Error analysis method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of data analysis, and provides an error analysis method and system, electronic equipment and a storage medium, and the method comprises the steps: carrying out the preprocessing of to-be-analyzed data, and obtaining an attribute vector; based on the analysis path, identifying errors in the attribute vector to obtain an error identification result; carrying out error classification on the error identification result, analyzing true errors in an error classification result, and determining error causes; and displaying the error reasons to obtain an error map, and displaying the analysis process of the true error in the form of a tree diagram to obtain an analysis map. The method supports a user to customize an analysis path, can improve the flexibility of error analysis and meet the requirements of the user, can improve the efficiency and accuracy of error analysis by classifying error recognition results and focusing on analysis of true errors, and can improve the accuracy of error analysis by displaying an error map and an analysis map. A user can more intuitively and quickly position the error reason, so that the analysis efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to an error analysis method, system, electronic device and storage medium. Background Art

[0002] At present, the error analysis process in the recognition field (such as speech recognition, text recognition, image recognition, etc.) mainly relies on manual review and analysis. For example, in the field of speech recognition, the error analysis method is usually to manually monitor the audio, transcribe it into text, and then compare the transcribed text with the reference text to identify and analyze the differences and errors in the speech recognition process.

[0003] Error analysis methods that rely on manual review, relying on human hearing and language comprehension, can guarantee the accuracy of analysis to a certain extent. However, this analysis method is not only inefficient and costly, but also easily affected by subjective judgment, making it difficult to guarantee the objectivity and consistency of the analysis results. In addition, in the face of the need to process massive data in the big data era, manual review and analysis methods are obviously inadequate and difficult to adapt to efficient, large-scale data set error analysis tasks. Summary of the invention

[0004] The present invention provides an error analysis method, system, electronic device and storage medium to solve the defects of the related art in that the error analysis is inefficient, costly and difficult to expand to large-scale data sets.

[0005] The present invention provides an error analysis method, comprising: Determine the data to be analyzed, and pre-process the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process; Based on an analysis path, an error in the attribute vector is identified to obtain an error identification result, wherein the analysis path is pre-defined by a user; Error classification is performed on the error recognition results to obtain error classification results, and true errors in the error classification results are analyzed to determine the causes of the errors; The cause of the error is displayed to obtain an error map, and the analysis process of the true error is displayed in the form of a tree diagram to obtain an analysis map.

[0006] According to an error analysis method provided by the present invention, the step of determining the analysis path includes: Receive task configuration operations; In response to the task configuration operation, determining the task to be analyzed and each analysis function point corresponding to the task to be analyzed; receiving a selection operation, and determining, in response to the selection operation, analysis tools corresponding to the analysis function points and an execution order of the analysis tools; Based on the analysis tools corresponding to the analysis function points and the execution order of the analysis tools, the analysis path corresponding to the task to be analyzed is determined. The task to be analyzed includes one or more tasks, and each task to be analyzed corresponds to an analysis path.

[0007] According to an error analysis method provided by the present invention, the analysis function points include at least one of short board mining, decoding analysis, language model analysis, acoustic analysis, annotation correctness analysis, correction effect analysis, resource correctness analysis, and voice activity detection analysis.

[0008] According to an error analysis method provided by the present invention, the error in the attribute vector is identified based on the analysis path to obtain an error identification result, including: Based on the analysis tools in the analysis path and the execution order of the analysis tools, errors in the attribute vector are identified to obtain error identification results, which include at least one of audio error information, labeling error information, common error information, word error information, and decoding error information.

[0009] According to an error analysis method provided by the present invention, the error classification of the error recognition result to obtain the error classification result includes: A decision tree is applied to classify the error recognition results to obtain error classification results, wherein the error classification results include at least one of false errors, undeterminable errors, negligible errors, and true errors.

[0010] According to an error analysis method provided by the present invention, the step of determining the data to be analyzed includes: receiving a data upload operation, and determining data to be analyzed in response to the data upload operation; or, A data selection operation is received, and in response to the data selection operation, data to be analyzed is determined from the historical data set.

[0011] According to an error analysis method provided by the present invention, the receiving data selection operation, in response to the data selection operation, determining the data to be analyzed from the historical data set, includes: Receive user input operations; In response to the user input operation, determining a query keyword, and determining initial analysis data from the historical data set based on the query keyword; Receive data selection operation; In response to the data selection operation, the initial analysis data is filtered to obtain data to be analyzed.

[0012] According to an error analysis method provided by the present invention, the attribute vector includes audio, annotation, recognition result, post-processing information, voice activity detection information, decoding process information, and streaming recognition information.

[0013] According to an error analysis method provided by the present invention, the error cause includes at least one of a data problem, a voice activity detection problem, an acoustic model problem, a language model problem, and an engine parameter problem.

[0014] An error analysis method provided by the present invention further includes: Based on the cause of the error, a speech recognition optimization plan is generated and displayed.

[0015] The present invention also provides an error analysis system, comprising: A data preprocessing module, used to determine the data to be analyzed, and preprocess the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process; An error recognition module, used to recognize errors in the attribute vector based on an analysis path to obtain an error recognition result, wherein the analysis path is pre-defined by a user; An error classification module is used to classify the error recognition result to obtain an error classification result; An error cause analysis module is used to analyze the true errors in the error classification results and determine the error causes; The user interaction interface is used to display the cause of the error to obtain an error map, and to display the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

[0016] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements any of the above-mentioned error analysis methods when executing the computer program.

[0017] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the error analysis method described above is implemented.

[0018] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the error analysis method described above is implemented.

[0019] The error analysis method, system, electronic device and storage medium provided by the present invention, by determining the data to be analyzed and preprocessing, generate an attribute vector containing key information in the speech recognition process, which can lay a solid foundation for subsequent error analysis. Subsequently, according to the analysis path pre-defined by the user, the errors in the attribute vector can be accurately identified, and the error recognition results can be generated. This step not only enhances the flexibility of error analysis, but also ensures that the analysis process can meet the personalized needs of the user. On this basis, by carefully classifying the error recognition results, focusing on analyzing true errors, and deeply analyzing the causes of errors, the error analysis process is made more efficient and targeted, thereby improving the efficiency and accuracy of error analysis. In addition, by intuitively displaying the error map, the complex causes of errors can be presented in a graphical manner, so that users can intuitively understand and analyze errors and their causes. At the same time, the analysis map in the form of a tree diagram shows the analysis process of true errors, so that the output error causes are interpretable, which can not only further help users quickly locate the causes of errors and the specific links of error generation, but also facilitate users to judge the reliability of the analysis process. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical 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 paying creative work.

[0021] Figure 1 It is a flow chart of the error analysis method provided by the present invention; Figure 2 It is a schematic flow chart of the analysis path determination step provided by the present invention; Figure 3 is a schematic diagram of an error map provided by the present invention; Figure 4 is a schematic diagram of an analysis map provided by the present invention; Figure 5 It is a structural schematic diagram of the error analysis system provided by the present invention; Figure 6 It is an overall flow chart of the error analysis method for speech recognition provided by the present invention; Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0023] Error analysis refers to comparing the output of the recognition system with the actual standard answer or expected result, identifying the differences and errors, and further analyzing the causes of the errors. Its ultimate goal is to improve the accuracy and reliability of the recognition system and optimize the user experience. For example, speech recognition error analysis refers to identifying and classifying errors by analyzing the correspondence between the output of the speech recognition system and the original speech; text recognition error analysis refers to identifying and classifying errors in text recognition technologies such as OCR (Optical Character Recognition) by analyzing the correspondence between the recognition results and the original text.

[0024] At present, in the field of error analysis technology, especially in complex application scenarios such as speech recognition, it has long relied mainly on manual review and analysis to detect errors in the speech recognition process. The core of this method is to manually monitor the audio, transcribe it into text, and then compare the transcribed text with the reference text to identify and classify the differences and errors. However, the manual review and analysis method is inefficient, costly, and highly dependent on individual subjective judgment, making it difficult to guarantee the objectivity and consistency of the analysis results. In addition, the manual review and analysis method cannot adapt to the analysis tasks of efficient and large-scale data sets.

[0025] With the rapid development of artificial intelligence and machine learning technologies, automated error analysis tools have emerged, bringing new possibilities to the field of error analysis. These tools use advanced classification algorithms, deep learning models, and reinforcement learning strategies to try to automate and intelligentize error identification, thereby reducing the manual burden and improving analysis efficiency. However, although automated tools have alleviated the limitations of manual review to a certain extent, they still have shortcomings in error classification, cause location, and optimization solution generation. For example, some analysis tools cannot accurately distinguish between different types of errors, or it is difficult to accurately capture the root cause of the error. In addition, with the development of technology, users' expectations for error analysis tools are also increasing. They not only hope that the tools can provide accurate analysis results, but also hope that the tools can provide an intuitive user interaction interface to meet the personalized analysis needs of different users and different application scenarios.

[0026] The present invention comes into being under such technical background and demand, and aims to provide an intelligent error analysis method and system, which can automatically identify, classify, locate and optimize errors in the recognition process (such as speech recognition process), and provide a friendly user interaction interface and support customized analysis paths to meet the needs of different users and application scenarios, thereby overcoming the above defects and providing users with an efficient, accurate, flexible and friendly error analysis solution. The technical solution provided by the present invention will be introduced below by taking error analysis for speech recognition as an example.

[0027] Figure 1 It is a flow chart of the error analysis method provided by the present invention, such as Figure 1 As shown, the method includes: Step 110, determining the data to be analyzed, and preprocessing the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process.

[0028] Specifically, the method provided by the embodiment of the present invention can be applied to an error analysis system. The data to be analyzed refers to the original speech recognition data that needs to be analyzed for errors, and may include audio, annotation information (i.e., annotated reference text), etc. After obtaining the data to be analyzed, the data to be analyzed can be preprocessed by calling a speech recognition system, open source tools, etc. to obtain an attribute vector. Here, the preprocessing may include steps such as parsing of audio files, extraction of post-processing information, acquisition of Voice Activity Detection (VAD) results, recording of decoding process information, and organization of streaming recognition information. These steps are intended to convert the original speech recognition data into a structured, easy-to-analyze form.

[0029] It can be understood that the attribute vector refers to a vector obtained after preprocessing that contains various information about the data to be analyzed during the speech recognition process. Each attribute value in this vector represents a label of the data, which contains all the key information required for error analysis. For example, the attribute vector can include audio, annotated text, recognition result text, post-processing information (such as punctuation addition, case conversion, etc.), VAD information (such as which parts are speech and which parts are silence), decoding process information, and streaming recognition information. The construction of the attribute vector enables subsequent error analysis to be performed based on this rich information.

[0030] Step 120, based on the analysis path, the error in the attribute vector is identified to obtain an error identification result, wherein the analysis path is obtained based on user pre-definition.

[0031] Specifically, after obtaining the attribute vector of the data to be analyzed, the error in the attribute vector can be identified according to the analysis path to obtain an error identification result. Here, the error identification result refers to the specific information of the error identified from the attribute vector under the guidance of the analysis path. For example, the error identification result may include audio error information, labeling error information, word error information, decoding error information, etc.

[0032] An analysis path is a strategy or process used to guide the error identification process. It defines how to traverse and check each attribute value (i.e., label) in the attribute vector to identify errors. This path is designed based on an understanding of the types and characteristics of errors that may occur during speech recognition, and is intended to efficiently locate and analyze errors. The process of identifying errors in the attribute vector according to the analysis path is actually a process of traversing and checking the attribute vector according to predetermined rules. Specifically, the analysis path can define the order of checks, the conditions for checks, and how to handle the errors found. For example, the analysis path can first check for inconsistencies between audio and annotations, then check for grammatical errors in the recognition results, then analyze the problems that may be introduced in the post-processing information, and finally consider the errors that may be caused by voice activity detection and decoding. By checking along such a path, errors in the attribute vector can be systematically identified.

[0033] It is understandable that before executing step 120, the user can pre-customize the analysis path according to the actual analysis needs, which can be achieved in the following ways: First, the user needs to clarify the specific business scenario, data characteristics and analysis goals; then, based on the business scenario, data characteristics and analysis goals and other information, the user can create an analysis task through the interactive interface provided by the system, and configure the corresponding analysis path for each analysis task. It should be understood that the user can create a single error analysis task at a time, or can create multiple error analysis tasks in parallel. Each analysis task is configured with an analysis path to meet the user's diverse analysis needs. By customizing the analysis path, the user can flexibly adapt to different business scenarios and analysis needs to achieve efficient and accurate error analysis.

[0034] Step 130, error classification is performed on the error recognition result to obtain an error classification result, and the true errors in the error classification result are analyzed to determine the cause of the error.

[0035] Specifically, after obtaining the error recognition result, various error information in the result can be classified to obtain the error classification result. Here, the error classification can be performed based on a series of predefined rules or standards, which can be derived from a deep understanding of the characteristics of the speech recognition system, or can be obtained based on historical data and experience summary.

[0036] Specifically, first, it is necessary to define the error categories. For example, errors can be divided into several categories such as false errors, undeterminable errors, ignorable errors, and true errors. Then, according to the predefined rules or standards for each error category, each error message in the error recognition result is carefully analyzed to determine which category it belongs to, thereby obtaining the error classification result. It should be understood that a false error means that the speech recognition result is correct but the labeling is wrong; an undeterminable error means that the audio is unclear or the audio pronunciation can correspond to multiple sets of texts; ignorable errors refer to spelling errors, modal particle errors, pronoun errors, erroneous sounds, etc. This type of ignorable error can be decided by the user whether to include it in the classification criteria according to actual needs; a true error means that the audio is clear and the labeling is correct, but the speech recognition result is wrong and does not belong to the above-mentioned ignorable error category.

[0037] Here, the error classification result refers to the error category information obtained after the error classification process. It can be presented in a structured form, such as a list, table or database, which contains the category label of each error, the specific information of the error (such as audio error information, labeling error information, etc.), etc.

[0038] In the error classification results, true errors actually refer to those errors that are indeed caused by the speech recognition system and need to be corrected. Compared with false errors and negligible errors, true errors are the focus of system optimization and improvement. Therefore, after obtaining the error classification results, if there are true errors, you can continue to conduct in-depth analysis of the true errors to determine the cause of the error, so as to provide specific problem location. Here, the cause of the error refers to the root cause of the true error. Determining the cause of the error is one of the key steps to optimize and improve the speech recognition system, which helps system developers to solve problems in a targeted manner and improve the accuracy and reliability of speech recognition.

[0039] Step 140, the error cause is displayed to obtain an error map, and the analysis process of the true error is displayed in the form of a tree diagram to obtain an analysis map.

[0040] Specifically, after completing the analysis of the true error, the cause of the error can be displayed in the form of an error map. Here, the error map is a graphical display method for intuitively presenting the cause of the error and its distribution. On the error map, different colors, shapes, or sizes can be used to represent different types of errors or causes of errors, so that users can see at a glance which areas or links have problems. The error map can help users quickly locate the problem.

[0041] At the same time, the analysis process of the true error can also be displayed through an analysis map in the form of a tree diagram. Here, the analysis process of the true error refers to the process of in-depth analysis of the identified true error in step 130. This process aims to find out the root cause of the true error and evaluate its impact on the performance of the speech recognition system. The analysis process can be displayed through a tree diagram to show how much data passes through which branch and which conclusion (i.e., the cause of the error) is finally reached. The analysis path taken by the analysis conclusion is presented, which is convenient for users to judge the reliability of the analysis process.

[0042] Specifically, when displaying the analysis process of true errors to form an analysis map, you can first determine the root node of the tree diagram, which usually represents the starting point or overall goal of the true error analysis process. Then, according to the specific steps and content of the true error analysis process, decompose the root node into multiple sub-nodes, each of which represents a specific link or step in the analysis process. Next, establish a hierarchical relationship based on the logical relationship and data flow relationship between the sub-nodes. In other words, determine the order of the sub-nodes, the subordinate relationship, and the connection path between them. Finally, you can also add corresponding descriptive information to each node, such as step name, operation content, analysis results, etc. This information helps users to have a deeper understanding of the analysis process and its results.

[0043] It is understandable that, compared with the error map, the analysis map focuses more on presenting the details of the analysis process and the logical relationship between the steps. It is an expansion of the error map. Through the analysis map, users can clearly see how the real errors are identified and analyzed, as well as the dependencies and impacts between each step. The analysis map helps users to have a deeper understanding of the performance and problems of the speech recognition system, and provides strong support for subsequent improvements and optimizations.

[0044] The method provided by the embodiment of the present invention can lay a solid foundation for subsequent error analysis by determining the data to be analyzed and preprocessing it to generate an attribute vector containing key information in the speech recognition process. Subsequently, according to the analysis path pre-defined by the user, the errors in the attribute vector can be accurately identified and the error recognition results can be generated. This step not only enhances the flexibility of error analysis, but also ensures that the analysis process can meet the personalized needs of the user. On this basis, by carefully classifying the error recognition results, focusing on analyzing true errors, and deeply analyzing the causes of errors, the error analysis process is made more efficient and targeted, thereby improving the efficiency and accuracy of error analysis. In addition, by intuitively displaying the error map, the complex causes of errors can be presented in a graphical manner, so that users can intuitively understand and analyze errors and their causes. At the same time, the analysis map in the form of a tree diagram shows the analysis process of true errors, so that the output error causes are interpretable, which can not only further help users quickly locate the causes of errors and the specific links of error generation, but also facilitate users to judge the reliability of the analysis process.

[0045] Based on the above embodiment, in step 110, determining the data to be analyzed includes: Step 111, receiving a data upload operation, and determining data to be analyzed in response to the data upload operation; or, Step 112: receiving a data selection operation, and determining data to be analyzed from the historical data set in response to the data selection operation.

[0046] It should be noted that the data to be analyzed may be all or part of the data currently uploaded by the user, or may be data selected and determined by the user from a historical data set, which is not specifically limited in the embodiment of the present invention.

[0047] Specifically, data upload operation refers to the process in which users transfer local or remotely stored data files to a designated server or data storage system through the interactive interface provided by the system for subsequent processing, analysis or storage. Data selection operation refers to the process in which users select a specific part of the data from the stored data set through the interactive interface provided by the system for subsequent processing and analysis.

[0048] In one embodiment, according to the data upload operation, the object that needs to be analyzed for error, that is, the data to be analyzed, can be determined. For example, the user selects the data file to be uploaded through the interactive interface provided by the system and clicks the upload button. After the system receives the upload request, it starts to receive the data file and performs basic verification, such as file format, size, integrity, etc. After the verification is passed, the system stores the data file in a specified location or database and generates corresponding data identification or metadata; then, the system determines the data to be analyzed from the data file just uploaded according to the user configuration or default rules. It should be understood that each time a data set is uploaded to the system, it will be stored in the database for subsequent use.

[0049] In another embodiment, the user may also select and determine the data to be analyzed from the historical data set. For example, the user browses the stored data set list through the system's interactive interface and selects a specific data set or data subset to be analyzed. After the system receives the user's selection request, it filters out the corresponding data from the database according to the user's selection conditions (such as data set name, time range, data label, etc.), thereby obtaining the data to be analyzed.

[0050] Based on any of the above embodiments, step 112 specifically includes: Step 1121, receiving user input operation; Step 1122, in response to the user input operation, determining a query keyword, and determining initial analysis data from the historical data set based on the query keyword; Step 1123, receiving a data selection operation; Step 1124: In response to the data selection operation, the initial analysis data is filtered to obtain data to be analyzed.

[0051] It should be noted that the error analysis system of the embodiment of the present invention supports query function and filtering function. The query function refers to the ability to query the data to be analyzed by keywords, and the filtering function supports the user to independently select and filter redundant data. Through the query function and filtering function, the user can accurately determine the data to be analyzed from the historical data set or the currently uploaded data set.

[0052] Specifically, the user input operation refers to the user inputting keywords or conditions for querying the data set through the input box or search box on the system interface. When the user enters text or selects query conditions in the search box on the interface, the system captures the input content and uses it as the query keyword or condition. These keywords or conditions are usually related to the content, attributes or labels of the data set, and are used to guide the system to filter out initial analysis data that meets user needs from the historical data set. Here, the query keyword refers to a word, phrase or condition used to filter out initial analysis data that meets user needs from the historical data set. It should be understood that the query function can be implemented in a variety of ways to determine the initial analysis data. For example, it can be queried by specifying a data set, it can be queried by specifying a word, and it can also be queried by specifying an analysis attribute. The embodiment of the present invention does not specifically limit this.

[0053] In addition, through the query function provided by the system, users can also manually enter keywords, select task analysis status, select task analysis scenario, select task time range, etc. in the search box of the system interface to search for analysis tasks they have created or analysis tasks created by other users, in order to query analysis results or execute subsequent analysis tasks.

[0054] It is understandable that the initial analysis data refers to the data set that is filtered out from the historical data set by querying keywords and used for further analysis and processing. Considering that different users may focus on different points when performing analysis tasks, for example, a certain analysis task focuses on all actual scenarios to verify the overall effect and robustness, then the required analysis data must contain some interference such as labeling errors and poor audio quality, which are also considered to be the error categories that this analysis task focuses on; and in some analysis scenarios, the analysis task may be debugged and verified for a certain fixed-point scene, and does not pay attention to the interference of poor quality data, but needs to analyze and display the error categories that the task actually focuses on. Therefore, the embodiment of the present invention can also support a filtering function. After the initial analysis data is obtained by query, the user can filter the initial analysis data through the filtering function to remove redundant, erroneous or data that does not meet user needs, so as to obtain more accurate and useful data to be analyzed.

[0055] Specifically, users can select the data type or condition to be filtered through the options or buttons on the system interface, and the system will filter the initial analysis data according to the user's selection. The system provides a variety of filtering options: ① spelling errors; ② abnormal annotations, that is, special symbols in the annotations, such as percentage signs, Arabic numerals, asterisks, and pound signs; ③ suspicious annotations, that is, the correctness of the annotations is questionable; ④ low-quality audio, that is, the audio noise is too loud or the audio volume is too low; ⑤ modal particles; ⑥ erhua sounds. Users can select the corresponding filtering conditions according to their needs, so as to remove these interference factors from the initial analysis data and obtain purer and more useful data to be analyzed.

[0056] Based on any of the above embodiments, Figure 2 is a flow chart of the analysis path determination step provided by the present invention, such as Figure 2 As shown, the step of determining the analysis path includes: Step 210, receiving a task configuration operation; Step 220: In response to the task configuration operation, determine the task to be analyzed and each analysis function point corresponding to the task to be analyzed.

[0057] Specifically, users can create analysis tasks through the interactive interface provided by the system, and customize the analysis path for each task. For example, when users create analysis tasks in the interactive interface, they are supported to configure the analysis function points that the task focuses on, select the corresponding analysis tools for each analysis function point, the indicators to be presented, and specify the analysis priority (i.e., determine the execution order of the analysis tools), etc., so as to customize the analysis path.

[0058] Task configuration operations refer to a series of operations performed by the user on the system's interactive interface to set and configure the analysis task to be executed. The operations may include setting the name of the analysis task, baseline comparison of the analysis, configuring the machine used for the analysis task, selecting the analysis function points corresponding to the analysis task, etc. Based on the user's task configuration operations, the system can determine the task to be analyzed and determine the analysis function points corresponding to the task to be analyzed from the preset analysis function point set. These analysis function points are what the user believes need to be focused on in the task to be analyzed.

[0059] Here, the task to be analyzed refers to the specific error analysis work that the user wants to perform through the system. It can be an analysis of speech recognition errors, an evaluation analysis of audio quality, an analysis of language model performance, etc. The analysis function points corresponding to the task to be analyzed refer to the specific functions selected by the user from the analysis functions provided by the system in order to complete the task to be analyzed.

[0060] It is understandable that the preset function points supported by the system mainly include eight analysis functions, including short board mining, decoding analysis, language model analysis, acoustic analysis, annotation correctness analysis, correction effect analysis, resource correctness analysis, and voice activity detection (VAD) analysis. Users can choose from these function points. Each function point covers more detailed functions and corresponding execution tools, so that users can choose which aspects to analyze as needed each time. The analysis function points corresponding to the task to be analyzed are the specific function points determined from the above preset function points. For example, each analysis function point can include decoding analysis, language model analysis, acoustic analysis, VAD analysis, etc., which together constitute the analysis path of the task to be analyzed.

[0061] It should be noted that weakness mining refers to identifying and analyzing weak links or performance bottlenecks in the speech recognition system; decoding analysis refers to analyzing the decoding process information of speech recognition to determine whether there are errors in the decoding process; language model analysis refers to evaluating the accuracy and performance of the language model to identify problems and improvement directions in the model; acoustic analysis refers to analyzing audio signals to evaluate audio quality and speech characteristics; annotation correctness analysis refers to checking the accuracy and consistency of annotation data; patch effect analysis refers to evaluating the impact of local updates or corrections to one or more components in the speech recognition system (such as acoustic models, language models, etc.) on the overall performance of the system. These updates or corrections may be to fix known errors, improve system performance, adapt to new data distributions, or meet specific application requirements; resource correctness analysis refers to checking the correctness and integrity of system resources (such as files, databases, etc.); VAD analysis refers to analyzing the voice activity detection information in the speech recognition process to determine whether there are errors in the VAD algorithm.

[0062] Step 230, receiving a selection operation, and in response to the selection operation, determining the analysis tools corresponding to the analysis function points and the execution order of the analysis tools; Step 240, based on the analysis tools corresponding to the analysis function points and the execution order of the analysis tools, determine the analysis path corresponding to the task to be analyzed, wherein the task to be analyzed includes one or more tasks, and each task to be analyzed corresponds to an analysis path.

[0063] Specifically, for each analysis function point, the system provides multiple analysis tools, and users can choose according to actual needs. The selection operation refers to the selection and configuration of the analysis tools for each analysis function point provided by the system by clicking, checking, dragging, etc. on the system interaction interface. After selecting the corresponding analysis tool for each analysis function point, the user can also specify the execution order of these analysis tools, which can be achieved by dragging the tool icon on the interface, setting the priority, etc.

[0064] It is understandable that the analysis tools corresponding to each analysis function point refer to the special analysis software or modules provided by the system in order to complete the specific analysis function point. These tools have specific functions and algorithms, and can accurately analyze and evaluate the needs of the analysis function point. The execution order of each analysis tool refers to the order in which each analysis tool is called when executing the analysis task. This order is specified by the user based on the needs of the analysis task and the characteristics of the tool. The correct execution order can ensure that the analysis task can be carried out according to the expected logic and process, so as to obtain accurate analysis results.

[0065] After determining the analysis tools and execution order corresponding to each analysis function point, the system will build the analysis path corresponding to the task to be analyzed based on this information. The analysis path is an ordered sequence of tool calls, which describes the various steps and tools required to complete the analysis task. The system will add these tools to the analysis path in turn according to the analysis tools and execution order specified by the user, and set the corresponding parameters and configurations. Finally, the system will automatically execute each analysis tool according to the constructed analysis path, complete the analysis task, and generate the corresponding analysis results.

[0066] In the embodiment of the present invention, through task configuration operations and selection operations, users can create tasks to be analyzed, specify various analysis function points, corresponding analysis tools and execution order, so as to construct an analysis path that meets their needs, making error analysis more flexible.

[0067] Based on any of the above embodiments, step 120 specifically includes: Based on the analysis tools in the analysis path and the execution order of the analysis tools, errors in the attribute vector are identified to obtain error identification results, which include at least one of audio error information, labeling error information, common error information, word error information, and decoding error information.

[0068] Specifically, for each task to be analyzed, after the analysis path is defined, the preprocessed attribute vector can be transmitted to each analysis tool entry for calling and analysis during the actual analysis process. Each analysis tool is executed in sequence according to the execution order specified by the user. Different analysis tools will output different error messages after analysis and processing.

[0069] It is understood that the attribute vector may include audio, annotation, recognition results, post-processing information, voice activity detection information, decoding process information, streaming recognition information, etc. Here, audio refers to the original voice data to be analyzed, which is the input of the speech recognition system, usually a recording or a voice stream. Annotation refers to the process of manually or automatically annotating audio data, which may include timestamps and corresponding text labels. In error analysis, the annotation information can be used as a reference text to compare with the speech recognition results. Recognition results refer to the text output by the speech recognition system after processing the audio data. Post-processing information refers to information for further processing or optimization of the recognition results after the recognition results are generated. Voice activity detection (VAD) information refers to information used to distinguish between speech and non-speech parts in audio data. Decoding process information refers to information about the algorithms and parameters used by the speech recognition system in converting audio data into text. Streaming recognition information refers to information about how the system processes continuously input audio data and generates real-time recognition results when performing real-time speech recognition. The various components in the attribute vector together constitute a complete set of information in the speech recognition process, which provides basic data support for subsequent error recognition, classification, and analysis. By deeply analyzing this information, errors in the speech recognition system can be discovered more effectively and corresponding improvement measures can be taken.

[0070] During the analysis process, the analysis tools are used in turn to analyze the information in the attribute vector in a determined execution order. Each analysis tool can analyze a specific part or all of the information in the attribute vector to detect potential errors. If an analysis tool detects an error, it will output the corresponding error information. Finally, all the error information will be integrated into an error recognition result. Here, the error recognition result can include audio error information, labeling error information, common error information, word error information, decoding error information, etc.

[0071] Among them, audio error information refers to errors related to the audio data itself, such as poor audio quality (too much noise, too low volume, etc.), incompatible audio formats, etc. Annotation error information refers to errors related to audio annotation, such as incorrect annotation, incomplete annotation, etc. Common error information refers to errors that often occur in the speech recognition process. Such errors may be related to the limitations of the recognition algorithm or the system itself. For example, the recognition algorithm may not be able to accurately recognize certain specific pronunciations or vocabulary. Word error information refers to errors related to the vocabulary in the recognition results, such as vocabulary recognition errors, vocabulary omissions or additions, etc. Decoding error information refers to errors related to the decoding process, such as decoding failures, inaccurate decoding results, etc.

[0072] Based on any of the above embodiments, in step 130, the error classification of the error recognition result to obtain the error classification result includes: A decision tree is applied to classify the error recognition results to obtain error classification results, wherein the error classification results include at least one of false errors, undeterminable errors, negligible errors, and true errors.

[0073] Specifically, after obtaining the wrong recognition results, the decision tree technology can be used to classify the wrong recognition results to obtain the wrong classification results. Here, the decision tree is a commonly used machine learning algorithm that simulates the human decision-making process through a series of questions (i.e., decision nodes) and possible answers (i.e., branches).

[0074] Specifically, after obtaining the error recognition results (such as audio error information, labeling error information, common error information, word error information, decoding error information, etc.), the feature data that has the greatest impact on the error classification can be determined first, such as audio clarity, labeling correctness, comparison between recognition results and labels, etc. Here, the selection of these feature data can be determined based on information gain, Gini index and other criteria, which can be used to measure the importance of features to classification. Then, the above feature data set is extracted from the error recognition results, and the extracted feature data set is used to build a decision tree. At each node, a feature is selected for analysis, and the data set is divided into different subsets based on the analysis results. This process is repeated until the stop condition is met, such as the node purity reaches the set threshold, the number of node samples is insufficient (that is, the number of node samples is less than the minimum number of splits), the maximum depth is reached, or the gain after splitting is less than a certain threshold.

[0075] According to the output of the decision tree, the error recognition results can be classified into false errors, undeterminable errors, negligible errors, and true errors. For true errors, they can be further analyzed to determine the specific cause of the error. The system will display the analysis process of true errors through a tree diagram, showing how much data passes through which branch and ultimately reaches which error cause, which helps users understand the source of the error and take corresponding measures to improve it.

[0076] Based on the above embodiment, the error cause includes at least one of a data problem, a voice activity detection problem, an acoustic model problem, a language model problem, and an engine parameter problem.

[0077] Specifically, based on the error analysis experience and user feedback, we can summarize the following error reasons that users need to pay attention to and display them in the error map for users to view and analyze. The first is data problems, including audio anomalies and annotation anomalies, such as poor audio quality or incorrect annotations. If these problems are not clarified, they will cause confusion in subsequent analysis results, so they are more critical; the second is the voice activity detection problem. Unreasonable VAD segmentation of audio will lead to reduced speech recognition effect. For example, VAD segmentation of audio leads to sentence loss, or useful speech frames are lost at the beginning and end of the audio; the third is the acoustic model problem, that is, due to the problem of the acoustic model itself, a certain pronunciation cannot be accurately recognized; the fourth is the language model problem, such as the lack of certain words in the dictionary, or the pronunciation score of a certain word is much lower than that of another homophonic word; the fifth is the hot word problem, which is mainly caused by the false triggering of hot words; the sixth is the patch (i.e. correction) problem, which may cause problems if the patch is falsely triggered or not triggered; the seventh is the unreasonable parameter setting in the speech recognition system, such as the unreasonable setting of the beamHist value. It should be understood that the hot word problems, patch problems, and unreasonable parameter settings here can be collectively classified as engine parameter problems.

[0078] Figure 3 is a schematic diagram of an error map provided by the present invention, such as Figure 3 As shown in the figure, by further analyzing the true errors in the error classification results and obtaining the error causes, the error causes can be displayed through the error map. The error classifications specifically displayed in the error map are the specific error causes that lead to the true errors introduced above. Figure 3 There are several major categories shown in the figure. The darker the color, the more attention it needs to pay to the error. It represents the key error and is convenient for users to analyze the error. Figure 3 ED in it refers to End-to-End, and CE corresponds to Connectionist Temporal Classification. The two are used to represent different training criteria or loss functions of the speech recognition system.

[0079] Figure 4 is a schematic diagram of the analysis map provided by the present invention, such as Figure 4 As shown in the figure, after completing the in-depth analysis of the true error, the analysis process of the true error can be displayed through the analysis map. Figure 4The analysis map shown in is a directed graph composed of various analysis paths. The circles without color markings in the middle (i.e. nodes) represent the analysis tools in the analysis path, and the edges are the data connection flow directions. Path 1 and Path 2 represent the analysis paths composed of multiple analysis tools selected by the user. Each path from the starting point to the end point in the figure is an analysis path, which can be expanded by the system based on the user-defined analysis path.

[0080] Understandably, Figure 4 The analysis map shown in has a navigation function. It shows the path of the data through the decision tree, which steps are taken in the middle, how much data is excluded because it does not meet the conditions, and how much data eventually reaches the specified error cause. The analysis map will show the analysis details of the true error, which is an expansion of the error map. For each node of each analysis path shown in the analysis map, each choice will be given, how much data, and how much data finally reached the final error cause. In other words, the main content presented in the analysis map includes the classification process and the number of data passing through each branch. By displaying the analysis process of the true error on the analysis map, users can understand the process corresponding to the error cause more clearly, making the output error cause more explainable.

[0081] Based on any of the above embodiments, the method further includes: Based on the cause of the error, a speech recognition optimization plan is generated and displayed.

[0082] Specifically, based on the reasons for the errors obtained through analysis, the system can automatically generate targeted speech recognition optimization solutions based on preset rules or algorithms. Here, the speech recognition optimization solution refers to a series of improvement measures or suggestions for problems and errors in the speech recognition process. These solutions are intended to improve the accuracy and efficiency of speech recognition, reduce the error rate, and thus improve the user experience. For example, the speech recognition optimization solution may include adjusting the data set, modifying model parameters, optimizing the engine configuration, etc. It should be understood that the system can generate multiple optional optimization solutions so that users can choose the most appropriate solution according to the specific situation.

[0083] After the optimization plan is generated, the system can display it through the user interface so that users can view the detailed description of each plan, the expected effect, the difficulty of implementation, etc. In addition, users can also use the analysis navigation function to quickly browse multiple analysis processes and choose to perform specific analysis actions to generate more detailed analysis reports.

[0084] Based on any of the above embodiments, Figure 5 is a schematic diagram of the structure of the error analysis system provided by the present invention, such as Figure 5 As shown, the system includes: The data preprocessing module 510 is used to determine the data to be analyzed, and preprocess the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process; An error identification module 520, configured to identify errors in the attribute vector based on an analysis path to obtain an error identification result, wherein the analysis path is pre-defined by a user; An error classification module 530 is used to classify the error recognition result to obtain an error classification result; An error cause analysis module 540 is used to analyze the true errors in the error classification results to determine the error causes; The user interaction interface 550 is used to display the cause of the error to obtain an error map, and to display the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

[0085] Specifically, the system provided by the embodiment of the present invention integrates multiple modules such as data preprocessing, error identification, error classification, error cause analysis, user interaction, and optimization solution generation, which can achieve comprehensive analysis and optimization of errors. The goal of the error analysis map (i.e., error map and analysis map) proposed and designed by the embodiment of the present invention is to locate errors and their causes, and provide solutions to the errors. This requires analyzing the data in detail, analyzing the errors that can be optimized, analyzing the causes of the errors, and finally outputting the optimization solution.

[0086] The reason why the error analysis map is called a map is that it can present the analysis path, the distribution of errors and the causes of the errors, and allow users to plan the analysis path. It is an interactive analysis function. The system provided by the embodiment of the present invention has three characteristics: interactive analysis, learnability, and flexibility. The system will remember the user's analysis conclusions and form a more accurate analysis report. The system specifically includes the following functions: Interactive Analysis: The system provides interactive analysis functions, allowing users to intuitively view and analyze errors and their causes through error maps.

[0087] Multi-dimensional error classification: The system can classify errors into false errors, undeterminable errors, negligible errors, and true errors based on the characteristics of the data to be analyzed.

[0088] Accurately locate the cause of errors: The system conducts in-depth analysis of real errors and directly points out the specific points that need to be adjusted, such as data problems, acoustic model problems, language model problems, etc.

[0089] Customized analysis path: Users can customize analysis tools and indicators according to their needs, and the system supports user-defined analysis paths.

[0090] Automatic generation of optimization plans: The system automatically generates targeted optimization plans based on the error cause analysis results.

[0091] Analysis Navigation: The system provides analysis navigation function to help users quickly locate the cause of the error and generate analysis reports.

[0092] It can be understood that the system provided by the embodiment of the present invention mainly includes the following key components: a data processing module, which is responsible for preprocessing the data to obtain an attribute vector; an error recognition module, which is used to automatically identify errors in the attribute vector using advanced algorithms, and to perform preliminary classification to obtain error recognition results, that is, one or more information including audio data errors, labeling errors, common errors, word errors, and decoding errors; an error classification module, which is used to distinguish errors into false errors, undeterminable errors, negligible errors, and true errors according to the nature of the errors; an error cause analysis module, which is used to conduct in-depth analysis of true errors and determine the causes of the errors, such as data problems, VAD problems, acoustic model problems, language model problems, engine parameter problems, etc.; a user interaction interface, which is used to provide a friendly interface so that users can customize analysis paths, select analysis tools, configure parameters, and intuitively view error maps and analysis maps. In addition, the system can also include an optimization solution generation module, which is used to automatically generate optimization solutions according to the causes of the errors, such as adjusting data sets, modifying model parameters, optimizing engine configuration, etc. It should be understood that when classifying errors and analyzing true errors, decision trees and analysis map technologies can be used to display the analysis process and results, and provide multi-angle error display and error cause analysis.

[0093] Figure 6 is an overall flow chart of the error analysis method for speech recognition provided by the present invention, such as Figure 6 As shown, the error analysis system provided by the embodiment of the present invention can be a software system, which can run on multiple operating systems and be accessed by users through a Web interface or a desktop application. The method for implementing the system specifically includes the following steps: Step S1, the user determines the data to be analyzed through the user interaction interface. The data may be uploaded by the user, or may be screened from historical data sets by the user through query and filter functions.

[0094] In step S2, the user configures the analysis task through the user interaction interface, selects the analysis function point of this analysis task, selects the corresponding analysis tool and parameter configuration, and specifies the execution order of each analysis tool, thereby customizing the analysis path.

[0095] Step S3, preprocessing the data to be analyzed through the data preprocessing module to obtain an attribute vector, which includes various types of information of the audio during the speech recognition process, such as audio, annotation, recognition results, post-processing information, VAD information, decoding process information, streaming recognition information, etc.

[0096] Step S4, using an error recognition module, identifies errors in the attribute vector to obtain error recognition results. Here, the error recognition results may include audio error information, labeling error information, common error information, word error information, decoding error information, etc.

[0097] Step S5, using the error classification module, a decision tree is used to classify the error recognition results to distinguish between false errors, undeterminable errors, negligible errors and true errors.

[0098] Step S6: The error cause analysis module is used to conduct an in-depth analysis of the true error to determine the error cause and provide specific problem location.

[0099] Step S7, displaying the cause of the error through an error map, and displaying the analysis process of the true error through an analysis map in the form of a tree diagram.

[0100] Step S8, generating an optimization solution based on the error cause obtained by analysis, and presenting it to the user through a user interaction interface.

[0101] The embodiment of the present invention provides an automated and efficient error analysis tool that can significantly improve the efficiency and accuracy of error analysis. By supporting user-defined analysis paths, the flexibility and adaptability of the system can be improved; by displaying error maps, users can intuitively understand and analyze errors and their causes; through the analysis navigation function, users can quickly locate the cause of the error and improve the efficiency of the analysis; by automatically generating optimization solutions, users are provided with clear improvement directions.

[0102] Figure 7 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the error analysis method, which includes: determining the data to be analyzed, and preprocessing the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed in the speech recognition process; based on the analysis path, identifying the error in the attribute vector to obtain an error recognition result, wherein the analysis path is obtained based on user pre-definition; performing error classification on the error recognition result to obtain an error classification result, and analyzing the true error in the error classification result to determine the error cause; displaying the error cause to obtain an error map, and displaying the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

[0103] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the relevant technology or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0104] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the error analysis method provided by the above methods, which method includes: determining the data to be analyzed, and preprocessing the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information about the data to be analyzed during a speech recognition process; based on an analysis path, identifying errors in the attribute vector to obtain an error recognition result, wherein the analysis path is obtained based on user pre-customization; classifying the error recognition result to obtain an error classification result, and analyzing the true error in the error classification result to determine the cause of the error; displaying the cause of the error to obtain an error map, and displaying the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

[0105] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the error analysis method provided by the above-mentioned methods, the method comprising: determining the data to be analyzed, and preprocessing the data to be analyzed to obtain an attribute vector, the attribute vector comprising information of the data to be analyzed in a speech recognition process; based on an analysis path, identifying errors in the attribute vector to obtain an error recognition result, the analysis path being obtained based on user pre-customization; error classifying the error recognition result to obtain an error classification result, and analyzing the true error in the error classification result to determine the cause of the error; displaying the cause of the error to obtain an error map, and displaying the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

[0106] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0107] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An error analysis method, characterized in that: include: Determine the data to be analyzed, and pre-process the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process; Based on an analysis path, an error in the attribute vector is identified to obtain an error identification result, wherein the analysis path is pre-defined by a user; Error classification is performed on the error recognition results to obtain error classification results, and true errors in the error classification results are analyzed to determine the causes of the errors; The cause of the error is displayed to obtain an error map, and the analysis process of the true error is displayed in the form of a tree diagram to obtain an analysis map.

2. The error analysis method according to claim 1, characterized in that: The step of determining the analysis path includes: Receive task configuration operations; In response to the task configuration operation, determining the task to be analyzed and each analysis function point corresponding to the task to be analyzed; receiving a selection operation, and determining, in response to the selection operation, analysis tools corresponding to the analysis function points and an execution order of the analysis tools; Based on the analysis tools corresponding to the analysis function points and the execution order of the analysis tools, the analysis path corresponding to the task to be analyzed is determined. The task to be analyzed includes one or more tasks, and each task to be analyzed corresponds to an analysis path.

3. The error analysis method according to claim 2, characterized in that: The analysis function points include at least one of short board mining, decoding analysis, language model analysis, acoustic analysis, annotation correctness analysis, correction effect analysis, resource correctness analysis, and voice activity detection analysis.

4. The error analysis method according to claim 2, characterized in that: The step of identifying errors in the attribute vector based on the analysis path to obtain an error identification result includes: Based on the analysis tools in the analysis path and the execution order of the analysis tools, errors in the attribute vector are identified to obtain error identification results, which include at least one of audio error information, labeling error information, common error information, word error information, and decoding error information.

5. The error analysis method according to claim 1, characterized in that: The error classification of the error recognition result to obtain the error classification result includes: A decision tree is applied to classify the error recognition results to obtain error classification results, wherein the error classification results include at least one of false errors, undeterminable errors, negligible errors, and true errors.

6. The error analysis method according to claim 1, characterized in that: The step of determining the data to be analyzed includes: receiving a data upload operation, and determining data to be analyzed in response to the data upload operation; or, A data selection operation is received, and in response to the data selection operation, data to be analyzed is determined from the historical data set.

7. The error analysis method according to claim 6, characterized in that: The receiving of the data selection operation and determining the data to be analyzed from the historical data set in response to the data selection operation include: Receive user input operations; In response to the user input operation, determining a query keyword, and determining initial analysis data from the historical data set based on the query keyword; Receive data selection operation; In response to the data selection operation, the initial analysis data is filtered to obtain data to be analyzed.

8. The error analysis method according to any one of claims 1 to 7, characterized in that: The attribute vector includes audio, annotation, recognition result, post-processing information, voice activity detection information, decoding process information, and streaming recognition information.

9. The error analysis method according to any one of claims 1 to 7, characterized in that: The error cause includes at least one of a data problem, a voice activity detection problem, an acoustic model problem, a language model problem, and an engine parameter problem.

10. The error analysis method according to any one of claims 1 to 7, characterized in that: Also includes: Based on the cause of the error, a speech recognition optimization plan is generated and displayed.

11. An error analysis system, characterized in that: include: A data preprocessing module, used to determine the data to be analyzed, and preprocess the data to be analyzed to obtain an attribute vector, wherein the attribute vector includes information of the data to be analyzed during the speech recognition process; An error recognition module, used to recognize errors in the attribute vector based on an analysis path to obtain an error recognition result, wherein the analysis path is pre-defined by a user; An error classification module is used to classify the error recognition result to obtain an error classification result; An error cause analysis module is used to analyze the true errors in the error classification results and determine the error causes; The user interaction interface is used to display the cause of the error to obtain an error map, and to display the analysis process of the true error in the form of a tree diagram to obtain an analysis map.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the error analysis method according to any one of claims 1 to 10 is implemented.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the error analysis method according to any one of claims 1 to 10 is implemented.

Images