Multi-service voice navigation method and device based on voice intent tracking micro agent

Abstract

The application relates to the technical field of voice data processing, and discloses a multi-service voice navigation method and device based on a voice intention tracking micro agent. According to the application, voice recognition is performed on a voice navigation request to obtain recognized text, intention recognition and keyword extraction are then performed, the core intention and key information in the voice request of a user can be accurately grasped, and the accuracy of voice navigation can be improved; emotion recognition and emotion score calculation operations can also be performed on a first voice processing result, and the current emotion state of a user can be intelligently perceived; based on this, the emotion factor of a user can be considered when a target intention node is determined and a jump is performed, more humanized and intelligent interaction is realized, the efficiency and accuracy of guiding a user to enter a corresponding service process are improved, the waiting time and operation steps of a user in voice navigation are reduced, and therefore the processing efficiency of multi-service voice navigation is improved and the experience of a user in the process of using voice navigation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice data processing, and in particular to a multi-service voice navigation method and device based on voice intent tracking micro-agents. BACKGROUND

[0002] In customer service systems in the telecommunications, government affairs, and financial industries, an automatic voice navigation (Interactive Voice Response, IVR) system, as a key self-service guidance tool, has been widely used. It provides users with a convenient way to handle business through automated voice interaction, effectively reducing the workload of human customer service, and plays an important role in improving service efficiency and reducing operating costs.

[0003] However, the use of traditional IVR systems has many defects:

[0004] First, the interactive experience is poor. Traditional IVR systems have fixed paths, and when faced with complex, ambiguous, or cross-process business requirements from users, it is easy for users to lose their way, repeat operations, or even be transferred to human customer service inefficiently, wasting users' time and effort and greatly reducing the users' interactive experience.

[0005] Second, the intelligent understanding ability is weak. Most traditional IVR systems use static menu navigation or shallow intent recognition based on keywords. The static menu lacks flexibility and cannot dynamically adjust the navigation function according to user needs. Keyword recognition only recognizes shallow intent and ignores context, resulting in low accuracy of intent recognition.

[0006] Third, there is a lack of user emotion and trajectory analysis. When processing user voice, traditional IVR systems only focus on a single and one-sided dimension of text information, ignoring user emotional state and conversation trajectory. This results in an inability to timely perceive negative emotions from users and difficulty in dynamically and intelligently jumping and connecting business processes based on conversation trajectories, increasing the user's operational burden. SUMMARY

[0007] The present application provides a multi-service voice navigation method and device based on voice intent tracking micro-agents, which can record and track user intent changes and emotional states during multi-round interactions, implement dynamic navigation and precise transfer across business processes, and improve user experience.

[0008] The first aspect of the present application discloses a multi-service voice navigation method based on voice intent tracking micro-agents, the method comprising:

[0009] When a user triggered voice navigation request is detected, a speech recognition operation is performed on the voice navigation request to obtain recognized text for the voice navigation request as a speech recognition result;

[0010] An intent recognition operation and a keyword extraction operation are respectively performed on the speech recognition result to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first speech processing result;

[0011] An emotion recognition operation and an emotion score calculation operation are respectively performed on the first speech processing result to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second speech processing result;

[0012] According to the first speech processing result and the second speech processing result, a determination and jump operation for a target intent node are performed to obtain a jump result for the target intent node as a final response result responding to the voice navigation request.

[0013] As an optional implementation, in the first aspect of the present application, the speech recognition operation performed on the voice navigation request to obtain recognized text for the voice navigation request comprises:

[0014] Obtaining voice stream data corresponding to the voice navigation request;

[0015] According to a preset signal processing algorithm, a speech feature extraction operation is performed on the voice stream data to obtain a speech feature for the voice stream data;

[0016] Determining a speech text recognition model matched with the voice stream data;

[0017] Inputting the speech feature into the speech text recognition model to sequentially perform a text reasoning operation and a text output operation on the speech feature by the speech text recognition model to obtain output text for the speech feature;

[0018] Performing a text post-processing operation on the output text to obtain recognized text for the voice navigation request, and the text post-processing operation is used to correct the output text.

[0019] As an optional implementation, in the first aspect of the present application, the intent recognition operation and the keyword extraction operation are respectively performed on the speech recognition result to obtain the voice navigation intent for the intent recognition operation and the keyword set for the keyword extraction operation as the first speech processing result, comprising:

[0020] perform a text preprocessing operation on the speech recognition result, and generate a semantic vector embedding for a text preprocessing result; the text preprocessing operation includes a text content cleaning operation and a word segmentation operation;

[0021] perform an intent recognition operation on the semantic vector embedding according to a preset large language model in combination with context window content of the semantic vector embedding, to obtain a voice navigation intent for the semantic vector embedding, the voice navigation intent including at least one main intent category, all of the main intent categories including at least one of query, reservation, complaint, and request for human;

[0022] perform a keyword extraction operation on the semantic vector embedding according to a preset keyword recognition model in combination with a historical keyword recognized for the semantic vector embedding, to obtain a keyword set for the semantic vector embedding, the keyword set including at least one keyword;

[0023] perform a fusion operation on the voice navigation intent and the keyword set according to a preset fusion module, to obtain a fusion result for the voice navigation intent and the keyword set as a first voice processing result.

[0024] As an optional implementation, in the first aspect of the present application, the voice features include at least one of pitch, speech rate, and sound intensity;

[0025] perform an emotion recognition operation and an emotion score calculation operation on the first voice processing result respectively, to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result, including:

[0026] perform an emotion recognition operation on the recognized text and the voice features according to a text emotion recognition model, to obtain an emotion recognition result for the recognized text and the voice features; the emotion recognition result including a confidence degree for a plurality of preset emotion dimensions;

[0027] generate an emotion snapshot for the emotion recognition result, and update an emotion trajectory for the user according to the emotion snapshot, the emotion trajectory being used to record dimension change information of the user for one or more of the preset emotion dimensions; the emotion snapshot including a round associated with the voice navigation request, a timestamp, and an emotion label distribution;

[0028] perform a weighted summation operation on all target emotion snapshots included in the emotion trajectory according to a preset emotion accumulation algorithm, to obtain a weighted summation result for the emotion trajectory; the weighted summation result including an emotion score value for one or more of the preset emotion dimensions;

[0029] The emotion snapshot, the updated emotion track, and the weighted sum result are determined as a second voice processing result.

[0030] As an optional implementation, in the first aspect of the present application, the determining and jumping operation for the target intent node according to the first voice processing result and the second voice processing result to obtain a jumping result for the target intent node comprises:

[0031] According to the voice navigation intent, an intent state jump graph is constructed, which includes a state node corresponding to the current location and at least one intent node corresponding to the voice navigation intent; a node edge is set between the state node and each intent node;

[0032] According to the keyword set, an edge attribute configuration operation is performed on each node edge to update the intent state jump graph;

[0033] According to the fusion result, in combination with the intent state jump graph, a comprehensive jump score corresponding to each intent node is calculated; according to the comprehensive jump score corresponding to each intent node, a target intent node to be jumped is determined from all the intent nodes, and the state node is jumped to the target intent node as a jumping result for the target intent node.

[0034] As an optional implementation, in the first aspect of the present application, each main intent category corresponds to a category probability; each keyword corresponds to a vocabulary probability;

[0035] The fusion operation on the voice navigation intent and the keyword set according to a preset fusion module to obtain a fusion result for the voice navigation intent and the keyword set comprises:

[0036] Using a preset probability threshold as a data screening reference, a screening operation is performed on the voice navigation intent whose category probability is lower than the probability threshold, and the keyword set whose vocabulary probability is lower than the probability threshold, to update the voice navigation intent and the keyword set;

[0037] According to a preset triple data structure, a first triple corresponding to the voice navigation intent and a second triple corresponding to the keyword set are respectively constructed; the first triple includes a first sub-group corresponding to each main intent category; the second triple includes a second sub-group corresponding to each keyword;

[0038] According to a preset voting mechanism, a first voting operation is performed on all the first subgroups, and a second voting operation is performed on all the second subgroups, to obtain a first voting result corresponding to all the first subgroups and a second voting result corresponding to all the second subgroups;

[0039] According to the first module weight set for the voice navigation intention and the second module weight set for the keyword set, a weighted calculation operation and a semantic correction operation are respectively performed on the first voting result and the second voting result, to obtain a fusion result of the first voting result and the second voting result as a fusion result of the voice navigation intention and the keyword set.

[0040] As an optional implementation, in the first aspect of the present application, the method further comprises:

[0041] According to the intention state jump graph, in combination with the state node and the target intention node, it is judged whether the current satisfies the supplementary question condition, when it is judged that the current satisfies the supplementary question condition, the supplementary question information for the voice navigation request is generated, and the supplementary question information is fed back to the user to prompt the user to feed back a new voice navigation request;

[0042] When it is judged that the current does not satisfy the supplementary question condition, according to the state node, the target intention node, the jump path corresponding to the state node and the target intention node, the user jump track recorded for the user is updated, and the user jump track is used to perform a path rollback operation for the jump path according to the path rollback demand when detecting the path rollback demand of the user;

[0043] Wherein, satisfying the supplementary question condition includes:

[0044] There is no legal jump path between the state node and the target intention node, or the user current emotion value calculated for the user in the second voice processing result is lower than the node jump threshold value set for the two nodes of the state node and the target intention node.

[0045] The second aspect of the present application discloses a multi-service voice navigation device based on voice intention tracking micro agent, the device comprises:

[0046] The voice recognition module is used for performing a voice recognition operation on the voice navigation request when detecting a voice navigation request triggered by the user, to obtain a recognized text for the voice navigation request as a voice recognition result;

[0047] The first speech processing module is configured to perform an intent recognition operation and a keyword extraction operation on the speech recognition result respectively, to obtain a speech navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first speech processing result.

[0048] The second speech processing module is configured to perform an emotion recognition operation and an emotion score calculation operation on the first speech processing result respectively, to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second speech processing result.

[0049] The node jump module is configured to perform a determination and a jump operation for a target intent node according to the first speech processing result and the second speech processing result, to obtain a jump result for the target intent node as a final response result responding to the speech navigation request.

[0050] As an optional implementation, in the second aspect of the present application, the manner in which the speech recognition module performs a speech recognition operation on the speech navigation request to obtain a recognized text for the speech navigation request specifically comprises:

[0051] Obtaining speech stream data corresponding to the speech navigation request;

[0052] Performing a speech feature extraction operation on the speech stream data according to a preset signal processing algorithm to obtain a speech feature for the speech stream data;

[0053] Determining a speech text recognition model matched with the speech stream data;

[0054] Inputting the speech feature into the speech text recognition model to sequentially perform a text reasoning operation and a text output operation on the speech feature by the speech text recognition model to obtain an output text for the speech feature;

[0055] Performing a text post-processing operation on the output text to obtain a recognized text for the speech navigation request, wherein the text post-processing operation is configured to correct the output text.

[0056] As an optional implementation, in the second aspect of the present application, the manner in which the first speech processing module performs an intent recognition operation and a keyword extraction operation on the speech recognition result to obtain a speech navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first speech processing result specifically comprises:

[0057] Perform a text preprocessing operation on the speech recognition result, and generate a semantic vector embedding for the text preprocessing result; the text preprocessing operation includes a text content cleaning operation and a word segmentation operation;

[0058] According to a preset large language model, in combination with the context window content of the semantic vector embedding, perform an intent recognition operation on the semantic vector embedding to obtain a voice navigation intent for the semantic vector embedding, the voice navigation intent including at least one main intent category, all of the main intent categories including at least one of query, reservation, complaint, and request for human;

[0059] According to a preset keyword recognition model, in combination with the historical keywords recognized for the semantic vector embedding, perform a keyword extraction operation on the semantic vector embedding to obtain a keyword set for the semantic vector embedding, the keyword set including at least one keyword;

[0060] According to a preset fusion module, perform a fusion operation on the voice navigation intent and the keyword set to obtain a fusion result for the voice navigation intent and the keyword set as a first voice processing result.

[0061] As an optional implementation, in the second aspect of the application, the voice features include at least one of pitch, speech rate, and sound intensity;

[0062] The second voice processing module performs an emotion recognition operation and an emotion score calculation operation on the first voice processing result, respectively, to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result. The manner includes:

[0063] According to a text emotion recognition model, perform an emotion recognition operation on the recognized text and the voice features to obtain an emotion recognition result for the recognized text and the voice features; the emotion recognition result includes a confidence level for a plurality of preset emotion dimensions;

[0064] Generate an emotion snapshot for the emotion recognition result, and update an emotion trajectory for the user according to the emotion snapshot, the emotion trajectory being used to record dimension change information of the user for one or more of the preset emotion dimensions; the emotion snapshot includes a round associated with the voice navigation request, a timestamp, and an emotion label distribution;

[0065] According to a preset emotion accumulation algorithm, perform a weighted summation operation on all target emotion snapshots included in the emotion trajectory to obtain a weighted summation result for the emotion trajectory; the weighted summation result includes an emotion score value for one or more of the preset emotion dimensions;

[0066] The emotion snapshot, the updated emotion track, and the weighted sum result are determined as a second speech processing result.

[0067] As an optional implementation, in the second aspect of the present application, the node jumping module performs determination and jumping operation on a target intent node according to the first speech processing result and the second speech processing result, and obtains a jumping result of the target intent node in the following manner:

[0068] An intent state jumping graph is constructed according to the speech navigation intent, the intent state jumping graph including a state node corresponding to a current location and at least one intent node corresponding to the speech navigation intent, and a node edge being arranged between the state node and each intent node;

[0069] According to the keyword set, an edge attribute configuration operation is performed on each node edge to update the intent state jumping graph;

[0070] According to the fusion result, a comprehensive jumping score corresponding to each intent node is calculated in combination with the intent state jumping graph, a target intent node to be jumped is determined from all the intent nodes according to the comprehensive jumping score corresponding to each intent node, and the state node is jumped to the target intent node as a jumping result of the target intent node.

[0071] As an optional implementation, in the second aspect of the present application, each main intent category corresponds to a category probability, and each keyword corresponds to a vocabulary probability.

[0072] The first speech processing module performs fusion operation on the speech navigation intent and the keyword set according to a preset fusion module to obtain a fusion result of the speech navigation intent and the keyword set in the following manner:

[0073] The data to be screened out, in which the category probability of the speech navigation intent is lower than a preset probability threshold and the vocabulary probability of the keyword set is lower than the probability threshold, is subjected to a screening operation with the probability threshold as a data screening benchmark, so as to update the speech navigation intent and the keyword set;

[0074] A first triple corresponding to the speech navigation intent and a second triple corresponding to the keyword set are respectively constructed according to a preset triple data structure, the first triple including a first sub-group corresponding to each main intent category, and the second triple including a second sub-group corresponding to each keyword;

[0075] According to a preset voting mechanism, a first voting operation is performed on all the first subgroups, and a second voting operation is performed on all the second subgroups, to obtain a first voting result corresponding to all the first subgroups and a second voting result corresponding to all the second subgroups;

[0076] According to the first module weight set for the voice navigation intention and the second module weight set for the keyword set, a weighted calculation operation and a semantic correction operation are respectively performed on the first voting result and the second voting result, to obtain a fusion result of the first voting result and the second voting result as a fusion result of the voice navigation intention and the keyword set.

[0077] As an optional implementation, in the second aspect of the present application, the device further comprises:

[0078] A judgment module is configured to judge whether a supplementary question condition is met according to the intention state jump graph, in combination with the state node and the target intention node;

[0079] A supplementary question module is configured to generate supplementary question information for the voice navigation request and feed back the supplementary question information to the user to prompt the user to feed back a new voice navigation request when the judgment module judges that the supplementary question condition is met.

[0080] An update module is configured to update a user jump track recorded for the user according to the state node, the target intention node, and a jump path corresponding to the state node and the target intention node when the judgment module judges that the supplementary question condition is not met, the user jump track being used to perform a path backtracking operation for the jump path according to a path backtracking demand of the user.

[0081] The supplementary question condition is met when:

[0082] There is no legal jump path between the state node and the target intention node, or a user current emotion value calculated for the user in the second voice processing result is lower than a node jump threshold value set for the two nodes of the state node and the target intention node.

[0083] The third aspect of the present application discloses another multi-service voice navigation device based on a voice intention tracking micro-agent, which comprises:

[0084] A memory storing executable program codes;

[0085] A processor coupled with the memory;

[0086] The processor invokes the executable program code stored in the memory to execute part or all of the steps in the multi-service voice navigation method based on voice intention tracking micro-agents according to any one of the first aspect of the present application.

[0087] The fourth aspect of the present application discloses a computer storage medium storing computer instructions, which are invoked to execute part or all of the steps in the multi-service voice navigation method based on voice intention tracking micro-agents according to any one of the first aspect of the present application.

[0088] Compared with the prior art, the present application has the following beneficial effects:

[0089] In the embodiment of the present application, a multi-service voice navigation method based on voice intention tracking micro-agents is provided, which includes: when a voice navigation request triggered by a user is detected, performing a voice recognition operation on the voice navigation request to obtain a recognition text for the voice navigation request as a voice recognition result; performing an intention recognition operation and a keyword extraction operation on the voice recognition result to obtain a voice navigation intention for the intention recognition operation and a keyword set for the keyword extraction operation as a first voice processing result; performing an emotion recognition operation and an emotion score calculation operation on the first voice processing result to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result; and performing a determination and a jump operation for a target intention node according to the first voice processing result and the second voice processing result to obtain a jump result for the target intention node as a final response result for the response voice navigation request. As can be seen, by performing voice recognition on a voice navigation request to obtain a recognition text, and then performing intention recognition and keyword extraction, the core intention and key information in the user's voice request can be accurately grasped, navigation errors caused by understanding bias can be effectively avoided, and the accuracy of subsequent voice navigation can be improved. Emotion recognition and emotion score calculation operations are also performed on the first voice processing result to realize intelligent perception of the user's current emotional state. Based on this, the user's emotional factors can be considered when determining the target intention node and jumping, more personalized and intelligent interaction is realized, the efficiency and accuracy of guiding the user to enter the corresponding business process are improved, the waiting time and operation steps of the user in voice navigation are reduced, and thus the processing efficiency of multi-service voice navigation is improved and the experience and satisfaction of the user in the voice navigation process are improved. BRIEF DESCRIPTION OF DRAWINGS

[0090] In order to make the technical solutions in the embodiments of the present application clearer, the accompanying drawings needed in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art without any creative work on the basis of these drawings are within the protection scope of the present application.

[0091] Figure 1 is a flow diagram of a multi-service voice navigation method based on a voice intention tracking micro-agent according to an embodiment of the present application;

[0092] Figure 2 is a flow diagram of another multi-service voice navigation method based on a voice intention tracking micro-agent according to an embodiment of the present application;

[0093] Figure 3 is a structural diagram of a multi-service voice navigation device based on a voice intention tracking micro-agent according to an embodiment of the present application;

[0094] Figure 4 is a structural diagram of another multi-service voice navigation device based on a voice intention tracking micro-agent according to an embodiment of the present application;

[0095] Figure 5 is a structural diagram of another multi-service voice navigation device based on a voice intention tracking micro-agent according to an embodiment of the present application. DETAILED DESCRIPTION

[0096] In order to make the technical solutions in the embodiments of the present application clearer, the accompanying drawings needed in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art without any creative work on the basis of these drawings are within the protection scope of the present application.

[0097] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned accompanying drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or end.

[0098] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. As will be apparent to those of ordinary skill in the art, embodiments described herein can be combinable with other embodiments.

[0099] The application discloses a multi-service voice navigation method and device based on a voice intention tracking micro-agent, which can accurately grasp the core intention and key information in the user voice request by performing voice recognition on the voice navigation request, performing intention recognition and keyword extraction, effectively avoiding navigation errors caused by understanding deviation, and being beneficial to improving the accuracy of subsequent voice navigation; can also perform emotion recognition and emotion score calculation operations on the first voice processing result, realizing intelligent perception of the current emotional state of the user; based on this, the user emotional factors can be considered when determining the target intention node and jumping, realizing more personalized and intelligent interaction, improving the efficiency and accuracy of guiding the user to enter the corresponding business process, reducing the waiting time and operation steps of the user in the voice navigation, thereby improving the processing efficiency of multi-service voice navigation and improving the experience and satisfaction of the user in the voice navigation process. The following will be described in detail.

[0100] Embodiment one

[0101] Please refer to Figure 1 , Figure 1 is a flowchart of a multi-service voice navigation method based on a voice intention tracking micro-agent disclosed by the embodiments of the application. Among them, Figure 1 The multi-service voice navigation method based on a voice intention tracking micro-agent described above can be applied to a multi-service voice navigation device based on a voice intention tracking micro-agent, and the embodiments of the application are not limited. As shown in Figure 1 The multi-service voice navigation method based on a voice intention tracking micro-agent can include the following operations:

[0102] 101、When detecting a voice navigation request triggered by a user, performing a voice recognition operation on the voice navigation request to obtain a recognition text for the voice navigation request as a voice recognition result.

[0103] 102、Performing an intention recognition operation and a keyword extraction operation on the voice recognition result to obtain a voice navigation intention for the intention recognition operation and a keyword set for the keyword extraction operation as a first voice processing result.

[0104] 103. Perform emotion recognition and emotion score calculation operations on the first speech processing result to obtain the target recognized emotion for the emotion recognition operation and the emotion score value for the emotion score calculation operation, which are used as the second speech processing result.

[0105] 104. Based on the first speech processing result and the second speech processing result, perform the determination and jump operation for the target intent node to obtain the jump result for the target intent node, which serves as the final response result to the voice navigation request.

[0106] It is evident that implementation Figure 1 The described multi-service voice navigation method based on a micro-agent for voice intent tracking accurately grasps the core intent and key information in a user's voice request by performing speech recognition on the navigation request, obtaining the recognized text, and then performing intent recognition and keyword extraction. This effectively avoids navigation errors caused by misunderstandings and improves the accuracy of subsequent voice navigation. Furthermore, it performs emotion recognition and emotion scoring on the first speech processing result, enabling intelligent perception of the user's current emotional state. Based on this, user emotional factors can be considered when determining target intent nodes and jumps, achieving a more humanized and intelligent interaction. This improves the efficiency and accuracy of guiding users into corresponding business processes while reducing user waiting time and operation steps during voice navigation, thereby improving the processing efficiency of multi-service voice navigation and enhancing the user experience and satisfaction.

[0107] In an optional embodiment, step 101 above performs a speech recognition operation on the voice navigation request to obtain the recognized text for the voice navigation request, specifically including:

[0108] Obtain the voice stream data corresponding to the voice navigation request;

[0109] Based on a preset signal processing algorithm, a speech feature extraction operation is performed on the speech stream data to obtain speech features specific to the speech stream data.

[0110] Determine the speech-to-text recognition model that matches the speech stream data;

[0111] The speech features are input into the speech-text recognition model, and the speech-text recognition model performs text inference and text output operations on the speech features in sequence to obtain the output text for the speech features.

[0112] Perform text post-processing on the output text to obtain the recognized text for the voice navigation request. The text post-processing operation is used to correct the output text.

[0113] In the optional embodiment, the preset signal processing algorithm can realize voice feature extraction based on a conventional spectrogram or a mel-frequency cepstral coefficient (MFCC).

[0114] In the optional embodiment, performing voice feature extraction on the voice stream data according to the preset signal processing algorithm can accurately capture key feature information in the voice stream data, remove noise and other interference factors, and avoid recognition errors caused by subsequent quality problems of the voice signal.

[0115] In the optional embodiment, different voice scenes, accents, etc. can have different characteristics, and the matched model can better adapt to these differences to improve the recognition accuracy and adaptability of various voice data. Specifically, the voice text recognition model can use existing general Whisper, Wav2Vec 2.0, etc. multilingual pre-training model, which can support real-time transcription of Mandarin, dialect, and mixed sentences.

[0116] As can be seen, in the optional embodiment, by performing voice feature extraction on the voice stream data to remove noise interference and improve the quality of the voice stream data, by matching the voice text recognition model to improve the recognition accuracy of the voice stream data, by performing text reasoning and text output on the voice features based on the voice text recognition model to improve the conversion efficiency of the output text, and by performing text post-processing on the output text to further improve the quality of the output text.

[0117] In another optional embodiment, the manner in which the step 102 performs an intent recognition operation and a keyword extraction operation on the voice recognition result to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as the first voice processing result specifically includes:

[0118] performing a text preprocessing operation on the voice recognition result and generating a semantic vector embedding for the text preprocessing result; the text preprocessing operation includes a text content cleaning operation and a word segmentation operation;

[0119] performing an intent recognition operation on the semantic vector embedding according to a preset large language model in combination with context window content of the semantic vector embedding to obtain a voice navigation intent for the semantic vector embedding, the voice navigation intent including at least one main intent category, all main intent categories including at least one of query, reservation, complaint, and request for human;

[0120] performing a keyword extraction operation on the semantic vector embedding according to a preset keyword recognition model in combination with historical keywords recognized for the semantic vector embedding to obtain a keyword set for the semantic vector embedding, the keyword set including at least one keyword;

[0121] According to the preset fusion module, a fusion operation is performed on the voice navigation intention and the keyword set to obtain a fusion result for the voice navigation intention and the keyword set as the first voice processing result.

[0122] In this optional embodiment, text content cleaning and word segmentation operations are performed on the voice recognition result, which can effectively remove noise information in the text, such as special symbols, irrelevant characters, etc., and the text can be reasonably divided into meaningful word units by using the BERT tokenizer. On this basis, a pre-trained language model (such as BERT or ChatGLM) can be used to generate semantic vector embedding to convert the text into a numerical vector form that can be efficiently processed by a computer, accurately capturing the semantic information of the text and providing a reliable semantic basis for subsequent accurate intent recognition and keyword extraction.

[0123] In this optional embodiment, the above-mentioned preset large language model can use ChatGLM. In addition, the context window content of the semantic vector embedding can be the context window content before the current dialogue turn, such as the past 2-3 rounds of user / system dialogue text. By fusing the historical round content (the context window content), it can avoid misjudgment of information due to the current user's too brief expression, and also achieve accurate information recognition when the user omits information or the semantic is ambiguous.

[0124] In this optional embodiment, the above-mentioned keyword recognition model can use the BERT-CRF or BERT + SpanExtraction model. In addition, the above-mentioned recognized keyword set can be classified and recognized based on information such as time, place, object, service name, etc.

[0125] In this optional embodiment, if the current round does not contain keywords, but the voice navigation intention points to a specific operation (such as "I want to order food" but does not specify the location), the keyword completion mechanism is triggered, which queries historical keywords by calling and referencing historical keywords with high frequency, high confidence and strong semantic association in the previous text. Among them, the relevance between the current semantic and the historical keywords can be identified by attention mechanism or word embedding similarity, so as to determine whether to reuse or prompt for additional questions.

[0126] In this optional embodiment, the keyword extraction operation is performed by using the preset keyword recognition model in combination with the historical keywords identified by the semantic vector embedding. By combining historical information, the keyword recognition model can intelligently identify the key information in the current voice navigation request according to the association between the context and the historical keywords, which is beneficial to improve the accuracy and pertinence of keyword extraction.

[0127] It can be seen that in the optional embodiment, through accurate text preprocessing and semantic representation, efficient and accurate intent recognition, intelligent adaptive keyword extraction, and deep fusion processing, the accuracy, pertinence, and comprehensiveness of intent recognition and keyword extraction in the voice navigation request are significantly improved, and high-quality first voice processing results are provided for subsequent voice navigation processes.

[0128] In yet another optional embodiment, the voice features include at least one of pitch, speech rate, and sound intensity.

[0129] The manner of performing the emotion recognition operation and the emotion score calculation operation on the first voice processing result respectively to obtain the target recognition emotion for the emotion recognition operation and the emotion score value for the emotion score calculation operation as the second voice processing result specifically includes:

[0130] According to the text emotion recognition model, an emotion recognition operation is performed on the recognized text and the voice features to obtain an emotion recognition result for the recognized text and the voice features; the emotion recognition result includes confidence for a plurality of preset emotion dimensions.

[0131] An emotion snapshot for the emotion recognition result is generated, and an emotion trajectory for the user is updated according to the emotion snapshot; the emotion trajectory is used to record dimension change information of the user for one or more preset emotion dimensions; the emotion snapshot includes a round associated with the voice navigation request, a timestamp, and an emotion label distribution.

[0132] According to a preset emotion accumulation algorithm, a weighted summation operation is performed on all target emotion snapshots included in the emotion trajectory to obtain a weighted summation result for the emotion trajectory; the weighted summation result includes an emotion score value for one or more preset emotion dimensions.

[0133] The emotion snapshot, the updated emotion trajectory, and the weighted summation result are determined as the second voice processing result.

[0134] In the optional embodiment, the voice features such as pitch, speech rate, and sound intensity can reflect the emotional state of the user from the voice level, for example, a rising pitch may indicate excitement, and a faster speech rate may mean nervousness; and the recognized text provides emotional clues from the semantic perspective, and the combination of the two provides a more comprehensive and accurate basis for emotion recognition, which is conducive to improving the reliability of the emotion recognition result.

[0135] In the optional embodiment, the plurality of preset emotion dimensions can be emotion dimensions such as anger, complaint, anxiety, positive emotion, neutral emotion, and negative emotion. The confidence for the preset emotion dimension is used to indicate the probability that the user is currently in the preset emotion dimension, and the value range of the confidence is [0, 1], that is, the higher the confidence of a certain preset emotion dimension, the greater the probability that the user is in the certain preset emotion dimension. Optionally, the confidence can also be used to indicate the emotional intensity of the user in the preset emotion dimension, that is, the higher the confidence of a certain preset emotion dimension, the greater the emotional intensity of the user in the certain preset emotion dimension, for example, for the same anger emotion, a confidence of 0.8 indicates that the user's anger emotion is high, and a confidence of 0.3 indicates that the user's anger emotion is general.

[0136] In the optional embodiment, the emotion snapshot can be generated according to the dialogue turn, that is, one emotion snapshot is generated for one round of dialogue. Then, the generation of the emotion snapshot record associated with the turn, the timestamp, and the emotion label distribution of the voice navigation request can accurately locate the time and the dialogue turn at which the emotion occurs, and clearly show the performance of the emotion in the specific interaction scenario. At the same time, the user emotion trajectory is updated according to the emotion snapshot, and the dimension change information of the user for a plurality of preset emotion dimensions is recorded, so that the evolution process of the user emotion can be dynamically tracked, and the fluctuation of the user emotion at different stages can be deeply understood, for example, if a certain emotion continuously appears and intensifies (for example, anger from 0.4→0.6→0.75), an emotion trend marker is triggered.

[0137] In the optional embodiment, the weighted sum operation performed on all target emotion snapshots included in the emotion trajectory according to the preset emotion accumulation algorithm can be as follows:

[0138] For each preset emotion dimension of each target emotion snapshot, the product of the confidence corresponding to the preset emotion dimension in the target emotion snapshot, the turn weight corresponding to the target emotion snapshot, and the preset dimension weight corresponding to the preset emotion dimension is obtained.

[0139] For each preset emotion dimension, the sum of all product results corresponding to the preset emotion dimension is calculated to obtain the weighted sum result for the preset emotion dimension.

[0140] Optionally, the weighted sum result corresponding to each preset emotion dimension can also be normalized.

[0141] In the optional embodiment, the weighted sum operation performed above takes into account the importance and influence degree of different emotion snapshots in the emotion trajectory, so that the calculated emotion score can more scientifically and accurately reflect the overall emotional state of the user, and avoid one-sidedness of a single emotion snapshot.

[0142] It can be seen that in this optional embodiment, by performing emotion recognition on multi-dimensional data, emotion dynamic tracking based on emotion snapshots and emotion trajectories, and scientific and accurate emotion score calculation, the accuracy, comprehensiveness and intelligence level of emotion recognition and processing in the voice navigation system are significantly improved.

[0143] In another optional embodiment, the manner of performing determination and jump operation for the target intent node according to the first voice processing result and the second voice processing result to obtain the jump result for the target intent node specifically comprises:

[0144] An intent state jump graph is constructed according to the voice navigation intent, the intent state jump graph comprising a state node corresponding to the current position and at least one intent node corresponding to the voice navigation intent, and a node edge being arranged between the state node and each intent node;

[0145] According to the keyword set, an edge attribute configuration operation is performed on each node edge to update the intent state jump graph;

[0146] According to the fusion result, a comprehensive jump score corresponding to each intent node is calculated in combination with the intent state jump graph, a target intent node to be jumped is determined from all intent nodes according to the comprehensive jump score corresponding to each intent node, and the state node is jumped to the target intent node as the jump result for the target intent node.

[0147] In this optional embodiment, the node edge arranged between the state node and each intent node is used to indicate the possibility / probability of the user currently jumping from the state node to each intent node. Further, each node edge is provided with a plurality of edge configuration parameters, the plurality of edge configuration parameters comprising an intent keyword weight corresponding to the keyword set, an emotion adjustment weight, and a jump confidence threshold.

[0148] In this optional embodiment, the intent keyword weight is used to represent semantic similarity of jump, for example, “complaint”→“consultation”, and if the semantics is not smooth, the corresponding intent keyword weight is low. Optionally, if the user continuously expresses keywords close to the target intent node, the intent keyword weight can be increased. Optionally, if there is intent ambiguity or cross-domain jump such as “consultation”→“complaint”, similar to the above processing of semantics not being smooth, the intent keyword weight can also be reduced.

[0149] In this optional embodiment, the setting of the emotion adjustment weight increases the consideration of the influence factor of emotion on node jump. For example, if it is identified that the user is in an anxious / angry emotion, the weight of jumping to the “human transfer” or “calming module” is increased. If the user is in a positive / stable emotion, the emotion adjustment weight can be reduced.

[0150] In the optional embodiment, the setting of the jump confidence threshold is used to control whether the jump is triggered to avoid false jumps. Specifically, the comprehensive jump score calculated for the target intent node needs to be higher than the jump confidence threshold set for the node edge corresponding to the state node and the target intent node, so that the jump operation from the state node to the target intent node can be performed.

[0151] In the optional embodiment, on the basis of the setting of the plurality of edge configuration parameters, the calculation formula corresponding to the comprehensive jump score of each intent node can be: comprehensive jump score = intent similarity x intent keyword weight + emotion value x emotion adjustment weight.

[0152] In the optional embodiment, by constructing the intent state jump graph, the jump relationship between different intents is intuitively presented in a graphical representation, which provides a clear structural framework for subsequent jump operations, and is conducive to improving the efficiency of processing various voice navigation intents.

[0153] In the optional embodiment, the keyword set reflects the key information in the user's voice, and by associating it with the node edge attribute, the jump conditions and weights between different intent nodes can be more accurately defined.

[0154] As can be seen, in the optional embodiment, by constructing an intuitive intent jump framework, accurately configuring node edge attributes, scientifically calculating comprehensive jump scores, and efficiently determining and executing node jumps, the accuracy, scientificity and efficiency of target intent node determination and jump in the voice navigation system are significantly improved, that is, the running smoothness of voice navigation is improved, and the user's voice navigation experience is improved.

[0155] Embodiment Two

[0156] Please refer to Figure 2 , Figure 2 is another flowchart of a multi-service voice navigation method based on a voice intent tracking micro-agent disclosed in the embodiments of the present application. Among them, Figure 2 The multi-service voice navigation method based on a voice intent tracking micro-agent described can be applied to a multi-service voice navigation device based on a voice intent tracking micro-agent, and the embodiments of the present application are not limited. As Figure 2 shown, the multi-service voice navigation method based on a voice intent tracking micro-agent can include the following operations:

[0157] 201、When a voice navigation request triggered by a user is detected, a speech recognition operation is performed on the voice navigation request to obtain a recognized text for the voice navigation request as a speech recognition result.

[0158] 202. Perform intent recognition and keyword extraction operations on the speech recognition results to obtain the speech navigation intent for intent recognition and the keyword set for keyword extraction, which are used as the first speech processing result.

[0159] 203. Perform emotion recognition operation and emotion score calculation operation on the first speech processing result respectively to obtain the target recognized emotion for the emotion recognition operation and the emotion score value for the emotion score calculation operation, which are used as the second speech processing result.

[0160] 204. Based on the first speech processing result and the second speech processing result, perform the determination and jump operation for the target intent node to obtain the jump result for the target intent node, which serves as the final response result to the voice navigation request.

[0161] For further descriptions of steps 201-204 in this embodiment of the invention, please refer to the other specific descriptions of steps 101-104 in Embodiment 1. These descriptions will not be repeated in this embodiment of the invention.

[0162] 205. Based on the intent state transition diagram, and combining the state nodes and the target intent node, determine whether the current condition for supplementing the question is met.

[0163] In this embodiment of the invention, when the judgment result of step 205 is yes, step 206 is triggered; when the judgment result of step 205 is no, step 207 is triggered.

[0164] 206. Generate supplementary questions for the voice navigation request and send the supplementary questions back to the user to prompt the user to submit a new voice navigation request.

[0165] 207. Update the user jump trajectory recorded for the user based on the status node, target intent node, and the jump path corresponding to the status node and target intent node.

[0166] In this embodiment of the invention, the user jump trajectory is used to perform a path rollback operation on the jump path according to the path rollback requirement when the user's path rollback requirement is detected.

[0167] In this embodiment of the invention, the conditions for satisfying the supplementary question include: there is no legal jump path between the state node and the target intent node, or the current emotion value of the user calculated for the user in the second speech processing result is lower than the node jump threshold set for the two nodes, the state node and the target intent node.

[0168] In the embodiment of the present application, the condition of supplementing the question also includes that the deviation value between the jump path corresponding to the state node and the target intent node and the recommended jump path recorded in the preset recommended service process is greater than the preset deviation threshold. For example, the jump path is A→B→D; the originally predetermined recommended jump path is A→B→C→D. At this time, there is a deviation of 1 node between the two paths, so it can be determined that the condition of supplementing the question is met, and in this case, the user can be asked whether to jump back to the missing node.

[0169] In the embodiment of the present application, when there is no legal jump path between the state node and the target intent node, it means that the system cannot directly process the user's request according to the existing logic, at this time, the supplement question information is generated and fed back to the user, which can guide the user to provide more explicit or related information, and help the system accurately understand the user's intent.

[0170] In the embodiment of the present application, when the user's current emotion value in the second voice processing result is lower than the node jump threshold, it indicates that the user may be in a state of confusion, dissatisfaction, and other emotions that are not conducive to normal interaction. The supplement question information can further communicate with the user to understand their real needs, avoid interaction failure due to user emotional problems, and improve the accuracy and success rate of interaction.

[0171] In the embodiment of the present application, for the case that does not meet the condition of supplementing the question, by performing the operation of the above-mentioned step 207, the path information of the user in the voice navigation process can be automatically recorded, providing a basis for the possible path rollback demand in the future. When detecting that the user has a path rollback demand, the system can quickly perform the path rollback operation according to the recorded user jump trajectory, so that the user can conveniently return to the previous interaction node to reselect or operate, enhancing the flexibility of the system and the convenience of user interaction, and being conducive to improving the user experience.

[0172] It can be seen that the implementation Figure 2 The multi-service voice navigation method based on the voice intent tracking micro-agent described above realizes accurate supplement questions by means of the intent state jump graph, guides the user to clearly express the demand when normal jump is not possible or the user's emotion is not good, which is conducive to improving the interaction accuracy; at the same time, it can also realize intelligent updating of the user jump trajectory and support path rollback, enhancing the flexibility of the system and the convenience of user interaction, and being conducive to improving the user experience and interaction effect of multi-service voice navigation.

[0173] In an optional embodiment, each main intent category corresponds to a category probability; each keyword corresponds to a word probability;

[0174] The manner of performing the fusion operation on the voice navigation intent and the keyword set according to the preset fusion module to obtain the fusion result for the voice navigation intent and the keyword set specifically includes:

[0175] The preset probability threshold is used as a data screening criterion. The data to be screened, whose category probability in the voice navigation intent is lower than the probability threshold and whose keyword probability in the keyword set is lower than the probability threshold, is subjected to a screening operation, so as to update the voice navigation intent and update the keyword set.

[0176] According to a preset ternary data structure, a first ternary group corresponding to the voice navigation intent and a second ternary group corresponding to the keyword set are constructed respectively. The first ternary group includes a first sub-group corresponding to each main intent category. The second ternary group includes a second sub-group corresponding to each keyword.

[0177] According to a preset voting mechanism, a first voting operation is performed on all the first sub-groups, and a second voting operation is performed on all the second sub-groups, so as to obtain a first voting result corresponding to all the first sub-groups and a second voting result corresponding to all the second sub-groups.

[0178] According to a first module weight set for the voice navigation intent and a second module weight set for the keyword set, a weighted calculation operation and a semantic correction operation are respectively performed on the first voting result and the second voting result, so as to obtain a fusion result of the first voting result and the second voting result as a fusion result of the voice navigation intent and the keyword set.

[0179] In this optional embodiment, by setting the probability threshold, the main intent category whose category probability in the voice navigation intent is lower than the threshold and the keyword whose keyword probability in the keyword set is lower than the threshold are screened out. This operation effectively removes low-probability and low-reliability data, reduces noise interference, and makes the data for subsequent processing more accurate and representative.

[0180] In this optional embodiment, the ternary data structure can be (semantic type, candidate value, confidence). In the first ternary group, the semantic type corresponds to the voice navigation intent, and in the second ternary group, the semantic type corresponds to the keyword set. Specifically, in actual application, the semantic type can be "intent", "keyword", and optionally, the semantic type can also be "emotion". In the first ternary group, the candidate value is a specific main intent category corresponding to the voice navigation intent, and the confidence is the category probability of the main intent category. In the second ternary group, the candidate value is a specific keyword corresponding to the keyword set, and the confidence is the keyword probability of the keyword. If the semantic type includes "emotion", in the third ternary group corresponding to "emotion", the candidate value is a specific preset emotion dimension, and the confidence is the emotion probability of the preset emotion dimension. Specifically, a certain first sub-group in the first ternary group can be (intent, unsubscribe, 0.9). A certain second sub-group in the second ternary group can be (keyword, future time, 0.85). A certain third sub-group in the third ternary group can be (emotion, dissatisfaction, 0.75).

[0181] In the optional embodiment, the information of multiple subgroups can be integrated through a voting mechanism, and a more reasonable decision can be made according to the majority principle or specific rules, which is beneficial to enhancing the accuracy and stability of the system in judging the voice navigation intent and keywords.

[0182] In the optional embodiment, the above-mentioned weighted fusion method fully considers the different importance of the voice navigation intent and keyword set in the overall judgment, and further optimizes the result through semantic correction, so that the final fusion result is more in line with the actual semantics and business requirements.

[0183] As can be seen, in the optional embodiment, noise interference is reduced through data screening, a standard data structure is constructed using triplets, reasonable decisions are made based on a voting mechanism, and the importance of different factors is comprehensively considered by using weighted semantic fusion, thereby effectively improving the accuracy, stability and rationality of the judgment of voice intent and keywords in multi-service voice navigation.

[0184] Embodiment Three

[0185] Please refer to Figure 3 , Figure 3 is a structural schematic diagram of a multi-service voice navigation device based on a voice intent tracking micro-agent disclosed by the embodiments of the present application. The multi-service voice navigation device based on the voice intent tracking micro-agent can be a multi-service voice navigation terminal, device, system or server based on the voice intent tracking micro-agent. The server can be a local server, a remote server, or a cloud server (also known as a cloud server). When the server is a non-cloud server, the non-cloud server can be in communication connection with the cloud server, and the embodiments of the present application do not limit it. As shown in Figure 3 The multi-service voice navigation device based on the voice intent tracking micro-agent can include a voice recognition module 301, a first voice processing module 302, a second voice processing module 303, and a node jump module 304, wherein:

[0186] The voice recognition module 301 is configured to perform a voice recognition operation on the voice navigation request when detecting a voice navigation request triggered by a user, and obtain a recognition text for the voice navigation request as a voice recognition result.

[0187] The first voice processing module 302 is configured to perform an intent recognition operation and a keyword extraction operation on the voice recognition result, and obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first voice processing result.

[0188] The second speech processing module 303 is used to perform emotion recognition operation and emotion score calculation operation on the first speech processing result respectively, and obtain the target recognized emotion for the emotion recognition operation and the emotion score value for the emotion score calculation operation as the second speech processing result.

[0189] The node jump module 304 is used to perform the determination and jump operation for the target intent node based on the first voice processing result and the second voice processing result, and obtain the jump result for the target intent node as the final response result in response to the voice navigation request.

[0190] It is evident that implementation Figure 3 The described multi-service voice navigation device based on a micro-agent for voice intent tracking accurately grasps the core intent and key information in a user's voice request by performing speech recognition on the navigation request, obtaining the recognized text, and then performing intent recognition and keyword extraction. This effectively avoids navigation errors caused by misunderstandings and improves the accuracy of subsequent voice navigation. Furthermore, it performs emotion recognition and emotion scoring calculation on the first speech processing result, enabling intelligent perception of the user's current emotional state. Based on this, user emotional factors can be considered when determining target intent nodes and jumps, achieving a more humanized and intelligent interaction. This improves the efficiency and accuracy of guiding users into corresponding business processes while reducing user waiting time and operation steps during voice navigation, thereby improving the processing efficiency of multi-service voice navigation and enhancing the user experience and satisfaction.

[0191] In an optional embodiment, the speech recognition module 301 performs a speech recognition operation on the voice navigation request, and the specific method for obtaining the recognized text for the voice navigation request includes:

[0192] Obtain the voice stream data corresponding to the voice navigation request;

[0193] Based on a preset signal processing algorithm, a speech feature extraction operation is performed on the speech stream data to obtain speech features specific to the speech stream data.

[0194] Determine the speech-to-text recognition model that matches the speech stream data;

[0195] The speech features are input into the speech-text recognition model, and the speech-text recognition model performs text inference and text output operations on the speech features in sequence to obtain the output text for the speech features.

[0196] Perform text post-processing on the output text to obtain the recognized text for the voice navigation request. The text post-processing operation is used to correct the output text.

[0197] It can be seen that in the optional embodiment, the quality of the voice stream data is improved by performing voice feature extraction on the voice stream data to remove noise interference; the recognition accuracy of the voice stream data is improved by matching the voice text recognition model; the conversion efficiency of the output text is improved by performing text reasoning and text output on the voice features based on the voice text recognition model; and the quality of the output text is further improved by performing text post-processing operation on the output text.

[0198] In another optional embodiment, the first voice processing module 302 performs an intent recognition operation and a keyword extraction operation on the voice recognition result respectively to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as the first voice processing result. The manner specifically includes:

[0199] performing a text preprocessing operation on the voice recognition result and generating a semantic vector embedding for the text preprocessing result; the text preprocessing operation includes a text content cleaning operation and a word segmentation operation;

[0200] performing an intent recognition operation on the semantic vector embedding according to a preset large language model in combination with context window content of the semantic vector embedding to obtain a voice navigation intent for the semantic vector embedding, the voice navigation intent including at least one main intent category, all main intent categories including at least one of query, reservation, complaint, and request for human;

[0201] performing a keyword extraction operation on the semantic vector embedding according to a preset keyword recognition model in combination with historical keywords recognized for the semantic vector embedding to obtain a keyword set for the semantic vector embedding, the keyword set including at least one keyword;

[0202] performing a fusion operation on the voice navigation intent and the keyword set according to a preset fusion module to obtain a fusion result for the voice navigation intent and the keyword set as the first voice processing result.

[0203] It can be seen that in the optional embodiment, the accuracy, relevance and comprehensiveness of intent recognition and keyword extraction in the voice navigation request are significantly improved through precise text preprocessing and semantic representation, efficient and accurate intent recognition, intelligent and adaptive keyword extraction, and deep fusion processing, and a high-quality first voice processing result is provided for the subsequent voice navigation process.

[0204] In yet another optional embodiment, the voice features include at least one of pitch, speech rate, and sound intensity;

[0205] The second voice processing module 303 performs emotion recognition operation and emotion score calculation operation on the first voice processing result respectively, obtains a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as the second voice processing result, and the manner of obtaining the second voice processing result specifically includes:

[0206] According to the text emotion recognition model, perform emotion recognition operation on the recognized text and the voice feature to obtain an emotion recognition result for the recognized text and the voice feature; the emotion recognition result includes a confidence for a plurality of preset emotion dimensions;

[0207] Generate an emotion snapshot for the emotion recognition result, and update an emotion track for the user according to the emotion snapshot, the emotion track being used to record dimension change information of the user for one or more preset emotion dimensions; the emotion snapshot includes a round associated with the voice navigation request, a timestamp, and an emotion label distribution;

[0208] According to a preset emotion accumulation algorithm, perform weighted summation operation on all target emotion snapshots included in the emotion track to obtain a weighted summation result for the emotion track; the weighted summation result includes an emotion score value for one or more preset emotion dimensions;

[0209] Determine the emotion snapshot, the updated emotion track, and the weighted summation result as the second voice processing result.

[0210] As can be seen, in this optional embodiment, through emotion recognition on multi-dimensional data, emotion dynamic tracking based on emotion snapshot and emotion track, and scientific and accurate emotion score calculation, the accuracy, comprehensiveness and intelligence level of emotion recognition and processing in the voice navigation system are significantly improved.

[0211] In another optional embodiment, the node jump module 304 performs determination and jump operation on the target intent node according to the first voice processing result and the second voice processing result to obtain a jump result for the target intent node, and the manner of obtaining the jump result for the target intent node specifically includes:

[0212] Construct an intent state jump graph according to the voice navigation intent, the intent state jump graph including a state node corresponding to the current location and at least one intent node corresponding to the voice navigation intent; a node edge is set between the state node and each intent node;

[0213] According to the keyword set, perform edge attribute configuration operation on each node edge to update the intent state jump graph;

[0214] According to the fusion result, the comprehensive jump score corresponding to each intent node is calculated in combination with the intent state jump graph; the target intent node to be jumped from is determined from all intent nodes according to the comprehensive jump score corresponding to each intent node, and the state node is jumped to the target intent node as the jump result for the target intent node.

[0215] It can be seen that in this optional embodiment, by constructing an intuitive intent jump framework, accurately configuring node edge attributes, scientifically calculating comprehensive jump scores, and efficiently determining and executing node jumps, the accuracy, scientificity and efficiency of target intent node determination and jump in the voice navigation system are significantly improved, that is, the running fluency of voice navigation is improved, and the user's voice navigation experience is improved.

[0216] In yet another optional embodiment, each main intent category corresponds to a category probability; each keyword corresponds to a vocabulary probability;

[0217] The first voice processing module 302 performs a fusion operation on the voice navigation intent and the keyword set according to a preset fusion module to obtain a fusion result for the voice navigation intent and the keyword set. The manner specifically includes:

[0218] The probability threshold is used as a data screening benchmark, and a screening operation is performed on the to-be-screened data in the voice navigation intent whose category probability is lower than the probability threshold and in the keyword set whose vocabulary probability is lower than the probability threshold, so as to update the voice navigation intent and update the keyword set;

[0219] According to a preset ternary data structure, a first ternary group corresponding to the voice navigation intent and a second ternary group corresponding to the keyword set are respectively constructed; the first ternary group includes a first sub-group corresponding to each main intent category; the second ternary group includes a second sub-group corresponding to each keyword;

[0220] According to a preset voting mechanism, a first voting operation is performed on all first sub-groups, and a second voting operation is performed on all second sub-groups, to obtain a first voting result corresponding to all first sub-groups and a second voting result corresponding to all second sub-groups;

[0221] According to the first module weight set for the voice navigation intent and the second module weight set for the keyword set, a weighted calculation operation and a semantic correction operation are respectively performed on the first voting result and the second voting result to obtain a fusion result for the first voting result and the second voting result, which is used as the fusion result for the voice navigation intent and the keyword set.

[0222] It can be seen that in the optional embodiment, the noise interference is reduced by data screening, the normative data structure is constructed by using triplets, the reasonable decision is realized based on the voting mechanism, and the weighted semantic fusion is adopted to comprehensively consider the importance of different factors, so that the accuracy, stability and rationality of the judgment of the voice intention and the key word in the multi-service voice navigation are effectively improved.

[0223] In another optional embodiment, please refer to Figure 4 , Figure 4 is another structure diagram of a multi-service voice navigation device based on a voice intention tracking micro-agent disclosed by the embodiment of the present application. As shown in Figure 4 , the device further includes a judgment module 305, a follow-up question module 306 and an updating module 307, wherein:

[0224] The judgment module 305 is configured to judge whether the follow-up question condition is met according to the intention state jump graph, in combination with the state node and the target intention node.

[0225] The follow-up question module 306 is configured to generate follow-up question information for the voice navigation request and feed back the follow-up question information to the user to prompt the user to feed back a new voice navigation request when the judgment module 305 judges that the follow-up question condition is met.

[0226] The updating module 307 is configured to update the user jump track recorded for the user according to the state node, the target intention node and the jump path corresponding to the state node and the target intention node when the judgment module 305 judges that the follow-up question condition is not met, and the user jump track is configured to perform a path backtracking operation for the jump path according to the path backtracking demand when detecting the path backtracking demand of the user.

[0227] In the optional embodiment, the follow-up question condition includes:

[0228] There is no legal jump path between the state node and the target intention node, or the user current emotion value calculated for the user in the second voice processing result is lower than the node jump threshold set for the two nodes corresponding to the state node and the target intention node.

[0229] It can be seen that in the optional embodiment, the precise follow-up question is realized by means of the intention state jump graph, the user is guided to explicitly demand when the normal jump cannot be realized or the user emotion is not good, which is conducive to improving the interaction accuracy; at the same time, the user jump track can be intelligently updated and the path backtracking can be supported, the system flexibility and the user interaction convenience are enhanced, which is conducive to improving the user experience and the interaction effect of the multi-service voice navigation.

[0230] Embodiment Four

[0231] Please refer to Figure 5 , Figure 5is a structural schematic view of still another multi-service voice navigation device based on voice intention tracking micro-agent disclosed by the embodiment of the present application. As shown in Figure 5 The multi-service voice navigation device based on voice intention tracking micro-agent can include:

[0232] a memory 401 storing executable program codes;

[0233] a processor 402 coupled with the memory 401;

[0234] The processor 402 invokes the executable program codes stored in the memory 401 to execute part or all steps of any one of the multi-service voice navigation based on voice intention tracking micro-agent described in the embodiment one or the embodiment two of the present application.

[0235] Embodiment five

[0236] The embodiment of the present application discloses a computer storage medium storing computer instructions, which when invoked, is used to execute part or all steps of any one of the multi-service voice navigation based on voice intention tracking micro-agent described in the embodiment one or the embodiment two of the present application.

[0237] The device embodiments described above are only schematic, wherein the modules illustrated as separate components can or can not be physically separated, and the components illustrated as modules can or can not be physical modules, i.e. can be located in one place, or can be distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0238] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above specific description of the embodiments, and the various embodiments can be realized by means of software and necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in terms of contribution to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, including a Read-Only Memory (ROM), a Random Access Memory (RAM), a Programmable Read-only Memory (PROM), an Erasable Programmable Read Only Memory (EPROM), a One-time Programmable Read-Only Memory (OTPROM), an Electrically-Erasable Programmable Read-Only Memory (EEPROM), a Compact Disc Read-Only Memory (CD-ROM), or other optical disk storage, magnetic disk storage, magnetic tape storage, or any other medium that can be used to carry or store computer readable instructions.

[0239] Finally, it should be noted that: the above-mentioned embodiments disclosed only the preferred embodiments of the present application, only for the description of the technical solutions of the present application, and not limited; although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand; it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-service voice navigation method based on voice intent tracking micro-agents, characterized in that, The method comprises: When a user-triggered voice navigation request is detected, performing a voice recognition operation on the voice navigation request to obtain recognized text for the voice navigation request as a voice recognition result; Performing an intent recognition operation and a keyword extraction operation on the voice recognition result, respectively, to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first voice processing result; Performing an emotion recognition operation and an emotion score calculation operation on the first voice processing result, respectively, to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result; According to the first voice processing result and the second voice processing result, performing a determination and jump operation for a target intent node to obtain a jump result for the target intent node as a final response result in response to the voice navigation request; The emotion recognition operation and the emotion score calculation operation on the first voice processing result, respectively, to obtain a target recognition emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result, comprises: According to a text emotion recognition model, performing an emotion recognition operation on the recognized text and the determined voice features to obtain an emotion recognition result for the recognized text and the voice features; the emotion recognition result comprises confidence degrees for a plurality of preset emotion dimensions; the voice features comprise at least one feature of pitch, speech rate, and sound intensity; and the voice features are obtained after performing the voice recognition operation on the voice navigation request; Generating an emotion snapshot for the emotion recognition result, and updating an emotion trajectory for the user according to the emotion snapshot, the emotion trajectory being used to record dimension change information of the user for one or more of the preset emotion dimensions; the emotion snapshot comprises a round, a timestamp, and an emotion label distribution associated with the voice navigation request; According to a preset emotion accumulation algorithm, performing a weighted summation operation on all target emotion snapshots included in the emotion trajectory to obtain a weighted summation result for the emotion trajectory; the weighted summation result comprises emotion score values for one or more of the preset emotion dimensions; The emotion snapshot, the updated emotion trajectory, and the weighted summation result are determined as the second voice processing result. 2.The multi-service voice navigation method based on voice intent tracking micro agent according to claim 1, wherein, The voice recognition operation on the voice navigation request to obtain recognized text for the voice navigation request comprises: Obtaining voice stream data corresponding to the voice navigation request; According to a preset signal processing algorithm, performing a voice feature extraction operation on the voice stream data to obtain voice features for the voice stream data; Determining a voice text recognition model matching the voice stream data; Inputting the voice features into the voice text recognition model to sequentially perform a text reasoning operation and a text output operation on the voice features by the voice text recognition model to obtain output text for the voice features; Perform a text post-processing operation on the output text to obtain recognized text for the voice navigation request, the text post-processing operation being used to correct the output text. 3.The multi-service voice navigation method based on voice intent tracking micro agent according to claim 1 or 2, characterized in that, The intent recognition operation and the keyword extraction operation are performed on the voice recognition result respectively to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first voice processing result, which includes: Perform a text preprocessing operation on the voice recognition result and generate a semantic vector embedding for the text preprocessing result; the text preprocessing operation includes a text content cleaning operation and a word segmentation operation; According to a preset large language model, the context window content of the semantic vector embedding is combined to perform an intent recognition operation on the semantic vector embedding to obtain a voice navigation intent for the semantic vector embedding, the voice navigation intent including at least one main intent category, all the main intent categories including at least one of query, reservation, complaint, and request for manual; According to a preset keyword recognition model, the historical keywords recognized for the semantic vector embedding are combined to perform a keyword extraction operation on the semantic vector embedding to obtain a keyword set for the semantic vector embedding, the keyword set including at least one keyword; According to a preset fusion module, a fusion operation is performed on the voice navigation intent and the keyword set to obtain a fusion result for the voice navigation intent and the keyword set as a first voice processing result. 4.The multi-service voice navigation method based on voice intent tracking micro agent according to claim 3, wherein, The determination and jump operation for the target intent node is performed according to the first voice processing result and the second voice processing result to obtain a jump result for the target intent node, which includes: An intent state jump graph is constructed according to the voice navigation intent, the intent state jump graph including a state node corresponding to the current location and at least one intent node corresponding to the voice navigation intent; a node edge is provided between the state node and each intent node; According to the keyword set, an edge attribute configuration operation is performed on each node edge to update the intent state jump graph; According to the fusion result, the intent state jump graph is combined to calculate a comprehensive jump score corresponding to each intent node; according to the comprehensive jump score corresponding to each intent node, a target intent node to be jumped is determined from all the intent nodes, and the state node is jumped to the target intent node as the jump result for the target intent node. 5.The multi-service voice navigation method based on voice intent tracking micro agent according to claim 3, wherein, Each main intent category corresponds to a category probability; each keyword corresponds to a vocabulary probability; The fusion operation is performed on the voice navigation intent and the keyword set according to the preset fusion module to obtain a fusion result for the voice navigation intent and the keyword set, which includes: The preset probability threshold is used as a data screening criterion. Data to be screened, whose category probability in the voice navigation intent is lower than the probability threshold and whose vocabulary probability in the keyword set is lower than the probability threshold, is subjected to a screening operation to update the voice navigation intent and update the keyword set. According to a preset ternary data structure, a first ternary tuple corresponding to the voice navigation intent and a second ternary tuple corresponding to the keyword set are constructed respectively. The first ternary tuple includes a first sub-group corresponding to each main intent category. The second ternary tuple includes a second sub-group corresponding to each keyword. According to a preset voting mechanism, a first voting operation is performed on all the first sub-groups, and a second voting operation is performed on all the second sub-groups, to obtain a first voting result corresponding to all the first sub-groups and a second voting result corresponding to all the second sub-groups. According to a first module weight set for the voice navigation intent and a second module weight set for the keyword set, a weighted calculation operation and a semantic correction operation are performed on the first voting result and the second voting result respectively, to obtain a fusion result for the first voting result and the second voting result as a fusion result for the voice navigation intent and the keyword set. 6.The multi-service voice navigation method based on voice intent tracking micro agent according to claim 4, wherein, The method further comprises: According to the intent state jump graph, in combination with the state node and the target intent node, it is judged whether the current satisfies the supplementary question condition. When it is judged that the current satisfies the supplementary question condition, supplementary question information for the voice navigation request is generated, and the supplementary question information is fed back to the user to prompt the user to feed back a new voice navigation request. When it is judged that the current does not satisfy the supplementary question condition, according to the state node, the target intent node, the jump path corresponding to the state node and the target intent node, a user jump trajectory recorded for the user is updated. The user jump trajectory is used to perform a path rollback operation for the jump path according to a path rollback demand when the path rollback demand is detected. Wherein, the supplementary question condition is satisfied, which includes: There is no legal jump path between the state node and the target intent node, or the user's current emotion value calculated for the user in the second voice processing result is lower than the node jump threshold set for the two nodes of the state node and the target intent node.

7. A multi-service voice navigation device based on voice intent tracking micro-agents, characterized by, The device is used to execute the voice intent tracking micro-agent based multi-service voice navigation method according to any one of claims 1-6, and the device comprises: A voice recognition module is configured to perform a voice recognition operation on a voice navigation request triggered by a user when the voice navigation request is detected, to obtain a recognized text for the voice navigation request as a voice recognition result. A first voice processing module is configured to perform an intent recognition operation and a keyword extraction operation on the voice recognition result, to obtain a voice navigation intent for the intent recognition operation and a keyword set for the keyword extraction operation as a first voice processing result. a second voice processing module, configured to perform emotion recognition operation and emotion score calculation operation on the first voice processing result respectively, to obtain a target recognized emotion for the emotion recognition operation and an emotion score value for the emotion score calculation operation as a second voice processing result; a node jump module, configured to perform determination and jump operation for a target intent node according to the first voice processing result and the second voice processing result, to obtain a jump result for the target intent node as a final response result responding to the voice navigation request.

8. A multi-service voice navigation device based on voice intent tracking micro-agents, characterized by, The device comprises: a memory storing executable program codes; a processor coupled with the memory; the processor invokes the executable program codes stored in the memory to execute the multi-service voice navigation method based on the voice intent tracking micro agent according to any one of claims 1-6.

9. A computer storage medium, characterized in that The computer storage medium stores computer instructions, which are invoked to execute the multi-service voice navigation method based on the voice intent tracking micro agent according to any one of claims 1-6.

Images