Intelligent call shorthand method and system, terminal and medium

Abstract

The invention discloses an intelligent call shorthand method and system, a terminal and a medium, and relates to the technical field of voice processing. According to the main technical scheme, entity features are recognized and extracted from an initial call shorthand record; constructing a context dynamic knowledge graph based on the entity features; and when it is detected that the indication feature exists in the initial call shorthand record, analyzing the indication feature according to the context dynamic knowledge graph at the current moment to obtain a target call shorthand record. And constructing a context dynamic knowledge graph by using entity features identified and extracted from the initial call shorthand record, and when detecting that indication features exist in the initial call shorthand record, analyzing the indication features according to the context dynamic knowledge graph at the current moment. In order to expect to achieve the purposes that the generated conversation shorthand record is clear in semantics, coherent in logic and convenient for a reader to understand by digesting indication features and replacing fuzzy pronouns with clear entity names.

Description

Technical Field

[0001] This invention relates to the field of voice processing technology, specifically to an intelligent call transcription method, system, terminal, and medium. Background Technology

[0002] Call transcription is a voice processing function that automatically converts the voice content of a call into text in real time or afterward, extracts key information, and helps users quickly grasp the core content of the call.

[0003] In modern communication environments, whether it's business meetings, customer service, legal evidence collection, or important daily calls, there's a high demand for accurate recording of call content. Traditional manual recording methods are not only inefficient but also prone to errors due to negligence or subjective interpretation, failing to meet the needs for efficient and accurate recording. With the continuous advancement of speech recognition technology, AI-based intelligent call transcription has become a reality.

[0004] However, in natural conversations, participants frequently use pronouns (such as "it," "this," "that," "the product," "that company," etc.) to refer to previously mentioned entities. Existing call transcription methods generally lack effective contextual association and referential resolution mechanisms, making it difficult to accurately identify and reconstruct the specific objects referred to by pronouns. This results in semantic ambiguity and insufficient logical coherence in the generated text records, severely impacting readability and practical value, and posing significant obstacles to subsequent information retrieval and understanding. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent call shorthand method, system, terminal, and medium, which solves the problem that existing call shorthand methods generally lack effective context association and referential resolution mechanisms, making it difficult to accurately identify and restore the specific objects referred to by pronouns. As a result, the generated text records are semantically ambiguous and lack logical coherence, which seriously affects the readability and practical value of the records and also brings significant obstacles to subsequent information retrieval and understanding.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] Firstly, a smart call transcription method is provided, including the following operations:

[0008] The audio data is processed based on voiceprint features to obtain the audio stream corresponding to each speaker and the timestamp corresponding to each audio stream.

[0009] Speech recognition is performed on each audio stream to generate an initial text stream carrying the speaker and timestamp;

[0010] Based on the speakers and timestamps in the initial text streams, the initial text streams are integrated and processed to generate an initial shorthand record of the call.

[0011] Identify and extract entity features from initial call stenography;

[0012] Constructing a dynamic knowledge graph based on entity features;

[0013] When an indicative feature is detected in the initial call transcript, the indicative feature is parsed based on the current context dynamic knowledge graph to obtain the target call transcript.

[0014] A further solution is that the process of obtaining the audio streams corresponding to each speaker and the timestamps corresponding to each audio stream includes:

[0015] Speech activity detection is performed on the audio data to obtain several valid speech segments and the timestamp of each valid speech segment;

[0016] Voiceprint features are extracted from each valid speech segment to obtain the speaker of each valid speech segment;

[0017] Merge valid speech segments of the same speaker to obtain audio streams corresponding to each speaker;

[0018] Assign a start timestamp and an end timestamp to each valid speech segment in the audio stream to obtain the timestamp corresponding to each audio stream.

[0019] A further approach is that the process of generating the initial text stream carrying the speaker and timestamp includes:

[0020] Speech recognition is performed on each valid speech segment in the audio stream to obtain the text unit corresponding to each valid speech segment;

[0021] Assign each text unit a start timestamp and an end timestamp corresponding to a valid speech segment;

[0022] Based on the start and end timestamps of each text unit, the text units are concatenated to obtain the initial text stream;

[0023] By associating each text unit in the initial text stream with the speaker, start timestamp, and end timestamp corresponding to each text unit, an initial text stream carrying the speaker and timestamp is obtained.

[0024] A further proposed solution is that the process of constructing a context-based dynamic knowledge graph based on entity features includes:

[0025] Based on the grammatical structure and semantic relationships of the initial call stenography, the relationships between the features of each entity are identified;

[0026] A knowledge graph is constructed using the extracted entity features as nodes and the relationships between entity features as edges.

[0027] The timestamps and context information of the entity features in the initial call stenography are attached to the corresponding nodes as the first attribute information; and the context information of the relationships between the entity features in the initial call stenography is attached to the corresponding edges as the second attribute information, thus obtaining the context knowledge graph.

[0028] Based on the initial call transcript, the nodes, edges, first attribute information, and second attribute information in the context knowledge graph are dynamically updated to obtain the context dynamic knowledge graph.

[0029] A further approach is: the process of parsing the indicator features based on the current context dynamic knowledge graph includes:

[0030] Construct a query vector; wherein the query vector includes indicative features and the local context of the indicative features in the initial call shorthand record;

[0031] Calculate the semantic similarity between the query vector and the corresponding mapping vectors of each node in the context dynamic knowledge graph;

[0032] The node with the highest semantic similarity is marked as the parsing result indicating the feature;

[0033] After attaching the entity features corresponding to the parsing results to the indicator features, the target call shorthand record is obtained.

[0034] A further approach is: the process of marking the node with the highest semantic similarity as the parsing result indicating the feature includes:

[0035] S6031, mark the node with the highest semantic similarity as a suspected node;

[0036] S6032, determine if the number of suspected nodes is greater than 1; if yes, execute S6033; if no, execute S6035.

[0037] S6033, Calculate the time interval between the indicator feature and each suspected node based on the timestamp of the suspected node and the timestamp of the indicator feature;

[0038] S6034, mark the suspected node with the smallest time interval as the parsing result indicating the feature;

[0039] S6035 marks suspected nodes as parsing results indicating features.

[0040] In a second aspect, an intelligent call transcription system is provided, which is applicable to the intelligent call transcription method described in the first aspect. The intelligent call transcription system includes a processing module, an identification module, an integration module, an extraction module, a construction module, and a parsing module.

[0041] The processing module processes audio data based on voiceprint features to obtain audio streams corresponding to each speaker and timestamps corresponding to each audio stream. The recognition module performs speech recognition on each audio stream to generate an initial text stream carrying the speaker and timestamp. The integration module integrates the initial text streams based on the speaker and timestamp in the initial text stream to generate an initial call transcript. The extraction module identifies and extracts entity features from the initial call transcript. The construction module constructs a contextual dynamic knowledge graph based on the entity features. The parsing module parses the indicator features based on the current contextual dynamic knowledge graph when an indicator feature is detected in the initial call transcript to obtain the target call transcript.

[0042] Thirdly, a terminal is provided, including a processor and a memory, the memory being used to store processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the intelligent call transcription method as described in the first aspect.

[0043] Fourthly, a computer-readable storage medium is provided having computer program instructions stored thereon, which, when executed by a processor, implement the intelligent call shorthand method as described in the first aspect.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] A contextual dynamic knowledge graph is constructed using entity features identified and extracted from the initial call transcript. When an indicative feature is detected in the initial call transcript, it is parsed based on the current contextual dynamic knowledge graph. The aim is to resolve ambiguous pronouns by replacing them with explicit entity names, thereby making the generated call transcript semantically clear, logically coherent, and easier for readers to understand. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating an intelligent call shorthand method in this embodiment. Detailed Implementation

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] Example 1: This example provides an intelligent call transcription method, such as... Figure 1 As shown, the following operations are included:

[0049] S10. Process the audio data based on voiceprint features to obtain the audio stream corresponding to each speaker and the timestamp corresponding to each audio stream;

[0050] For example, during implementation, voiceprint recognition technology is used to analyze different speech features in the original audio data, and speaker separation is performed based on the voiceprint features to generate an independent audio stream for each speaker. Simultaneously, each audio stream is marked with a timestamp corresponding to the audio data.

[0051] S20. Perform speech recognition on each audio stream to generate an initial text stream carrying the speaker and timestamp;

[0052] For example, during implementation, the audio stream of each speaker is sent to an automatic speech recognition engine for transcription, resulting in an initial text stream corresponding to each audio stream. Each initial text stream is bound to a corresponding speaker and timestamp. For example, [10:01:02] Customer: I would like to inquire about the logistics status of order number 12345.

[0053] S30. Based on the speakers and timestamps in the initial text streams, integrate the initial text streams to generate an initial shorthand record of the conversation;

[0054] For example, during implementation, the initial text streams from different speakers are interwoven and arranged according to the order of their timestamps. This forms a complete, time-series initial call transcript containing the speakers and timestamps. For example, [10:01:02] Customer: I would like to inquire about the logistics status of order number 12345. [10:01:08] Customer Service: Okay, please wait a moment, I will check for you. [10:01:20] Customer Service: Found it, it is currently at the Shanghai transit station.

[0055] S40. Identify and extract entity features from the initial call stenography;

[0056] For example, during implementation, named entity recognition technology is used to scan the integrated initial call transcript to identify and extract entity features from the initial call transcript. These entity features include people, organizations, locations, items, and times. For instance, the extracted entity features include: item: order number 12345, location: Shanghai transit station.

[0057] S50. Construct a dynamic knowledge graph based on entity features;

[0058] For example, during implementation, the extracted entity features are used as nodes in the knowledge graph.

[0059] For example, the nodes added in sequence are Order 12345 and Shanghai Transfer Station.

[0060] Analyze the relationships between entity features in the dialogue context, establish edges (relationships) between nodes, and obtain a dynamic contextual knowledge graph. For example, establish an edge between the node "Order 12345" and the node "Shanghai Transfer Station": Order 12345 is located at Shanghai Transfer Station.

[0061] S60. When an indicative feature is detected in the initial call shorthand record, the indicative feature is parsed according to the context dynamic knowledge graph at the current moment to obtain the target call shorthand record.

[0062] For example, during implementation, the initial call transcript is monitored in real time, and indicative features (such as "it," "this," "that," "this product," "that company," etc.) are identified in the initial call transcript. When an indicative feature is detected, the node that best matches and is most relevant to the indicative feature is found from the current context dynamic knowledge graph. The entity feature corresponding to the node is then appended to the indicative feature to obtain the optimized target call transcript. For example, when the indicative feature "it" is detected in "[10:01:20] Customer Service: Found it, it is currently at the Shanghai transit station.", the "order 12345" that best matches and is most relevant to "it" is found from the current context dynamic knowledge graph, and the initial call transcript "It is currently at the Shanghai transit station." is optimized to "It (order 12345) is currently at the Shanghai transit station." The resulting target call transcript is: "[10:01:02] Customer: I would like to inquire about the logistics status of order number 12345." [10:01:08] Customer Service: Okay, please wait a moment while I check for you. [10:01:20] Customer Service: Found it, it (order 12345) is currently at the Shanghai transit station.

[0063] The intelligent call transcription method in this embodiment constructs a contextual dynamic knowledge graph using entity features identified and extracted from the initial call transcription record. When an indicative feature is detected in the initial call transcription record, it is parsed based on the current contextual dynamic knowledge graph. The aim is to resolve the indicative feature, replacing ambiguous pronouns with explicit entity names, thereby making the generated call transcription record semantically clear, logically coherent, and easy for the reader to understand.

[0064] In step S10, the process of obtaining the audio stream corresponding to each speaker and the timestamp corresponding to each audio stream includes:

[0065] S101. Perform speech activity detection on the audio data to obtain several valid speech segments and the timestamp of each valid speech segment;

[0066] For example, during implementation, speech activity detection is performed on the original audio data to identify segments containing human speech, and non-speech parts such as silence and background noise are filtered out to obtain several valid speech segments. The start and end timestamps of the valid speech segments in the original audio data are recorded.

[0067] S102. Extract voiceprint features from each valid speech segment to obtain the speaker of each valid speech segment;

[0068] S103. Merge the valid speech segments of the same speaker to obtain the audio stream corresponding to each speaker;

[0069] For example, during implementation, speaker clustering or speaker verification models are used to determine whether each valid speech segment comes from the same speaker. All valid speech segments identified as belonging to the same speaker are then concatenated or logically aggregated in chronological order to form the audio stream corresponding to each speaker.

[0070] S104. Assign each valid speech segment in the audio stream a start timestamp and an end timestamp corresponding to the audio data, and obtain the timestamp corresponding to each audio stream.

[0071] For example, during implementation, the start and end timestamps of each valid speech segment in the audio stream are preserved in the original audio.

[0072] In this embodiment, on the one hand, speech activity detection is first performed on the audio data, and then valid speech segments from the same speaker are merged. This aims to filter out a large amount of invalid noise and silent segments using speech activity detection, allowing the subsequent, more computationally intensive voiceprint recognition model to only process pure human voice segments, thereby significantly reducing computational resource consumption and improving the accuracy of voiceprint clustering. On the other hand, each valid speech segment retains its start and end timestamps from the original audio, enabling the final call transcript to accurately trace back to the original audio position, facilitating manual review, editing, or evidence collection.

[0073] In step S20, the process of generating an initial text stream carrying the speaker and timestamp includes:

[0074] S201. Perform speech recognition on each valid speech segment in the audio stream to obtain the text unit corresponding to each valid speech segment;

[0075] S202. Assign each text unit a start timestamp and an end timestamp corresponding to a valid speech segment;

[0076] S203. Based on the start and end timestamps of each text unit, concatenate the text units to obtain the initial text stream;

[0077] For example, during implementation, all text units of the same speaker are sorted in ascending order by the start timestamp to form a coherent initial text stream that retains the time gap between the start and end timestamps.

[0078] S204. Associate each text unit in the initial text stream with the speaker, start timestamp and end timestamp corresponding to each text unit to obtain the initial text stream carrying the speaker and timestamp.

[0079] For example, during implementation, each text unit in the initial text stream, as well as the speaker, start timestamp, and end timestamp corresponding to each text unit, are structurally associated to obtain an initial text stream carrying the speaker and timestamp.

[0080] In this embodiment, each text unit in the initial text stream, along with the corresponding speaker, start timestamp, and end timestamp, is structurally associated, providing a precise temporal basis for subsequently integrating the initial text stream into an initial call stenography record. The aim is to facilitate the accurate interweaving and arrangement of text units from different speakers, thereby realistically recreating the rhythm of the dialogue interaction.

[0081] In step S50, the process of constructing a contextual dynamic knowledge graph based on entity features includes:

[0082] S501. Based on the grammatical structure and semantic relationships of the initial call shorthand record, identify the relationships between the features of each entity;

[0083] For example, during implementation, the relationships between entity features are identified by analyzing the grammatical structure and semantic information of the initial call transcript containing entity features. For instance, from "[10:01:02] Customer: I would like to inquire about the logistics status of order number 12345.", the entity feature "order12345" is extracted; from "[10:01:20] Customer Service: Found it, it is currently at the Shanghai transit station.", the entity feature "Shanghai transit station" is extracted, and the relationship between the entity features "order12345" and "Shanghai transit station" is identified as "currently", resulting in the triple (order12345, currently, Shanghai transit station).

[0084] S502. Construct a knowledge graph using the extracted entity features as nodes and the relationships between entity features as edges;

[0085] For example, during implementation, each entity feature (such as "order12345", "Shanghai transit station", "customer", "customer service") is extracted as a node in the knowledge graph, and the identified relationships (such as "currently" and "inquiry") are used as edges connecting the corresponding nodes to obtain a preliminary, static knowledge graph.

[0086] S503. The timestamp and context information of the entity features in the initial call stenography record are attached to the corresponding nodes as the first attribute information; and the context information of the relationship between each entity feature in the initial call stenography record is attached to the corresponding edges as the second attribute information to obtain the context knowledge graph.

[0087] For example, during implementation, the timestamps of the first appearance and most recent mention of the entity feature in the initial call transcript, along with contextual information, are appended as the first attribute information to the corresponding node. The contextual information includes the original sentence containing the entity feature, the turn of the conversation, and the speaker. For instance, the first attribute information "{First mention: [10:01:02], Original text: "I would like to inquire about the logistics status of order number 12345"} is appended to the node "Order 12345".

[0088] The contextual information of the relationships between entity features in the initial call transcript is appended as second attribute information to the corresponding edges. For example, the second attribute information "is currently" is appended to the edge "is currently" {based on: "It is currently at the Shanghai transit station"}. This results in a contextual knowledge graph that carries rich dialogue history and specific evidence.

[0089] S504. Based on the initial call transcript, dynamically update the nodes, edges, first attribute information, and second attribute information in the context knowledge graph to obtain the context dynamic knowledge graph.

[0090] For example, during implementation, as the conversation progresses, new initial call transcripts are continuously generated and executed in steps S40 and S501-S503. When a new entity feature is discovered, it is added as a new node to the context knowledge graph. When a new relationship is discovered, it is added as a new edge to the context knowledge graph. The timestamps and context information of nodes are updated, and new context information is added to edges. This results in a constantly evolving, dynamic context knowledge graph that remains synchronized with the current conversation content.

[0091] In this embodiment, the nodes, edges, first attribute information, and second attribute information in the context knowledge graph are dynamically updated to obtain a dynamic context knowledge graph. The aim is to provide the entity features most likely to be referred to in the current context, thereby significantly improving the accuracy of parsing indicative features such as "it," "this," and "that."

[0092] In step S60, the process of parsing the indicator features based on the current context dynamic knowledge graph includes:

[0093] S601. Construct a query vector; wherein the query vector includes an indicator feature and the local context of the indicator feature in the initial call shorthand record;

[0094] For example, during implementation, when an indicative feature (such as "it") is detected, the local context in which the indicative feature is located is captured. This local context typically includes the entire sentence containing the indicative feature, and even adjacent sentences. Example: For the sentence "The query found it; it is currently at the Shanghai transit station.", the local context is the entire sentence.

[0095] The indicative features and local context (such as "found, it is currently at the Shanghai transit station") are input into a pre-trained deep learning model to obtain a high-dimensional, numerical vector, which is the query vector of the indicative features.

[0096] S602. Calculate the semantic similarity between the query vector and the corresponding mapping vectors of each node in the context dynamic knowledge graph;

[0097] For example, during implementation, each node in the contextual dynamic knowledge graph (i.e., each entity feature, such as "order12345" or "Shanghai transit station") is input into a pre-trained deep learning model to obtain a high-dimensional, numerical vector, i.e., the node's mapping vector. The semantic similarity (commonly cosine similarity, with values ​​between -1 and 1, the closer to 1, the more semantically similar) between the query vector ("It is currently at the Shanghai transit station") and the mapping vector of the node "order12345" is 0.88. The semantic similarity between the query vector ("It is currently at the Shanghai transit station") and the mapping vector of the node "Shanghai transit station" is 0.65.

[0098] S603. Mark the node with the highest semantic similarity as the parsing result indicating the feature;

[0099] For example, during implementation, the node with the highest semantic similarity is selected as the parsing result of the indicator feature. For instance, if the semantic similarity between the indicator feature "it" and the node "order12345" is the highest at 0.88, then the node "order12345" is selected as the parsing result of the indicator feature "it".

[0100] S604. After attaching the entity features corresponding to the parsing results to the indicator features, the target call shorthand record is obtained.

[0101] For example, during implementation, the parsing results (entity features) are directly inserted as footnotes after the indicator features in the initial call shorthand record to obtain the target call shorthand record. For instance, the sentence "Found, it is currently at the Shanghai transit station." is optimized to "Found, it (order 12345) is currently at the Shanghai transit station."

[0102] In this embodiment, deep semantic matching is used to more accurately handle complex referential scenarios such as cross-sentence, cross-speaker, and coexistence of multiple entity features, in order to further significantly improve the accuracy of parsing indicator features such as "it", "this", and "that".

[0103] To improve the accuracy of indicator feature parsing when there are multiple nodes with the highest semantic similarity, step S603, which involves marking the node with the highest semantic similarity as the parsing result of the indicator feature, includes:

[0104] S6031, mark the node with the highest semantic similarity as a suspected node;

[0105] S6032, determine if the number of suspected nodes is greater than 1; if yes, execute S6033; if no, execute S6035.

[0106] For example, during implementation, the number of suspected nodes is counted, and it is determined whether the number of suspected nodes is greater than 1. If the number of suspected nodes is greater than 1, S6033 is executed; if the number of suspected nodes is equal to 1, S6035 is executed.

[0107] S6033, Calculate the time interval between the indicator feature and each suspected node based on the timestamp of the suspected node and the timestamp of the indicator feature;

[0108] For example, during implementation, when the number of suspected nodes is greater than 1, the timestamp of the most recent mention of each suspected node and the timestamp of the statement containing the indicator feature are obtained, and the time interval between the indicator feature and each suspected node is calculated based on the timestamp of the indicator feature and the timestamp of the most recent mention of each suspected node.

[0109] S6034, mark the suspected node with the smallest time interval as the parsing result indicating the feature;

[0110] For example, during implementation, the suspected node with the smallest time interval is selected as the parsing result of the indicative feature.

[0111] S6035 marks suspected nodes as parsing results indicating features.

[0112] When the number of suspected nodes is equal to 1, the suspected node is used as the parsing result of the indicator feature.

[0113] In this embodiment, when ambiguous scenarios arise that cannot be handled by relying solely on semantic similarity, the "temporal proximity" principle is used to select the parsing result based on the indicative features. The aim is to make the parsing results more consistent with human conversational habits (people tend to use pronouns to refer to things that have just been mentioned), thereby reducing misreferences and making the output results more reliable.

[0114] Example 2: This example provides an intelligent call transcription system. The intelligent call transcription system is applicable to the intelligent call transcription method described in Example 1. The intelligent call transcription system includes a processing module, an identification module, an integration module, an extraction module, a construction module, and a parsing module.

[0115] The processing module processes audio data based on voiceprint features to obtain audio streams corresponding to each speaker and timestamps corresponding to each audio stream. The recognition module performs speech recognition on each audio stream to generate an initial text stream carrying the speaker and timestamp. The integration module integrates the initial text streams based on the speaker and timestamp in the initial text stream to generate an initial call transcript. The extraction module identifies and extracts entity features from the initial call transcript. The construction module constructs a contextual dynamic knowledge graph based on the entity features. The parsing module parses the indicator features based on the current contextual dynamic knowledge graph when an indicator feature is detected in the initial call transcript to obtain the target call transcript.

[0116] The intelligent call transcription system in this embodiment first uses an extraction module to identify and extract entity features from the initial call transcription record. Next, a construction module builds a contextual dynamic knowledge graph. Then, a parsing module, upon detecting indicative features in the initial call transcription record, parses the indicative features based on the current contextual dynamic knowledge graph. The aim is to resolve indicative features, replacing ambiguous pronouns with explicit entity names, thereby making the generated call transcription record semantically clear, logically coherent, and easy for the reader to understand.

[0117] In this embodiment, a terminal is also provided, including a processor and a memory, wherein the memory is used to store processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the intelligent call transcription method as described in Embodiment 1.

[0118] In this embodiment, a computer-readable storage medium is also provided, on which computer program instructions are stored, which, when executed by a processor, implement the intelligent call shorthand method as described in Embodiment 1.

[0119] Although the invention has been described herein with reference to several illustrative embodiments, it should be understood that many other modifications and implementations can be devised by those skilled in the art, which will fall within the scope and spirit of the principles disclosed herein. More specifically, various variations and modifications can be made to the components and / or layout of the subject matter arrangement within the scope of the disclosure, drawings, and claims. Besides variations and modifications to the components and / or layout, other uses will be apparent to those skilled in the art.

Claims

1. A method for intelligent call transcription, characterized in that, Includes the following operations: The audio data is processed based on voiceprint features to obtain the audio stream corresponding to each speaker and the timestamp corresponding to each audio stream. Speech recognition is performed on each audio stream to generate an initial text stream carrying the speaker and timestamp; Based on the speakers and timestamps in the initial text streams, the initial text streams are integrated and processed to generate an initial shorthand record of the call. Identify and extract entity features from initial call stenography; Constructing a dynamic knowledge graph based on entity features; When an indicative feature is detected in the initial call transcript, the indicative feature is parsed based on the current context dynamic knowledge graph to obtain the target call transcript.

2. The intelligent call shorthand method according to claim 1, characterized in that, The process of obtaining the audio streams corresponding to each speaker and the timestamps corresponding to each audio stream includes: Speech activity detection is performed on the audio data to obtain several valid speech segments and the timestamp of each valid speech segment; Voiceprint features are extracted from each valid speech segment to obtain the speaker of each valid speech segment; Merge valid speech segments of the same speaker to obtain audio streams corresponding to each speaker; Assign a start timestamp and an end timestamp to each valid speech segment in the audio stream to obtain the timestamp corresponding to each audio stream.

3. The intelligent call shorthand method according to claim 2, characterized in that, The process of generating an initial text stream carrying the speaker and timestamp includes: Speech recognition is performed on each valid speech segment in the audio stream to obtain the text unit corresponding to each valid speech segment; Assign each text unit a start timestamp and an end timestamp corresponding to a valid speech segment; Based on the start and end timestamps of each text unit, the text units are concatenated to obtain the initial text stream; By associating each text unit in the initial text stream with the speaker, start timestamp, and end timestamp corresponding to each text unit, an initial text stream carrying the speaker and timestamp is obtained.

4. The intelligent call shorthand method according to claim 1, characterized in that, The process of constructing a context-based dynamic knowledge graph based on entity features includes: Based on the grammatical structure and semantic relationships of the initial call stenography, the relationships between the features of each entity are identified; A knowledge graph is constructed using the extracted entity features as nodes and the relationships between entity features as edges. The timestamps and context information of the entity features in the initial call stenography are attached to the corresponding nodes as the first attribute information; and the context information of the relationships between the entity features in the initial call stenography is attached to the corresponding edges as the second attribute information, thus obtaining the context knowledge graph. Based on the initial call transcript, the nodes, edges, first attribute information, and second attribute information in the context knowledge graph are dynamically updated to obtain the context dynamic knowledge graph.

5. The intelligent call shorthand method according to claim 4, characterized in that, The process of parsing the indicator features based on the current context dynamic knowledge graph includes: Construct a query vector; wherein the query vector includes indicative features and the local context of the indicative features in the initial call shorthand record; Calculate the semantic similarity between the query vector and the corresponding mapping vectors of each node in the context dynamic knowledge graph; The node with the highest semantic similarity is marked as the parsing result indicating the feature; After attaching the entity features corresponding to the parsing results to the indicator features, the target call shorthand record is obtained.

6. The intelligent call shorthand method according to claim 5, characterized in that, The process of marking the node with the highest semantic similarity as the parsing result indicating the feature includes: S6031, mark the node with the highest semantic similarity as a suspected node; S6032, determine if the number of suspected nodes is greater than 1; if yes, execute S6033; if no, execute S6035. S6033, Calculate the time interval between the indicator feature and each suspected node based on the timestamp of the suspected node and the timestamp of the indicator feature; S6034, mark the suspected node with the smallest time interval as the parsing result indicating the feature; S6035 marks suspected nodes as parsing results indicating features.

7. An intelligent call transcription system, characterized in that, The intelligent call stenography system is applicable to the intelligent call stenography method as described in any one of claims 1-6, wherein the intelligent call stenography system comprises: The processing module is used to process audio data according to voiceprint features to obtain audio streams corresponding to each speaker and timestamps corresponding to each audio stream. The recognition module is used to perform speech recognition on each audio stream and generate an initial text stream carrying the speaker and timestamp. An integration module is used to integrate each initial text stream based on the speaker and timestamp in the initial text stream to generate an initial call transcript. Extraction module, which is used to identify and extract entity features from the initial call transcript; The construction module is used to construct a contextual dynamic knowledge graph based on entity features; The parsing module is used to parse the indicator features based on the current context dynamic knowledge graph when an indicator feature is detected in the initial call shorthand record, so as to obtain the target call shorthand record.

8. A terminal, characterized in that, include: A processor and a memory, wherein the memory is used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the intelligent call shorthand method as described in any one of claims 1-6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the intelligent call shorthand method as described in any one of claims 1-6.

Images