Training method of deep learning model, voiceprint recognition method, device and equipment

Abstract

The present disclosure provides a method for training a deep learning model, relates to the field of artificial intelligence, and particularly relates to the field of voiceprint recognition. The specific implementation scheme is: according to the characteristics of M speech samples, the M speech samples are classified into N sample groups respectively corresponding to N objects, wherein each sample group includes at least one speech sample, the object corresponding to each sample group is used as the pseudo label of the speech sample in the sample group, M and N are integers greater than 1, and M is greater than or equal to N; and the deep learning model is trained using the M speech samples and the pseudo label of each speech sample. The present disclosure also provides a voiceprint recognition method and device, an electronic device and a storage medium.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and in particular to a voiceprint recognition technology. More specifically, the present disclosure provides a deep learning model training method, a voiceprint recognition method, an apparatus, an electronic device and a storage medium. BACKGROUND

[0002] In the field of speech processing, the task of identifying a speaker is also referred to as a voiceprint recognition task. In the voiceprint recognition task, the training of a related model depends on a large amount of voiceprint data with speaker labels. However, due to the privacy involved in voiceprint data, it is difficult to collect a large amount of labeled data in actual application scenarios. SUMMARY

[0003] The present disclosure provides a deep learning model training method, a voiceprint recognition method, an apparatus, an electronic device and a storage medium.

[0004] According to a first aspect, a deep learning model training method is provided. The method comprises: classifying M speech samples into N sample groups each corresponding to an object according to features of the M speech samples, wherein each sample group includes at least one speech sample, the object corresponding to each sample group is a pseudo label of the speech sample in the sample group, M and N are integers greater than 1, and M is greater than or equal to N; and training a deep learning model using the M speech samples and the pseudo label of each speech sample.

[0005] According to a second aspect, a voiceprint recognition method is provided. The method comprises: obtaining a to-be-recognized speech; and inputting the to-be-recognized speech into a deep learning model to obtain a target object corresponding to the to-be-recognized speech sample; wherein the deep learning model is trained according to the deep learning model training method described above.

[0006] According to a third aspect, a deep learning model training apparatus is provided. The apparatus comprises: a classification module configured to classify M speech samples into N sample groups each corresponding to an object according to features of the M speech samples, wherein each sample group includes at least one speech sample, the object corresponding to each sample group is a pseudo label of the speech sample in the sample group, M and N are integers greater than 1, and M is greater than or equal to N; and a first training module configured to train a deep learning model using the M speech samples and the pseudo label of each speech sample.

[0007] According to a fourth aspect, a voiceprint recognition apparatus is provided. The apparatus comprises: an obtaining module configured to obtain a to-be-recognized speech; and a recognition module configured to input the to-be-recognized speech into a deep learning model to obtain a target object corresponding to the to-be-recognized speech sample; wherein the deep learning model is trained according to the deep learning model training apparatus described above.

[0008] According to a fifth aspect, an electronic device is provided, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method provided by the present disclosure.

[0009] According to a sixth aspect, a non-transitory computer-readable storage medium storing computer instructions is provided, the computer instructions being used to cause a computer to perform the method provided by the present disclosure.

[0010] According to a seventh aspect, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method provided by the present disclosure.

[0011] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0012] The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. Among them:

[0013] Figure 1 is an exemplary system architecture schematic diagram of a training method of a deep learning model and a voiceprint recognition method according to an embodiment of the present disclosure;

[0014] Figure 2 is a flowchart of a training method of a deep learning model according to an embodiment of the present disclosure;

[0015] Figure 3 is a schematic diagram of a method of classifying M speech samples into N sample groups respectively corresponding to N objects according to an embodiment of the present disclosure;

[0016] Figure 4 is a schematic diagram of a training method of a deep learning model according to an embodiment of the present disclosure;

[0017] Figure 5 is a flowchart of a voiceprint recognition method according to an embodiment of the present disclosure;

[0018] Figure 6 is a block diagram of a training device of a deep learning model according to an embodiment of the present disclosure;

[0019] Figure 7 is a block diagram of a voiceprint recognition device according to an embodiment of the present disclosure;

[0020] Figure 8 is a block diagram of an electronic device of a training method of a deep learning model and / or a voiceprint recognition method according to one embodiment of the disclosure. DETAILED DESCRIPTION

[0021] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, which should be considered in their context only. Thus, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of known functions and constructions are omitted in the following description for clarity and conciseness.

[0022] Audio data (voiceprint data) with speaker labels is relatively difficult to collect and has a high labeling cost, and in actual application scenarios, only limited speech samples (small samples) can often be collected. Under the condition of small samples, the speaker recognition algorithm based on the deep learning neural network model will have an overfitting phenomenon, lacking the ability to learn limited samples.

[0023] Under the condition of small samples, the data quantity can be increased by data enhancement on the speech samples. For example, the speech samples can be subjected to scrambling, denoising or changing the speech speed in the time domain, etc. preprocessing operations, to obtain speech samples with different noise levels or different speech speeds. The speech samples can also be preprocessed in the frequency domain, such as masking part of the area of the frequency spectrum of the speech samples to obtain the preprocessed frequency spectrum. Although the time domain and frequency domain data enhancement of the speech samples can increase the sample quantity, the problem of insufficient and too few class samples will still arise with the first appearance of a new class speaker.

[0024] A general voice pre-training model HuBERT (Hidden-Unit BERT, BERT-based self-supervised speech representation learning) is a neural network model trained for speech processing tasks, which can be used for speech recognition tasks, speech generation tasks and voiceprint recognition tasks. However, the training of voiceprint recognition tasks based on HuBERT still relies on a large number of samples, and there is still a risk of overfitting under the condition of small sample resources.

[0025] In the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the technical solutions comply with the relevant legal regulations and do not violate public order and good customs.

[0026] In the technical solutions of the present disclosure, the authorization or consent of the user is obtained before the user's personal information is acquired or collected.

[0027] Figure 1is an exemplary system architecture schematic diagram of a training method of a deep learning model and a voiceprint recognition method according to an embodiment of the present disclosure. It should be noted that Figure 1 The system architecture shown is only an example of a system architecture to which embodiments of the present disclosure can be applied, to help those skilled in the art understand the technical content of the present disclosure, but does not mean that embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios.

[0028] As Figure 1 shown, the system architecture 100 according to the embodiment can include a plurality of terminal devices 101, a network 102 and a server 103. The network 102 is a medium for providing a communication link between the terminal device 101 and the server 103. The network 102 can include various connection types, such as wired and / or wireless communication links, etc.

[0029] A user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. The terminal device 101 can be various electronic devices, including but not limited to smartphones, tablet computers, laptop computers, etc.

[0030] At least one of the training method of a deep learning model and the voiceprint recognition method provided by the embodiments of the present disclosure can generally be executed by the server 103. Accordingly, the training device of a deep learning model and the voiceprint recognition device provided by the embodiments of the present disclosure can generally be provided in the server 103. The training method of a deep learning model and the voiceprint recognition method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 103 and capable of communicating with the terminal device 101 and / or the server 103. Accordingly, the training device of a deep learning model and the voiceprint recognition device provided by the embodiments of the present disclosure can also be provided in a server or a server cluster different from the server 103 and capable of communicating with the terminal device 101 and / or the server 103.

[0031] Figure 2 is a flowchart of a training method of a deep learning model according to an embodiment of the present disclosure.

[0032] As Figure 2 shown, the training method 200 of a deep learning model can include operation S210 to operation S220.

[0033] At operation S210, M speech samples are classified into N sample groups each corresponding to N objects according to features of the M speech samples.

[0034] For example, the M speech samples can be voiceprint data from N objects (speakers), and the voiceprint data, like fingerprints, can uniquely identify a person, so the speaker identity can be identified using the voiceprint data.

[0035] The M speech samples can be a limited number of samples collected, for example, the M speech samples can be 2000 (M = 2000) speech samples, which can be from 50 (N = 50) objects. Before obtaining the 2000 speech samples of the 50 objects, the authorization or consent of the 50 objects is obtained.

[0036] The M speech samples are input into the deep learning model to obtain the features of the M speech samples. For example, for each speech sample, a speech spectrogram is generated, and MFCC (Mel-Frequency Cepstral Coefficients) feature extraction or FBank (for example, Log-mel Filter Bank) feature extraction is performed on the speech spectrogram to obtain the features of the speech sample.

[0037] The 2000 speech samples can obtain 2000 speech sample features, and according to the 2000 speech sample features, the 2000 speech samples can be classified into sample groups corresponding to the 50 objects respectively.

[0038] For example, a clustering algorithm or a classification model can be used to classify the features of the 2000 speech samples into 50 feature groups corresponding to the 50 objects respectively, and then divide the 2000 speech samples into 50 sample groups according to the 50 feature groups, and the object corresponding to each feature group is taken as the object label of the sample group corresponding to the feature group. After each sample group obtains the object label, the object label of the sample group can be taken as the pseudo-label of the speech sample in the sample group.

[0039] In operation S220, the deep learning model is trained using the M speech samples and the pseudo-labels of each speech sample.

[0040] For example, the M speech samples are input into the deep learning model, and the deep learning model predicts the object corresponding to each speech sample. For each speech sample, the mean square error or cross-entropy error between the predicted object and the object indicated by the pseudo-label of the speech sample can be calculated as the loss of the deep learning model, and the parameters of the deep learning model can be adjusted according to the loss of the deep learning model.

[0041] Embodiments of the present disclosure can improve the performance of model training under small sample resource conditions by classifying M speech samples into N sample groups corresponding to N objects respectively, taking the object corresponding to each sample group as the pseudo-label of the speech sample in the sample group, and training the deep learning model using the speech samples and their pseudo-labels.

[0042] It can be understood that the embodiments of the present disclosure can train the deep learning model in the case of small sample resources, and can solve the problems of difficult sample collection and high labeling cost.

[0043] Figure 3 is a schematic diagram of a method of classifying M speech samples into N sample groups respectively corresponding to N objects according to one embodiment of the present disclosure.

[0044] As shown in Figure 3 M speech samples 310 include sample 1, sample 2,..., sample 10,..., sample M. M speech samples 310 are input into a deep learning model to obtain M speech sample features 320, which include feature 1, feature 2, feature 10,..., feature M, where M is greater than 10, for example, M = 2000.

[0045] The M speech sample features 320 can be clustered by a clustering algorithm to obtain N clusters, each cluster containing at least one feature, which can be used as a feature group, to obtain N feature groups 330, which include feature group 1, feature group 2,..., feature group N, where N is an integer greater than 1, for example, N = 50. According to the relationship between the center feature of each cluster and the feature (such as age, gender, etc.) of each object in the N objects, the corresponding relationship between the N feature groups 330 and the N objects can be determined. For example, feature group 1 corresponds to object 1, feature group 2 corresponds to object 2,..., and feature group N corresponds to object N. The object corresponding to each feature group can be used as the object label of the feature group, for example, object 1 is used as the object label of feature group 1, object 2 is used as the object label of feature group 2,..., and object N is used as the object label of feature group N.

[0046] According to the N feature groups 330, the M speech samples 310 can be divided into N sample groups 340. For example, the N sample groups 340 obtained by division include sample group 1, sample group 2,..., and sample group N. The object label of each feature group can be passed to the sample group corresponding to the feature group, for example, feature group 1 passes object 1 to sample group 1, feature group 2 passes object 2 to sample group 2,..., and feature group N passes object N to sample group N. The object label obtained by each sample group can be used as the pseudo label of each speech sample in the sample group, for example, object 1 is used as the pseudo label of the speech sample in sample group 1, object 2 is used as the pseudo label of the speech sample in sample group 2,..., and object N is used as the pseudo label of the speech sample in sample group N.

[0047] Embodiments of the present disclosure can determine pseudo labels of each speech sample in a sample group by classifying M samples into sample groups corresponding to N objects respectively.

[0048] A general speech pre-training model HuBERT can be used for training of a voiceprint recognition task. For example, in a first stage (pre-training stage), the HuBERT is trained using a large number of speech samples with speaker labels to obtain a deep learning model that can be used for speaker identity recognition, wherein the large number of speech samples used can be obtained by data augmentation on small samples. In a second stage (fine-tuning stage), the HuBERT is fine-tuned using a small number of speech samples with speaker labels, or transfer learning is performed to achieve the generalization ability to small sample data. However, in actual scenarios, the voiceprint recognition training task usually starts from a cold start, and there is little labeled data in actual scenarios. In this case, the common practice of fine-tuning the pre-training model for the voiceprint recognition task is prone to poor performance.

[0049] Therefore, embodiments of the present disclosure provide a training method of a deep learning model, which adds an intermediate training stage between the pre-training stage and the fine-tuning stage, generates pseudo labels of small sample data in the intermediate training stage, and uses the small sample data and the pseudo labels to train the deep learning model, so as to improve the performance of the deep learning model.

[0050] Figure 4 is a schematic diagram of a training method of a deep learning model according to an embodiment of the present disclosure.

[0051] As shown in Figure 4 The training method of the deep learning model includes three stages, namely a pre-training stage 410, an intermediate training stage 420, and a fine-tuning stage 430.

[0052] In the pre-training stage 410, a speech pre-training model (for example, HuBERT) can be trained using a large number of speech samples to obtain a deep learning model 411 in a general field, which can be used for voiceprint recognition tasks in the general field. For example, the large number of speech samples used in this stage can come from voiceprint data in various fields (for example, web conferencing, voiceprint registration platform, etc.), and are provided with speaker labels. Each speech sample in the large number of speech samples can be input to the speech pre-training model to obtain a predicted object corresponding to each speech sample, and the parameters of the speech pre-training model are adjusted according to the mean square error, cross entropy or other error between the predicted object and the object indicated in the label, and when a certain condition is reached (for example, the speech pre-training model converges), the training is ended, and the deep learning model 411 is obtained.

[0053] In the intermediate training stage 420, speech samples in the actual application field can be obtained. For example, in the voiceprint registration business field, speech samples in the speaker recognition business under the new registration category can be obtained. The amount of samples is small, and the effect is not ideal if directly used for fine-tuning. Therefore, an unsupervised classification task, such as a clustering task, can be introduced in this stage. The speech samples are classified into sample groups corresponding to the respective objects in the preset multiple objects through the unsupervised classification task, so as to obtain the object labels of the sample groups as the pseudo labels of the speech samples in the sample groups. The speech samples and the pseudo labels are used to train the deep learning model 411, and the deep learning model 421 is obtained.

[0054] For example, in the intermediate training stage 420, 2000 speech samples can be obtained, which can come from 50 objects. The 2000 speech samples are input into the deep learning model 411 to obtain 2000 speech sample features. The 2000 speech sample features are clustered to obtain feature groups corresponding to the respective objects of the 50 objects. The object corresponding to each feature group is taken as the object label of the feature group. According to the 50 feature groups, the 2000 speech samples are divided into 50 sample groups. The object label corresponding to each feature group is transmitted to the sample group corresponding to the feature group. After obtaining the object label in each sample group, the object label is taken as the pseudo label of the speech sample in the sample group. The speech samples and the pseudo labels are used to train the deep learning model 411, and the deep learning model 421 is obtained.

[0055] For example, each speech sample in the small sample can be input into the deep learning model 411 to obtain the predicted object corresponding to each speech sample. The parameters of the deep learning model 411 are adjusted according to the mean square error, cross entropy or other errors between the predicted object and the object indicated in the pseudo label. When a certain condition is reached (for example, the deep learning model 411 converges), the training is completed, and the deep learning model 421 is obtained.

[0056] In the fine-tuning stage 430, the deep learning model 421 can be fine-tuned using the small sample and the actual label of the speech sample in the small sample to obtain the deep learning model 431. For example, each speech sample in the small sample is input into the deep learning model 421 to obtain the predicted object corresponding to each speech sample. The parameters of the deep learning model 421 are adjusted according to the mean square error, cross entropy or other errors between the predicted object and the object indicated in the actual label. When a certain condition is reached (for example, the deep learning model 421 converges), the training is completed, and the deep learning model 431 is obtained.

[0057] Compared with the training of only two stages, the three stages of the pre-training stage 410, the intermediate training stage 420 and the fine-tuning stage 430 can improve the performance of the deep learning model, and further improve the accuracy of the deep learning model in identifying the speaker.

[0058] Figure 5 is a flowchart of a voiceprint recognition method according to an embodiment of the present disclosure.

[0059] As shown in Figure 5 , the voiceprint recognition method 500 can include operation S510 to operation S520.

[0060] In operation S510, the voice to be identified is obtained.

[0061] In operation S520, the voice to be identified is input into the deep learning model to obtain a target object corresponding to the voice sample to be identified.

[0062] The deep learning model is obtained by training according to the training method of the deep learning model.

[0063] For example, the voice to be identified can include the voiceprint features of the speaker, and inputting the voice to be identified into the deep learning model can obtain a target object corresponding to the voice to be identified, which can be the speaker who produces the voice to be identified.

[0064] The embodiment of the present disclosure can identify the speaker identity information who produces the voice to be identified by inputting the voice to be identified into the deep learning model.

[0065] Figure 6 is a block diagram of a training device of a deep learning model according to an embodiment of the present disclosure.

[0066] As shown in Figure 6 , the training device 600 of the deep learning model includes a classification module 601 and a first training module 602.

[0067] The classification module 601 is configured to classify M voice samples into N sample groups each corresponding to an object according to the features of the M voice samples, wherein each sample group includes at least one voice sample, the object corresponding to each sample group is a pseudo label of the voice sample in the sample group, M and N are integers greater than 1, and M is greater than or equal to N.

[0068] The first training module 602 is configured to train the deep learning model using the M voice samples and the pseudo label of each voice sample.

[0069] According to an embodiment of the present disclosure, each of the M speech samples is provided with an actual label, the actual label of each speech sample representing an object that generates the speech sample; the training apparatus 600 of the deep learning model further comprises a second training module.

[0070] The second training module is configured to train the deep learning model using the M speech samples and the actual label of each speech sample.

[0071] The second training module comprises a first prediction unit and a first adjustment unit.

[0072] The first prediction unit is configured to input the M speech samples into the deep learning model to obtain a first prediction result corresponding to each speech sample.

[0073] The first adjustment unit is configured to adjust the parameters of the deep learning model according to the difference between the actual label and the first prediction result.

[0074] The first training module 602 comprises a second prediction unit and a second adjustment unit.

[0075] The second prediction unit is configured to input the M speech samples into the deep learning model to obtain a second prediction result corresponding to each speech sample.

[0076] The second adjustment unit is configured to adjust the parameters of the deep learning model according to the difference between the pseudo label and the second prediction result.

[0077] The classification module 601 comprises a clustering unit and a division unit.

[0078] The clustering unit is configured to cluster the features of the M speech samples to obtain N feature groups respectively corresponding to the N objects, wherein each feature group comprises the features of at least one speech sample; and

[0079] The division unit is configured to divide the M speech samples into N sample groups according to the N feature groups.

[0080] The training apparatus 600 of the deep learning model further comprises a feature extraction module.

[0081] The feature extraction module is configured to input the M speech samples into the deep learning model to obtain the features of the M speech samples.

[0082] Figure 7 is a block diagram of a voiceprint recognition apparatus according to an embodiment of the present disclosure.

[0083] As shown in Figure 7 , the voiceprint recognition apparatus 700 comprises an acquisition module 701 and a recognition module 702.

[0084] The acquisition module 701 is configured to acquire the to-be-recognized speech.

[0085] The recognition module 702 is configured to input the to-be-recognized speech into the deep learning model to obtain a target object corresponding to the to-be-recognized speech sample.

[0086] The deep learning model is obtained according to the training method of the deep learning model.

[0087] According to an embodiment of the present disclosure, the present disclosure further provides an electronic device, a readable storage medium, and a computer program product.

[0088] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present disclosure described and / or claimed in this document.

[0089] As shown in Figure 8 The device 800 includes a computing unit 801 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. Various programs and data required for the operation of the device 800 can also be stored in the RAM 803. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0090] Various components in the device 800 are connected to the I / O interface 805, including an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, a speaker, etc.; the storage unit 808, such as a magnetic disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0091] The computing unit 801 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs various methods and processes described above, such as the training method of a deep learning model and / or the voiceprint recognition method. For example, in some embodiments, the training method of a deep learning model and / or the voiceprint recognition method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded onto the RAM 803 and executed by the computing unit 801, one or more steps of the training method of a deep learning model and / or the voiceprint recognition method described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform the training method of a deep learning model and / or the voiceprint recognition method by any other suitable means, such as by means of firmware.

[0092] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0093] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces a means for implementing the functions / acts specified in the flowcharts and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0094] In the context of the present disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0095] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0096] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0097] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0098] It should be understood that the various forms of flow shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, in series, or in a different order, as long as the desired results of the technology disclosed in the present disclosure can be achieved, which is not limited herein.

[0099] The specific implementation described above does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a deep learning model, comprising: The speech pre-trained model is trained using speech samples from multiple domains to obtain a deep learning model; Based on the features of M speech samples from a specific domain, the M speech samples are classified into N sample groups corresponding to N objects. Each sample group includes at least one speech sample. The object corresponding to each sample group is used as a pseudo-label for the speech samples in that sample group. Each speech sample is also assigned an actual label, which represents the object that generated the speech sample. M and N are integers greater than 1, and M is greater than or equal to N. The deep learning model is trained using the M speech samples from a specific domain and the pseudo-labels for each speech sample to obtain a deep learning model trained with pseudo-labels. Using the M speech samples from a specific domain and the actual labels of each speech sample, the deep learning model trained with pseudo-labels is trained to obtain a deep learning model trained with actual labels. The step of classifying the M speech samples into N sample groups corresponding to N objects based on the features of the M speech samples from a specific domain includes: Cluster the features of the M speech samples to obtain N feature groups, wherein each feature group includes features of at least one speech sample; Based on the N feature groups, the M speech samples are divided into N sample groups; Based on the relationship between the central feature of each feature group and the features of each of the N objects, determine the correspondence between the N feature groups and the N objects; Based on the correspondence between the N feature groups and the N objects, the N objects are assigned as pseudo-labels to the N sample groups respectively.

2. The method according to claim 1, wherein, The step of training the deep learning model using the M speech samples and the actual labels of each speech sample includes: The M speech samples are input into the deep learning model to obtain a first prediction result corresponding to each speech sample; and The parameters of the deep learning model are adjusted based on the difference between the actual label and the first prediction result.

3. The method according to claim 1, wherein, The step of training the deep learning model using the M speech samples and the pseudo-labels for each speech sample includes: The M speech samples are input into the deep learning model to obtain a second prediction result corresponding to each speech sample; and The parameters of the deep learning model are adjusted based on the difference between the pseudo-label and the second prediction result.

4. The method according to claim 1, further comprising: The M speech samples are input into the deep learning model to obtain the features of the M speech samples.

5. A voiceprint recognition method, comprising: Acquire the speech to be recognized; as well as The speech to be recognized is input into a deep learning model to obtain the target object corresponding to the speech sample to be recognized; The deep learning model is trained using the method described in any one of claims 1 to 4.

6. A training device for a deep learning model, comprising: The first training module is used to train the speech pre-trained model using speech samples from multiple domains to obtain a deep learning model. The classification module is used to classify the M speech samples from a specific domain into N sample groups corresponding to N objects, based on the features of the M speech samples. Each sample group includes at least one speech sample, and the object corresponding to each sample group is used as a pseudo label for the speech sample in that sample group. Each speech sample is also assigned an actual label, which represents the object that generated the speech sample. M and N are integers greater than 1, and M is greater than or equal to N. The second training module is used to train the deep learning model using the M speech samples from a specific domain and the pseudo-labels of each speech sample, so as to obtain a deep learning model trained with pseudo-labels. The third training module is used to train the deep learning model trained with pseudo-labels using the M speech samples from a specific domain and the actual labels of each speech sample, so as to obtain a deep learning model trained with actual labels. The classification module includes: A clustering unit is used to cluster the features of the M speech samples to obtain N feature groups, wherein each feature group includes features of at least one speech sample. A partitioning unit is used to partition the M speech samples into N sample groups based on the N feature groups; The correspondence determination unit is used to determine the correspondence between the N feature groups and the N objects based on the relationship between the central feature of each feature group and the features of each of the N objects; The pseudo-label determination unit is used to assign the N objects as pseudo-labels to the N sample groups according to the correspondence between the N feature groups and the N objects.

7. The apparatus according to claim 6, wherein, The third training module includes: The first prediction unit is used to input the M speech samples into the deep learning model to obtain a first prediction result corresponding to each speech sample; and The first adjustment unit is used to adjust the parameters of the deep learning model based on the difference between the actual label and the first prediction result.

8. The apparatus according to claim 6, wherein, The second training module includes: The second prediction unit is used to input the M speech samples into the deep learning model to obtain a second prediction result corresponding to each speech sample; and The second adjustment unit is used to adjust the parameters of the deep learning model based on the difference between the pseudo-label and the second prediction result.

9. The apparatus according to claim 6, further comprising: The feature extraction module is used to input the M speech samples into the deep learning model to obtain the features of the M speech samples.

10. A voiceprint recognition device, comprising: The acquisition module is used to acquire the speech to be recognized; as well as The recognition module is used to input the speech to be recognized into a deep learning model to obtain the target object corresponding to the speech sample to be recognized; The deep learning model is trained using the apparatus according to any one of claims 6 to 9.

11. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 5.

12. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 5.

13. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images