A mutual information processing method for language recognition model

Abstract

The present application provides an interactive information processing method of a language recognition model, which relates to the field of language recognition technology. The method includes: obtaining an input target user, collecting language recognition samples, performing high-frequency recognition error phrase analysis, outputting a first confusion mapping phrase, obtaining a language conversion scene feature sample, establishing multiple confusion recognition modalities, constructing a confusion recognition mechanism, and when the target language recognition model triggers the confusion recognition mechanism, performing confusion vocabulary recognition, outputting conversion text, performing semantic recognition, and outputting semantic recognition results. The present application can solve the technical problem of low speech recognition accuracy due to individual differences in users and complex language environments in the prior art. By collecting speech recognition samples of target users, performing high-frequency recognition error phrase analysis, and establishing multiple confusion recognition modalities, the technical effect of improving speech recognition accuracy is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of language recognition, and in particular to an interactive information processing method of a language recognition model. Background Art

[0002] With the continuous development of artificial intelligence, speech recognition technology has been widely used in many fields such as classroom teaching, smart home, voice assistant, and in-vehicle system. Language recognition models can handle multiple languages ​​and maintain high recognition accuracy in different environments. Currently, existing speech recognition technologies perform poorly when processing complex or ambiguous language input.

[0003] In summary, the prior art has a technical problem of low accuracy of speech recognition due to individual differences among users and complex language environments. Summary of the invention

[0004] The purpose of this application is to provide an interactive information processing method for a language recognition model to solve the technical problem in the prior art of low speech recognition accuracy due to individual differences among users and complex language environments.

[0005] In view of the above problems, the present application provides an interactive information processing method for a language recognition model, wherein the method includes: obtaining an input target user of a target language recognition model, wherein the input target user includes at least one user; collecting a language recognition sample of the input target user, performing a high-frequency recognition error phrase analysis, and outputting a first confusion mapping phrase whose high-frequency confusion probability is greater than a preset probability, wherein the first confusion mapping phrase contains multiple confusion words; obtaining a language conversion scene feature sample of the language recognition sample; performing associated scene and associated context vocabulary recognition on the first confusion mapping phrase based on the language conversion scene feature sample and the language recognition sample, and establishing multiple confusion recognition modalities; constructing a confusion recognition mechanism based on the multiple confusion recognition modalities, and when the target language recognition model triggers the confusion recognition mechanism, performing confusion vocabulary recognition based on the multiple confusion recognition modalities, and outputting converted text; performing semantic recognition based on the converted text, and outputting a semantic recognition result.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] By acquiring the input target user of the target language recognition model, wherein the input target user includes at least one user; collecting the language recognition sample of the input target user, performing high-frequency recognition error phrase analysis, and outputting a first confusion mapping phrase with a high-frequency confusion probability greater than a preset probability, wherein the first confusion mapping phrase contains multiple confusion words; acquiring the language conversion scene feature sample of the language recognition sample; performing associated scene and associated context vocabulary recognition on the first confusion mapping phrase based on the language conversion scene feature sample and the language recognition sample, and establishing multiple confusion recognition modalities; constructing a confusion recognition mechanism based on the multiple confusion recognition modalities, when the target language recognition model triggers the confusion recognition mechanism, performing confusion vocabulary recognition based on the multiple confusion recognition modalities, and outputting conversion text; performing semantic recognition based on the conversion text, and outputting semantic recognition results. That is, by acquiring the speech recognition sample of the target user, performing high-frequency recognition error phrase analysis, and establishing multiple confusion recognition modalities, the technical effect of improving the accuracy of speech recognition is achieved.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0010] Figure 1 A flowchart of an interactive information processing method for a language recognition model of the present application;

[0011] Figure 2 A schematic diagram of a flow chart for establishing multiple confusion recognition modes in an interactive information processing method of a language recognition model of the present application. DETAILED DESCRIPTION

[0012] This application solves the technical problem of low speech recognition accuracy due to individual differences of users and complex language environment in the prior art by providing an interactive information processing method for a language recognition model. By collecting speech recognition samples of target users, performing high-frequency recognition error phrase analysis, and establishing multiple confusion recognition modes, the technical effect of improving speech recognition accuracy is achieved.

[0013] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0014] Please see attached Figure 1 The present application provides a method for processing interactive information of a language recognition model, wherein the method specifically comprises the following steps:

[0015] Step 1: Obtain an input target user of a target language recognition model, wherein the input target user includes at least one user.

[0016] Specifically, identify the user groups that the language recognition model is intended to serve. These users are the ones that the language recognition model hopes to serve. Select at least one user as the starting point, which can be a single user or a group of users with similar language usage habits. For example, if the model is designed to help people who speak a specific dialect, then the target users are all people who speak that dialect. Perform feature analysis on the target users, including but not limited to the target users' language habits, accents, common vocabulary, etc. By identifying the target users, the model can more accurately capture the user's language characteristics.

[0017] Step 2: Collect the input target user's language recognition sample, perform high-frequency recognition error phrase analysis, and output a first confusion mapping phrase whose high-frequency confusion probability is greater than a preset probability, wherein the first confusion mapping phrase contains multiple confusing words.

[0018] Specifically, data generated by the target user when actually using the language recognition system is collected, including input speech, output conversion text, and expected conversion text. The samples need to cover a variety of situations and contexts that users may encounter, such as different noise backgrounds, speech speeds, tones, etc. The collected samples are analyzed to identify phrases where the model frequently makes mistakes when converting text. These phrases make mistakes due to similar pronunciations, similar contexts, or other factors. The number of times each word or phrase appears in the error sample and the frequency of incorrect recognition are recorded. According to the analysis results, high-frequency error phrases whose confusion probability exceeds a preset threshold are screened out. The preset probability is a pre-set value used to define which error frequencies can be identified as high frequencies. A first confusion mapping phrase containing multiple confusion words is output, and all high-frequency confusion words whose recognition error frequencies are higher than the preset threshold are listed, including homophones, easily confused phrases, or errors caused by the unique pronunciation habits of individual users. By identifying the first confusion mapping phrase, it helps to identify phrases that frequently cause recognition errors in actual applications of the language recognition model, thereby improving the accuracy of the model when processing the target user's speech.

[0019] Step 3: Obtain a language conversion scene feature sample of the language recognition sample.

[0020] Specifically, language conversion scenarios include different communication environments, such as classroom teaching, formal meetings, daily conversations, telephone conversations, etc. Each scenario has its own specific language usage habits and expressions. In classroom teaching, language use is usually more formal and structured; in formal meetings, formal and professional vocabulary is more likely to be used; and in daily conversations, more colloquial and informal expressions are used. Obtaining these scene feature samples is crucial to improving the accuracy and adaptability of the language recognition model. For example, if the model is exposed to a large number of classroom teaching scene feature samples during training, when the user records the classroom teaching, the model can more accurately identify and understand the language input of teachers and students. By analyzing these samples, the model can learn the language usage patterns in different scenarios, so that when encountering similar scenarios, it can more accurately understand and convert the user's language input.

[0021] Step 4: Based on the language conversion scene feature samples and the language recognition samples, the first confusion mapping phrase is subjected to associated scene and associated context vocabulary recognition to establish multiple confusion recognition modalities.

[0022] Specifically, the language conversion scene feature samples include different environmental noises, speaker voice features, language style, etc.; the language recognition samples include input voice, output conversion text, and expected conversion text. By comparing the actual conversion text with the expected conversion text, find out the phrases that are easily confused by the model during the recognition process, and use these phrases as candidates for the first confusion mapping phrases. Using the language conversion scene feature samples, analyze which words or phrases are prone to confusion in specific scenarios. Identify the contextual vocabulary related to the confusing phrases, and establish multiple confusion recognition modes based on the scene features and contextual vocabulary. Each mode targets a specific confusion situation, including different acoustic models, language models, and context-aware mechanisms. The establishment of multiple confusion recognition modes helps the model to more accurately identify and distinguish easily confused words. For example, in classroom teaching scenarios, teachers’ explanations and students’ questions need to be quickly recorded for later review and organization. In the teaching process, words with the same pronunciation but different meanings need to be distinguished. Based on this, a confusion recognition modality for homophones is established. According to the different meanings of the same words in different subjects, a context-related vocabulary confusion modality is established. In the meeting record scenario, professional terms or names are easy to confuse. According to the context of the meeting content, a confusion recognition modality specifically for professional terms in the meeting is established, and another modality is established for the accent and word usage habits of specific speakers. In the driving scenario, navigation instructions or place names are easy to confuse. A confusion recognition modality specifically for navigation instructions is established, and another modality is established for short instructions issued by the driver during driving. By establishing multiple confusion recognition modalities, the adaptability of the model to different scenarios is improved, making the model more accurate and reliable when processing specific types of voice input.

[0023] Step 5: construct a confusion recognition mechanism based on the multiple confusion recognition modalities, and when the target language recognition model triggers the confusion recognition mechanism, confusion vocabulary recognition is performed based on the multiple confusion recognition modalities, and converted text is output.

[0024] Specifically, multiple confusion recognition modalities are established, each modality is targeted at different confusion situations. Different confusion phrases are identified and distinguished through different training data and optimization strategies. The confusion recognition mechanism is embedded in the target language recognition model. If the model detects a possible confusion situation, the confusion recognition mechanism is triggered. The triggering mechanism can be based on multiple factors, such as the confidence of the model recognition results, the similarity of contextual information, the matching degree of acoustic features, etc. When the target language recognition model triggers the confusion recognition mechanism, the previously established multiple sets of confusion recognition modalities are used to recognize and interpret the user's language input. Each modality will provide recognition results for the confusing words, and these results can be verified with each other to improve the overall recognition accuracy. The results of multiple confusion recognition modalities are combined to convert the recognized speech or confusing words into written text. Through the integration of multiple confusion recognition modalities, the system can more comprehensively consider the various possibilities and contextual information of confusing words, thereby reducing misrecognition.

[0025] Step 6: Perform semantic recognition based on the converted text and output the semantic recognition result.

[0026] Specifically, the converted text is analyzed, semantic features are extracted from it, and semantic recognition is performed on the converted text, including intent recognition, sentiment analysis, topic classification, etc. Based on the results of semantic recognition, the corresponding semantic recognition results are output, which can be in the form of text or structured data. These results include the understanding of the meaning of the text and the inference of the user's intention. Semantic recognition is a key area of ​​natural language processing, which involves understanding the meaning and context of language. The meaning of language is understood by using models and algorithms, including the meaning of vocabulary, sentence structure, contextual relationships, and implicit meaning. Through semantic recognition, it is ensured that the language recognition model can not only accurately convert speech into text, but also understand the meaning of these texts.

[0027] Further, step 2 of this application includes:

[0028] The language recognition samples include multiple groups of recognition samples, wherein any group of recognition samples includes input speech, output conversion text and expected conversion text; multiple groups of error samples are established based on the output conversion text and the expected conversion text; high-frequency confusion probability identification is performed on the multiple groups of error samples, and multiple groups of error phrases are output, and any error phrase has identification information identifying the high-frequency confusion probability; based on the identification information, an error phrase whose high-frequency confusion probability is greater than a preset probability is output, and the first confusion mapping phrase is established.

[0029] Specifically, the recognition sample usually includes input speech, output conversion text and expected conversion text, where the input speech refers to the speech actually spoken by the user and is the original input of the language recognition model; the output conversion text is the text generated by the model based on the input speech; and the expected conversion text is the correct text corresponding to the input speech. For each group of recognition samples, the output conversion text is compared with the expected conversion text. If the output conversion text is inconsistent with the expected conversion text, this group of samples is marked as error samples. These error samples include input speech, output conversion text, expected conversion text and specific circumstances of errors in the recognition process. Each group of error samples is analyzed to calculate the frequency of occurrence of each phrase. If a phrase appears frequently in multiple error samples, it is a high-frequency confusion phrase in model recognition. These high-frequency confusion phrases are areas where the model needs special attention and improvement. For example, assuming that in multiple error samples, the model often misidentifies "apple" as "pingguo", then "apple" and "pingguo" will be regarded as a high-frequency confusion phrase. After identifying these high-frequency confusion phrases, identification information is generated for each phrase. This information describes the high-frequency confusion probability of the phrase, reflecting the frequency of the model mistakenly identifying a phrase as another phrase during the recognition process. For each error phrase, the high-frequency confusion probability in its identification information is compared with the preset probability. This preset probability is a threshold used to determine whether the frequency of occurrence of the confusion phrase is high enough. If the high-frequency confusion probability of an error phrase is greater than this preset probability, it is part of the first confusion mapping phrase. For example, assuming that the preset probability is 0.01, that is, only when the high-frequency confusion probability of the error phrase is greater than 1%, it is considered as a high-frequency confusion phrase that requires special attention. If a phrase has more than 3 errors in 100 recognitions, the high-frequency confusion probability is 3%, which is greater than the preset probability, so it will be included in the first confusion mapping phrase. By analyzing a large number of recognition samples, the model can learn the correspondence between different speech features and text, thereby improving the accuracy and efficiency of recognition.

[0030] Further, step 2 of this application includes:

[0031] Taking the consistency of the input speech as a clustering constraint, clustering the multiple groups of error samples to generate multiple clusters of error samples; calculating the number of positive samples of the first cluster of error samples based on the language recognition samples, calculating the ratio of the number of samples of the first cluster of error samples to the number of positive samples, and generating a first high-frequency confusion probability; generating identification information based on the first high-frequency confusion probability, and establishing a mapping association with the first cluster of error samples to generate the multiple groups of error phrases.

[0032] Specifically, the consistency of input speech is used as a clustering constraint, and similar error samples are clustered together according to speech features. An unsupervised learning algorithm, such as K-means or hierarchical clustering, is used to perform cluster analysis on multiple groups of error samples, and error samples with similar input speech are classified into the same cluster. After the clustering process is completed, multiple clusters of error samples will be generated. Each cluster represents a group of words that are similar in speech features and easily confused. For each clustering result, that is, each cluster of error samples, the number of positive samples of the first cluster of error samples, that is, the number of correctly recognized samples, is calculated based on the language recognition samples. The total number of samples of the first cluster of error samples is calculated, including correctly and incorrectly recognized samples. The ratio of the number of samples of the first cluster of error samples to the number of positive samples is calculated, that is, the first high-frequency confusion probability, which reflects the frequency and severity of misrecognition in the cluster. Based on the calculated high-frequency confusion probability, identification information is generated. This information is a quantitative description of the confusion probability of each cluster error sample, which can be the probability value itself or a label set according to the probability value, such as high confusion, medium confusion, and low confusion. The generated identification information is mapped and associated with the first cluster error samples to generate multiple groups of error phrases, so that each cluster has its corresponding high-frequency confusion probability information. Finally, each cluster will generate a group of error phrases, each group includes speech and expected conversion text. These phrases contain easily confused words and confusion probabilities.

[0033] Further, as attached Figure 2 As shown, step 4 of this application includes:

[0034] The first confusion mapping phrase is segmented into recognition modalities to generate multiple recognition modalities; the language conversion scene feature samples and the language recognition samples are traversed, and recognition scene association and context-associated vocabulary association are performed on the first recognition modality among the multiple recognition modalities to generate a first modality scene and a first modality key context; a first confusion recognition modality is constructed with the first recognition modality, the first modality scene and the first modality key context; and the first confusion recognition modality is added to the multiple confusion recognition modalities.

[0035] Specifically, the first confusion mapping phrase is divided into multiple recognition modes according to different features or attributes, and each recognition mode represents the specific usage of the phrase in different scenarios or contexts. Each phrase in the first confusion mapping phrase is analyzed and classified according to factors such as language features, usage scenarios, and contextual information. Phrases with similar features are classified into the same recognition mode, and each recognition mode represents the specific usage of the phrase in different scenarios or contexts. The first recognition mode in each recognition mode is associated with the recognition scene, and the mode is associated with a specific usage scenario, such as classroom teaching, meetings, driving, daily conversations, etc. For example, if a recognition mode mainly processes the language in the classroom teaching scene, it will focus on identifying professional terms, teaching instructions, interactive language between students and teachers, etc.; if a recognition mode mainly processes the language in the meeting scene, it will focus on identifying and understanding the language features in the meeting environment. Secondly, the mode is associated with specific contextual vocabulary, such as professional terms, spoken expressions, etc. Through these associations, the first modal scene and the first modal key context are generated.

[0036] The first recognition modality refers to a modality selected as a basis from multiple recognition modalities, representing a basic recognition method of the model when processing language recognition tasks. The first modality scenario refers to a specific scenario where this recognition modality is applicable, such as classroom teaching, meetings, driving, etc. The first modality key context refers to the vocabulary and expressions that the model needs to pay special attention to in this scenario. The first confusion recognition modality is constructed by integrating the recognition features, usage scenarios and key context information of the phrase. The constructed first confusion recognition modality is added to multiple confusion recognition modalities. The first confusion recognition modality is constructed based on the association analysis of the language conversion scenario feature samples and the language recognition samples. It is a model for identifying and processing easily confused vocabulary, which is constructed based on specific scenarios and context information after analyzing the language recognition samples. The model selects the most appropriate recognition modality to process the user's language input according to different scenarios and contexts.

[0037] Modal segmentation refers to the classification of phrases or speech fragments in language recognition tasks according to their characteristics or attributes, thereby forming multiple different recognition modalities, each of which represents a specific language usage scenario or vocabulary usage method. For example, in a complex language recognition model, multiple recognition modalities are included, such as the daily conversation modality that focuses on recognizing spoken language and informal expressions in daily communication; the modality specifically used to recognize and understand professional terminology in specific fields or industries; and the dialect recognition modality optimized for dialects in specific regions. By performing modal segmentation, the model can more accurately recognize and understand language input in different scenarios.

[0038] Furthermore, step five of this application includes:

[0039] The confusion recognition mechanism is embedded in the target language recognition model; the input speech of the target language recognition model is segmented and recognized, and the multiple similarities between the segmented speech and the confused speech samples of the multiple confusion recognition modalities are compared; when any similarity among the multiple similarities is greater than a preset similarity, the confusion recognition mechanism is triggered; the real-time recognition scene and the real-time context features of the input speech are collected; based on the real-time recognition scene and the real-time context features, matching is performed in the multiple confusion recognition modalities, and a matching modality is generated to perform confusion vocabulary recognition.

[0040] Specifically, multiple confusion recognition modalities are combined with the original language recognition model so that these modalities can be called in real time to assist recognition when the model processes speech input. When the target language recognition model receives the input speech, it first performs speech segmentation recognition on the speech, and divides the input continuous speech signal into a series of short-term speech segments, such as words, phrases or syllables. The segmented speech units are easier to accurately recognize and analyze. The model compares the multiple similarities between the segmented speech and the confused speech samples of multiple confusion recognition modalities, that is, compares the similarity between the segmented speech and the known confused speech samples. By calculating these similarities, the model can determine whether the input speech is similar to the known confused speech samples, thereby identifying possible confusion situations. When any similarity in the comparison result is greater than the preset similarity, the confusion recognition mechanism is triggered. The preset similarity is a threshold used to determine whether two speech units are similar enough to cause confusion. For example, the similarity threshold of the model is set to 0.8. When the model compares a segmented unit of the input speech with a known confused speech sample, if the similarity is greater than 0.8, it is considered that the input speech is similar to the confused speech sample, thereby triggering the confusion recognition mechanism.

[0041] The real-time recognition scenario refers to the specific environment or situation in which the model is processing the user's language input, including classroom teaching, meetings, driving, daily conversations, etc. Real-time context features refer to other information related to the real-time recognition scenario, such as the volume, speaking speed, and intonation of the voice. By matching the real-time recognition scenario and context features with multiple confusion recognition modalities, the model selects the most appropriate recognition modality to process the user's language input. Each confusion recognition modality is optimized for specific confusion situations. The matching model can more accurately recognize and understand the user's language input, especially when faced with easily confused words. Once the confusion recognition modality that best suits the current scenario is determined, the system will use this modality to further process the voice segment to generate the final confusion vocabulary recognition result, and convert the recognized confusion vocabulary into written text through the confusion recognition mechanism. For example, when a teacher is explaining the word "factor" and needs to distinguish it from the word "factory" which may be confused with it, the model collects real-time recognition scene and context features, such as the teacher's language pattern, class topic, etc., or the model receives text and image information (words or pictures on the slide) at the same time, the system will not mistakenly recognize it as "factory", so that the teacher's explanation can be understood more accurately; when the driver says "navigate to the nearest gas station" while driving, the system will not mistakenly recognize it as "navigate to the nearest supermarket", thus reducing the error of navigation instructions. By establishing multiple confusion recognition modes and trigger mechanisms, the accuracy of the speech recognition system in specific scenarios can be improved.

[0042] Furthermore, the present application also includes the following steps:

[0043] If the input target user is multiple users, voiceprint feature recognition is performed on the multiple users to generate multiple user voiceprint features; multiple confusion recognition mechanisms are established for the multiple users, wherein the multiple confusion recognition mechanisms have unique user identification codes; a multi-user confusion activation mechanism is constructed according to the multiple user voiceprint features, and the startup control of the multiple confusion recognition mechanisms is performed.

[0044] Specifically, voice samples of each user are collected, and the voiceprint features in these samples are respectively identified. Voiceprint feature recognition is a technology that distinguishes different individuals by analyzing the unique features of individual voices. The voiceprint features of each individual are unique, including the pitch, timbre, speaking speed, intonation, etc. of the voice, as well as the user's voice habits and personalized language style. By identifying and analyzing the voiceprint features of multiple users, a voiceprint feature model for each user can be generated. A unique confusion recognition mechanism is established for each user, and each mechanism is optimized for the specific language features and recognition requirements of the user. By assigning a unique user identification code to each user, the model can more accurately identify and understand the voice features and language habits of different users. Based on the voiceprint features of multiple users, a multi-user confusion activation mechanism is constructed, and the corresponding user confusion recognition mechanism is activated according to the features of the input voice. During the voice recognition process, when the system detects that the features of the input voice match the voiceprint features of a certain user, the multi-user confusion activation mechanism will start the corresponding user confusion recognition mechanism. When the system receives inquiries from multiple users about the same question, it can provide personalized answers and suggestions based on the voiceprint features and confusion recognition mechanisms of each user. For example, suppose there are three users in the system, namely user A, user B, and user C. The system collects voice samples from each user and analyzes the voiceprint features in these samples. By extracting and comparing these features, the model can generate three different voiceprint feature models, representing the voice features of user A, user B, and user C respectively. The system establishes a unique confusion recognition mechanism for each user and assigns a unique user identification code to each user. When user A, user B, and user C use the language recognition system, the system selects the appropriate confusion recognition mechanism for processing based on the activation codes of these users. If the voiceprint features of the user match the voiceprint features of user A, the system activates the confusion recognition mechanism associated with user A and gives a personalized answer suitable for user A. By collecting the voiceprint features of multiple users and establishing a personalized confusion recognition mechanism, the model's adaptability to different user groups is enhanced.

[0045] Furthermore, the present application also includes the following steps:

[0046] The input speech of the target language recognition model is compared with the multiple user voiceprint features for similarity to determine the user to be identified; a target confusion recognition mechanism is matched among the multiple confusion recognition mechanisms based on the user to be identified; and confusion vocabulary recognition is performed according to the target confusion recognition mechanism.

[0047] Specifically, the input speech of the target language recognition model is compared with the voiceprint features of multiple users that have been collected. The comparison process involves matching algorithms for voiceprint features, such as dynamic time warping, hidden Markov models, etc. Based on the results of the similarity comparison, it is determined which user the input speech is most likely to belong to. If there is more than one user with the highest similarity, the system needs further confirmation, such as requesting user confirmation or making a judgment based on contextual information. Based on the determined user to be identified, a matching target confusion recognition mechanism is found among the multiple confusion recognition mechanisms that have been established. The input speech is processed using the matching target confusion recognition mechanism to identify and distinguish confusing words. Through voiceprint feature comparison, the language recognition model can more accurately identify and understand the language input of different users.

[0048] Furthermore, the present application also includes the following steps:

[0049] If the input target user is a plurality of users, similarity recognition of confusion mapping phrases is performed between any two of the plurality of users to generate a plurality of user similarities; a plurality of users are combined based on the plurality of user similarities to generate a first user combination; and a unified confusion recognition mechanism is established for the first user combination.

[0050] Specifically, a confusion mapping phrase is extracted from the language recognition sample of each user, and the confusion mapping phrases between any two users among the multiple users are similarly identified. By calculating the similarity between these phrases, it is possible to identify whether the confusion patterns of the two users in the model are similar. Based on the results of similarity identification, the similarity between multiple users is generated to reflect the degree of similarity between different users in the confusion mapping phrases. Based on the similarities of multiple users, multiple users are combined to generate a first user combination. The combination process involves setting a threshold for user similarity, and users with higher similarity are combined together. These users have similarities in language habits and confusion patterns. A unified confusion identification mechanism is established for the first user combination, and a confusion identification mechanism applicable to all these users is established based on the similarity of these users in the confusion mapping phrases. Based on the combination of user similarities, the system can use resources more efficiently, merge users with higher similarity, and reduce resource consumption.

[0051] In summary, the interactive information processing method of a language recognition model provided by this application has the following technical effects:

[0052] By acquiring the input target user of the target language recognition model, wherein the input target user includes at least one user; collecting the language recognition sample of the input target user, performing high-frequency recognition error phrase analysis, and outputting a first confusion mapping phrase with a high-frequency confusion probability greater than a preset probability, wherein the first confusion mapping phrase contains multiple confusion words; acquiring the language conversion scene feature sample of the language recognition sample; performing associated scene and associated context vocabulary recognition on the first confusion mapping phrase based on the language conversion scene feature sample and the language recognition sample, and establishing multiple confusion recognition modalities; constructing a confusion recognition mechanism based on the multiple confusion recognition modalities, when the target language recognition model triggers the confusion recognition mechanism, performing confusion vocabulary recognition based on the multiple confusion recognition modalities, and outputting conversion text; performing semantic recognition based on the conversion text, and outputting semantic recognition results. That is, by acquiring the speech recognition sample of the target user, performing high-frequency recognition error phrase analysis, and establishing multiple confusion recognition modalities, the technical effect of improving the accuracy of speech recognition is achieved.

[0053] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0054] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A method for processing interactive information of a language recognition model, characterized in that: The method comprises: Acquire an input target user of a target language recognition model, wherein the input target user includes at least one user; Collecting the input target user's language recognition sample, performing high-frequency recognition error phrase analysis, and outputting a first confusion mapping phrase whose high-frequency confusion probability is greater than a preset probability, wherein the first confusion mapping phrase includes a plurality of confusing words; Acquire a language conversion scene feature sample of the language recognition sample; Based on the language conversion scene feature sample and the language recognition sample, the first confusion mapping phrase is identified with associated scenes and associated context words, and multiple confusion recognition modes are established, including: Performing recognition modality segmentation on the first confusion mapping phrase to generate multiple recognition modalities; Traversing the language conversion scene feature samples and the language recognition samples, performing recognition scene association and context association vocabulary association on a first recognition modality among the multiple recognition modalities, and generating a first modality scene and a first modality key context; Constructing a first confusion recognition modality with the first recognition modality, the first modality scenario, and the first modality key context; adding the first confusion identification modality to the plurality of confusion identification modalities; A confusion recognition mechanism is constructed based on the multiple confusion recognition modalities. When the target language recognition model triggers the confusion recognition mechanism, confusion vocabulary recognition is performed based on the multiple confusion recognition modalities, and converted text is output, including: embedding the confusion identification mechanism into the target language recognition model; Performing speech segmentation and recognition on the input speech of the target language recognition model, and comparing multiple similarities between the segmented speech and the confused speech samples of the multiple confused recognition modalities; When any similarity among the multiple similarities is greater than a preset similarity, triggering the confusion identification mechanism; Collecting real-time recognition scenes and real-time context features of the input speech; Based on the real-time recognition scenario and the real-time context feature, matching is performed on the multiple confusion recognition modalities to generate a matching modality for confusing vocabulary recognition; Semantic recognition is performed based on the converted text, and a semantic recognition result is output.

2. The interactive information processing method of a language recognition model according to claim 1, characterized in that: Collecting the input target user's language recognition sample, performing high-frequency recognition error phrase analysis, and outputting a first confusion mapping phrase with a high-frequency confusion probability greater than a preset probability, including: The language recognition samples include multiple groups of recognition samples, wherein any group of recognition samples includes input speech, output conversion text and expected conversion text; Establishing a plurality of groups of error samples based on the output conversion text and the expected conversion text; Performing high-frequency confusion probability identification on the multiple groups of error samples, and outputting multiple groups of error phrases, wherein any error phrase has identification information identifying the high-frequency confusion probability; The first confusion mapping phrase is established based on the error phrase whose high-frequency confusion probability is greater than a preset probability output by the identification information.

3. The interactive information processing method of a language recognition model according to claim 2, characterized in that: The high-frequency confusion probability is identified for the multiple groups of error samples, and multiple groups of error phrases are output, and any error phrase has identification information identifying the high-frequency confusion probability, including: Taking the consistency of the input speech as a clustering constraint, clustering the multiple groups of error samples to generate multiple clusters of error samples; Calculating the number of positive samples of the first cluster of error samples based on the language recognition samples, calculating the ratio of the number of samples of the first cluster of error samples to the number of positive samples, and generating a first high-frequency confusion probability; Identification information is generated based on the first high-frequency confusion probability, and a mapping association is established with the first cluster of error samples to generate the multiple groups of error phrases.

4. The interactive information processing method of a language recognition model according to claim 1, characterized in that: The method further comprises: If the input target user is multiple users, voiceprint feature recognition is performed on the multiple users to generate multiple user voiceprint features; Establishing multiple obfuscation identification mechanisms for the multiple users, wherein the multiple obfuscation identification mechanisms have unique user identification codes; A multi-user confusion activation mechanism is constructed according to the multiple user voiceprint features, and the startup control of the multiple confusion identification mechanisms is performed.

5. The interactive information processing method of a language recognition model according to claim 4, characterized in that: Constructing a multi-user confusion activation mechanism according to the multiple user voiceprint features, and performing startup control of the multiple confusion identification mechanisms, including: Performing a similarity comparison between the input speech of the target language recognition model and the multiple user voiceprint features to determine the user to be recognized; Matching a target obfuscation identification mechanism among the multiple obfuscation identification mechanisms based on the user to be identified; Confusion vocabulary recognition is performed according to the target confusion recognition mechanism.

6. The interactive information processing method of a language recognition model according to claim 1, characterized in that: The method further comprises: If the input target user is a plurality of users, similarity recognition of the confusion mapping phrases between any two of the plurality of users is performed to generate a plurality of user similarities; Performing a multi-user combination based on the multiple user similarities to generate a first user combination; A unified confusion identification mechanism is established for the first user combination.

Images