Voice data annotation method and device, and storage medium
By using multiple pre-trained voiceprint models to perform voiceprint recognition and clustering on vehicle business voice data, the problem of voice annotation data acquisition efficiency and accuracy in vehicle business scenarios is solved, and efficient and low-cost voice annotation data acquisition is achieved. .
Patent Information
- Application Number
- CN202411441724.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-01-21
AI Technical Summary
It is difficult for the prior art to obtain highly accurate voice label data efficiently and at low cost, especially in vehicle business scenarios.
By obtaining vehicle business voice data, construct training data, and use multiple pre-trained voiceprint models to perform voiceprint recognition. After obtaining the voiceprint matching score, clustering is performed to determine the voice data of the same speaker and perform speaker annotation.
It achieves efficient and low-cost acquisition of high-accuracy speech annotation data, reduces the time cost and economic cost of manual annotation, and improves the training effect of the speech recognition model.
Smart Images
Figure CN119339727A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of speech processing technology, and in particular to a speech data annotation method, device and storage medium. Background Art
[0002] At present, voiceprint recognition based on deep learning has been rapidly developed and widely used. In voiceprint recognition based on deep learning, in order to identify voiceprint data through a neural network model, it is first necessary to improve the accuracy of the voiceprint recognition model. For the training of the voiceprint recognition model, massive data and correct labels are particularly important. For example, in voiceprint recognition in vehicle-mounted business scenarios, a large amount of business-related voice annotation data is required to train the voiceprint recognition model. There are usually the following methods for obtaining voice annotation data: 1. Use open source data. This method has low data acquisition costs, but open source data and actual business scenarios do not fully match, and the voiceprint recognition effect is limited; 2. Collection or recording. This method requires a large number of people to record voice. The data obtained by this method has low accuracy, and the voice channel matches the scene and the actual business, but it requires a lot of manpower and economic costs. In addition, the recorded voice data is not natural enough.
[0003] Therefore, this proposal addresses the problem of how to obtain high-accuracy speech annotation data efficiently and at low cost. Summary of the invention
[0004] In order to overcome the above-mentioned defects, a speech data annotation method, device and storage medium proposed in the present application can solve the technical problem of obtaining high-accuracy speech annotation data efficiently and at low cost.
[0005] In a first aspect, the present application provides a voice data labeling method, the method comprising: obtaining in-vehicle business voice data, and constructing training data based on the in-vehicle business voice data; performing voiceprint recognition on the training data based on multiple pre-trained voiceprint models, and obtaining the voiceprint matching scores of the training data corresponding to each pre-trained voiceprint model; clustering the voiceprint matching scores of the training data to obtain the voiceprint clustering results corresponding to each pre-trained voiceprint model; determining the voice data corresponding to the same speaker based on the voiceprint clustering results corresponding to each pre-trained voiceprint model, and performing speaker labeling.
[0006] In some implementation schemes, the in-vehicle service voice data is specifically a plurality of voice data classified according to vehicle information; and constructing training data based on the in-vehicle service voice data specifically includes:
[0007] For each voice data corresponding to each vehicle information, construct a positive sample based on the voice data corresponding to the same vehicle information, and construct a negative sample based on the voice data corresponding to the vehicle information different from the positive sample;
[0008] The training data is obtained based on the constructed positive samples and negative samples, and the training data includes at least one positive sample and at least one negative sample.
[0009] Furthermore, the voiceprint matching score includes a score of a positive sample and a score of a negative sample; the voiceprint matching scores of the training data are clustered to obtain the voiceprint clustering result corresponding to each pre-trained voiceprint model:
[0010] According to the same speaker as the clustering label, unsupervised clustering is performed on the scores of all positive samples and all negative samples, and the speech data corresponding to the training data is divided into at least one cluster, each cluster corresponding to a voiceprint clustering result; based on the at least one cluster, the voiceprint clustering result corresponding to each pre-trained voiceprint model is determined.
[0011] Furthermore, the unsupervised clustering of the scores of all positive samples and the scores of all negative samples includes:
[0012] The highest score among all negative samples is determined as the threshold value;
[0013] Determine, according to the threshold value, positive samples having scores higher than the threshold value;
[0014] Recognize the speech data corresponding to the determined positive samples as speech data of the same speaker;
[0015] The speech data recognized as the same speaker are divided into the same cluster.
[0016] In some implementation schemes, the method of determining the voice data corresponding to the same speaker based on the voiceprint clustering results corresponding to each pre-trained voiceprint model and performing speaker labeling is specifically as follows: voting on the voiceprint clustering results corresponding to each pre-trained voiceprint model, selecting the voice data identified as the same speaker that reaches a preset number of votes, and labeling the selected voice data with the same speaker identifier.
[0017] In some implementation schemes, the above-mentioned voice data labeling method further includes: based on the vehicle information, merging the labeled voice data; the labeled voice data refers to the voice data that has been labeled with the speaker.
[0018] Furthermore, the merging of the labeled voice data based on the vehicle information specifically includes:
[0019] Take the feature average value of the labeled voice data corresponding to each vehicle information;
[0020] Determine the similarity between any two users corresponding to the vehicle information based on the feature average value;
[0021] The annotated voice data corresponding to different vehicle information are merged based on the similarity.
[0022] In a second aspect, the present application provides a voiceprint recognition method, comprising: training a pre-trained voiceprint model to obtain a trained voiceprint model based on the speaker-annotated voice data obtained by the above-mentioned voice data annotation method; and performing voiceprint recognition based on the trained voiceprint model.
[0023] In a third aspect, the present application provides an intelligent device comprising a processor and a memory, wherein the memory is suitable for storing multiple program codes, and the program codes are suitable for being loaded and run by the processor to execute the speech data labeling method described in any one of the technical solutions of the above-mentioned speech data labeling method.
[0024] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a plurality of program codes, wherein the program codes are suitable for being loaded and run by a processor to execute the speech data labeling method described in any one of the technical solutions of the above-mentioned speech data labeling method.
[0025] One or more of the above technical solutions of this application have the following beneficial effects:
[0026] In the technical solution of the present application, the in-vehicle service voice data is obtained, and training data is constructed based on the in-vehicle service voice data; voiceprint recognition is performed on the training data based on multiple pre-trained voiceprint models to obtain the voiceprint matching score of the training data corresponding to each pre-trained voiceprint model; the voiceprint matching scores of the training data are clustered to obtain the voiceprint clustering results corresponding to each pre-trained voiceprint model; based on the voiceprint clustering results corresponding to each pre-trained voiceprint model, the voice data corresponding to the same speaker is determined, and the speaker is labeled. The method of the present application obtains voice annotation data based on a machine learning model, and has a high accuracy rate of annotation, which reduces the error probability of manual annotation, reduces the time cost and economic cost of manual annotation, and realizes efficient and low-cost acquisition of high-accuracy voice annotation data, which can be widely used in the training of voiceprint recognition models to improve voiceprint recognition effects. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] The disclosure of the present application will become more easily understood with reference to the accompanying drawings. It is easy for those skilled in the art to understand that these drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present application. In addition, similar numbers in the drawings are used to represent similar components, among which:
[0028] Figure 1 This is a flow chart of the main steps of a method for annotating speech data according to an embodiment of the present application;
[0029] Figure 2 is based on Figure 1 A schematic diagram of the steps of unsupervised clustering implemented by the method shown;
[0030] Figure 3 is based on Figure 1 A schematic flow chart of the steps of voice data merging processing implemented by the method shown;
[0031] Figure 4 It is a schematic diagram of an implementation method of a voice data annotation method in a specific application scenario of the present application;
[0032] Figure 5 It is a schematic block diagram of a voice data annotation device according to an embodiment of the present application. DETAILED DESCRIPTION
[0033] Some embodiments of the present application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application and are not intended to limit the protection scope of the present application.
[0034] In the description of the present application, "module" and "processor" may include hardware, software or a combination of the two. A module may include hardware circuits, various suitable sensors, communication ports, memory, and may also include software parts, such as program code, or a combination of software and hardware. The processor may be a central processing unit, a microprocessor, an image processor, a digital signal processor or any other suitable processor. The processor has data and / or signal processing functions. The processor may be implemented in software, hardware or a combination of the two. Non-temporary computer-readable storage media include any suitable medium that can store program code, such as a disk, a hard disk, an optical disk, a flash memory, a read-only memory, a random access memory, etc. The term "A and / or B" means all possible combinations of A and B, such as only A, only B or A and B. The term "at least one A or B" or "at least one of A and B" has a similar meaning to "A and / or B" and may include only A, only B or A and B. The singular terms "one" and "the" may also include plural forms.
[0035] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application may be combined with each other if there is no conflict.
[0036] Combine the following Figure 1-Figure 4The voice data annotation method of the present application is described. The voice data annotation method can be applied to intelligent devices such as servers, mobile phones, computers, and vehicles, and can also be applied to voice data annotation devices set in electronic devices such as servers, mobile phones, computers, and vehicles. The voice data annotation device can be implemented by software, hardware, or a combination of both. The voice data in the embodiment of the present application can be voice data collected in real time by a voice acquisition device (such as a microphone), or it can be recorded offline voice data. The intelligent device can obtain real-time voice data from the voice acquisition device, or read recorded offline voice data from the memory.
[0037] like Figure 1 FIG. 1 is a flow chart of the main steps of a method for annotating speech data according to an embodiment of the present application, and as shown in the figure, it mainly includes the following steps:
[0038] Step S11: acquiring in-vehicle service voice data, and constructing training data based on the in-vehicle service voice data;
[0039] Specifically in this embodiment, the in-vehicle service voice data is a plurality of voice data classified according to vehicle information. In a specific application scenario, different vehicle information corresponds to different vehicles. Acquiring the in-vehicle service voice data can be acquiring user voice data collected in real time by a vehicle microphone or acquiring pre-recorded voice data corresponding to each vehicle. Exemplarily, the vehicle information can be a vehicle ID, i.e., a unique vehicle identification code. It should be understood that the vehicle information can also be other information used to identify the vehicle in addition to the vehicle ID.
[0040] Specifically in this embodiment, constructing training data based on the in-vehicle service voice data specifically includes:
[0041] For each voice data corresponding to each vehicle information, construct a positive sample based on the voice data corresponding to the same vehicle information, and construct a negative sample based on the voice data corresponding to the vehicle information different from the positive sample;
[0042] The training data is obtained based on the constructed positive samples and negative samples, and the training data includes at least one positive sample and at least one negative sample.
[0043] Exemplarily, the voice data corresponding to vehicle a are a1, a2, a3, ..., the voice data corresponding to vehicle b are b1, b2, b3, ..., and the voice data corresponding to vehicle c are c1, c2, c3, ...; the positive samples constructed for voice data a1 can be a1 a2, a1a3, and so on, and the negative samples can be a1 b1, a1 b2, a1 c2, a1 c3, and so on.
[0044] Step S12: performing voiceprint recognition on the training data based on multiple pre-trained voiceprint models to obtain a voiceprint matching score of the training data corresponding to each pre-trained voiceprint model; clustering the voiceprint matching scores of the training data to obtain a voiceprint clustering result corresponding to each pre-trained voiceprint model;
[0045] Specifically, in this embodiment, the number of the multiple pre-trained voiceprint models is M, and the multiple pre-trained voiceprint models can be multiple different voiceprint models trained using open source data. The pre-trained voiceprint model is used to perform voiceprint recognition on the training data, i.e., all positive samples and negative samples, to obtain a voiceprint matching score, which includes the score of the positive sample and the score of the negative sample; wherein the score of the positive sample or the score of the negative sample can be understood as referring to the similarity of the voiceprint features between the two voice data used to construct the positive sample or the negative sample. For example, the higher the similarity of the voiceprint features of the two voice data, the greater the possibility that the two voice data are from the same speaker.
[0046] Exemplarily, the value of M is 4. In this embodiment, the number of pre-trained voiceprint models selected is 4, and the 4 pre-trained voiceprint models are Chinese ECAPA-TDNN model, Chinese CAM++ model, English ECAPA-TDNN model and English CAM++ model.
[0047] Based on the clustering of the voiceprint matching scores of the training data in step S12, the voiceprint clustering result corresponding to each pre-trained voiceprint model can be specifically obtained as follows:
[0048] According to the same speaker as the clustering label, unsupervised clustering is performed on the scores of all positive samples and all negative samples, and the speech data corresponding to the training data is divided into at least one cluster, each cluster corresponding to a voiceprint clustering result; based on the at least one cluster, the voiceprint clustering result corresponding to each pre-trained voiceprint model is determined.
[0049] Furthermore, the specific implementation method of performing unsupervised clustering on the scores of all positive samples and all negative samples is as follows: Figure 2 As shown, the following steps 121 to 124 may be included:
[0050] Step 121: Determine the highest score among the scores of all negative samples as the threshold value;
[0051] Step 122: determining, according to the threshold value, positive samples having scores higher than the threshold value;
[0052] Step 123: recognizing the speech data corresponding to the determined positive samples as speech data of the same speaker;
[0053] Step 124: Divide the speech data recognized as the same speaker into the same cluster.
[0054] Step S13: Determine the speech data corresponding to the same speaker based on the voiceprint clustering results corresponding to each pre-trained voiceprint model, and perform speaker labeling.
[0055] Specifically, in this embodiment, the voiceprint clustering results corresponding to each pre-trained voiceprint model are voted, and the voice data identified as the same speaker that reaches a preset number of votes are selected, and the selected voice data are labeled with the same speaker identifier. For example, the number of pre-trained voiceprint models is M, the preset number of votes is k, and k<=M.
[0056] For example, if M=4, k=3, and the four pre-trained voiceprint models are model one, model two, model three and model four, if the clustering result of model one determines that voices a, b, and c are from the same person, the clustering result of model two determines that voices a, b, and c are from the same person, the clustering result of model three determines that voices a, b, and d are from the same person, and the clustering result of model four determines that voices b and d are from the same person, because k models (model one, two, and three) all think that a and b are the same person, by voting on the voiceprint clustering results of the four models, a and b are identified as voice data corresponding to the same speaker.
[0057] Furthermore, step S14 may be included after the above step S13.
[0058] Step S14: Based on the vehicle information, the annotated voice data is merged.
[0059] Specifically, for the annotated voice data corresponding to different vehicle IDs, the voice data corresponding to different vehicle IDs suspected to be of the same owner are merged by calculating feature similarity. In this embodiment, the annotated voice data refers to the voice data that has been annotated by the speaker.
[0060] It should be understood that step S14 is an optional implementation step. After step S14 is processed, the labeled speech data obtained based on clustering and voting can be further combined with vehicle information to perform speaker merging.
[0061] In some embodiments, the specific implementation of step S14 is as follows: Figure 3 As shown, the following steps 141 to 143 may be included.
[0062] Step 141: Take the feature average value of the labeled voice data corresponding to each piece of vehicle information.
[0063] It should be understood that based on the existing technology, voiceprint features can be extracted from speech data. The processing of speech data in this embodiment can be understood as the processing of voiceprint features extracted from speech data. The feature average here should be understood as the average of the voiceprint features extracted from each speech data.
[0064] Step 142: determining the similarity between any two users corresponding to the vehicle information based on the feature average value;
[0065] Step 143: Based on the similarity, the annotated voice data corresponding to different vehicle information are merged into voice data corresponding to the same speaker.
[0066] Among them, vehicles with high similarity are suspected to have the same owner. When merging voice data, the labeled voice data of different vehicles can be merged into the same speaker or the labeled voice data of one of the vehicles can be deleted.
[0067] Exemplarily, for the voice data of the same car owner, the annotated voice data obtained based on multiple vehicles of the car owner may be retained at the same time, or only the annotated voice data obtained based on any one of the vehicles of the car owner may be retained.
[0068] Further, such as Figure 4 FIG. 1 is a schematic diagram of the main implementation steps of the voice data annotation method in a specific application scenario of the present application. As shown in the figure, in this application scenario, the voice data processed by the voice data annotation method provided in the embodiment of the present application is specifically vehicle-borne service voice data, and the specific process of annotating the vehicle-borne service voice data is as follows:
[0069] 1) Classification by vehicle ID;
[0070] Specifically, the in-vehicle service voice data is first classified according to the vehicle ID; the voice data corresponding to the same vehicle ID should be understood as the in-vehicle service voice data obtained based on the same vehicle.
[0071] 2) Construct a voiceprint trial pair;
[0072] Specifically, target t ial pairs are constructed based on the voice data of the same vehicle, and at least one other vehicle data is randomly selected from each voice data corresponding to each vehicle to construct a non-target t ial pair. In this embodiment, the voiceprint t ial pair can be understood as a voiceprint test pair or a voiceprint training pair, the target t ial pair can be understood as a positive sample, the non-target t ial pair can be understood as a negative sample, and the other vehicle data refers to the voice data of the vehicle different from the target t ial pair.
[0073] 3) Scoring of pre-trained voiceprint models;
[0074] Specifically, M pre-trained voiceprint models are selected, and the models can be trained using open source data, and the pre-trained voiceprint models are used to score the voiceprint trial pairs in step 2). In this embodiment, M is a positive integer, and the selected pre-trained voiceprint models can be Chinese ECAPA-TDNN model, Chinese CAM++ model, English ECAPA-TDNN model, and English CAM++ model.
[0075] 4) Score clustering;
[0076] Specifically, unsupervised speaker clustering is performed according to the score of each pre-trained voiceprint model. Any clustering method can be used. One clustering method is: the trail score of each voice data is sorted from high to low, the maximum non-target score is used as the threshold, and the audio corresponding to the target score with a score higher than the threshold is considered to be the same speaker, and the voiceprint clustering result of each pre-trained voiceprint model is obtained.
[0077] 5) Results voting;
[0078] Specifically, the voiceprint clustering results obtained by the M pre-trained voiceprint models are voted, and at least k (k<=M) voice data with consistent clustering results are selected and labeled as the same speaker to obtain labeled voice data.
[0079] 6) Speaker merging;
[0080] Specifically, the feature average value is taken for the labeled voice data obtained by clustering and voting for each vehicle, and the features of each vehicle are scored pairwise based on the feature average value. Vehicles with high similarity are suspected to have the same owner. The labeled voice data corresponding to the two vehicles are merged as the same speaker or the labeled voice data of one of the vehicles is discarded. That is, the speaker-labeled voice data obtained after the speaker merging process is the final voice labeling data.
[0081] In a specific application scenario, a large amount of voice data with speaker annotation can be obtained after batches of vehicle-mounted business voice data are annotated according to the above process. A voice data annotation method proposed in the embodiment of the present application combines vehicle information and performs unsupervised clustering and voting scoring on batches of data through multiple groups of pre-trained voiceprint recognition models, thereby obtaining labeled voice annotation data. The method obtains voice annotation data through a machine learning method, and the annotation accuracy is greater than 98%, which effectively improves the accuracy of speaker annotation. The voice annotation data is obtained based on the method of the embodiment of the present application, which reduces the error probability of manual annotation, and also reduces the time cost and economic cost of manual annotation, and realizes efficient and low-cost acquisition of a large amount of business-related voice annotation data for voiceprint recognition model training.
[0082] Furthermore, the present application also provides a voiceprint recognition method, the method comprising: obtaining speaker-annotated voice data based on the voice data annotation method provided in the above method embodiment, training a pre-trained voiceprint model to obtain a trained voiceprint model, and performing voiceprint recognition based on the trained voiceprint model. It should be understood that the pre-trained voiceprint model can be any model used for voiceprint recognition in the prior art, and is not limited to any one of the M pre-trained voiceprint models mentioned in the method embodiment of the present application.
[0083] Furthermore, the present application also provides a speech data annotation device.
[0084] See attached Figure 5 , Figure 5 : is a main structural block diagram of a speech data annotation device according to an embodiment of the present application, and the device mainly includes:
[0085] Acquisition module 201: used to acquire in-vehicle service voice data, and construct training data based on the in-vehicle service voice data;
[0086] Scoring module 202: used to perform voiceprint recognition on the training data constructed by the acquisition module 201 based on multiple pre-trained voiceprint models, and obtain the voiceprint matching score of the training data corresponding to each pre-trained voiceprint model;
[0087] Clustering module 203: used to cluster the voiceprint matching scores of the training data obtained by the scoring module 202 to obtain the voiceprint clustering results corresponding to each pre-trained voiceprint model;
[0088] The labeling module 204 is used to determine the speech data corresponding to the same speaker based on the voiceprint clustering results corresponding to each pre-trained voiceprint model obtained by the clustering module 203, and label the speaker. Specifically, the voiceprint clustering results corresponding to each pre-trained voiceprint model are voted, and the speech data identified as the same speaker that reaches a preset number of votes are selected, and the same speaker identification is labeled for the selected speech data.
[0089] Furthermore, the labeling module 204 can also be used to merge the labeled voice data based on the vehicle information.
[0090] For ease of explanation, the introduction to the above-mentioned speech data annotation device only shows the part related to the embodiment of the present application. For specific technical details not disclosed, please refer to the method part of the embodiment of the present application.
[0091] It should be understood that since the setting of each module is only for illustrating the functional units of the present application, the physical devices corresponding to these modules may be the processor itself, or a part of the software in the processor, a part of the hardware, or a part of the combination of software and hardware. Therefore, the number of each module in the figure is only schematic.
[0092] It is understood by those skilled in the art that each module in the device can be adaptively split or merged. Such splitting or merging of specific modules will not cause the technical solution to deviate from the principle of the present application, and therefore, the technical solutions after splitting or merging will fall within the protection scope of the present application.
[0093] It is understood by those skilled in the art that all or part of the processes in the method for implementing the above-mentioned embodiment of the present application can also be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of each of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device, medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory, random access memory, electric carrier signal, telecommunication signal and software distribution medium that can carry the computer program code. It should be noted that the content contained in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable storage media do not include electric carrier signals and telecommunication signals.
[0094] Furthermore, the present application also provides a smart device.
[0095] In an embodiment of an intelligent device according to the present application, the intelligent device mainly includes a processor and a memory, the memory can be configured to store a program for executing the voice data annotation method of the above method embodiment, and the processor can be configured to execute the program in the memory, which includes but is not limited to the program for executing the voice data annotation method of the above method embodiment. For ease of explanation, only the part related to the embodiment of the present application is shown. For specific technical details not disclosed, please refer to the method part of the embodiment of the present application.
[0096] In the embodiment of the present application, the intelligent device may be a control device device including various electronic devices. In some possible implementations, the intelligent device may include multiple memories and multiple processors. The program for executing the voice data annotation method of the above method embodiment can be divided into multiple subprograms, each of which can be loaded and run by a processor to execute different steps of the voice data annotation method of the above method embodiment. Specifically, each subprogram can be stored in different memories, and each processor can be configured to execute programs in one or more memories to jointly implement the voice data annotation method of the above method embodiment, that is, each processor executes different steps of the voice data annotation method of the above method embodiment, to jointly implement the voice data annotation method of the above method embodiment.
[0097] The above-mentioned multiple processors may be processors deployed on the same device. For example, the above-mentioned smart device may be a high-performance device composed of multiple processors, and the above-mentioned multiple processors may be processors configured on the high-performance device. In addition, the above-mentioned multiple processors may also be processors deployed on different devices. For example, the above-mentioned smart device may be a server cluster, and the above-mentioned multiple processors may be processors on different servers in the server cluster.
[0098] Furthermore, the present application also provides a computer-readable storage medium.
[0099] In a computer-readable storage medium embodiment according to the present application, the computer-readable storage medium may be configured to store a program for executing the voice data annotation method of the above method embodiment, and the program may be loaded and run by the processor to implement the above voice data annotation method. For ease of explanation, only the parts related to the embodiment of the present application are shown. For specific technical details not disclosed, please refer to the method part of the embodiment of the present application. The computer-readable storage medium may be a memory device formed by various electronic devices. Optionally, the computer-readable storage medium in the embodiment of the present application is a non-temporary computer-readable storage medium.
[0100] The relevant user personal information that may be involved in the various embodiments of this application is strictly in accordance with the requirements of laws and regulations, following the principles of legality, legitimacy and necessity, based on the reasonable purposes of business scenarios, to process the personal information that users actively provide during the use of products / services or generated due to the use of products / services, as well as the personal information obtained with the user's authorization.
[0101] The user personal information processed by this application will vary depending on the specific product / service scenario, and will be based on the specific scenario in which the user uses the product / service. It may involve the user's account information, device information, driving information, vehicle information or other related information. The applicant will treat the user's personal information and its processing with a high degree of diligence.
[0102] This application attaches great importance to the security of user personal information and has taken reasonable and feasible security protection measures that meet industry standards to protect user information and prevent personal information from being accessed, disclosed, used, modified, damaged or lost without authorization.
[0103] So far, the technical solutions of the present application have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Without departing from the principles of the present application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present application.
Claims
1. A speech data annotation method, characterized in that: The method comprises: Acquire in-vehicle service voice data, and construct training data based on the in-vehicle service voice data; Perform voiceprint recognition on the training data based on multiple pre-trained voiceprint models to obtain a voiceprint matching score of the training data corresponding to each pre-trained voiceprint model; cluster the voiceprint matching scores of the training data to obtain a voiceprint clustering result corresponding to each pre-trained voiceprint model; Based on the voiceprint clustering results corresponding to each pre-trained voiceprint model, the speech data corresponding to the same speaker is determined and the speaker is labeled.
2. The method according to claim 1, characterized in that The vehicle-borne service voice data specifically includes a plurality of voice data classified according to vehicle information; The constructing of training data based on the vehicle-borne service voice data specifically includes: For each voice data corresponding to each vehicle information, construct a positive sample based on the voice data corresponding to the same vehicle information, and construct a negative sample based on the voice data corresponding to the vehicle information different from the positive sample; The training data is obtained based on the constructed positive samples and negative samples, and the training data includes at least one positive sample and at least one negative sample.
3. The method according to claim 2, characterized in that The voiceprint matching score includes a score of a positive sample and a score of a negative sample; The voiceprint matching scores of the training data are clustered to obtain the voiceprint clustering results corresponding to each pre-trained voiceprint model: According to the same speaker as the clustering label, unsupervised clustering is performed on the scores of all positive samples and all negative samples, and the speech data corresponding to the training data is divided into at least one cluster, each cluster corresponding to a voiceprint clustering result; based on the at least one cluster, the voiceprint clustering result corresponding to each pre-trained voiceprint model is determined.
4. The method according to claim 3, characterized in that: The unsupervised clustering of the scores of all positive samples and the scores of all negative samples includes: The highest score among all negative samples is determined as the threshold value; Determine, according to the threshold value, positive samples having scores higher than the threshold value; Recognize the speech data corresponding to the determined positive samples as speech data of the same speaker; The speech data recognized as the same speaker are divided into the same cluster.
5. The method according to claim 1, characterized in that The method of determining the speech data corresponding to the same speaker based on the voiceprint clustering results corresponding to each pre-trained voiceprint model and labeling the speaker is specifically as follows: Voting is performed on the voiceprint clustering results corresponding to each pre-trained voiceprint model, and voice data identified as the same speaker that reaches a preset number of votes is selected, and the selected voice data is labeled with the same speaker identifier.
6. The method according to claim 2, characterized in that The method further includes: merging the annotated voice data based on the vehicle information; The annotated speech data refers to speech data that has been annotated by the speaker.
7. The method according to claim 6, characterized in that The merging process of the labeled voice data based on the vehicle information specifically includes: Take the feature average value of the labeled voice data corresponding to each vehicle information; Determine the similarity between any two users corresponding to the vehicle information based on the feature average value; The annotated voice data corresponding to different vehicle information are merged based on the similarity.
8. A voiceprint recognition method, characterized in that: The method comprises: Based on the speech data annotated with speakers obtained by the speech data annotation method described in any one of claims 1 to 7, a pre-trained voiceprint model is trained to obtain a trained voiceprint model; and voiceprint recognition is performed based on the trained voiceprint model.
9. An intelligent device, comprising a processor and a memory, wherein the memory is suitable for storing a plurality of program codes, characterized in that: The program code is suitable for being loaded and run by the processor to execute the speech data annotation method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a plurality of program codes, characterized in that: The program code is suitable for being loaded and run by a processor to execute the speech data annotation method according to any one of claims 1 to 7.