End-to-end voiceprint recognition method and voiceprint recognition device

Abstract

Disclosed are an end-to-end voiceprint recognition method and a voiceprint recognition device. The voiceprint recognition method comprises: based on received input speech, using a speaker voice extraction module of an end-to-end deep learning network to perform a speaker voice extraction task to extract speech features of a target speaker; and based on the speech features of the target speaker, using a speaker recognition module of the end-to-end deep learning network to perform a speaker recognition task to recognize the target speaker in the received input speech.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech, and more particularly, to a voiceprint recognition method and a voiceprint recognition device. BACKGROUND

[0002] With the popularity of various electronic devices, the security protection of the electronic devices becomes particularly important. In particular, it is desirable that the electronic device is authenticated and / or unlocked only when a registered user uses the electronic device, thereby preventing the electronic device from being used by other unauthorized persons.

[0003] In order to achieve the security protection of the electronic devices, the voice can be used as unique information of the user. More specifically, a voiceprint identification (also known as speaker verification) technology can be used. Through the voiceprint identification technology, the voice features can be extracted from the voice of the speaker, and the identity of the voice features is verified. Generally, the voiceprint identification is divided into two processes of registration and verification. In the registration process, the user stores the voiceprint information of himself as a registered user; in the verification process, the voiceprint features extracted from the voice of the speaker are compared with the voiceprint features registered in advance, the similarity score of the features is obtained, and the score is compared with the threshold value. If the score is greater than the threshold value, it is judged that the verification and the registration are the same user, and if the score is less than or equal to the threshold value, it is judged that the verification and the registration are different users.

[0004] However, in the actual application scenario, the voice environment of the user is relatively complex. The existing voiceprint identification technology can often only cope with a relatively single application scenario, and cannot effectively meet the needs of the user. SUMMARY

[0005] The purpose of the present application is to provide an end-to-end voiceprint recognition method and a voiceprint recognition device.

[0006] In an aspect of the present disclosure, an end-to-end voiceprint recognition method is provided, which includes: based on the received input voice, using a speaker voice extraction module of an end-to-end deep learning network to perform a speaker voice extraction task to extract the voice features of a target speaker; based on the voice features of the target speaker, using a speaker recognition module of the end-to-end deep learning network to perform a speaker recognition task to identify the target speaker in the received input voice.

[0007] The voiceprint recognition method can further include: based on the received input voice, extracting the original voice features of the input voice; wherein the step of performing the speaker voice extraction task includes: based on the original voice features and the pre-saved intermediate embedding voice features of the registered person, performing the speaker voice extraction task.

[0008] The voiceprint recognition method can further include determining a speaking scenario of the input voice based on the voice feature of the target speaker and the original voice feature, the speaking scenario including a single-person scenario and a multi-person scenario, wherein the step of performing the speaker recognition task includes performing the speaker recognition task based on the speaking scenario of the input voice.

[0009] The step of determining whether the speaking scenario of the input voice is a single-person scenario or a multi-person scenario can include determining the speaking scenario of the input voice as a single-speaker scenario when a mean square error between the voice feature of the target speaker and the original voice feature is less than a threshold value, and determining the speaking scenario of the input voice as a multi-speaker scenario when the mean square error between the voice feature of the target speaker and the original voice feature is greater than or equal to the threshold value, wherein the step of performing the speaker recognition task includes inputting the original voice feature to the speaker recognition module to recognize the target speaker in the received input voice using the pre-stored final voice feature of the registered person when the speaking scenario of the input voice is the single-speaker scenario, and inputting the voice feature of the target speaker to the speaker recognition module to recognize the target speaker in the received input voice using the final voice feature of the registered person when the speaking scenario of the input voice is the multi-speaker scenario.

[0010] In a voiceprint registration phase of the registered person, the voiceprint recognition method can further include extracting an original voice feature of the registered person's voice based on the received registered person's voice, extracting an intermediate embedded voice feature of the registered person using the speaker recognition module based on the original voice feature of the registered person's voice, and saving the intermediate embedded voice feature of the registered person, and extracting a final voice feature of the registered person using the speaker recognition module based on the intermediate embedded voice feature of the registered person, and saving the final voice feature of the registered person.

[0011] The step of performing the speaker voice extraction task can include extracting an intermediate embedded voice feature of the target speaker from the original voice feature using a first convolutional layer, concatenating the intermediate embedded voice feature of the target speaker and the intermediate embedded voice feature of the registered person into a concatenated feature using a concatenation layer, generating a mask from the input concatenated feature using a full convolutional layer, and multiplying the mask with the intermediate embedded voice feature of the target speaker using a multiplier to obtain the voice feature of the target speaker.

[0012] In an end-to-end deep learning network training phase, the end-to-end deep learning network can further include a speaker consistency module configured to perform a voice conversion task on the intermediate embedded voice feature output based on the speaker recognition module.

[0013] The speaker voice extraction module and the speaker recognition module can be trained by supervised learning, and the speaker consistency module can be trained by self-supervised learning to obtain a first loss value, a second loss value and a third loss value that are weighted and minimized, wherein the first loss value is a loss value of the speaker voice extraction module, the second loss value is a loss value of the speaker recognition module, and the third loss value is a loss value of the speaker consistency module.

[0014] In an aspect of the present disclosure, an end-to-end voiceprint recognition apparatus is provided, which includes: a speaker voice extraction module of an end-to-end deep learning network, configured to perform a speaker voice extraction task based on received input speech to extract speech features of a target speaker; and a speaker recognition module of the end-to-end deep learning network, configured to perform a speaker recognition task based on the speech features of the target speaker to recognize the target speaker in the received input speech.

[0015] The voiceprint recognition apparatus can further include: a feature extraction module, configured to extract original speech features of the input speech based on the received input speech, wherein the speaker voice extraction module is configured to perform the speaker voice extraction task based on the original speech features and pre-stored intermediate embedding speech features of registered persons.

[0016] The voiceprint recognition apparatus can further include: a speaking scene determination module, configured to determine a speaking scene of the input speech based on the speech features of the target speaker and the original speech features, wherein the speaking scene includes a single-person scene and a multi-person scene, and the speaker recognition module is configured to perform the speaker recognition task based on the speaking scene of the input speech.

[0017] The speaking scene determination module can be configured to: when a mean square error between the speech features of the target speaker and the original speech features is less than a threshold value, determine the speaking scene of the input speech as a single-speaker scene; and when the mean square error between the speech features of the target speaker and the original speech features is greater than or equal to the threshold value, determine the speaking scene of the input speech as a multi-speaker scene, wherein the speaker recognition module is configured to: when the speaking scene of the input speech is the single-speaker scene, input the original speech features to the speaker recognition module to recognize the target speaker in the received input speech by using pre-stored final speech features of registered persons; and when the speaking scene of the input speech is the multi-speaker scene, input the speech features of the target speaker to the speaker recognition module to recognize the target speaker in the received input speech by using the final speech features of the registered persons.

[0018] In the voiceprint registration phase of the registrant, the voiceprint recognition device can be further configured to: based on the received registrant voice, extract original speech features of the registrant voice; based on the original speech features of the registrant voice, extract intermediate embedding speech features of the registrant using the speaker recognition module and save; based on the intermediate embedding speech features of the registrant, extract final speech features of the registrant using the speaker recognition module and save.

[0019] The speaker voice extraction module can be configured to: extract target speaker intermediate embedding speech features from the original speech features using a first convolutional layer; concatenate the target speaker intermediate embedding speech features and the registrant intermediate embedding speech features into concatenated features using a concatenation layer; generate a mask for the input concatenated features using a full convolutional layer; and multiply the mask with the target speaker intermediate embedding speech features to obtain the target speaker speech features using a multiplier.

[0020] In the end-to-end deep learning network training phase, the end-to-end deep learning network can further include a speaker consistency module configured to perform a voice conversion task on the intermediate embedding speech features output by the speaker recognition module.

[0021] The speaker voice extraction module and the speaker recognition module can be trained through supervised learning, and the speaker consistency module can be trained through self-supervised learning to obtain a weighted and minimized first loss value, a second loss value, and a third loss value, wherein the first loss value is the loss value of the speaker voice extraction module, the second loss value is the loss value of the speaker recognition module, and the third loss value is the loss value of the speaker consistency module.

[0022] According to an aspect of the present disclosure, a computer readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, any voiceprint recognition method is implemented.

[0023] According to the example embodiments of the present disclosure, the end-to-end voiceprint recognition can be implemented through the end-to-end deep learning network that links the speaker voice extraction module and the speaker recognition module using the speech features of the target speaker, so that even in the face of complex scenarios, the end-to-end voiceprint recognition method of the present disclosure can achieve high recognition performance and / or generalization performance.

[0024] According to the example embodiments of the present disclosure, the speaking scenario of the input voice can be determined according to the speech features of the target speaker based on the speaker voice extraction module, and the speaker recognition task can be performed using the speaker recognition module based on the speaking scenario of the input voice, so that by using the associated speaker voice extraction module and speaker recognition module, the recognition needs of single speaker scenarios can be met, and the needs of multiple speaker scenarios can also be met.

[0025] According to the example embodiments of the present application, since the speaking scene of the input speech can be determined based on the difference between the speech feature of the target speaker and the original speech feature, the speaking scene of the input speech can be easily determined.

[0026] According to the example embodiments of the present application, by determining whether the speaking scene of the input speech is a single speaker scene or a multi-speaker scene, the input corresponding to the speaker scene suitable for the speaker recognition module is selected, thus the performance of the voiceprint recognition in the multi-speaker scene and the performance of the voiceprint recognition in the single speaker scene can be guaranteed, and thus a higher performance voiceprint recognition service can be provided to the user.

[0027] According to the example embodiments of the present application, the end-to-end deep learning network can be trained using a combination of supervised learning and self-supervised learning, so that the end-to-end deep learning network learns the relationship and difference between the three tasks (i.e., the speaker speech extraction task, the speaker recognition task, and the speaker consistency task) by using the speaker intermediate embedding speech feature to link the speaker speech extraction module, the speaker recognition module, and the speaker consistency module to each other, and thus an end-to-end deep learning network with high performance (e.g., high recognition rate and high generalization ability) can be obtained.

[0028] According to the example embodiments of the present application, the speech data in the original data set can be fused to generate a self-supervised learning training sample set, thus the learning of the end-to-end deep learning network can be performed without increasing the amount of data and annotations.

[0029] According to the example embodiments of the present application, since the end-to-end deep learning network is updated by considering the first loss of the speaker speech extraction module, the second loss of the speaker recognition module, and the third loss of the speaker consistency module during training, the trained end-to-end deep learning network can have high recognition performance and generalization performance. BRIEF DESCRIPTION OF DRAWINGS

[0030] The above and other objects and features of the present application will become clearer from the following description of an example, given freely in conjunction with the accompanying drawings, wherein:

[0031] Figure 1 A flowchart showing an end-to-end voiceprint recognition method according to an example embodiment is shown.

[0032] Figure 2 A schematic diagram of a speaker speech extraction module according to an example embodiment is shown.

[0033] Figure 3 A schematic diagram of a speaker recognition module according to an example embodiment is shown.

[0034] Figure 4A schematic diagram showing training of a speaker voice extraction module and a speaker recognition module according to example embodiments.

[0035] Figure 5 A schematic diagram showing training of a speaker voice extraction module and a speaker recognition module according to example embodiments.

[0036] Figure 6 A block diagram showing an end-to-end voiceprint recognition apparatus according to example embodiments.

[0037] Figure 7 A block diagram showing a computing apparatus according to example embodiments.

[0038] Figure 8 A flowchart showing a method of performing a voiceprint recognition task according to example embodiments. DETAILED DESCRIPTION

[0039] The following detailed description is presented to aid the reader in gaining a thorough understanding of the methods, devices, and / or systems described herein. However, the following detailed description is not intended to limit the methods, devices, and / or systems described herein, as various changes, modifications, and equivalents thereof will be clear to those of ordinary skill in the art from the disclosure herein. For example, the order in which operations are described is not intended to be a limitation unless otherwise specified, and any number of additional operations can be performed and / or described operations can be eliminated or modified. Additionally, the description set forth herein describes particular features, and it is understood that the features can be combined in order to form various aspects of the methods, devices, and / or systems described herein.

[0040] The features described herein can be implemented in different forms and should not be construed as limited to the examples described herein. Rather, these examples are provided as illustrative of only a few of the many possible implementations of the methods, devices, and / or systems described herein.

[0041] As used herein, the term "and / or" includes any one of the associated listed items, as well as any combination of any two or more of the associated listed items.

[0042] Although terms such as "first", "second", and "third" can be used herein to describe various elements, components, regions, layers or sections, these elements, components, regions, layers or sections should not be limited by these terms. Rather, these terms are only used to distinguish one element, component, region, layer or section from another element, component, region, layer or section. Thus, the first element, the first component, the first region, the first layer or the first section referred to in the examples described herein can also be referred to as the second element, the second component, the second region, the second layer or the second section without departing from the teachings of the examples.

[0043] In the description, when an element (such as a layer, a region, or a substrate) is referred to as being "on" another element, "connected to" or "coupled to" another element, it can be directly on, directly connected to, or directly coupled to the other element, or one or more other elements can be interposed therebetween. In contrast, when an element is referred to as being "directly on," "directly connected to," or "directly coupled to" another element, there are no other elements interposed therebetween.

[0044] The terms used herein are merely used to describe various examples and are not intended to limit the disclosure. Singular forms are intended to include plural forms unless the context clearly indicates otherwise. The terms "comprise," "include," and "have" indicate the presence of the described features, numbers, operations, components, elements, and / or combinations thereof, but do not exclude the presence or addition of one or more other features, numbers, operations, components, elements, and / or combinations thereof.

[0045] Unless otherwise defined, all terms used herein (including technical terms and scientific terms) have the same meaning as those commonly understood by one of ordinary skill in the art to which this disclosure belongs when the disclosure is understood in light of the present disclosure. Unless explicitly defined otherwise herein, terms such as those defined in a generally used dictionary should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and the disclosure, and should not be interpreted ideally or overly formally.

[0046] In addition, in the description of the examples, when it is considered that a detailed description of a related structure or function that is well known will cause a vague interpretation of the disclosure, such a detailed description will be omitted.

[0047] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, the embodiments can be implemented in various forms, and are not limited to the embodiments described herein.

[0048] Figure 1 A flowchart of an end-to-end voiceprint recognition method according to an example embodiment is shown.

[0049] Referring to Figure 1 In operation S110, a speaker voice extraction task is performed using a speaker voice extraction module of an end-to-end deep learning network based on the received input speech to extract speech features of a target speaker.

[0050] Here, the input speech can be speech to be recognized that is input in order to use an electronic device. The input speech can be generated and input in different scenarios. In one example, the input speech can be speech input by a single person in a single-person scenario. In another example, the input speech can be speech input by multiple people (e.g., two or more people) in a multi-person scenario.

[0051] The target speaker can be a speaker to be recognized and / or verified in the input speech. For example, in a single-person scenario, the target speaker can be the single person. For another example, in a multi-person scenario, the target speaker can be any one of the multiple persons or a specified person.

[0052] In one embodiment, the speaker voice extraction module can utilize a first convolutional layer to extract the intermediate-embedded speech feature of the target speaker from the original speech feature. The speaker voice extraction module can utilize a concatenation layer to concatenate the intermediate-embedded speech feature of the target speaker and the intermediate-embedded speech feature of the registered person into a concatenated feature. The speaker voice extraction module can utilize a full convolutional layer to generate a mask from the input concatenated feature. The speaker voice extraction module can utilize a multiplier to multiply the mask with the intermediate-embedded speech feature of the target speaker to obtain the speech feature of the target speaker. Details will be described later in connection with Figure 2 The speaker voice extraction module is described in more detail.

[0053] In operation S120, the speaker recognition module using the end-to-end deep learning network performs a speaker recognition task based on the speech feature of the target speaker to recognize the target speaker in the received input speech.

[0054] That is, in the end-to-end voiceprint recognition method of the present disclosure, the end-to-end voiceprint recognition can be achieved by using the end-to-end deep learning network in which the speaker voice extraction module and the speaker recognition module are linked by using the speech feature of the target speaker. Therefore, even in the face of complex scenarios, the end-to-end voiceprint recognition method of the present disclosure can achieve high recognition performance and / or generalization performance.

[0055] In contrast, in the prior art, the voiceprint recognition is usually performed using a two-stage voiceprint recognition method in which two models are independent of each other. More specifically, the two-stage voiceprint recognition method first extracts the speech of the target speaker using one model, and then inputs the extracted speech into a voiceprint recognition model to obtain a speaker similarity score. Since the two-stage voiceprint recognition method uses two models that are independent of each other, the two models obtained fail to utilize the direct relationship between the two tasks performed thereby to be optimized, and the models obtained are only suboptimal, rather than optimal.

[0056] The speaker recognition module is described in more detail. Figure 3 The speaker recognition module is described in more detail.

[0057] In addition, optionally, raw speech features of the input speech are extracted based on the received input speech. The raw speech features can be features extracted using a feature extraction module based on the received input speech. In one example, the raw speech features can be Mel-scale Frequency Cepstral Coefficients (MFCC). In another example, the raw speech features can be Filter Bank features. However, the above examples are merely exemplary and the present disclosure is not limited thereto. The raw speech features can also be other commonly used speech features.

[0058] The raw speech features can be used for a speaker speech extraction task. More specifically, the speaker speech extraction module can perform the speaker speech extraction task based on the raw speech features and the pre-stored embedded speech features of the enrolled speakers.

[0059] In addition, in one embodiment, the speech features of the target speaker can also be used to determine the speaking scenario of the input speech. Thereafter, the speaker recognition task can be performed based on the speaking scenario of the input speech. More specifically, the speaking scenario of the input speech can be determined based on the speech features of the target speaker and the raw speech features. Here, the speaking scenario includes a single person scenario and a multiple person scenario.

[0060] In other words, in this embodiment, the speaking scenario of the input speech can be determined based on the speech features of the target speaker based on the speaker speech extraction module, and the speaker recognition task can be performed based on the speaking scenario of the input speech using the speaker recognition module, so that by using the associated speaker speech extraction module and speaker recognition module, the recognition needs of a single speaker scenario can be met, and the needs of a multiple speaker scenario can also be met.

[0061] According to one example embodiment of the present disclosure, the speaking scenario of the input speech (i.e., whether the speaking scenario of the input speech is a single person scenario or a multiple person scenario) can be determined based on a difference between the speech features of the target speaker and the raw speech features.

[0062] In addition, the speaker speech extraction module can perform the speaker speech extraction task based on the raw speech features to extract the speech features of the target speaker. In other words, the speaker speech extraction module can perform the speaker speech extraction task by receiving the raw speech features as input from the feature extraction module to extract the speech features of the target speaker. Here, the raw speech features can be the same type of features as the speech features of the target speaker.

[0063] Since the speaking scenario of the input speech can be determined based on a difference between the speech features of the target speaker and the raw speech features, the speaking scenario of the input speech can be easily determined.

[0064] For example, when the difference between the voice feature of the target speaker and the original voice feature is large, the speaking scenario of the input voice can be determined as a multi-speaker scenario. When the difference between the voice feature of the target speaker and the original voice feature is small, the speaking scenario of the input voice can be determined as a single-speaker scenario.

[0065] In an embodiment, when the mean square error between the voice feature of the target speaker and the original voice feature is less than a threshold, the speaking scenario of the input voice can be determined as a single-speaker scenario. In addition, when the mean square error between the voice feature of the target speaker and the original voice feature is greater than or equal to the threshold, the speaking scenario of the input voice can be determined as a multi-speaker scenario.

[0066] When the speaking scenario of the input voice is a single-speaker scenario, the original voice feature can be input to the speaker recognition module to identify the target speaker in the received input voice using the pre-stored final voice feature of the registered person. When the speaking scenario of the input voice is a multi-speaker scenario, the voice feature of the target speaker can be input to the speaker recognition module to identify the target speaker in the received input voice using the final voice feature of the registered person. The final voice feature of the registered person can be generated in a voiceprint registration phase of the registered person. More specifically, in the voiceprint registration phase of the registered person, the original voice feature of the registered person's voice can be extracted based on the received voice of the registered person. Then, the intermediate embedded voice feature of the registered person can be extracted using the speaker recognition module based on the original voice feature of the registered person's voice and stored. Next, the final voice feature of the registered person can be extracted using the speaker recognition module based on the intermediate embedded voice feature of the registered person and stored.

[0067] In other words, by determining whether the speaking scenario of the input voice is a single-speaker scenario or a multi-speaker scenario, the input suitable for the speaker recognition module corresponding to the speaking scenario can be selected. Therefore, the end-to-end voiceprint recognition method according to the present disclosure can guarantee the performance of voiceprint recognition in both a multi-speaker scenario and a single-speaker scenario, thereby providing a user with a higher performance voiceprint recognition service.

[0068] In contrast, in the prior art, a two-stage voiceprint recognition method using two models independently of each other is generally used to perform voiceprint recognition in a multi-speaker scenario. In this two-stage voiceprint recognition method, the default speaking scenario is a multi-speaker scenario. Therefore, when the two-stage voiceprint recognition method is applied to a single-speaker scenario, the performance of voiceprint recognition in the single-speaker scenario can be reduced.

[0069] Optionally, in the end-to-end deep learning network training phase, the end-to-end deep learning network can further include a speaker consistency module. The speaker consistency module can perform a voice conversion task based on the intermediate embedded voice feature output by the speaker recognition module.

[0070] In one embodiment, the speaker consistency module can include an Adaptive Instance Normalization (AdaIN) network for converting the voice timbre. The Adaptive Instance Normalization network can receive two inputs, one being the voice feature to be converted and the other being the target timbre feature. The formula of AdaIN is:

[0071]

[0072] wherein μ(x) and σ(x) represent the mean and standard deviation of the voice feature to be converted, respectively, and μ(y) and σ(y) represent the mean and standard deviation of the target timbre feature, respectively.

[0073] will be described later in connection with Figure 4 to Figure 5 Training of the speaker voice extraction module and the speaker recognition module will be described in more detail.

[0074] Figure 2 A schematic diagram of a speaker voice extraction module according to an example embodiment is shown.

[0075] Referring to Figure 3 The speaker voice extraction module can include a first convolutional layer 210, a concatenation layer 220, a fully convolutional layer 230, and a multiplier 240.

[0076] The first convolutional layer 210 can receive the original voice feature extracted using the feature extraction module and generate a speaker extraction encoder embedding, i.e., a speaker intermediate embedding voice feature, based on the original voice feature. In one example, the first convolutional layer 210 can be a single 1-dimensional (1D) convolution.

[0077] The concatenation layer 220 can receive the speaker extraction encoder embedding feature from the first convolutional layer 210 and receive a registered speaker intermediate embedding voice feature from a storage that stores information of registered users. The concatenation layer 220 can concatenate the speaker extraction encoder embedding feature and the registered speaker intermediate embedding voice feature into a concatenated feature.

[0078] The fully convolutional layer 230 can receive the concatenated feature from the concatenation layer 220 and generate a mask based on the concatenated feature. For example, the fully convolutional layer 230 can include a plurality of 1D dilated convolutional layers. Optionally, a normalization layer (LayerNorm) can be added between the concatenation layer 220 and the fully convolutional layer 230 to normalize the concatenated feature.

[0079] The multiplier 240 can receive the mask from the full convolutional layer 230, and receive the speaker extraction encoder embedding feature from the first convolutional layer 210. The multiplier 210 can multiply the mask with the speaker extraction encoder embedding feature to obtain the speech feature of the target speaker.

[0080] As described above with reference to Figure 1 The speech feature of the target speaker can be used to determine the speaking scenario of the input speech.

[0081] In one example, the enrollment person intermediate embedding speech feature and the enrollment person final speech feature are obtained in the enrollment person voiceprint registration phase. The enrollment person intermediate embedding speech feature is used to extract the speech feature of the target speaker, and the enrollment person final speech feature is used to identify whether the input speech includes the speech of the target speaker. The voiceprint registration information of the enrollment person can be obtained by using a voiceprint recognition module, and the obtaining process can include: based on the received speech of the enrollment person, extracting the original speech feature of the speech of the enrollment person; based on the original speech feature of the speech of the enrollment person, extracting the enrollment person intermediate embedding speech feature by using a speaker recognition module and saving; based on the enrollment person intermediate embedding speech feature, extracting the enrollment person final speech feature by using a speaker recognition module and saving.

[0082] Figure 3 A schematic diagram of a speaker recognition module according to an example embodiment is shown.

[0083] With reference to Figure 3 The speaker recognition module 300 according to the example embodiment can include a speaker embedder 310 and an attention statistic pooling layer 320.

[0084] In the registration phase, after the original speech feature of the speech of the enrollment person is extracted by the feature extraction module, the original speech feature is input to the speaker embedder 310 to generate the enrollment person intermediate embedding speech feature and save; then the enrollment person intermediate embedding speech feature is input to the attention statistic pooling layer 320 to generate the enrollment person final speech feature and save.

[0085] In the identification phase requiring to distinguish the speech scenario, when the speaking scenario of the input speech is a single speaker scenario, the speaker recognition module can receive the original speech feature, and generate the speaker intermediate embedding speech feature based on the received original speech feature, that is, the input of the speaker embedder 310 is the original speech feature of the received speech. When the speaking scenario of the input speech is a multi-speaker scenario, the speaker recognition module can receive the speech feature of the target speaker, and generate the intermediate embedding speech feature based on the received speech feature of the target speaker, that is, the input of the speaker embedder 310 is the speech feature of the target speaker.

[0086] In one example, the speaker embedder 310 can be configured with a SE block (Sequeze and Excitation Block). In another example, the speaker embedder 310 can be configured with a ResNet Block. In yet another example, the speaker embedder 310 can be configured with a TDNN. However, the above examples are merely exemplary, and the present disclosure is not limited thereto. The speaker embedder 310 can also be configured with other commonly used neural networks or combinations thereof.

[0087] The attention statistics pooling layer 320 can generate speaker final speech features based on the speaker intermediate speech features. The speaker final speech features can also be referred to as speaker embeddings. In one example, the speaker final speech features can be one-dimensional vectors. Merely by way of example, the speaker final speech features can be set to have a size of 1x128, 1x256, or 1x512.

[0088] Subsequently, the speaker recognition module 300 can identify a target speaker in the received input speech based on a similarity between the speaker final speech features and pre-stored registered speaker final speech features. When the similarity (e.g., cosine similarity) is greater than a threshold value, the target speaker can be identified as being consistent with a registered speaker. When the similarity is less than or equal to the threshold value, the target speaker can be identified as being inconsistent with the registered speaker.

[0089] Figure 4 A diagram illustrating training of a speaker speech extraction module and a speaker recognition module according to an example embodiment is shown.

[0090] Referring to Figure 4 The speaker speech extraction module and the speaker recognition module can be trained by supervised learning, and the speaker consistency module can be trained by self-supervised learning to obtain a first loss value, a second loss value, and a third loss value that are weighted and minimized. The first loss value is a loss value of the speaker speech extraction module, the second loss value is a loss value of the speaker recognition module, and the third loss value is a loss value of the speaker consistency module. Accordingly, according to the example embodiment of the present disclosure, the speaker speech extraction module, the speaker recognition module, and the speaker consistency module can be jointly trained using supervised learning and self-supervised learning, such that an end-to-end deep learning network learns a connection and a difference between three tasks (i.e., a speaker speech extraction task, a speaker recognition task, and a speaker consistency task) by linking the speaker speech extraction module, the speaker recognition module, and the speaker consistency module to each other using the speaker intermediate speech features, and thus a network having high performance (e.g., high recognition rate and high generalization ability) can be obtained.

[0091] In Figure 4In some embodiments, the original dataset includes a plurality of speeches and corresponding speaker labels. Each speech in the plurality of speeches can include speech of a single speaker. In other words, each speech in the original dataset can correspond to speech in a single speaker scenario.

[0092] The self-supervised learning training sample set module can generate the self-supervised learning training sample set by fusing speech data in the original dataset. More specifically, the self-supervised learning training sample set module can select speech of 2 or more different speakers to fuse. For example, taking the case of selecting a combination of two speakers, the self-supervised learning training sample set module can select one speech C of speaker 1 and one speech D of speaker 2 in the original dataset, and combine the two speeches to obtain speech M. For example, speech M can be a direct addition of speech C and speech D, and speech C can be used as the label of speech M. This process does not require manual labeling of each speech, and achieves automatic generation of samples in self-supervised learning. Therefore, according to embodiments, network learning can be performed without increasing the amount of data and labeling.

[0093] Optionally, the fused speech can be preprocessed by speech enhancement (e.g., random cropping, noise addition, reverberation addition, volume enhancement, etc.). For example, only as an example, the random cropping time can be set to 2 seconds. In the present application, the speaker speech extraction task can be understood as inputting speech M and outputting speech C, and the speaker consistency task can be understood as inputting speech C and outputting speech C.

[0094] Optionally, the speaker speech extraction module and the speaker recognition module can receive data from the feature extraction module. For example, the feature extraction module can extract original speech features of the original dataset based on the original dataset, and input the original speech features of the original dataset to the speaker recognition module. In addition, the feature extraction module can extract original speech features of the self-supervised learning training sample set based on the self-supervised learning training sample set, and input the original speech features of the self-supervised learning training sample set to the speaker speech extraction module. Here, the feature extraction module can be a pre-trained module.

[0095] In one embodiment, the speaker voice extraction module, the speaker recognition module, and the speaker consistency module can be trained such that a weighted sum of the first loss of the speaker voice extraction module, the second loss of the speaker recognition module, and the third loss of the speaker consistency module is minimized. Since the end-to-end deep learning network is updated during training taking into account the first loss of the speaker voice extraction module, the second loss of the speaker recognition module, and the third loss of the speaker consistency module, the trained speaker voice extraction module and the speaker recognition module can have high recognition performance and generalization performance. It should be understood that although the speaker consistency module is used during training to improve the performance of the speaker voice extraction module and the speaker recognition module, the speaker consistency module is not used during recognition.

[0096] The training of the speaker voice extraction module, the speaker recognition module, and the speaker consistency network will be described below in connection with an example, however the training of the end-to-end deep learning network is not limited to the example below.

[0097] In this example, for the speaker recognition task, the training data comes from the original dataset, each min-batch takes a predetermined number (e.g., 200 people) of speakers, each speaker selects multiple different voices (e.g., two different voices A and B), and randomly selects the voice of other speakers (e.g., one voice F), and the output is the embedding representation of the speaker voice. The speaker recognition loss function can use a metric learning loss to reduce the intra-class distance, increase the inter-class distance, and calculate the error (i.e., the second loss of the speaker recognition module) to update the speaker recognition module (SV). For the speaker voice extraction task, the voice of speaker 1 synthesized with the voice of other speakers can be input into the speaker voice extraction module (SE), at the same time, the speaker recognition module (SV) generates the speaker representation vector of speaker 1, the speaker representation vector of speaker 1 is input into the speaker voice extraction module, and the extracted predicted speaker 1 voice The speaker voice extraction loss function calculates the error (i.e., the first loss of the speaker voice extraction module) according to the label voice Y s and the predicted voice The error is calculated (i.e., the first loss of the speaker voice extraction module), which updates the parameters of the speaker recognition module (SV) and the speaker voice extraction module (SE) through backpropagation. For the speaker consistency module, the voice C of speaker 1 is input into the speaker consistency module (SC), at the same time, the speaker recognition module (SV) generates the speaker intermediate embedding voice feature of speaker 1, the speaker intermediate embedding voice feature of speaker 1 is input into the speaker consistency module (SC), and the synthesized speaker 1 voice The speaker consistency module loss function is calculated based on the labeled speech Y s and the predicted speech An error (i.e., a third loss of the speaker consistency module) is calculated, and the parameters of the speaker recognition module (SV) and the speaker consistency module (SC) are updated by backpropagation.

[0098] Figure 5 A schematic diagram of training the speaker voice extraction module and the speaker recognition module according to an example embodiment is shown.

[0099] Referring to Figure 4 Each network described can be substantially similar to each network of Figure 5 In other words, the description of each network of Figure 4 can also be applied to each network of Figure 5 Therefore, to avoid redundancy, the repeated description will be omitted.

[0100] In addition, Figure 5 The main difference between the embodiment of Figure 4 and the embodiment of is that one input of the speaker consistency module is not from the speaker extraction module, but from the feature extraction module.

[0101] Figure 5 Here, the network structure of Figure 4 can be referred to as a network structure in parallel form, and the network structure of can be referred to as a network structure in series form.

[0102] Figure 4 Although Figure 5 shows an example composition of the speaker recognition module (SV), the speaker voice extraction module (SE), and the speaker consistency module (SC) (i.e., series and parallel), the present application is not limited thereto, and the speaker recognition module (SV), the speaker voice extraction module (SE), and the speaker consistency module (SC) can also be connected and / or composed in other ways.

[0103] Figure 6 A block diagram of an end-to-end voiceprint recognition apparatus according to an example embodiment is shown.

[0104] Referring to Figure 6 , the voiceprint recognition apparatus 600 can include a speaker voice extraction module 610 of an end-to-end deep learning network and a speaker recognition module 620 of an end-to-end deep learning network.

[0105] In other words, the end-to-end deep learning network can include the speaker voice extraction module 610 and the speaker recognition module 620. The speaker voice extraction module 610 can be as described with reference to Figure 1 to Figure 5The speaker voice extraction module 610 can be implemented as any of the speaker voice extraction modules described in reference toIn other words, the speaker voice extraction module 610 can perform a speaker voice extraction task based on the received input speech to extract the speech feature of the target speaker.

[0106] The speaker recognition module 620 can also be implemented as any of the speaker recognition modules described in reference to Figure 1 to Figure 5 In other words, the speaker recognition module 620 can perform a speaker recognition task based on the speaking scenario of the input speech to recognize the target speaker in the received input speech.

[0107] Optionally, the voiceprint recognition apparatus 600 can further comprise a speaking scenario determination module (not shown). The speaking scenario determination module can determine the speaking scenario of the input speech based on the speech feature of the target speaker as described in reference to Figure 1

[0108] Optionally, the voiceprint recognition apparatus 600 can further comprise a feature extraction module (not shown). The feature extraction module can extract the original speech feature of the received speech based on the received speech as described in reference to Figure 1

[0109] The speaker voice extraction task performed by the speaker voice extraction network 610, the speaker recognition task performed by the speaker recognition module 620, the speaker consistency task performed by the speaker consistency module during the training, the method of extracting the original speech feature, and the method of determining the speaking scenario of the input speech have been described above in reference to at least one of Figure 1 to Figure 5 Therefore, for the sake of brevity and unnecessary redundancy description, the specific description of the methods performed by the speaker voice extraction module 610 and the speaker recognition module 620 will be omitted.

[0110] Figure 7 A block diagram of an electronic apparatus according to an example embodiment is shown.

[0111] Referring to Figure 7 , the electronic apparatus 700 according to an embodiment of the present application can comprise a processor 710 and a memory 720. Here, the memory 720 stores a computer program which, when executed by the processor 710, implements any of the voiceprint recognition methods described in reference to Figure 1 to Figure 5 For the sake of brevity, the description of any of the voiceprint recognition methods described in reference to Figure 1 to Figure 5 by the processor 710 will not be repeated here.

[0112] ​Further, a method according to an exemplary embodiment of the present application can be implemented as a computer program in a computer readable recording medium. A person skilled in the art can implement the computer program according to the description of the above method. When the computer program is executed in a computer, the end-to-end voiceprint recognition method of the present application is implemented.

[0113] According to an exemplary embodiment of the present application, a computer readable storage medium having stored thereon a computer program which, when executed by a processor, causes the processor to implement any of the methods disclosed in the present application. For example, the computer program, when executed by the processor, causes the processor to perform the following steps: performing, based on the received input speech, a speaker voice extraction task using a speaker voice extraction module of an end-to-end deep learning network to extract speech features of a target speaker; and performing, based on the speech features of the target speaker, a speaker recognition task using a speaker recognition module of the end-to-end deep learning network to recognize the target speaker in the received input speech.

[0114] Figure 8 A flowchart illustrating a method of performing a voiceprint recognition task according to an exemplary embodiment is shown.

[0115] Referring to Figure 8 In step ①, the input speech can be received by the feature extraction module. In step ②, the original speech features of the input speech can be extracted by the feature extraction module based on the input speech.

[0116] In step ③, the original speech features of the input speech can be received by the speaker voice extraction module. More specifically, the original speech features of the input speech can be received by the first convolutional layer of the speaker voice extraction module. The first convolutional layer can generate speaker extraction encoder embedding features (i.e., speaker intermediate embedding speech features) based on the original speech features and input the speaker intermediate embedding speech features to a concatenation layer. In addition, the first convolutional layer can also input the speaker intermediate embedding speech features to a multiplier

[0117] In step ④, the registered speaker intermediate embedding speech features of the registered speaker can be obtained from a storage device, and in step ⑤, the registered speaker intermediate embedding speech features can be input to the concatenation layer of the speaker voice extraction module. The concatenation layer can concatenate the speaker extraction encoder embedding features and the registered speaker intermediate embedding speech features into concatenated features. Thereafter, the concatenated features can be input to a full convolutional layer which can include a plurality of 1D dilated convolutional layers (1-D Conv). Optionally, a normalization layer (LayerNorm) can be added between the concatenation layer and the full convolutional layer to normalize the concatenated features. The full convolutional layer can generate a mask based on the concatenated features and output the mask to a multiplier The multiplier The mask can be multiplied with the speaker intermediate-embedded speech feature to obtain a speech feature of the target speaker. In step ⑥, a multiplier The speech feature of the target speaker can be outputted.

[0118] In step ⑦, the original speech feature and the speech feature of the target speaker can be compared by a speaking scene determination module. For example, the speaking scene determination module can compare the mean square error (MSE) between the original speech feature and the speech feature of the target speaker.

[0119] In step ⑧, when the mean square error between the original speech feature and the target speaker is less than a threshold, the original speech feature can be inputted to the speaker recognition module; when the mean square error between the original speech feature and the target speaker is greater than or equal to the threshold, the speech feature of the target speaker can be inputted to the speaker recognition module.

[0120] The speaker recognition module can perform a speaker recognition task according to the received input. More specifically, a speaker embedder of the speaker recognition module can receive the corresponding input and generate a speaker intermediate-embedded speech feature based on the input. The speaker embedder can input the speaker intermediate-embedded speech feature to an attention statistic pooling layer of the speaker recognition module. In step ⑨, the attention statistic pooling layer can generate a speaker final speech feature based on the speaker intermediate-embedded speech feature.

[0121] In step ⑩, the registered speaker final speech feature of the registered speaker can be obtained from a storage device. In step A similarity score can be determined based on the similarity between the speaker final speech feature and the registered speaker final speech feature to identify the target speaker in the received input speech. When the similarity (e.g., cosine similarity) is greater than a threshold, the target speaker can be identified as consistent with the registered speaker. When the similarity is less than or equal to the threshold, the target speaker can be identified as inconsistent with the registered speaker.

[0122] Note that at least one step and / or module described in Figure 8 may correspond to at least one step and / or module described with reference to Figure 1 to Figure 7 . Therefore, the description of at least one step and / or module in Figure 1 to Figure 7 may also apply to at least one step and / or module in Figure 8 .

[0123] In addition, it should be understood that each unit in the device according to the exemplary embodiments of the present application can be implemented by hardware components and / or software components. Those skilled in the art can implement each unit using, for example, a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC) according to the processing performed by the defined each unit.

[0124] According to the example embodiments of the present application, the end-to-end deep learning network that links the speaker voice extraction module and the speaker recognition module by using the voice features of the target speaker can be used to implement the end-to-end voiceprint recognition, so that the end-to-end voiceprint recognition method of the present application can achieve high recognition performance and / or generalization performance even in the face of complex scenarios.

[0125] According to the example embodiments of the present application, the speaking scenario of the input voice can be determined according to the voice features of the target speaker based on the speaker voice extraction module, and the speaker recognition task can be performed based on the speaking scenario of the input voice using the speaker recognition module, so that by using the associated speaker voice extraction module and speaker recognition module, the recognition needs of the single speaker scenario can be met, and the needs of the multi-speaker scenario can also be met.

[0126] According to the example embodiments of the present application, since the speaking scenario of the input voice can be determined based on the difference between the voice features of the target speaker and the original voice features, the speaking scenario of the input voice can be easily determined.

[0127] According to the example embodiments of the present application, the input suitable for the speaker recognition module corresponding to the speaking scenario can be selected by determining whether the speaking scenario of the input voice is a single speaker scenario or a multi-speaker scenario, so that the performance of voiceprint recognition in the multi-speaker scenario can be guaranteed, and the performance of voiceprint recognition in the single speaker scenario can also be guaranteed, thereby providing the user with a voiceprint recognition service with higher performance.

[0128] According to the example embodiments of the present application, the end-to-end deep learning network can be trained using a combination of supervised learning and self-supervised learning, so that the end-to-end deep learning network learns the relationship and difference between the three tasks (i.e., the speaker voice extraction task, the speaker recognition task, and the speaker consistency task) by linking the speaker voice extraction module, the speaker recognition module, and the speaker consistency module to each other using the speaker intermediate embedding voice features, thereby obtaining an end-to-end deep learning network with high performance (e.g., high recognition rate and high generalization ability).

[0129] According to the example embodiments of the present application, the self-supervised learning training sample set can be generated by fusing the voice data in the original data set, so that the learning of the end-to-end deep learning network can be performed without increasing the amount of data and annotations.

[0130] According to the example embodiments of the present application, the end-to-end deep learning network is updated by considering the first loss of the speaker voice extraction module, the second loss of the speaker recognition module, and the third loss of the speaker consistency module during training, so that the trained end-to-end deep learning network can have high recognition performance and generalization performance.

[0131] While the disclosure includes certain examples, it will be clear to one of ordinary skill in the art that various changes in form and details can be made without departing from the spirit and scope of the claims and their equivalents. The examples described herein should be considered in a descriptive sense only and not for purposes of limitation. Descriptions of features or aspects in each example should be considered as being applicable to similar features or aspects in other examples. Suitable results can be achieved if the described techniques are performed in a different order, and / or if components in the described systems, architectures, devices, or circuits are combined in a different manner, and / or replaced or supplemented by other components or their equivalents. Therefore, the scope of the disclosure is defined not by the specific embodiments described herein, but by the appended claims and their equivalents, and all variations within the scope of the claims and their equivalents are to be construed as being included in the disclosure.

Claims

1. An end-to-end voiceprint recognition method, the voiceprint recognition method comprising: Based on the received input speech, the speaker speech extraction module of the end-to-end deep learning network is used to perform the speaker speech extraction task to extract the speech features of the target speaker. Based on the speech features of the target speaker, a speaker recognition module using an end-to-end deep learning network is used to perform a speaker recognition task in order to identify the target speaker in the received input speech. In this system, the end-to-end deep learning network uses the speech features of the target speaker to link the speaker speech extraction module and the speaker recognition module. During the training phase of the end-to-end deep learning network, the network also includes a speaker consistency module, configured to perform a speech conversion task based on the intermediate embedded speech features output by the speaker recognition module. Specifically, the speaker speech extraction module and speaker recognition module are trained through supervised learning, while the speaker consistency module is trained through self-supervised learning, in order to obtain weighted and minimized first, second, and third loss values. The first loss value is the loss value of the speaker speech extraction module, the second loss value is the loss value of the speaker recognition module, and the third loss value is the loss value of the speaker consistency module.

2. The voiceprint recognition method as described in claim 1, wherein, The voiceprint recognition method further includes: Based on the received input speech, extract the original speech features of the input speech; The steps involved in performing the speaker speech extraction task include: Based on the original speech features and the pre-saved registrant embedded speech features, the speaker speech extraction task is performed.

3. The voiceprint recognition method as described in claim 2, wherein, The voiceprint recognition method further includes: Based on the speech features of the target speaker and the original speech features, the speaking scenario of the input speech is determined. The speaking scenario includes single-person scenarios and multi-person scenarios. The steps for performing the speaker recognition task include: performing the speaker recognition task based on the speaking scenario of the input speech.

4. The voiceprint recognition method as described in claim 3, wherein, The steps to determine whether the input speech is in a one-person or multi-person scenario include: When the mean square error between the target speaker's speech features and the original speech features is less than the threshold, the speaking scenario of the input speech is determined as a single speaker scenario. When the mean square error between the target speaker's speech features and the original speech features is greater than or equal to a threshold, the speaking scenario of the input speech is determined to be a multi-speaker scenario. The steps involved in performing the speaker recognition task include: When the input speech is a single-speaker scenario, the original speech features are input into the speaker recognition module. The target speaker is identified in the received input speech using the pre-saved final speech features of the registrant. The final speech features of the registrant are speech features extracted by the speaker recognition module based on the intermediate embedded speech features of the registrant. When the input speech is in a multi-speaker scenario, the speech features of the target speaker are input into the speaker recognition module, and the target speaker is identified in the received input speech using the registrant's final speech features.

5. The voiceprint recognition method as described in claim 1, wherein, During the voiceprint registration phase for the registrant, the voiceprint recognition method further includes: Based on the received registrant's voice, extract the original speech features of the registrant's voice; Based on the original speech features of the registrant, the speaker recognition module is used to extract and save the embedded speech features of the registrant. Based on the registrant's embedded speech features, the final speech features of the registrant are extracted and saved using the speaker recognition module.

6. The voiceprint recognition method as described in claim 5, wherein, The steps involved in performing speaker speech extraction include: The first convolutional layer is used to extract embedded speech features of the target speaker from the original speech features; Using a splicing layer, the embedded speech features of the target speaker and the embedded speech features of the registrant are spliced ​​together to form a spliced ​​feature; A mask is generated from the concatenated input features using fully convolutional layers; The target speaker's speech features are obtained by multiplying the mask with the embedded speech features of the target speaker using a multiplier.

7. An end-to-end voiceprint recognition device, the voiceprint recognition device comprising: The speaker speech extraction module of the end-to-end deep learning network performs a speaker speech extraction task based on the received input speech to extract the speech features of the target speaker. The speaker recognition module of the end-to-end deep learning network performs a speaker recognition task based on the speech features of the target speaker, in order to identify the target speaker in the received input speech. In this system, the end-to-end deep learning network uses the speech features of the target speaker to link the speaker speech extraction module and the speaker recognition module. During the training phase of the end-to-end deep learning network, the network also includes a speaker consistency module, configured to perform a speech conversion task based on the intermediate embedded speech features output by the speaker recognition module. Specifically, the speaker speech extraction module and speaker recognition module are trained through supervised learning, while the speaker consistency module is trained through self-supervised learning, in order to obtain weighted and minimized first, second, and third loss values. The first loss value is the loss value of the speaker speech extraction module, the second loss value is the loss value of the speaker recognition module, and the third loss value is the loss value of the speaker consistency module.

8. The voiceprint recognition device as described in claim 7, wherein, The voiceprint recognition device further includes a feature extraction module, which extracts the original voice features of the received input voice based on the received input voice. The speaker speech extraction module is configured to perform speaker speech extraction tasks based on the original speech features and the pre-saved registrant embedded speech features.

9. The voiceprint recognition device as described in claim 8, wherein, The voiceprint recognition device further includes a speaking scene determination module, which determines the speaking scene of the input voice based on the voice features of the target speaker and the original voice features. The speaking scene includes single-person scenes and multi-person scenes. The speaker recognition module is configured to perform speaker recognition tasks based on the speaking scenario of the input speech.

10. The voiceprint recognition device as claimed in claim 9, wherein, The speaking scenario determination module is configured as follows: When the mean square error between the target speaker's speech features and the original speech features is less than the threshold, the speaking scenario of the input speech is determined as a single speaker scenario. When the mean square error between the target speaker's speech features and the original speech features is greater than or equal to a threshold, the speaking scenario of the input speech is determined to be a multi-speaker scenario. The speaker recognition module is configured as follows: When the input speech is a single-speaker scenario, the original speech features are input into the speaker recognition module. The target speaker is identified in the received input speech using the pre-saved final speech features of the registrant. The final speech features of the registrant are speech features extracted by the speaker recognition module based on the intermediate embedded speech features of the registrant. When the input speech is in a multi-speaker scenario, the speech features of the target speaker are input into the speaker recognition module, and the target speaker is identified in the received input speech using the registrant's final speech features.

11. The voiceprint recognition device as claimed in claim 8, wherein, During the voiceprint registration phase, the voiceprint recognition device is further configured to: Based on the received registrant's voice, extract the original speech features of the registrant's voice; Based on the original speech features of the registrant, the speaker recognition module is used to extract and save the embedded speech features of the registrant. Based on the registrant's embedded speech features, the final speech features of the registrant are extracted and saved using the speaker recognition module.

12. The voiceprint recognition device as claimed in claim 11, wherein, The speaker speech extraction module is configured as follows: The first convolutional layer is used to extract embedded speech features of the target speaker from the original speech features; Using a splicing layer, the embedded speech features of the target speaker and the embedded speech features of the registrant are spliced ​​together to form a spliced ​​feature. The embedded speech features of the registrant are speech features extracted by the speaker recognition module based on the original speech features of the registrant. A mask is generated from the concatenated input features using fully convolutional layers; The target speaker's speech features are obtained by multiplying the mask with the embedded speech features of the target speaker using a multiplier.

13. A computer-readable storage medium storing a computer program, wherein, When the computer program is executed by a processor, it implements the voiceprint recognition method according to any one of claims 1 to 6.

Images