Audio annotation method and device

Abstract

The present invention discloses an audio annotation method and device, which can be used in the financial field or other technical fields. The method comprises: inputting text data corresponding to the audio data to be annotated into a first annotation model and a second annotation model respectively, and obtaining the annotation results output by the two models; if the annotation result output by the first annotation model is the same as the annotation result output by the second annotation model, then using the same annotation result to perform incremental training on the first annotation model and the second annotation model respectively; if the recognition accuracy of the first annotation model increases after the incremental training, then the same annotation result is used as the correct annotation result; if the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, then the annotation result output by the expert annotation unit is used as the correct annotation result. The present invention helps to improve the efficiency and accuracy of audio annotation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an audio annotation method and device. Background Art

[0002] With the development of artificial intelligence technology, the demand for annotated data is increasing, and annotation has become a bottleneck for the rapid development of artificial intelligence. The existing mainstream audio annotation technology still uses more traditional manual annotation, which identifies the audio content as understandable content and subjectively annotates it based on experience and intention judgment. This annotation method has the disadvantages of consuming a lot of manpower and time, low efficiency, high cost, and unstable accuracy. The existing technology lacks a more efficient audio annotation solution. Summary of the invention

[0003] In order to solve the technical problem of low efficiency of the existing audio annotation method, the present invention proposes an audio annotation method and device.

[0004] In order to achieve the above object, according to one aspect of the present invention, a method for audio annotation is provided, the method comprising:

[0005] Convert the audio data to be annotated into text data;

[0006] Inputting the text data into a first annotation model and a second annotation model respectively, to obtain an annotation result output by the first annotation model and an annotation result output by the second annotation model;

[0007] If the labeling result output by the first labeling model is the same as the labeling result output by the second labeling model, using the same labeling result to perform incremental training on the first labeling model and the second labeling model respectively;

[0008] If the recognition accuracy of the first annotation model remains unchanged or decreases after the incremental training, the audio data to be annotated and / or the text data are sent to the expert annotation unit, and the annotation result output by the expert annotation unit is used as the correct annotation result; if the recognition accuracy of the first annotation model increases after the incremental training, the same annotation result is used as the correct annotation result;

[0009] If the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, the audio data to be annotated and / or the text data are sent to the expert annotation unit, and the annotation result output by the expert annotation unit is taken as the correct annotation result.

[0010] Optionally, the audio annotation method further includes:

[0011] If the recognition accuracy of the first annotation model increases after incremental training, the second annotation model is replaced by the first annotation model after incremental training, and the first annotation model is rolled back to the version before incremental training.

[0012] Optionally, the audio annotation method further includes:

[0013] If the recognition accuracy of the first labeling model remains unchanged or decreases after the incremental training, the first labeling model is rolled back to the state before the incremental training.

[0014] Optionally, the audio annotation method further includes:

[0015] After the labeling result output by the expert labeling unit is taken as the correct labeling result, the second labeling model is incrementally trained according to the labeling result output by the expert labeling unit.

[0016] Optionally, the audio annotation method further includes:

[0017] After taking the annotation result output by the expert annotation unit as the correct annotation result, if the annotation result output by the expert annotation unit is different from the annotation result output by the second annotation model, the text data is recorded to perform tuning analysis on the second annotation model based on the text data.

[0018] Optionally, the first annotation model includes semantic recognition models that are common to multiple fields, and the second annotation model includes a semantic recognition model that is specific to the field to which the audio data to be annotated belongs.

[0019] In order to achieve the above object, according to another aspect of the present invention, an audio tagging device is provided, the device comprising:

[0020] A text conversion module, used to convert the audio data to be annotated into text data;

[0021] A model output result acquisition module, used to input the text data into the first annotation model and the second annotation model respectively, to obtain the annotation result output by the first annotation model and the annotation result output by the second annotation model;

[0022] A first incremental training module, configured to perform incremental training on the first annotation model and the second annotation model respectively using the same annotation result if the annotation result output by the first annotation model is the same as the annotation result output by the second annotation model;

[0023] A first labeling result determination module, configured to send the audio data to be labeled and / or the text data to an expert labeling unit if the recognition accuracy of the first labeling model remains unchanged or decreases after incremental training, and then use the labeling result output by the expert labeling unit as a correct labeling result; and to use the same labeling result as a correct labeling result if the recognition accuracy of the first labeling model increases after incremental training;

[0024] The second annotation result determination module is used to send the audio data and / or the text data to be annotated to the expert annotation unit if the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, and then use the annotation result output by the expert annotation unit as the correct annotation result.

[0025] To achieve the above objective, according to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the above audio tagging method when executing the computer program.

[0026] In order to achieve the above objective, according to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed in a computer processor, the steps in the above audio tagging method are implemented.

[0027] In order to achieve the above objective, according to another aspect of the present invention, a computer program product is provided, comprising a computer program / instruction, which implements the steps in the above audio tagging method when executed by a processor.

[0028] The beneficial effects of the present invention are:

[0029] The embodiment of the present invention automatically performs annotation through the annotation model, thereby improving the efficiency of audio annotation. In addition, the present invention combines expert annotation while automatically performing annotation using the annotation model, thereby improving the accuracy of annotation by utilizing expert experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0031] Figure 1 is a flow chart of an audio tagging method according to an embodiment of the present invention;

[0032] Figure 2 is an overall flow chart of an audio tagging method according to an embodiment of the present invention;

[0033] Figure 3 is a structural block diagram of an audio tagging device according to an embodiment of the present invention;

[0034] Figure 4 It is a schematic diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0036] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0037] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0038] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] It should be noted that the audio annotation method of the present invention can be applied to the financial field as well as other technical fields.

[0040] Figure 1 is a flow chart of an audio tagging method according to an embodiment of the present invention. Figure 1 As shown, in one embodiment of the present invention, the audio tagging method of the present invention includes steps S101 to S105.

[0041] Step S101, converting the audio data to be annotated into text data.

[0042] In one embodiment of the present invention, this step specifically first uses a low-pass filtering algorithm to filter the audio data to be labeled, thereby improving the recognizability of the audio, obtaining filtered audio data, and then using the existing speech recognition (ASR) model to convert the audio data into text data.

[0043] Step S102: input the text data into a first annotation model and a second annotation model respectively, to obtain an annotation result output by the first annotation model and an annotation result output by the second annotation model.

[0044] In one embodiment of the present invention, the first annotation model includes semantic recognition models common to multiple fields, and the second annotation model includes a semantic recognition model specific to the field to which the audio data to be annotated belongs.

[0045] In the present invention, when the training data of the second annotation model is less or incomplete, a more general first annotation model that can be pre-trained has better recognition and annotation effects than the second annotation model, but as the second annotation model is continuously trained with annotation data, the second annotation model will have higher recognition accuracy for the corpus in the corresponding field than the first annotation model. In one embodiment of the present invention, the first annotation model and the second annotation model can select the same algorithm during configuration.

[0046] The present invention adopts two annotation models for annotation, so as to avoid the problem of inaccurate annotation results caused by single model annotation.

[0047] In one embodiment of the present invention, the annotation model is used to identify the semantics corresponding to the text data, and then annotate according to the identified semantics, and the annotation can be performed on the text data or on the audio data. In one embodiment of the present invention, the same annotation results can mean that the semantics recognized by the two models for the text data are the same.

[0048] In one embodiment of the present invention, the annotation result may be text data annotated with semantics. In another embodiment of the present invention, the annotation result may be audio data annotated with semantics.

[0049] Step S103: If the labeling result output by the first labeling model is the same as the labeling result output by the second labeling model, the first labeling model and the second labeling model are incrementally trained respectively using the same labeling result.

[0050] Step S104: if the recognition accuracy of the first annotation model remains unchanged or decreases after incremental training, the audio data and / or the text data to be annotated are sent to an expert annotation unit, and the annotation result output by the expert annotation unit is used as the correct annotation result; if the recognition accuracy of the first annotation model increases after incremental training, the same annotation result is used as the correct annotation result.

[0051] In one embodiment of the present invention, the expert annotation unit specifically performs manual annotation by experts based on their experience. The present invention combines expert annotation while using the annotation model to automatically perform annotation, and utilizes expert experience to improve the accuracy of annotation.

[0052] In one embodiment of the present invention, in this step, if the recognition accuracy of the first labeling model remains unchanged or decreases after incremental training, the first labeling model is rolled back to the state before incremental training. In the present invention, if the recognition accuracy of the first labeling model remains unchanged or decreases after incremental training, it means that the model after incremental training is not better than the model before training, so it is necessary to roll back to the state before incremental training.

[0053] In one embodiment of the present invention, if the recognition accuracy of the first annotation model increases after incremental training in this step, the second annotation model is replaced by the first annotation model after incremental training. In one embodiment of the present invention, while replacing the second annotation model with the first annotation model after incremental training, the first annotation model is also rolled back to the version before incremental training.

[0054] In an embodiment of the present invention, the first annotation model is a trained model, and the second annotation model is a model to be trained. The purpose of the replacement is to quickly improve the recognition accuracy of the second annotation model. Because it takes too long to slowly train the second annotation model using data, directly replacing the second annotation model with the first annotation model can quickly make the recognition accuracy of the second annotation model reach a level close to that of the first annotation model.

[0055] It should be noted that the purpose of the present invention is not to train the first labeling model and the second labeling model to a very high level of recognition accuracy, but to let the first labeling model and the second labeling model constrain each other to avoid the problem that the accuracy of single model labeling is difficult to ensure. Therefore, it is only necessary that the recognition accuracy of the second labeling model and the first labeling model is similar.

[0056] In one embodiment of the present invention, if the annotation result output by the expert annotation unit is obtained in this step, the second annotation model is incrementally trained using the annotation result output by the expert annotation unit.

[0057] In one embodiment of the present invention, if the annotation result output by the expert annotation unit is obtained in this step, the annotation result output by the expert annotation unit is also compared with the annotation result output by the second annotation model. If the annotation result output by the expert annotation unit is different from the annotation result output by the second annotation model, the text data is recorded to perform tuning analysis on the second annotation model based on the text data.

[0058] Step S105, if the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, the audio data to be annotated and / or the text data are sent to the expert annotation unit, and the annotation result output by the expert annotation unit is taken as the correct annotation result.

[0059] In one embodiment of the present invention, after the labeling result output by the expert labeling unit is taken as the correct labeling result, this step further uses the labeling result output by the expert labeling unit to perform incremental training on the second labeling model.

[0060] In one embodiment of the present invention, after taking the annotation results output by the expert annotation unit as the correct annotation results, this step also compares the annotation results output by the expert annotation unit with the annotation results output by the second annotation model. If the annotation results output by the expert annotation unit are different from the annotation results output by the second annotation model, the text data is recorded to perform tuning analysis on the second annotation model based on the text data.

[0061] In the present invention, expert annotation is to minimize the possible errors in the automatic annotation of the model. When the data is entrusted to experts for supervised annotation, the expert's suggestions have a higher priority and the flag bits of the data are clearly marked. This can ensure the accuracy of the corresponding corpus and reduce the risks caused by machine learning model annotation.

[0062] At the same time, the efficiency advantage brought by machine learning model annotation is unique to the overflow and waste of manpower caused by the use of all expert annotations. The accuracy of human ear recognition of text is not necessarily much better than the relatively mature ASR technology. Annotating the text content translated by ASR technology can greatly reduce the time required by experts. According to experiments, the accuracy of the final annotation result is as high as 75% when the method of the present invention is used for annotation, which is only slightly lower than 3 to 4 percentage points of full expert annotation, and the corresponding optimization of annotation time is only 40% of full expert annotation, which greatly improves efficiency.

[0063] Figure 2 FIG. 1 is an overall flow chart of an audio tagging method according to an embodiment of the present invention. Figure 2As shown, in one embodiment of the present invention, the present invention first uses a filtering algorithm to filter the audio data to be annotated, improves the recognizability of the audio, and obtains filtered audio. Then, the ASR model is used to translate the filtered audio to obtain the translated audio, which is the text corpus. Then, the holdout method is used to divide the text corpus, and the divided corpus is sent to the model for annotation.

[0064] like Figure 2 As shown, in one embodiment of the present invention, the first annotation model is a general model, which includes a trained semantic recognition model applicable to multiple fields, and the second annotation model is a model to be trained, that is, the second annotation model includes an immature semantic recognition model that still needs to be trained. The model to be trained can be continuously trained and optimized by the method of the present invention, so that the model to be trained has a certain degree of recognition accuracy for the field to which the audio data to be annotated belongs. In one embodiment of the present invention, after sufficient training, the recognition accuracy of the model to be trained for the field to which the audio data to be annotated belongs can be higher than the recognition accuracy of the general model for the field to which the audio data to be annotated belongs.

[0065] like Figure 2 As shown, in one embodiment of the present invention, the present invention simultaneously feeds data (i.e., text corpus) into the general model and the model to be trained, and then determines whether the annotation results of the general model and the model to be trained are the same. If the annotation results are the same, the annotation results are simultaneously sent to the two models for incremental training, and the changes in the prediction accuracy of the general model for its standard training set are monitored. If the accuracy of the general model remains unchanged or increases, the general model is used to replace the model to be trained to continue iterative training. If the accuracy of the general model decreases, the general model is rolled back to the previous version, and the expert annotation unit is used to annotate the data. If the annotation results of the general model and the model to be trained are different, the expert annotation unit is directly used to annotate the data.

[0066] In the present invention, if an expert labeling unit is used for labeling, the labeling result output by the expert labeling unit is regarded as the correct labeling result. Figure 2 As shown, in one embodiment of the present invention, after obtaining the labeling result output by the expert labeling unit, the labeling result output by the expert labeling unit is used to train the model to be trained. The labeling result of the expert is considered to be correct. The present invention uses the labeling result output by the expert labeling unit to train the model to be trained, thereby improving the recognition accuracy of the model to be trained. In one embodiment of the present invention, after obtaining the labeling result output by the expert labeling unit, the labeling result output by the expert labeling unit is compared with the labeling result of the model to be trained ( Figure 2The annotation results in 2) are compared. If the comparison results are consistent, it is considered that the recognition of the model to be trained is accurate, and the processing of the next data to be labeled can be continued. If the comparison results are inconsistent, it is considered that the current recognition of the model to be trained is wrong, and the text corpus with the recognition error can be marked for subsequent tuning analysis of the model to be trained.

[0067] It can be seen from the above embodiments that the present invention makes full use of the advantages of machine learning and combines the existing mature ASR technology to greatly improve the annotation efficiency of speech corpus, while the accuracy rate sacrificed is within the allowable error range. The use of ASR to translate speech compensates for the error problem caused by individual differences in expert annotation. At the same time, the use of expert annotation for compensatory annotation improves the accuracy of pure model annotation and avoids the problem of inaccurate model due to systematic problems. The present invention achieves at least the following beneficial effects:

[0068] 1. The present invention fully preprocesses the audio data to be annotated, converts the unstructured data that is difficult to analyze into easy-to-understand text data, greatly reducing the pressure of data cleaning under traditional methods.

[0069] 2. After using ASR to pre-process the audio data to be annotated, the present invention uses the annotation model to annotate the newly added corpus, which solves the pain points of traditional annotation methods and significantly improves the efficiency of annotation.

[0070] 3. The present invention adds an expert supervision mechanism to the process of labeling using the labeling model, thereby ensuring that the erroneous data automatically labeled by the model will not affect the model training effect, thereby avoiding the possibility of a vicious cycle.

[0071] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0072] Based on the same inventive concept, an embodiment of the present invention further provides an audio annotation device, which can be used to implement the audio annotation method described in the above embodiment, as described in the following embodiment. Since the principle of solving the problem by the audio annotation device is similar to that of the audio annotation method, the embodiment of the audio annotation device can refer to the embodiment of the audio annotation method, and the repeated parts will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiment is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0073] Figure 3is a structural block diagram of an audio tagging device according to an embodiment of the present invention. Figure 3 As shown, in one embodiment of the present invention, the audio tagging device of the present invention includes:

[0074] A text conversion module 1, used to convert the audio data to be annotated into text data;

[0075] Model output result acquisition module 2, used for inputting the text data into the first annotation model and the second annotation model respectively, to obtain the annotation result output by the first annotation model and the annotation result output by the second annotation model;

[0076] A first incremental training module 3, configured to perform incremental training on the first annotation model and the second annotation model respectively using the same annotation result if the annotation result output by the first annotation model is the same as the annotation result output by the second annotation model;

[0077] A first labeling result determination module 4 is used for sending the audio data to be labeled and / or the text data to an expert labeling unit if the recognition accuracy of the first labeling model remains unchanged or decreases after the incremental training, and then taking the labeling result output by the expert labeling unit as the correct labeling result; and for taking the same labeling result as the correct labeling result if the recognition accuracy of the first labeling model increases after the incremental training;

[0078] The second annotation result determination module 5 is used to send the audio data to be annotated and / or the text data to the expert annotation unit if the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, and then use the annotation result output by the expert annotation unit as the correct annotation result.

[0079] In one embodiment of the present invention, the audio annotation device further includes:

[0080] A model replacement module is used to replace the second labeling model with the first labeling model after incremental training if the recognition accuracy of the first labeling model increases after incremental training.

[0081] In one embodiment of the present invention, the audio annotation device further includes:

[0082] The second incremental training module is used to perform incremental training on the second annotation model according to the annotation results output by the expert annotation unit.

[0083] In one embodiment of the present invention, the audio annotation device further includes:

[0084] The model rollback module is used to roll back the first labeling model to the state before the incremental training if the recognition accuracy of the first labeling model remains unchanged or decreases after the incremental training.

[0085] In one embodiment of the present invention, the audio annotation device further includes:

[0086] The model tuning and analysis module is used to record the text data if the tagging result output by the expert tagging unit is different from the tagging result output by the second tagging model, so as to perform tuning and analysis on the second tagging model according to the text data.

[0087] In order to achieve the above object, according to another aspect of the present application, a computer device is also provided. Figure 4 As shown, the computer device includes a memory, a processor, a communication interface and a communication bus. A computer program that can be executed on the processor is stored in the memory. When the processor executes the computer program, the steps in the above-mentioned embodiment method are implemented.

[0088] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0089] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and units, such as the corresponding program units in the above method embodiments of the present invention. The processor executes various functional applications of the processor and works data processing by running the non-transitory software programs, instructions and modules stored in the memory, that is, implementing the method in the above method embodiments.

[0090] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0091] The one or more units are stored in the memory, and when executed by the processor, the method in the above embodiment is performed.

[0092] The specific details of the above-mentioned computer device can be understood by referring to the corresponding related descriptions and effects in the above-mentioned embodiments, and will not be repeated here.

[0093] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer-readable storage medium is also provided, wherein the computer-readable storage medium stores a computer program, and the computer program implements the steps in the above-mentioned audio tagging method when executed in a computer processor. Those skilled in the art can understand that the implementation of all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and the program can include the processes of the embodiments of the above-mentioned methods when executed. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk (Hard Disk Drive, abbreviated as: HDD) or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above-mentioned types of memory.

[0094] In order to achieve the above objective, according to another aspect of the present application, a computer program product is provided, including a computer program / instruction, which implements the steps of the above audio tagging method when executed by a processor.

[0095] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An audio tagging method, characterized in that: include: Convert the audio data to be annotated into text data; Inputting the text data into a first annotation model and a second annotation model respectively, to obtain an annotation result output by the first annotation model and an annotation result output by the second annotation model; If the labeling result output by the first labeling model is the same as the labeling result output by the second labeling model, using the same labeling result to perform incremental training on the first labeling model and the second labeling model respectively; If the recognition accuracy of the first annotation model remains unchanged or decreases after incremental training, the audio data to be annotated and / or the text data are sent to the expert annotation unit, and the annotation result output by the expert annotation unit is used as the correct annotation result; if the recognition accuracy of the first annotation model increases after incremental training, the same annotation result is used as the correct annotation result, wherein the second annotation model is the model to be trained, and if the recognition accuracy of the first annotation model increases after incremental training, the second annotation model is replaced by the first annotation model after incremental training, and the first annotation model is rolled back to the version before incremental training; If the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, the audio data to be annotated and / or the text data are sent to the expert annotation unit, and the annotation result output by the expert annotation unit is taken as the correct annotation result.

2. The audio tagging method according to claim 1, characterized in that: Also includes: If the recognition accuracy of the first labeling model remains unchanged or decreases after the incremental training, the first labeling model is rolled back to the state before the incremental training.

3. The audio tagging method according to claim 1, characterized in that: Also includes: After the labeling result output by the expert labeling unit is taken as the correct labeling result, the second labeling model is incrementally trained according to the labeling result output by the expert labeling unit.

4. The audio tagging method according to claim 1, characterized in that: Also includes: After taking the annotation result output by the expert annotation unit as the correct annotation result, if the annotation result output by the expert annotation unit is different from the annotation result output by the second annotation model, the text data is recorded to perform model tuning on the second annotation model according to the text data.

5. The audio tagging method according to any one of claims 1 to 4, characterized in that: The first annotation model includes semantic recognition models common to multiple fields, and the second annotation model includes a semantic recognition model dedicated to the field to which the audio data to be annotated belongs.

6. An audio tagging device, characterized in that: include: A text conversion module, used to convert the audio data to be annotated into text data; A model output result acquisition module, used to input the text data into the first annotation model and the second annotation model respectively, to obtain the annotation result output by the first annotation model and the annotation result output by the second annotation model; A first incremental training module, configured to perform incremental training on the first annotation model and the second annotation model respectively using the same annotation result if the annotation result output by the first annotation model is the same as the annotation result output by the second annotation model; A first labeling result determination module, configured to send the audio data to be labeled and / or the text data to an expert labeling unit if the recognition accuracy of the first labeling model remains unchanged or decreases after incremental training, and then use the labeling result output by the expert labeling unit as a correct labeling result; and for, if the recognition accuracy of the first annotation model increases after incremental training, taking the same annotation result as the correct annotation result, wherein the second annotation model is a model to be trained, and if the recognition accuracy of the first annotation model increases after incremental training, replacing the second annotation model with the first annotation model after incremental training, and rolling back the first annotation model to the version before incremental training; The second annotation result determination module is used to send the audio data and / or the text data to be annotated to the expert annotation unit if the annotation result output by the first annotation model is different from the annotation result output by the second annotation model, and then use the annotation result output by the expert annotation unit as the correct annotation result.

7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed in a computer processor, the computer program implements the method according to any one of claims 1 to 5.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images