method and a purely offline speaker separation method are proposed in the embodiments of the present invention, and the application of the target personalized voiceprint recognition model in the speaker separation task is described. The offline personalized voiceprint learning method in the embodiments of the present invention fully considers the limitations of offline device conditions, and performs personalized training locally through the sample voice accumulated during user use, which can continuously iterate and improve the effect of voiceprint recognition while ensuring confidentiality. The offline personalized voiceprint learning method has low deployment cost, low maintenance cost, and is user privacy-friendly. It is a good solution to the problem of "difficulty in upgrading offline algorithms on the end side", and is also a good supplement and extension to the cloud separation solution. In addition, this offline personalized voiceprint learning method is not only applicable to the offline voiceprint field, but can be expanded to many directions such as images, videos and texts, and has good research prospects and practical application value. Based on the offline personalized voiceprint learning method, the embodiments of the present invention also provide a purely offline speaker separation method, which can realize real-time and non-real-time speaker separation functions, and also has functions such as "multi-function registration" and "conversation continuation".

[0117] like Fig.18 As shown, based on the above embodiment, an offline personalized voiceprint learning device is provided in an embodiment of the present invention, which is applied to a device end and includes: The built-in information acquisition module 181 is used to determine the built-in universal voiceprint recognition model and target universal voiceprint learning data, and obtain personalized voiceprint learning data; the target universal voiceprint learning data includes training data and test data; A first meta-learning module 182, configured to perform meta-learning on the general voiceprint recognition model based on the training data and the personalized voiceprint learning data to obtain an initial personalized voiceprint recognition model; The first testing module 183 is used to test the universal voiceprint recognition model and the initial personalized voiceprint recognition model based on the test data, obtain a first test result, and determine a target personalized voiceprint recognition model based on the first test result.

[0118] Specifically, the functions of each module in the offline personalized voiceprint learning device provided in the embodiment of the present invention correspond one-to-one to the operation flow of each step in the above-mentioned method embodiment with the device end as the execution subject, and the effects achieved are also consistent. Please refer to the above-mentioned embodiment for details, and no further description will be given in the embodiment of the present invention.

[0119] like Fig.19 As shown, based on the above embodiment, an offline personalized voiceprint learning device is provided in the embodiment of the present invention, which is applied in the cloud, including: The second meta-learning module 191 is used to perform meta-learning on the neural network model based on the first initial universal voiceprint learning data to obtain a pre-trained model, and to perform meta-learning on the pre-trained models based on multiple single-task universal voiceprint learning data to obtain multiple first training models; A discrete index calculation module 192, used to calculate discrete indexes of structural parameters at the same position in the plurality of first training models, and based on the discrete indexes, select target parameters in the pre-training model for freezing; A third meta-learning module 193, configured to perform meta-learning on the frozen model based on the first initial universal voiceprint learning data to obtain the universal voiceprint recognition model; The device configuration module 194 is used to determine the target universal voiceprint learning data based on the multiple single-task universal voiceprint learning data, and configure the universal voiceprint recognition model and the target universal voiceprint learning data to the device end.

[0120] Specifically, the functions of each module in the offline personalized voiceprint learning device provided in the embodiment of the present invention correspond one-to-one to the operation procedures of each step in the above-mentioned method embodiment with the cloud as the execution subject, and the effects achieved are also consistent. Please refer to the above-mentioned embodiment for details, which will not be repeated in the embodiment of the present invention.

[0121] like Fig. 20 As shown, based on the above embodiment, an embodiment of the present invention provides a speaker separation system, which is applied to a device end and includes: The configuration information determination module 201 is used to determine user separation configuration information; the user separation configuration information includes at least one of a real-time separation mode and a non-real-time separation mode; The first speaker separation module 202 is used to receive a real-time voice stream in a real-time separation mode, and perform voiceprint recognition on the real-time voice stream based on the target personalized voiceprint recognition model obtained by the above-mentioned offline personalized voiceprint learning method to obtain a second voiceprint recognition result, and perform speaker separation on the real-time voice stream based on the second voiceprint recognition result or based on the speaker registration information and the second voiceprint recognition result; The second speaker separation module 203 is used to obtain the global speech in the non-real-time separation mode, and perform voiceprint recognition on the global speech based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result, and perform speaker separation on the global speech based on the third voiceprint recognition result or based on the speaker registration information and the third voiceprint recognition result.

[0122] Fig.21 An example of a physical structure diagram of a speaker separation integrated machine is shown in FIG. Fig.21As shown, the speaker separation integrated machine may include: a processor 210, a communications interface 220, a memory 230 and a communication bus 240, wherein the processor 210, the communications interface 220 and the memory 230 communicate with each other through the communication bus 240. The memory 230 may store a target personalized voiceprint recognition model, and the processor 210 may store a computer program that can be run on the processor 210.

[0123] The processor 210 may call the target personalized voiceprint recognition model in the memory 230 to execute the speaker separation method provided in the above embodiments.

[0124] In addition, the logic instructions in the above-mentioned memory 230 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the relevant technology or the 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 a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0125] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the offline personalized voiceprint learning method or speaker separation method provided by the above methods.

[0126] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the offline personalized voiceprint learning method or speaker separation method provided by the above methods.

[0127] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0128] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An offline personalized voiceprint learning method, characterized in that: Applied to the device side, including: Determine a built-in universal voiceprint recognition model and target universal voiceprint learning data, and obtain personalized voiceprint learning data; the target universal voiceprint learning data includes training data and test data; Based on the training data and the personalized voiceprint learning data, meta-learning the general voiceprint recognition model is performed to obtain an initial personalized voiceprint recognition model; Based on the test data, the universal voiceprint recognition model and the initial personalized voiceprint recognition model are tested respectively to obtain a first test result, and based on the first test result, a target personalized voiceprint recognition model is determined.

2. The offline personalized voiceprint learning method according to claim 1, characterized in that: The first test result includes a second test result corresponding to the universal voiceprint recognition model and a third test result corresponding to the initial personalized voiceprint recognition model; The step of determining a target personalized voiceprint recognition model based on the first test result includes: If the second test result is better than the third test result, determining the universal voiceprint recognition model as the target personalized voiceprint recognition model; If the third test result is better than the second test result, the initial personalized voiceprint recognition model is determined to be the target personalized voiceprint recognition model.

3. The offline personalized voiceprint learning method according to claim 1, characterized in that: The obtaining of personalized voiceprint learning data includes: Get sample speech; Based on the target personalized voiceprint recognition model, voiceprint recognition is performed on the sample speech to determine a first voiceprint recognition result, and based on the first voiceprint recognition result, speaker separation is performed on the sample speech to obtain a speaker separation result; Receive feedback information from a user regarding the speaker separation result, and construct the personalized voiceprint learning data based on the feedback information and the sample speech.

4. The offline personalized voiceprint learning method according to claim 3, characterized in that: The feedback information includes the interested speaker in the sample speech and the modification data of the user when an error occurs in the first voiceprint recognition result.

5. An offline personalized voiceprint learning method, characterized in that: Applied in the cloud, including: Based on the first initial universal voiceprint learning data, meta-learning is performed on the neural network model to obtain a pre-trained model, and based on multiple single-task universal voiceprint learning data, meta-learning is performed on the pre-trained models respectively to obtain multiple first training models; Calculating discrete indicators of structural parameters at the same position in the plurality of first training models, and selecting target parameters in the pre-training model for freezing based on the discrete indicators; Based on the first initial universal voiceprint learning data, meta-learning is performed on the frozen model to obtain the universal voiceprint recognition model; Based on the multiple single-task universal voiceprint learning data, target universal voiceprint learning data is determined, and the universal voiceprint recognition model and the target universal voiceprint learning data are configured to the device end.

6. The offline personalized voiceprint learning method according to claim 5, characterized in that: The target parameter is a structural parameter whose discrete index is less than a preset threshold.

7. The offline personalized voiceprint learning method according to claim 5, characterized in that: The determining target universal voiceprint learning data based on the plurality of single-task universal voiceprint learning data comprises: Based on the multiple single-task universal voiceprint learning data, meta-learning is performed on the universal voiceprint recognition model to obtain a second training model, and based on the second initial universal voiceprint learning data, the universal voiceprint recognition model and the second training model are tested respectively to obtain a fourth test result; Based on the accuracy of the universal voiceprint recognition model and the second training model for the multiple single-task universal voiceprint learning data, and / or the fourth test result, the multiple single-task universal voiceprint learning data are screened to obtain the target universal voiceprint learning data.

8. A method for speaker separation, characterized in that: Applied to the device side, including: Determine user separation configuration information; the user separation configuration information includes at least one of a real-time separation mode and a non-real-time separation mode; In the real-time separation mode, a real-time voice stream is received, and based on the target personalized voiceprint recognition model obtained by the offline personalized voiceprint learning method according to any one of claims 1 to 4, voiceprint recognition is performed on the real-time voice stream to obtain a second voiceprint recognition result, and speaker separation is performed on the real-time voice stream based on the second voiceprint recognition result or based on the speaker registration information and the second voiceprint recognition result; In the non-real-time separation mode, a global speech is obtained, and voiceprint recognition is performed on the global speech based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result, and speaker separation is performed on the global speech based on the third voiceprint recognition result, or based on the speaker registration information and the third voiceprint recognition result.

9. The speaker separation method according to claim 8, characterized in that: The speaker registration information includes offline speaker voiceprint registration information and / or dynamic speaker marking information.

10. The speaker separation method according to claim 8, characterized in that: The method further comprises performing speaker separation on the real-time voice stream based on the speaker registration information and the second voiceprint recognition result, and then further comprising: storing the speaker separation result and the second voiceprint recognition result in a temporary voiceprint database in correspondence; If the real-time voice stream is interrupted, receiving the voice after the interruption, and performing voiceprint recognition on the voice after the interruption based on the target personalized voiceprint recognition model to obtain a third voiceprint recognition result; Based on the third voiceprint recognition result, the temporary voiceprint library is applied to perform speaker separation on the interrupted speech.

11. A speaker separation integrated machine, characterized in that: It includes a memory, a processor, a target personalized voiceprint recognition model stored in the memory, and a computer program stored in and executable on the processor; When the processor executes the computer program, the target personalized voiceprint recognition model is called to implement the speaker separation method according to any one of claims 8 to 10.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the offline personalized voiceprint learning method according to any one of claims 1 to 7 or the speaker separation method according to any one of claims 8 to 10 is implemented.

Images