Method, system for training voiceprint recognition model, and voiceprint recognition method and system

Abstract

The training method and system of the voiceprint recognition model and the voiceprint recognition method and system provided in the specification propose a scheme for distinguishing similar speakers. In the voiceprint recognition model training process, the residual value of the similar category similar to the target sample voiceprint feature is adaptively adjusted, so that the weight of the similar category in the training process is increased, and the weight of the similar speaker gradient in the training process is increased, so that the network pays more attention to the distinction of similar speakers, and the voiceprint recognition model obtained by training can accurately distinguish the voiceprint features of similar speakers, and improve the security of the voiceprint recognition model.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of voiceprint recognition, and in particular, to a voiceprint recognition model training method and system, and a voiceprint recognition method and system. BACKGROUND

[0002] Speech is the most natural way of communication for human beings, and because of the differences in vocal organs such as vocal tract and vocal cords of different people, the voice of each person is unique. Therefore, voiceprint, like fingerprint, can be used for identity verification and recognition. With the development of computer and Internet technology, many scenarios use voiceprint recognition to verify the identity of users, such as voice unlocking, voice verification, voice payment, etc. Voiceprint recognition technology refers to using voiceprint recognition models to identify the identity of a speaker based on voiceprint. With the wide application of voiceprint recognition technology, the demand for the security of voiceprint recognition technology is increasing. In particular, when there are some speakers whose voices and voiceprint features are very similar, the existence of these similar speakers requires the voiceprint recognition model to have better discriminability to accurately distinguish the target speaker from the similar speakers.

[0003] Therefore, it is necessary to provide a voiceprint recognition model training method and system with higher precision, and a voiceprint recognition method and system, which can accurately identify similar speakers. SUMMARY

[0004] The present specification provides a voiceprint recognition model training method and system with higher precision, and a voiceprint recognition method and system, which can accurately identify similar speakers.

[0005] In a first aspect, the present specification provides a voiceprint recognition model training method, comprising: obtaining a sample audio data set, the sample audio data set comprising a plurality of sample audio data and respective categories to which the sample audio data belong; inputting the sample audio data set into a pre-trained first voiceprint recognition model to obtain a plurality of predicted values; determining a target sample data from the plurality of sample audio data, wherein a difference between a predicted value corresponding to at least one similar category in a category other than the category to which the target sample data belongs and the predicted value of the category to which the target sample data belongs is greater than a preset threshold value; adjusting the predicted value corresponding to the target sample data based on a margin mechanism to increase the proportion of the similar category; and training the first voiceprint recognition model based on the plurality of sample audio data and the adjusted predicted value to obtain a final voiceprint recognition model.

[0006] In some embodiments, the first voiceprint recognition model comprises a model trained based on the sample audio data set.

[0007] In some embodiments, the plurality of sample audio data comprises M categories, where M is an integer greater than 1, and the plurality of prediction values correspond to the plurality of sample audio data one-to-one, each of the plurality of prediction values comprises an M-dimensional feature vector corresponding to the M categories respectively.

[0008] In some embodiments, determining the target sample data from the plurality of sample audio data comprises performing the following on each prediction value: determining a prediction value of a category to which a current sample audio data belongs and a prediction value of a category other than the category to which the current sample audio data belongs from the M-dimensional feature vector; comparing the prediction value of the category other than the category to which the current sample audio data belongs with the prediction value of the category to which the current sample audio data belongs, and determining whether the current sample audio data is the target sample data based on the comparison result.

[0009] In some embodiments, determining whether the current sample audio data is the target sample data based on the comparison result comprises one of the following: determining that there is at least one prediction value of the prediction value of the category other than the category to which the current sample audio data belongs that has a difference greater than the threshold value with the prediction value of the category to which the current sample audio data belongs, determining that the category corresponding to the at least one prediction value is at least one similar category of the current sample audio data, and determining that the current sample audio data is the target sample data; and determining that there is no prediction value of the prediction value of the category other than the category to which the current sample audio data belongs that has a difference greater than the threshold value with the prediction value of the category to which the current sample audio data belongs, determining that the current sample audio data is not the target sample data.

[0010] In some embodiments, adjusting the prediction value corresponding to the target sample data based on the margin mechanism to increase the proportion of the similar category comprises one of the following: adjusting the prediction value of the similar category corresponding to the target sample data based on a positive margin; adjusting the prediction value of the similar category corresponding to the target sample data and the prediction value of the remaining categories other than the category to which the target sample data belongs based on a negative margin; and adjusting the prediction value of the similar category corresponding to the target sample data based on a positive margin, and adjusting the prediction value of the similar category corresponding to the target sample data and the prediction value of the remaining categories other than the category to which the target sample data belongs based on a negative margin.

[0011] In some embodiments, the training of the first voiceprint recognition model based on the plurality of sample audio data and the adjusted prediction values to obtain a final voiceprint recognition model comprises: determining a target loss function corresponding to the target sample data based on the adjusted prediction value corresponding to the target sample data; determining other loss functions corresponding to other sample audio data other than the target sample data based on prediction values of the other sample audio data; and training the first voiceprint recognition model based on the target loss function and the other loss functions to obtain the final voiceprint recognition model.

[0012] In a second aspect, the present specification provides a voiceprint recognition model training system, comprising at least one storage medium and at least one processor, the at least one storage medium storing at least one instruction set for training a voiceprint recognition model; the at least one processor is in communication connection with the at least one storage medium, wherein when the voiceprint recognition model training system is running, the at least one processor reads the at least one instruction set, and executes the voiceprint recognition model training method of the first aspect of the present specification according to the indication of the at least one instruction set.

[0013] In a third aspect, the present specification also provides a voiceprint recognition method, comprising: obtaining target audio data to be recognized; inputting the target audio data into a voiceprint recognition model to obtain target voiceprint features corresponding to the target audio data, the voiceprint recognition model being trained based on the voiceprint recognition model training method of the first aspect of the present specification; and determining identity information of a target speaker in the target audio data based on the target voiceprint features.

[0014] In a fourth aspect, the present specification also provides a voiceprint recognition system, comprising at least one storage medium and at least one processor, the at least one storage medium storing at least one instruction set for performing voiceprint recognition; the at least one processor is in communication connection with the at least one storage medium, wherein when the voiceprint recognition system is running, the at least one processor reads the at least one instruction set, and executes the voiceprint recognition method of the third aspect of the present specification according to the indication of the at least one instruction set.

[0015] According to the technical solution, the training method and system of the voiceprint recognition model and the voiceprint recognition method and system provided in the specification propose a scheme for distinguishing similar speakers. In the training process of the voiceprint recognition model, the residual value of the similar category similar to the voiceprint feature of the target sample is adaptively adjusted, so that the weight of the similar category in the training process is increased, and the weight of the gradient of the similar speaker in the training process is increased, so that the network pays more attention to the distinction of similar speakers, and the voiceprint recognition model obtained by training can accurately distinguish the voiceprint features of similar speakers, and the security of the voiceprint recognition model is improved.

[0016] Other functions of the training method and system of the voiceprint recognition model and the voiceprint recognition method and system provided in the specification will be partially listed in the following description. According to the description, the following numbers and examples will be apparent to those skilled in the art. The creative aspects of the training method and system of the voiceprint recognition model and the voiceprint recognition method and system provided in the specification can be fully explained by practicing or using the methods, devices and combinations described in the following detailed examples. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the specification, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the specification, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0018] Figure 1 A schematic diagram of a voiceprint recognition model training system according to an embodiment of the specification is shown;

[0019] Figure 2 A schematic diagram of a voiceprint recognition system according to an embodiment of the specification is shown;

[0020] Figure 3 A hardware structure diagram of a computing device according to an embodiment of the specification is shown;

[0021] Figure 4 A flowchart of a voiceprint recognition model training method according to an embodiment of the specification is shown;

[0022] Figure 5 A flowchart of determining target sample data according to an embodiment of the specification is shown; and

[0023] Figure 6 A flowchart of a voiceprint recognition method according to an embodiment of the specification is shown. DETAILED DESCRIPTION

[0024] The following description provides specific examples and requirements for the context of the present specification, in order to enable a person skilled in the art to make and use the content of the present specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of the present specification. Therefore, the present specification is not limited to the embodiments shown, but is consistent with the widest scope of the claims.

[0025] The terms used herein are only used for the purpose of describing specific example embodiments, and are not limiting. For example, unless the context clearly indicates otherwise, as used herein, the singular forms "a", "an", and "the" can also include the plural forms. When used in the present specification, the terms "comprise", "include" and / or "contain" mean that the associated integer, step, operation, element and / or component exists, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components and / or groups.

[0026] These features of the present specification and other features, as well as the operation and function of related elements of the structure, and the combination and economy of manufacture of components can be significantly improved in view of the following description. Referring to the drawings, all of which form part of the present specification. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of the present specification. It should also be understood that the drawings are not drawn to scale.

[0027] The flowchart used in the present specification shows the operation of system implementation according to some embodiments of the present specification. It should be clearly understood that the operations of the flowchart can not be implemented in sequence. On the contrary, the operations can be implemented in reverse order or simultaneously. In addition, one or more other operations can be added to the flowchart. One or more operations can be removed from the flowchart.

[0028] For the convenience of description, the present specification will explain the terms that will appear in the following description as follows:

[0029] Voiceprint recognition: a biometric technology for identifying the identity of a speaker based on voice;

[0030] Similar speaker: a speaker with very similar voiceprint characteristics;

[0031] Deep learning: a machine learning method that uses artificial neural networks for learning inference;

[0032] Loss function: A variable used to evaluate the degree of inconsistency between the model's predicted value and the true value (also known as error). The better the loss function, the better the model performance. Different models generally use different loss functions.

[0033] Margin: It is the object's outer margin interval attribute, the difference between the intra-class boundary interval and the inter-class boundary interval.

[0034] In the prior art, the additive angular margin mechanism is often used to learn speaker differentiation when training voiceprint recognition models. The basic idea is to introduce a certain margin into the predicted values ​​of sample data calculated by the voiceprint recognition model when designing the loss function. This introduces a certain margin between class boundaries, allowing samples belonging to the same class to be more compactly distributed around the class, while also increasing the spacing between samples from different classes. However, the margin introduced in the prior art does not distinguish between different non-classes other than the class to which the sample belongs. In other words, when training voiceprint recognition models, the prior art only considers increasing the difference between samples belonging to the sample class and non-classes using the margin mechanism, without considering the differences between samples from multiple different non-classes. During model training, for a target speaker, the proportion of speakers with similar voices to the target speaker is often small in the entire training sample set. As a result, the gradient information of similar speakers is often obscured during training, preventing the network from effectively learning to distinguish between similar speakers, thus limiting the security of the voiceprint recognition system.

[0035] This specification proposes a training method and system for a voiceprint recognition model, as well as a voiceprint recognition method and system. During the voiceprint recognition model training process, the residual value of the prediction value of similar categories that are close to the target sample voiceprint characteristics is adaptively adjusted, so that the weight of similar categories in the training process is increased, so that the trained voiceprint recognition model can accurately distinguish the voiceprint characteristics of similar speakers, thereby improving the security of the voiceprint recognition model.

[0036] Figure 1 FIG1 shows a schematic diagram of a training system 001 for a voiceprint recognition model provided in accordance with an embodiment of this specification. Figure 1 As shown, the voiceprint recognition model training system 001 is a deep learning processor that can include a control module 110, a storage module 120, and a computing module 130. The control module 110 is used to control the computing module 130 and the storage module 120 to complete the deep learning task. The computing module 130 is configured to process the deep learning computing task. The storage module 120 is used to store or transfer related computing data.

[0037] like Figure 1As shown, the control module 110 may include an instruction fetch unit 111 and an instruction decoding unit 112. The instruction fetch unit 111 is used to fetch instructions from an external memory (e.g., a DRAM, commonly known as a memory stick). The instruction decoding unit 112 decodes the instructions and sends them to the operation module 130 and storage module 120 for execution.

[0038] like Figure 1 As shown, the above-mentioned operation module 130 may include a vector operation unit 131 and a matrix operation unit 132. The vector operation unit 131 performs vector operations and can support complex operations such as vector multiplication, addition, and nonlinear transformation. It is applicable to various operation modes of various data types and can support pre-processing of input neurons and post-processing of output neurons such as table lookup, pooling, edge expansion, vector comparison, vector maximum value, and data format conversion. The matrix operation unit 132 can be responsible for the core calculations of deep learning algorithms, such as matrix multiplication and convolution, and can implement the operation functions of convolutional layers and fully connected layers.

[0039] like Figure 1 As shown, the storage module 120 may include a direct memory access unit 121, a neuron storage unit 122, and a weight storage unit 123. The direct memory access unit 121 may coordinate data interaction between the neuron storage unit 122 and the weight storage unit 123 and the external memory. The neuron storage unit 122 is used to store data such as input neurons, output neurons, and intermediate results of the deep learning network. The weight storage unit 123 is used to store the weights of the deep learning network. The direct memory access unit 121 connects the external memory to the neuron storage unit 122 and the weight storage unit 123 via a storage bus Bus, and is responsible for data transmission between the external memory and the neuron storage unit 122 and the weight storage unit 123.

[0040] like Figure 1As shown, when the deep learning processor starts a deep learning operation, the instruction acquisition unit 111 reads program instructions from the external memory through the direct memory access unit 121; the program instructions are decoded by the instruction decoding unit 112 and distributed to the direct memory access unit 121, the vector operation unit 131 and the matrix operation unit 132. After receiving the distributed instructions, the direct memory access unit 121 sends the memory access instructions to the external memory through the storage bus Bus, so as to read the neuron data stored in the external memory into the neuron storage unit 122 and read the weights stored in the external memory into the weight storage unit 123 by using the instructions; after receiving the instructions, the vector operation unit 131 reads the neuron data in the neuron storage unit 122 and pre-processes the neuron data (such as boundary expansion, etc.), and then sends the pre-processed neuron data to the matrix operation unit 132; after receiving the instructions, the matrix operation unit 132 receives the pre-processed neuron data from the vector operation unit 131 and reads the weight data from the weight storage unit 123, and sends the result to the vector operation unit 131 after completing the matrix operation; the vector operation unit 131 post-processes the output neurons, such as activation or pooling, etc., and then saves the post-processing result in the neuron storage unit 122; the direct memory access unit 121 writes the output neurons from the neuron storage unit 122 back to the external memory.

[0041] Figure 2 A schematic diagram of a voiceprint recognition system 002 provided by an embodiment of the present specification is shown. The voiceprint recognition system 002 can be used for voiceprint recognition in any scenario, such as voiceprint recognition in a door access scenario, voiceprint recognition in a payment scenario, voiceprint recognition in an unlocking scenario, etc. As shown in the figure, Figure 2 As shown, the voiceprint recognition system 002 can include a target speaker 210, a client 220, a server 230 and a network 240.

[0042] The target speaker 210 can be a user waiting for voiceprint recognition or a user undergoing voiceprint recognition. The target speaker 210 can perform voiceprint recognition on the client 220.

[0043] The client 220 can be a device capable of audio acquisition of the voice of the target speaker 210. In some embodiments, the voiceprint recognition method can be executed on the client 220. At this time, the client 220 can store data or instructions for executing the voiceprint recognition method described in the present specification, and can execute or be used to execute the data or instructions. In some embodiments, the client 220 can include a hardware device with data information processing function and necessary programs required to drive the hardware device to work. As shown in the figure, Figure 2As shown, the client 220 can be communicatively connected with the server 230. In some embodiments, the server 230 can be communicatively connected with a plurality of clients 220. In some embodiments, the client 220 can interact with the server 230 through the network 240 to receive or send messages, such as receiving or sending audio data. In some embodiments, the client 220 can include a mobile device, a tablet, a notebook, a built-in device of a motor vehicle, or the like, or any combination thereof. In some embodiments, the mobile device can include a smart home device, a smart mobile device, a virtual reality device, an augmented reality device, or the like, or any combination thereof. In some embodiments, the smart home device can include a smart television, a desktop computer, or the like, or any combination thereof. In some embodiments, the smart mobile device can include a smartphone, a personal digital assistant, a gaming device, a navigation device, or the like, or any combination thereof. In some embodiments, the virtual reality device or the augmented reality device can include a virtual reality headset, a virtual reality glasses, a virtual reality patch, an augmented reality headset, an augmented reality glasses, an augmented reality patch, or the like, or any combination thereof. For example, the virtual reality device or the augmented reality device can include Google glasses, a head-mounted display, a VR, or the like. In some embodiments, the built-in device in the motor vehicle can include an on-board computer, an on-board television, or the like. In some embodiments, the client 220 can include an audio acquisition device for acquiring speech audio of the target speaker 210. In some embodiments, the audio acquisition device can be a microphone (such as a bone conduction microphone, an air conduction microphone, or the like), or a microphone array, or the like. In some embodiments, the client 220 can be a device with positioning technology for positioning the location of the client 220.

[0044] In some embodiments, the client 220 can be installed with one or more applications (APPs). The APPs can provide the target speaker 110 with the ability to interact with the outside world through the network 240 and an interface. The APPs can include, but are not limited to, a web browser type APP, a search type APP, a chat type APP, a shopping type APP, a video type APP, a financial management type APP, an instant messaging tool, a mailbox client, a social platform software, or the like. In some embodiments, the client 220 can be installed with a target APP. The target APP can be used by the client 220 to acquire audio data. In some embodiments, the target APP can also perform voiceprint recognition on the target speaker 210. The target speaker 210 can trigger a voiceprint recognition request through the target APP. The target APP can respond to the voiceprint recognition request by performing the voiceprint recognition method described in this specification. The voiceprint recognition method will be described in detail in the following content.

[0045] Server 230 can be a server that provides various services, such as a backend server that supports the pages displayed on client 220. In some embodiments, the voiceprint recognition method can be executed on server 230. In this case, server 230 can store data or instructions for executing the voiceprint recognition method described in this specification and can execute or be used to execute such data or instructions. In some embodiments, server 230 can include hardware devices with data information processing capabilities and the necessary programs to drive the operation of such hardware devices. Server 230 can communicate with multiple clients 220 and receive data sent by clients 220.

[0046] The network 240 is used to provide a medium for communication connection between the client 220 and the server 230. The network 240 can facilitate the exchange of information or data. Figure 1 As shown, client 220 and server 230 can be connected to network 240 and transmit information or data to each other through network 240. In some embodiments, network 240 can be any type of wired or wireless network, or a combination thereof. For example, network 240 can include a cable network, a wired network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a wide area network (WAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, a near field communication (NFC) network, or the like. In some embodiments, network 240 can include one or more network access points. For example, network 240 can include a wired or wireless network access point, such as a base station or an Internet exchange point, through which one or more components of client 220 and server 230 can connect to network 240 to exchange data or information.

[0047] It should be understood that Figure 1 The number of clients 220, servers 230, and networks 240 in FIG. 2 is merely illustrative. Any number of clients 220, servers 230, and networks 240 may be provided as needed.

[0048] It should be noted that the voiceprint recognition method can be completely executed on the client 220 , can be completely executed on the server 230 , or can be partially executed on the client 220 and partially executed on the server 230 .

[0049] Figure 3A hardware structure diagram of a computing device 600 is shown according to an embodiment provided by the present specification. In some embodiments, the training method of the voiceprint recognition model described in the present specification can be executed on the computing device 600. The training method of the voiceprint recognition model is introduced in other parts of the present specification. In some embodiments, the voiceprint recognition method described in the present specification can be executed on the computing device 600. The voiceprint recognition method is introduced in other parts of the present specification.

[0050] As shown in Figure 3 The computing device 600 can include at least one storage medium 630 and at least one processor 620. In some embodiments, the computing device 600 can also include a communication port 650 and an internal communication bus 610. Meanwhile, the computing device 600 can also include an I / O component 660.

[0051] The internal communication bus 610 can connect different system components, including the storage medium 630, the processor 620 and the communication port 650.

[0052] The I / O component 660 supports input / output between the computing device 600 and other components.

[0053] The communication port 650 is used for data communication between the computing device 600 and the outside world, for example, the communication port 650 can be used for data communication between the computing device 600 and the network 240. The communication port 650 can be a wired communication port or a wireless communication port.

[0054] The storage medium 630 can include a data storage device. The data storage device can be a non-transitory storage medium or a transitory storage medium. For example, the data storage device can include one or more of a magnetic disk 632, a read-only memory (ROM) 634 or a random access memory (RAM) 636. The storage medium 630 also includes at least one instruction set stored in the data storage device. The instructions are computer program codes, which can include programs, routines, objects, components, data structures, processes, modules, etc. that execute the training method and / or the voiceprint recognition method of the voiceprint recognition model provided by the present specification.

[0055] The at least one processor 620 can be communicatively connected with the at least one storage medium 630 and the communication port 650 through the internal communication bus 610. The at least one processor 620 is configured to execute the at least one instruction set. When the computing device 600 is running, the at least one processor 620 reads the at least one instruction set and executes the training method of the voiceprint recognition model and / or the voiceprint recognition method provided in the present specification according to the instructions of the at least one instruction set. The processor 620 can execute all steps included in the training method of the voiceprint recognition model and / or the voiceprint recognition method. The processor 620 can be in the form of one or more processors, and in some embodiments, the processor 620 can include one or more hardware processors, such as a microcontroller, a microprocessor, a reduced instruction set computer (RISC), an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a central processing unit (CPU), a graphics processing unit (GPU), a physics processing unit (PPU), a microcontroller unit, a digital signal processor (DSP), a field programmable gate array (FPGA), an advanced RISC machine (ARM), a programmable logic device (PLD), any circuit or processor capable of executing one or more functions, or the like, or any combination thereof. For the sake of illustration only, only one processor 620 is described in the computing device 600 in the present specification. However, it should be noted that the computing device 600 in the present specification can also include multiple processors, and thus, the operations and / or method steps disclosed in the present specification can be executed by one processor as described in the present specification, or jointly executed by multiple processors. For example, if the processor 620 of the computing device 600 in the present specification executes step A and step B, it should be understood that step A and step B can also be executed jointly or separately by two different processors 620 (for example, a first processor executes step A, and a second processor executes step B, or the first and second processors jointly execute steps A and B).

[0056] The voiceprint recognition model can be used to extract the voiceprint features in the audio data to be recognized. When the voiceprint recognition model is used to extract the voiceprint features in the audio data to be recognized, the physical features (such as intensity, frequency distribution, frequency characteristics, etc.) in the frequency domain or time domain of the audio data to be recognized can be generally extracted, and the physical features are input into the voiceprint recognition model to obtain the voiceprint features of the audio data to be recognized, so as to extract as much as possible the features related to the speaker and exclude the interference information (also called noise information) irrelevant to the speaker from the input physical features. Different speakers have different voiceprint features.

[0057] The voiceprint recognition model can generally include two sub-models, i.e., a first sub-model and a second sub-model. The first sub-model and the second sub-model are connected in series. The first sub-model is a voiceprint feature extraction model, which is configured to extract voiceprint features from input features. The second sub-model is a classification model, which is configured to classify the extracted voiceprint features. The output features of the second sub-model are a feature vector. Each vector value in the feature vector corresponds to a class, and the vector value represents the score or probability of the current voiceprint feature in the corresponding class.

[0058] At present, the voiceprint recognition model is usually trained based on a loss function. Commonly used loss functions are some variants of the Softmax loss function and the loss function based on metric learning. The former usually has a sampling problem of mining difficult samples, which leads to unstable performance of the voiceprint recognition model and very time-consuming training, and the latter is easier to train and has performance not weaker than the former. For the Softmax loss function, the prior art usually optimizes the Softmax loss function by introducing an additive margin. The Softmax loss function corresponding to the i-th sample data can be expressed as the following formula:

[0059]

[0060] wherein N represents the number of sample data, M represents the number of classes to which the N sample data belong, x i represents the voiceprint feature corresponding to the i-th sample data, y i represents the class number or class label of the class to which the voiceprint feature corresponding to the i-th sample data belongs. represents the score or prediction value of the voiceprint feature corresponding to the i-th sample data in the class to which it belongs. k represents the class number or class label of the k-th class in the M classes. represents the score or prediction value of the voiceprint feature corresponding to the i-th sample data in the k-th class.

[0061] which can be expressed as the following formula:

[0062]

[0063] which can be expressed as the following formula:

[0064]

[0065] wherein, represents the angle between the voiceprint feature corresponding to the i-th sample data and the class center of the y i -th class to which the voiceprint feature belongs. represents the included angle between the voiceprint feature corresponding to the i-th sample data and the class center of the k-th class. m is the margin. s is a constant. Since the margin can be considered as a force pushing the voiceprint feature to the class center, by introducing the margin m in the loss function of the same class, the distance between the sample data in the same class is significantly lower than the inter-class distance between different classes, so that the learned features are more discriminative and can more accurately distinguish different classes.

[0066] However, the margin introduced in the prior art does not distinguish different non-sample belonging classes other than the class to which the sample belongs. That is, the prior art only considers increasing the difference between the sample belonging class and the non-sample belonging class through the margin mechanism when training the voiceprint recognition model, and does not consider the difference between the samples of multiple different non-sample belonging classes. In the model training process, for a target speaker, the proportion of similar speakers in the entire training sample set is often small. This leads to the gradient information of similar speakers being often masked in the training process, so that the network cannot effectively learn the ability to distinguish similar speakers, resulting in limited security of the voiceprint recognition system.

[0067] Figure 4 A flowchart of a voiceprint recognition model training method P100 according to an embodiment of the present specification is shown. As before, the computing device 600 can execute the voiceprint recognition model training method P100 of the present specification. Specifically, the processor 620 can read the instruction set stored in its local storage medium, and then execute the voiceprint recognition model training method P100 of the present specification according to the provisions of the instruction set. As shown in the figure, Figure 4 The method P100 can include:

[0068] S120: Obtain a sample audio data set, which includes a plurality of sample audio data and respective belonging classes.

[0069] The multiple sample audio data can come from different speakers or the same speaker. For the convenience of description, the number of the multiple sample audio data is defined as N, where N is a positive integer greater than 1. The N sample audio data can include M categories. M is a positive integer greater than 1. The M categories can be M different speakers. When the multiple sample data comes from different speakers, it can be that part of the sample audio data comes from different speakers, or all of the sample audio data comes from different speakers. The training system 001 of the voiceprint recognition model can obtain one or more pieces of speech audio data including one or more speakers, and preprocess the speech audio data to obtain the sample audio data set. The preprocessing includes but is not limited to frame processing, frequency domain conversion, frequency domain feature extraction, and category labeling, etc. The training system 001 of the voiceprint recognition model can obtain the physical features corresponding to each sample audio data through the preprocessing. The physical features can be frequency domain features or time domain features, such as amplitude features, frequency distribution features, frequency change features, etc. The category label can be the category number of the category to which each sample audio data belongs.

[0070] S140: inputting the sample audio data set into the pre-trained first voiceprint recognition model to obtain multiple prediction values.

[0071] The first voiceprint recognition model can have the same neural network structure as the voiceprint recognition model. The first voiceprint recognition model can include a model trained based on the sample audio data set. When training the first voiceprint recognition model, the training method described above can be used to train the sample audio data set as sample data. When training the first voiceprint recognition model, an additive margin can be introduced to define the loss function, or the additive margin can not be introduced to define the loss function.

[0072] The prediction value can be the output value of the first voiceprint recognition model. The number of the multiple prediction values is N. The multiple prediction values correspond one-to-one to the multiple sample audio data. Each sample audio data corresponds to a prediction value. As described above, the output result of the first voiceprint recognition model can be a feature vector. That is, each prediction value in the multiple prediction values can be an M-dimensional feature vector. The M-dimensional feature vector corresponds to the M categories respectively. Each vector value in the M-dimensional feature vector represents the score of the current sample audio data in its corresponding category. As described above, the voiceprint feature x i of the i-th sample audio data in its own category (y i category) can be represented as The voiceprint feature x iThe score or predicted value on the kth category can be expressed as k represents the category number or category label of the kth category in M ​​categories.

[0073] Taking the Softmax loss function as an example, for the yth i The voiceprint feature x of the i-th sample audio data of the category i , its loss function L i The gradient of the score or prediction value on the k-th category can be expressed as follows:

[0074]

[0075] It can be seen from formula (4) that the voiceprint feature x of the i-th sample audio data is i The corresponding loss function L i In a non-class category (i.e. i The gradient of the score on the other categories other than categories) and the voiceprint feature x of the i-th sample audio data i In this category (i.e. i categories), and is positively correlated with the voiceprint feature x of the i-th sample audio data. i In this category (i.e. i The scores on categories other than the given category are negatively correlated.

[0076] like Figure 4 As shown, the method P100 may further include:

[0077] S160: Determine target sample data from the plurality of sample audio data.

[0078] Among them, there is at least one similar category in the category to which the target sample data belongs, and the difference between the predicted value corresponding to the category to which the target sample data belongs and the predicted value is greater than a preset threshold. For the convenience of description, we mark the target sample data as t. The voiceprint feature corresponding to the target sample data is marked as x t The voiceprint feature x corresponding to the target sample data t The category is marked as y t The category to which the non-target sample data belongs is marked as y k , where y k ≠y t For the convenience of description, we mark the similar category of the target sample data as y s Similar category y s The category y to which the target sample data belongs k For the convenience of description, we put the target sample data after the prediction value calculated by the first voiceprint recognition model in the category yt The predicted value in the category y For convenience of description, we define as the self-category predicted value or self-person predicted value. We mark the predicted value of the target sample data after being calculated by the first voiceprint recognition model in the category y k The predicted value in the category y For convenience of description, we define as the non-self-category predicted value or non-self-person predicted value. We mark the predicted value of the target sample data after being calculated by the first voiceprint recognition model in the similar category y s The predicted value in the category y For convenience of description, we define as the similar-category predicted value or similar-person predicted value.

[0079] The similar-person predicted value and the self-person predicted value satisfy the following formula:

[0080]

[0081] In formula (5), λ is a preset threshold value. The threshold value λ can be any number, including negative numbers, positive numbers, natural numbers, etc. The threshold value λ can be obtained based on a statistical method, or obtained based on a machine learning method, or obtained based on an experience method. In some embodiments, the threshold value λ can be a negative number. In some embodiments, the threshold value λ can be a positive number.

[0082] When the similar-person predicted value and the self-person predicted value satisfy formula (5), it represents that there is a different person with similar or close voiceprint features to the category (target person) to which the target sample data belongs in the M categories (i.e. M persons). For convenience of description, we define the different person with similar or close voiceprint features to the target person as the similar person of the target person.

[0083] Figure 5 A flowchart for determining target sample data is shown according to an embodiment of the present specification. Figure 5 The flowchart corresponds to step S160. As Figure 5 shown, step S160 can include performing the following on each predicted value:

[0084] S162: determining the predicted value of the category (i.e. y i th category) to which the current sample audio data (i.e. the ith sample audio data) belongs from the M-dimensional feature vector and the prediction value of the other category other than the category (i.e., the y i th category) to which the current sample audio data (the i

[0085] As mentioned above, each prediction value corresponds to an M-dimensional feature vector. Each vector parameter in the M-dimensional feature vector corresponds to one of the M categories. The category (i.e., the y i th category) to which the current sample audio data (the i

[0086] S164: comparing the prediction value of the other category other than the category (i.e., the y i th category) to which the current sample audio data (the i th category) to which the current sample audio data (the i i th category) to which the current sample audio data (the i th category) to which the current sample audio data (the i

[0087] Step S164 can include one of the following cases:

[0088] S164-2: determining that there is at least one prediction value i th category) to which the current sample audio data (the i th category) to which the current sample audio data (the i th category) to which the current sample audio data (the i i th category) to which the current sample audio data (the i th category) to which the current sample audio data (the i corresponds to at least one similar category of the current sample audio data (the i s th category), and determining that the current sample audio data (the i t th category) to which the current sample audio data (the i t th category) to which the current sample audio data (the i s th category); and

[0089] S164-4: determining that the prediction value of the other category other than the category (i.e., the y i th category) to which the current sample audio data (the i There is no prediction value that is consistent with the category to which the current sample audio data (i-th sample audio data) belongs (i.e., the y-th i categories) If the difference between the two is greater than the threshold λ, it is determined that the current sample audio data (the i-th sample audio data) is not the target sample data.

[0090] like Figure 4 As shown, the method P100 may further include:

[0091] S180: Adjust the predicted value corresponding to the target sample data t based on the margin mechanism to increase the similarity category y s The proportion of.

[0092] Since the number of similar speakers in the entire sample audio dataset is small, the proportion of similar speakers in all categories of speakers in the entire sample audio dataset is small, which results in that the gradient information of similar speakers may be masked during the training of the voiceprint recognition model, making it impossible for the model to effectively learn the ability to distinguish similar speakers, resulting in limited security of the voiceprint recognition model. In order to enable the voiceprint recognition model to better distinguish similar speakers, the method P100 can increase the number of similar speakers (i.e., similar category y s ) in all categories, so that similar speakers have larger gradient weights in the model learning process, making the network more focused on distinguishing similar speakers, and thus improving the model's ability to distinguish similar speakers.

[0093] In some embodiments, the method P100 may be performed by adding similarity categories y s The predicted value of To increase the similarity category y s In some embodiments, the method P100 can reduce the similarity of the category y s The prediction value of other categories outside the similar category y is used to increase the similar category y s In some embodiments, the method P100 can also be performed by increasing the similarity category y s The predicted value of and reduce the similarity category y s The prediction value of other categories outside the similar category y is used to increase the similar category y s Specifically, step S180 may include one of the following methods:

[0094] S182: Adjust the similarity category y corresponding to the target sample data based on the positive margin (m) s The predicted value of

[0095] S184: adjusting the predicted value of the similar class y s and the class y t to which the target sample data belongs; and

[0096] S186: adjusting the predicted value of the similar class y s based on a positive margin (m) while adjusting the predicted value of the similar class y s and the class y t to which the target sample data belongs based on a negative margin (-m).

[0097] In step S182, for the convenience of description, we mark the adjusted predicted value of the similar class y s as the adjusted predicted value of the similar class y s The adjusted predicted value of the similar class y s can be represented by the following formula:

[0098]

[0099] where m is the margin. m is any number greater than 0.

[0100] In step S184, for the convenience of description, we mark the adjusted predicted value of the similar class y s and the class y t to which the target sample data belongs (y k ≠y t and y k ≠y s ) as (y k ≠y t and y k ≠y s ). The adjusted predicted value of the similar class y s and the class y t to which the target sample data belongs (y k ≠y t and y k ≠y s ) can be represented by the following formula:

[0101]

[0102] where m is the margin. m is any number greater than 0.

[0103] In step S186, for the convenience of description, we will adjust the category y to which the non-target sample data belongs k The predicted value of Marked as The category y to which the adjusted non-target sample data belongs k The predicted value of It can be expressed as the following formula:

[0104]

[0105] in,

[0106] In step S180, by increasing the similar speaker category y s The predicted value of and / or reduce the number of similar speaker categories y s The predicted values ​​of the other non-person categories (y k ≠y t and y k ≠y s ) to increase the similar speaker category y s The proportion of all categories, thus avoiding similar speaker categories y s The gradient information of is masked during the training process, so that similar speakers have larger gradient weights during the learning process, thereby improving the network's ability to distinguish similar speakers.

[0107] S190: Based on the multiple sample audio data and the adjusted prediction value, train the first voiceprint recognition model to obtain a final voiceprint recognition model.

[0108] Specifically, step S190 may include: determining the target loss function corresponding to the target sample data based on the adjusted prediction value corresponding to the target sample data; determining other loss functions corresponding to the other sample audio data based on the prediction values ​​of other sample audio data other than the target sample data in the multiple sample audio data; and training the first voiceprint recognition model based on the target loss function and the other loss functions to obtain the final voiceprint recognition model.

[0109] As previously described, the predicted value corresponding to the target sample data is adjusted in step S180 based on the margin mechanism to increase the proportion of the similar category in all categories. The target loss function of the target sample data should be calculated based on the adjusted predicted value to increase the gradient weight of the similar category. In step S180, only the predicted value corresponding to the target sample data is adjusted based on the margin mechanism, and for other sample audio data other than the target sample data, since there is no similar category similar or close to the current category in the M categories, there is no need to adjust the predicted value. Therefore, the other loss function corresponding to the other sample audio data other than the target sample data is still calculated based on the predicted value corresponding thereto.

[0110] After determining the target loss function and the other loss function, the method P100 can continue to train the first voiceprint recognition model based on the target loss function and the other loss function until the model converges. The converged model is used as the final voiceprint recognition model.

[0111] In summary, the present specification provides a voiceprint recognition model training method P100 and a voiceprint recognition model training system 001. In the voiceprint recognition model training process, the margin value of the predicted value of the similar category similar to the target sample voiceprint feature is adaptively adjusted, so that the weight of the similar category in the training process is increased, thereby enabling the voiceprint recognition model obtained by training to accurately distinguish the voiceprint features of similar speakers and improving the security of the voiceprint recognition model.

[0112] Figure 6 A flowchart of a voiceprint recognition method P200 according to an embodiment of the present specification is shown. As previously described, the computing device 600 can execute the voiceprint recognition method P200 of the present specification. Specifically, the processor 620 can read the instruction set stored in its local storage medium, and then execute the voiceprint recognition method P200 of the present specification according to the provisions of the instruction set. Take the execution of the method P200 on the server 230 as an example. It should be understood by those skilled in the art that the execution of the method P200 on the client 220 is also within the protection scope of the present specification. As shown in the figure, Figure 6 The method P200 can include:

[0113] S220: obtaining target audio data to be recognized.

[0114] The server 230 can collect the voice of the target speaker 210 through the client 220 and convert it into audio data as the target audio data to be recognized. In some embodiments, the client 220 can also perform noise reduction processing on the target audio data to be recognized in advance, or can not be subjected to noise reduction processing and directly uploaded to the server 230.

[0115] S240: input the target audio data into the voiceprint recognition model to obtain a target voiceprint feature corresponding to the target audio data.

[0116] The voiceprint recognition model is trained based on the training method P100 of the voiceprint recognition model provided in the specification. The server 230 can call the model parameters, neural source data, etc. saved by the storage medium 630, and extract the voiceprint feature of the target audio data to be recognized.

[0117] S260: based on the target voiceprint feature, determine the identity information of the target speaker 210 in the target audio data.

[0118] In some embodiments, the server 230 can pre-store voiceprint features of a plurality of different speakers. The voiceprint features can be extracted based on the voiceprint recognition model provided in the specification. The voiceprint features can be obtained by extracting the audio data of the user provided by the user in the process of registering the voiceprint recognition mechanism, and then storing the voiceprint features and the user identification of the user.

[0119] In step S260, the method P200 can calculate the similarity between the target voiceprint feature and each of the plurality of voiceprint features in the server 230, and select the user identification corresponding to the voiceprint feature with the largest similarity as the identity information of the target speaker 210. In some embodiments, according to the application scenario of the voiceprint recognition method, the pre-stored voiceprint features in the server 230 are also different.

[0120] Another aspect of the specification provides a non-transitory storage medium storing at least one set of executable instructions for training a voiceprint recognition model. When the executable instructions are executed by a processor, the executable instructions guide the processor to implement the steps of the training method P100 of the voiceprint recognition model described in the specification. In some possible implementations, various aspects of the specification can also be implemented in the form of a program product, which includes program code. When the program product is run on the computing device 600, the program code is used to make the computing device 600 execute the steps of the training method P100 of the voiceprint recognition model described in the specification.

[0121] Another aspect of the present specification provides a non-transitory storage medium storing at least one set of executable instructions for performing voiceprint recognition. When the executable instructions are executed by a processor, the executable instructions direct the processor to implement the steps of the voiceprint recognition method P200 described in the present specification. In some possible implementations, various aspects of the present specification can also be implemented in the form of a program product, which includes program codes. When the program product is run on the computing device 600, the program codes are used to cause the computing device 600 to perform the steps of the voiceprint recognition method P200 described in the present specification.

[0122] A program product for implementing the above-described methods can include a program code and can be executed on the computing device 600. However, the program product of the present specification is not limited thereto, and in the present specification, a readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system. The program product can employ any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium, for example, can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination thereof. More specific examples of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disc, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. The computer readable storage medium can include a data signal carried in a baseband or as part of a carrier wave, in which readable program codes are borne. Such a propagated data signal can take on various forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination thereof. The readable storage medium can also be any readable medium that is not a readable storage medium, which can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device. The program codes contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or the like, or any suitable combination thereof. The program codes for performing the operations of the present specification can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, or the like, and a conventional procedural programming language such as "C" language or a similar programming language. The program codes can be executed entirely on the computing device 600, partially on the computing device 600, as an independent software package, partially on the computing device 600 and partially on a remote computing device, or entirely on a remote computing device.

[0123] The above-described embodiments of the present specification have been described. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order and still accomplish desirable results. Additionally, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order to achieve desirable results. In certain implementations, multitasking and parallel processing can be advantageous.

[0124] In light of the foregoing disclosure, those skilled in the art will appreciate that the foregoing detailed description of the embodiments of the present application are presented for purposes of illustration and description. It is not intended to limit the

[0125] Moreover, certain terms have been used herein for the purpose of reference only and as such should not be construed as limiting. For example, "one embodiment", "an embodiment", and / or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Thus, use of the terms "in one embodiment" or "in an embodiment" or "in some embodiments" appears more than once in various examples is not necessarily referring to the same embodiment.

[0126] It should be understood that in the foregoing description of embodiments of the present application, various features are combined in a single embodiment for purposes of convenience and understanding. However, one of skill in the art will recognize that each of the various features described herein can be implemented in a different embodiment of the present application. Thus, it should be apparent that the various embodiments of the present application can include a different combination of features than those specifically described herein. Furthermore, while the present application has been described in detail with respect to specific embodiments, it will be apparent that various modifications and changes can be made without departing from the spirit and scope of the present application. For example, the various features of the present application can be combined in a single embodiment, or in different embodiments, without departing from the scope of the present application. Thus, the scope of the present application should not be limited by the specific embodiments described herein, but should be defined in accordance with the following claims, and their equivalents.

[0127] Each patent, patent application, publication, and other material cited in this document is incorporated herein by reference in its entirety for all purposes. Except to the extent necessary or required to be inconsistent with this document as a whole, the citation of any document herein is not to be construed as an admission that the document is at the priority date or prior to the priority date of this document, or that it alone, or along with some combination of other documents, teaches, discusses or otherwise mentions the contents of this document. To the extent that any meaning or definition of a term in this document conflicts with the meaning or definition of the same term in a document incorporated by reference, the meaning or definition assigned to the term in this document shall govern.

[0128] Finally, it should be understood that the embodiments of the application disclosed herein are illustrative of the principles of the application. Other modifications that fall within the scope of the present description can also be made. Accordingly, the disclosure of embodiments is intended to be illustrative, but not limiting, of the scope of the application. Those of ordinary skill in the art can readily devise alternative arrangements, which fall within the scope of the application as defined by the description herein. Thus, the embodiments disclosed in this description are not intended to be limited to the embodiments described herein but are meant to be illustrative of the principles of the application.

Claims

1. A method for training a voiceprint recognition model, comprising: obtaining a sample audio data set, the sample audio data set comprising a plurality of sample audio data and respective categories to which the sample audio data belong; the plurality of sample audio data comprising target sample data and a target category corresponding to the target sample data, and non-target sample data other than the target sample data; inputting the sample audio data set into a pre-trained first voiceprint recognition model to obtain a plurality of prediction values; determining the target sample data from the plurality of sample audio data, wherein at least one similar category exists in a category to which the non-target sample data belongs; the similar category is a category to which sample audio of a different speaker similar in voiceprint feature to the target sample audio data belongs, and a difference between a prediction value corresponding to the similar category and a prediction value of the category to which the target sample data belongs is greater than a preset threshold value; adjusting the prediction value corresponding to the target sample data based on a margin mechanism to increase a proportion of the similar category; and training the first voiceprint recognition model based on the plurality of sample audio data and the adjusted prediction value to obtain a final voiceprint recognition model. 2.The voiceprint recognition model training method of claim 1, wherein, The first voiceprint recognition model comprises a model trained based on the sample audio data set. 3.The voiceprint recognition model training method of claim 1, wherein, The plurality of sample audio data comprises M categories, M being an integer greater than 1, the plurality of prediction values correspond one-to-one to the plurality of sample audio data, and each prediction value in the plurality of prediction values comprises an M-dimensional feature vector corresponding to the M categories respectively.

4. The voiceprint recognition model training method of claim 3, wherein, The determining of the target sample data from the plurality of sample audio data comprises performing the following on each prediction value: determining, from the M-dimensional feature vector, a prediction value of a category to which a current sample audio data belongs and prediction values of categories other than the category to which the current sample audio data belongs; comparing the prediction values of the categories other than the category to which the current sample audio data belongs with the prediction value of the category to which the current sample audio data belongs, and determining whether the current sample audio data is the target sample data based on a comparison result.

5. The voiceprint recognition model training method of claim 4, wherein, The determining of whether the current sample audio data is the target sample data based on the comparison result comprises one of the following cases: determining that at least one prediction value of the prediction values of the categories other than the category to which the current sample audio data belongs has a difference greater than the threshold value with the prediction value of the category to which the current sample audio data belongs, determining that the at least one prediction value corresponds to at least one similar category of the current sample audio data, and determining that the current sample audio data is the target sample data; and determining that no prediction value of the prediction values of the categories other than the category to which the current sample audio data belongs has a difference greater than the threshold value with the prediction value of the category to which the current sample audio data belongs, and determining that the current sample audio data is not the target sample data. The adjusting of the prediction value corresponding to the target sample data based on the margin mechanism to increase the proportion of the similar category comprises one of the following ways:

6. The voiceprint recognition model training method of claim 3, wherein, adjusting the prediction value of the similar category corresponding to the target sample data based on a positive margin; and adjusting the prediction value of the similar category corresponding to the target sample data based on a negative margin. adjust the prediction value of the similar category corresponding to the target sample data based on a positive margin, and adjust the prediction value of the similar category corresponding to the target sample data and the prediction value of the remaining categories other than the category to which the target sample data belongs based on a negative margin. and adjust the prediction value of the similar category corresponding to the target sample data based on a positive margin, and adjust the prediction value of the similar category corresponding to the target sample data and the prediction value of the remaining categories other than the category to which the target sample data belongs based on a negative margin.

7. The voiceprint recognition model training method of claim 1, wherein, The training of the first voiceprint recognition model based on the plurality of sample audio data and the adjusted prediction values to obtain a final voiceprint recognition model comprises: determining a target loss function corresponding to the target sample data based on the adjusted prediction value corresponding to the target sample data; determining other loss functions corresponding to other sample audio data other than the target sample audio data based on the prediction values of the other sample audio data; and training the first voiceprint recognition model based on the target loss function and the other loss functions to obtain the final voiceprint recognition model.

8. A voiceprint recognition model training system comprising: at least one storage medium storing at least one instruction set for training a voiceprint recognition model; and at least one processor in communication connection with the at least one storage medium, wherein when the voiceprint recognition model training system is running, the at least one processor reads the at least one instruction set and executes the voiceprint recognition model training method according to any one of claims 1-7 as instructed by the at least one instruction set.

9. A voiceprint recognition method comprising: obtaining target audio data to be recognized; inputting the target audio data into a voiceprint recognition model to obtain target voiceprint features corresponding to the target audio data, the voiceprint recognition model being trained based on the voiceprint recognition model training method according to any one of claims 1-7; and determining identity information of a target speaker in the target audio data based on the target voiceprint features.

10. A voiceprint recognition system comprising: at least one storage medium storing at least one instruction set for performing voiceprint recognition; and at least one processor in communication connection with the at least one storage medium, wherein when the voiceprint recognition system is running, the at least one processor reads the at least one instruction set and executes the voiceprint recognition method according to claim 9 as instructed by the at least one instruction set.

Images