A method, system, and electronic device for providing event reminders based on call data.

By performing voiceprint recognition and text analysis on call data, combined with a database of speakers and contacts, and cross-modal association of sentences with target words, the system can automatically identify pending tasks and responsible persons, solving the problem that users need to manually check tasks in existing technologies and improving the automation and intelligence of the call assistant.

CN123123499APending Publication Date: 2026-11-17BOE TECHNOLOGY GROUP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611330909.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing call assistants require users to manually verify pending tasks after recording and transcribing, lacking automation and intelligence, and failing to accurately identify the task executor, resulting in a poor user experience.

Method used

By performing voiceprint recognition and text analysis on call data, and combining it with a database of speakers and contacts, the system can automatically identify tasks and responsible persons by cross-modal association of sentences and target words, eliminating the need for manual screening of recordings.

Benefits of technology

It achieves efficient and accurate task attribution identification in call scenarios, improves the automation and intelligence of task prompts, and reduces the user's operational burden.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123499A_ABST
    Figure CN123123499A_ABST
Patent Text Reader

Abstract

This application provides a method, system, and electronic device for providing task reminders based on call data, comprising: acquiring call data, including call audio and corresponding text; determining the speaker corresponding to multiple statements in the text based on the call audio and voiceprint features; analyzing the text and extracting target words representing the task to be done and the task executor; determining the target object referred to by the task executor based on the speakers and contact database corresponding to the multiple statements, and identifying the target object as the person responsible for the task to be done, wherein the target object includes at least one of the speakers corresponding to the multiple statements and other contacts in the contact database; and generating a reminder message including the task to be done and the person responsible for the task. In this way, accurate task attribution identification can be achieved without manual screening, significantly improving the automation and intelligence level of task reminders based on call data, and providing a better user experience.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of mobile communication technology, and in particular to a method, system, and electronic device for providing event notifications based on call data. Background Technology

[0002] With the widespread adoption of smart mobile devices, technologies such as call recording and transcription, and intelligent voice assistants have been widely applied. Existing call assistants can typically convert speech into text in real time during a call and allow users to create memos or schedule reminders after the call by manually selecting text or using voice commands. However, traditional voice assistants only support word-to-speech conversion, and users still need to manually check the information of tasks after the call, resulting in a low level of automation in the process. Summary of the Invention

[0003] This application provides a method, system, and electronic device for providing task reminders based on call data. It can achieve accurate task attribution identification without manual screening or listening to recordings, significantly improving the automation and intelligence level of task reminders based on call data, and helping to improve user experience.

[0004] The first aspect of this application provides a method for providing task reminders based on call data, comprising: acquiring call data, including call audio and corresponding text; determining the speaker corresponding to multiple statements in the text based on the call audio and voiceprint features; analyzing the text and extracting target words representing the task to be done and the task executor; determining the target object referred to by the task executor based on a database of speakers and contacts corresponding to the multiple statements, and identifying the target object as the person responsible for the task to be done, wherein the target object includes at least one of one or more speakers and other contacts corresponding to the multiple statements; and generating a reminder message, which includes the task to be done and the person responsible for the task to be done.

[0005] Based on the above scheme, by performing voiceprint recognition and text analysis on call data, target words containing the to-do tasks and task executors are obtained. By combining the speakers and other contacts involved in the call, the speakers corresponding to the statements are cross-modal associated with the target objects referred to by the target words. Accurate task attribution identification can be achieved without manual screening or listening to the recording, which significantly improves the automation and intelligence level of task prompts in call scenarios and helps to improve the user experience.

[0006] In conjunction with the first aspect, in one implementation, the target object also includes a pending contact person. The method further includes: when determining the target object based on the speakers and contact databases corresponding to multiple statements fails, marking the target object as a pending contact person. Outputting information to request user confirmation of the pending contact person's identity. Receiving the user's confirmation result of the pending contact person's identity and assigning the confirmation result as the person responsible for the pending task.

[0007] Based on the above scheme, when the actual identity of the target cannot be directly determined, the human-machine collaborative confirmation mechanism avoids misjudgment of the person responsible for the task and improves the robustness of determining responsibility in complex call environments.

[0008] In conjunction with the first aspect, in one implementation, the method further includes: extracting volume features and / or speech rate features of statements containing target words in the call audio; determining the urgency of the task in the target word based on the volume features and / or speech rate features; and adjusting the prompt parameters of the prompt message based on the urgency of the task, the prompt parameters including at least one of reminder priority, reminder frequency, and vibration frequency or ringtone volume of the terminal device.

[0009] Based on the above solution, by integrating acoustic physical characteristics such as volume and speech rate, the urgency of tasks to be done is accurately identified. Furthermore, by dynamically adjusting prompt parameters such as vibration frequency and volume according to the urgency level, the probability of users missing highly urgent tasks due to insufficient prompts is effectively reduced, thus improving the intelligence level of the task closure loop.

[0010] In conjunction with the first aspect, in one implementation, the target word also includes the time of the to-do task. Before generating the prompt information, the method further includes: comparing the time of the to-do task with existing schedules in the system calendar. When the time of the to-do task conflicts with an existing schedule, based on the available time slots in the system calendar, suggested information is output, including solutions for adjusting the time of the to-do task and / or the time of the existing schedule.

[0011] Based on the above solution, an intelligent time conflict retrieval and adaptive scheduling mechanism was established by comparing and dynamically adjusting the time of pending tasks with existing schedules, thereby improving the intelligence level of the task closure loop.

[0012] In conjunction with the first aspect, one implementation method further includes: analyzing the text and extracting target words that express the location of the task to be completed. When the distance between the user's current location and the location of the task to be completed is detected to be less than a distance threshold, a prompt message is proactively output.

[0013] Based on the above solution, the location of to-do tasks is automatically extracted from call data, and a reminder is proactively triggered when the user approaches the location of the to-do task. This further reduces the probability of the user missing to-do tasks and helps to improve the efficiency and intelligence level of the task loop.

[0014] In conjunction with the first aspect, one implementation method further includes: matching and generating a corresponding execution step chain from a preset step library based on the type of the task to be done; and automatically opening the local application software or document template required for the task on the terminal device based on the execution step chain.

[0015] Based on the above solution and combined with a preset step library, the system can automatically generate an execution step chain based on the task type identified by call data, and automatically launch local application software based on the execution step chain. This simplifies the user's operation process for executing pending tasks through the device and improves the device's operational efficiency and intelligent experience in complex office and call scenarios.

[0016] In conjunction with the first aspect, in one implementation, the method further includes: replacing the privacy information in the text with corresponding de-identified placeholders locally on the terminal device to obtain de-identified text; sending the de-identified text to a cloud server; receiving the structured reasoning result expressed by the de-identified placeholders returned by the cloud server locally on the terminal device, and backfilling the reasoning result according to the mapping relationship between the de-identified placeholders and the privacy information to obtain execution suggestions for the pending task.

[0017] Based on the above solution, terminal devices can work with cloud-deployed models for inference processing, and the text can be de-identified before uploading data. This allows for the use of the stronger inference capabilities of cloud models while ensuring user privacy, thus achieving intelligent services that combine privacy and security with an automated closed-loop experience.

[0018] In conjunction with the first aspect, one implementation method further includes: assigning task identifiers to pending tasks. Based on data associated with the task identifiers, the execution status corresponding to the task identifiers is dynamically updated. This data includes email data and instant messaging data.

[0019] Based on the above solution, by assigning a unique task identifier to the tasks extracted from call data and establishing a multimodal task fusion mechanism, fragmented information from multiple channels can be automatically aggregated into the same task chain, reducing the information retrieval and maintenance costs for users in multi-application collaborative scenarios and improving the intelligence level of task closed-loop processing.

[0020] A second aspect of this application provides a system for providing task reminders based on call data. This system includes a call interface module, a speaker separation module, a natural language understanding module, a responsible person determination module, and an output execution module. The call interface module acquires call data, including call audio and corresponding text. The speaker separation module determines the speakers corresponding to multiple statements in the text based on the call audio and voiceprint features. The natural language understanding module analyzes the text and extracts target words representing the task and its executor. The responsible person determination module determines the target object referred to by the task executor based on a database of speakers and contacts corresponding to the multiple statements, and identifies the target object as the responsible person for the task. The target object includes at least one of the speakers and other contacts corresponding to the multiple statements. The output execution module generates a reminder message, which includes the task and the responsible person for the task.

[0021] A third aspect of this application provides an electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program stored in the memory, such that the device performs the method of any of the first aspects described above.

[0022] The fourth aspect of this application provides a computer-readable storage medium storing program code that, when executed on a computer, causes the computer to perform any of the methods described in the first aspect above.

[0023] The fifth aspect of this application provides a computer program product comprising: computer program code that, when run on a computer, causes the computer to perform the method of any one of the first aspects described above.

[0024] The supplements and technical effects of the solutions provided in the second to fifth aspects above can be found in the corresponding descriptions in the first aspect, and will not be repeated here. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a method 100 for providing event reminders based on call data, provided in an embodiment of this application. Figure 2This is a first technical flowchart of method 100 provided in the embodiments of this application; Figure 3 This is a second technical flowchart of method 100 provided in the embodiments of this application; Figure 4 This is a third technical flowchart of method 100 provided in the embodiments of this application; Figure 5 This is a schematic diagram illustrating the presentation of prompt information on a terminal device according to an embodiment of this application; Figure 6 This is a fourth technical flowchart of method 100 provided in the embodiments of this application; Figure 7 This is the fifth technical flowchart of method 100 provided in the embodiments of this application; Figure 8 This is the sixth technical flowchart of method 100 provided in the embodiments of this application; Figure 9 This illustration shows a flowchart of a process for providing event reminders based on call data, according to an embodiment of this application. Figure 10 A schematic diagram of the architecture of a terminal intelligent agent provided in an embodiment of this application is shown. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0028] The terminology used in the following embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to also include expressions such as “one or more,” unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of this application, “at least one” and “one or more” refer to one, two, or more than two. The term “and / or” is used to describe the relationship between related objects, indicating that three relationships may exist; for example, A and / or B can indicate: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character “ / ” generally indicates that the preceding and following related objects are in an “or” relationship.

[0029] In the embodiments described in this application, references such as "one embodiment," "some embodiments," "one example," or "some examples" mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0030] Unless otherwise specified, all embodiments and optional embodiments of this application can be combined with each other to form new technical solutions.

[0031] Unless otherwise specified, all technical features and optional technical features of this application may be combined to form new technical solutions.

[0032] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements are represented by similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale. Furthermore, some well-known parts may not be shown in the drawings.

[0033] As mentioned in the background section above, current technologies have significant limitations in real-world call scenarios. On one hand, existing text analysis methods typically only perform isolated keyword recognition, lacking the ability to collaboratively extract and semantically correlate tasks and their executors. They cannot automatically pinpoint the correspondence between tasks and their executors during the text analysis phase, requiring users to manually review the transcribed text after the call, resulting in low efficiency in extracting key elements. On the other hand, the task executor in call text is usually presented only as static text. Because there are complex referential relationships between the task executor and the actual speaker in everyday call text, traditional text processing systems or voice assistants only support isolated text analysis. They lack the technical mechanism to cross-modal correlate the speaker's voiceprint features in the call audio with the task executor extracted from the text, failing to automatically map the executor in the text to the actual physical caller or contact. This forces users to manually verify the recording after the call to confirm the responsible party, resulting in a low level of intelligence in the task reminder process.

[0034] In view of this, embodiments of this application provide a method and electronic device for providing task reminders based on call data. By performing voiceprint recognition and text analysis on the call data, target words containing the to-do tasks and task executors are obtained. Combined with the speakers and other contacts involved in the call, the speakers corresponding to the statements are cross-modal associated with the target objects referred to by the target words. Accurate task attribution identification can be achieved without manual screening or listening to the recording, which significantly improves the automation and intelligence level of task reminders in call scenarios and helps to improve the user experience.

[0035] Figure 1 This illustration shows a schematic diagram of a method for providing event reminders based on call data, according to an embodiment of this application. Figure 1 As shown, method 100 includes the following steps: S110, acquire call data, which includes call audio and corresponding text.

[0036] In one example, the system can first acquire the call audio and then transcribe it into corresponding text. In this example, the initial input to the system is the call audio, which can be the raw digital acoustic signal during the call or a digital acoustic signal that has undergone noise reduction, voice amplification, and other processing. Furthermore, the text corresponding to the call audio is not external input to the system or pre-existing static text, but rather an intermediate product obtained by processing and converting the call audio in real time or offline on the terminal device.

[0037] In practice, the system can acquire call data in real time during a call or acquire it completely after the call ends, without limitation.

[0038] In one example, during a real-time call, the terminal device (such as a smartphone) acquires a digital audio stream converted from analog audio signals in real time through the system call interface or the microphone-earpiece data path. The terminal device uses this digital audio stream as initial input. Subsequently, the terminal device's local processor inputs the captured digital audio stream in real time to its local Automatic Speech Recognition (ASR) module. The ASR module is used to convert the original call audio stream into corresponding Chinese characters or character streams in real time, thereby generating the corresponding call text in real time during the call.

[0039] In another example, the terminal device directly reads the stored complete call audio file corresponding to the call from the system call recording storage or a designated local cache path, and uses this call audio file as input to the system. Similarly, the system can refer to the specific process described above to convert the call audio into a complete timestamped transcribed text through the ASR module or by calling the local offline speech recognition engine. Compared to real-time scenarios, offline acquisition can utilize richer contextual information to perform acoustic and semantic correction on homophones, slips of the tongue, or local dialects, thereby generating a more accurate complete transcribed text as an intermediate product obtained from offline processing.

[0040] In another example, the system can simultaneously acquire the call audio and the corresponding text, which is transcribed locally or in the cloud by other applications on the terminal device. In this example, the system's initial input is the call audio and the corresponding text.

[0041] Optionally, before acquiring call data, the system verifies whether the user has granted security authorization to the system. This authorization may include, but is not limited to, authorization for recording, transcription, event extraction, and cross-application linkage, or one or more of these. After successful authorization verification, the system executes step S110 as described above.

[0042] S120, based on the call audio and voiceprint characteristics, determines the speaker corresponding to each of the multiple sentences in the text.

[0043] like Figure 2 As shown, firstly, the system uses the Voice Activity Detection (VAD) algorithm to perform time-domain and frequency-domain energy analysis on the original call audio, filtering out background noise, silent segments, and call waiting tones to locate active audio segments containing human speech. Subsequently, by analyzing the spectral characteristics of these active audio segments, the system locates the boundary points where different speakers alternate speaking, thus segmenting the call audio and its corresponding text into multiple sets at the speaker level. Each set consists of one or more independent sentence segments spoken by a single speaker.

[0044] Furthermore, the system can extract and match voiceprint features from each obtained set to determine the identity of each speaker in the call audio. Specifically, the system extracts voiceprint features from each set or each segmented sentence fragment, and compares the global voiceprint embedding feature vector of each extracted sentence fragment with a locally pre-stored registered contact voiceprint feature database to obtain a similarity score between the global voiceprint embedding feature vector of the current sentence fragment and the voiceprint templates of each known contact in the database. This database can be stored locally on the terminal device or in the cloud, without restriction. Subsequently, the system compares the calculated similarity score with a preset similarity threshold. If the similarity score is greater than or equal to the similarity threshold, the registered contact corresponding to the highest score is identified as the speaker who uttered the sentence fragment. If the similarity score is lower than the similarity threshold, or if the current call audio fragment belongs to the other end of the call and the other end's contact is not registered in the database, the speaker of the sentence fragment is marked as an unknown speaker, thus classifying them as a pending contact for the user to manually select or input confirmation in subsequent steps.

[0045] Optionally, after locating the active audio segment, the terminal device uses an Overlapping Speech Detection (OSD) algorithm for multi-label classification to identify whether the current active audio segment contains overlapping sounds from two or more speakers speaking simultaneously. If no overlapping sounds are found, the system directly detects speaker change points based on the active audio segment. If overlapping sounds are found, the audio segment containing overlapping sounds is input into the local Blind Source Separation (BSS) module. The BSS module is used to reconstruct and decouple the single-channel mixed audio signal into multiple independent, single-speaker clean audio signals in the time or frequency domain, so that the system can perform speaker change point detection based on the clean audio signals.

[0046] S130, Analyze the text and extract target words that represent pending tasks and task performers.

[0047] like Figure 2 and Figure 3 As shown, after obtaining the text corresponding to the call audio, the system analyzes the text. Specifically, the Natural Language Understanding (NLU) module begins to perform text analysis and feature extraction on the local device to obtain target words. In this embodiment, the target words, in addition to including the task to be done and the task executor, may also include one or more constraints such as the time of the task to be done, the location of the task to be done, the duration of the task to be done, and the execution method of the task to be done. The specific methods of text analysis and feature extraction performed by the system are described below.

[0048] First, the system segments the continuous word stream into a sequence of words with independent semantics. Then, using a Part-of-Speech Tagging (POS) model, each segmented word is assigned a corresponding grammatical attribute label (e.g., noun, verb, pronoun, preposition, adverb). Next, the system runs a Dependency Parsing (DP) algorithm. By establishing a dependency tree relationship between words, it can locate the dependency path between the core governing word (e.g., verb) and its modifiers. Simultaneously, the system runs a Semantic Role Labeling (SRL) module. This module, centered on verbs, analyzes the semantic roles played by various nouns or entities in the sentence. For example, it determines the actor (task executor), the recipient (object or target word of the task), and the accompanying constraints (e.g., location, manner). In this way, the system can extract target words expressing the task and the executor from the text.

[0049] Optionally, the system calls the time expression parsing engine to perform regularization parsing and align the ambiguous or colloquial time words in the text (such as "before 3 p.m." or "tomorrow morning") with the base time, converting them into a machine-readable standard time format, thereby extracting the target words for the deadline of the task to be done, thus obtaining the target words for expressing the time of the task to be done.

[0050] Optionally, the system can fill the extracted time target words, location target words, task action target words, and task executor target words into a preset structured semantic frame to generate highly organized task candidate slots. For example, the task candidate slot can be represented by {action, object, time, location, responsible person candidate, constraint}, or by {task_id, title, action, object, responsible_party, due_time, priority, source_channel, evidence_span}, thereby completing the complete extraction of task elements.

[0051] Continue to refer to Figure 3In one example, if one party in a two-way audio call says, "Send the quote to Manager Wang before 3 PM today," the semantic parser performs dependency analysis and role labeling on the sentence. The action phrase extractor identifies "send quote," the object parser identifies "Manager Wang," and the time expression parser identifies "before 3 PM today." At this point, the system obtains the target words representing the task "send quote" and the person responsible for the task: "the other party in the audio call." Further, the system integrates the extracted target words into a task object. This task object includes at least the task name (send quote), the responsible person (the user), the deadline (3 PM today), the source call, and evidence fragments. The evidence fragments are used to preserve the original statement "Send the quote to Manager Wang before 3 PM today" for the user to trace the source later.

[0052] Continue to refer to Figure 2 In another example, if one party in a multi-party call says, "Xiao Li is responsible for organizing the contracts, and Xiao Wang is responsible for contacting the client to confirm the time," the semantic parser performs dependency analysis and role labeling on the sentence, extracting slots for "responsibility trigger words," "action phrases," "object objects," and "target deadlines." For example, from "Xiao Li is responsible for organizing the contracts," it extracts the task executor `responsible_party` = Xiao Li, the responsibility trigger `responsible_trigger` = responsible, the object object `object` = contract, and the action phrase `action` = organize. Ultimately, the system obtains two sets of target words representing the tasks to be done and the responsible parties, corresponding to the tasks assigned to Xiao Li and Xiao Wang in the call audio, respectively.

[0053] In some embodiments, in step S130, the system is further configured to extract the volume features and / or speech rate features of the statements containing the target words in the call audio, and determine the urgency of the task in the target words based on the volume features and / or speech rate features.

[0054] In practical implementation, the system can calculate the sum of squares of the signal amplitude within each frame of a call audio segment containing the target word, obtaining the root mean square energy value of that frame, which is used as a volume feature. Simultaneously, the system combines the results from the ASR module to calculate the speech rate feature of the call audio segment. The system can combine at least one of the following as quantitative indicators of speech rate features: the number of words spoken per second, the percentage of silent pauses, and the frequency of pauses.

[0055] Optionally, the system can also input the extracted volume features, speech rate features, and fundamental frequency fluctuations into a locally pre-trained or cloud-deployed speech emotion recognition classifier to perform multi-dimensional probability prediction of the speaker's emotional state and output probability distributions including emotional dimensions such as calm, anxiety, urging, and anger.

[0056] Optionally, the system can also use the NLU module to analyze whether there are time-urgent target words in the text, such as "immediately", "urgent", "within today", "right away", etc., and output the urgency weight at the text level.

[0057] Finally, the system weights and sums the emotion probability distribution output from the acoustic emotion analysis channel with the urgency weights output from the text semantic channel to obtain the final urgency level. For example, if the text is only a regular "Please send me a quote," but the acoustic features detect a significant increase in the speaker's volume, a significant increase in speaking speed, or the sum of the probabilities of "urging" and "anxiety" output by the voice emotion recognition classifier exceeds at least one of the preset reliability thresholds, the system will set the urgency level of the task to a high urgency level.

[0058] S140, based on the database of speakers and contacts corresponding to multiple statements, determine the target object referred to by the task executor, and identify the target object as the person responsible for the pending task. The target object includes at least one of one or more speakers and other contacts corresponding to multiple statements.

[0059] In practice, after obtaining the tasks to be done and the target words of the task executor expressed in the call text, and determining the speaker's identity for each statement by combining voiceprint features, the system's text analysis module and voiceprint recognition module collaborate across modalities to determine the actual person responsible for the tasks to be done.

[0060] In one example, if the target word for the task executor extracted in step S130 is a first-person pronoun, the system will identify the speaker who utters the statement corresponding to this target word as the person responsible for the task. First-person pronouns include, but are not limited to, "I" and "we". For example, if the two parties in a call are user Zhang San and colleague Li Si, and the system has already identified speaker A as Zhang San and speaker B as Li Si through voiceprint retrieval, during the call, if speaker A (Zhang San) utters the statement: "Then I'll send you the project proposal this afternoon," the system will identify the first-person pronoun "I" as the target word and, based on the aforementioned steps, confirm that the speaker who uttered this target word is speaker A. Therefore, the system will identify Zhang San as the person responsible for the task of "sending the project proposal."

[0061] In another example, if the target word for the task executor extracted in step S130 is a second-person pronoun, the system will automatically locate the other end contact in the current call who is speaking with the speaker who said the statement, and identify that other end contact as the person responsible for the task. For example, for a two-way call audio, the system directly identifies the other end speaker who said the statement as the person responsible for the task. As another example, for a multi-party call audio, the system, considering the context of the statement, identifies the speaker who is in conversation with the speaker who said the statement as the person responsible for the task. Second-person pronouns include, but are not limited to, "you" and "you all". Again, taking a two-way call between Zhang San and Li Si as an example, if speaker A (Zhang San) says the following statement: "Print out the contract tomorrow morning," the system identifies the second-person pronoun "you" as the target word, and based on the aforementioned steps, confirms that the speaker who said the target word is speaker A. Therefore, the system identifies the other end contact, Li Si, as the person responsible for the task of "printing the contract."

[0062] In another example, if the target word for the task executor extracted in step S130 is a specific name or a third-person pronoun, the system first compares the speakers corresponding to each statement, i.e., the current participants in the call. If the target object referred to by the target word matches the name of a speaker in the call, then that speaker is identified as the responsible person. If the target object referred to by the target word is not a current participant in the call, the system performs a quick search in the local contact database. If a matching contact is found, then that contact is identified as the responsible person. Taking a call between Zhang San and Li Si as an example, if speaker A (Zhang San) says the following statement during the call: "Have Wang Wu review the budget by tomorrow afternoon," the system identifies "Wang Wu" as the target word. The system first compares the current speaker A (Zhang San) with speaker B (Li Si) and determines that "Wang Wu" does not belong to the current call participants. Subsequently, the system searches and finds a contact record named "Wang Wu" in the local contact database. Finally, the system cross-modally maps the name target word "Wang Wu" in the text to the contact entity in the local contact database, identifies the entity as the target object, and identifies the contact "Wang Wu" in the database as the person responsible for the pending task of "reviewing the budget."

[0063] In some embodiments, when determining the target object based on multiple speaker and contact databases corresponding to different statements fails—for example, when the target object referred to by the target word does not exist in the local contact database, or when multiple contacts with the same name exist in the local contact database, making it impossible to accurately determine the target object referred to by the target word—the system marks the target word as a pending contact. Subsequently, the system outputs information requesting the user to confirm the identity of the pending contact. After receiving confirmation from the user or manual input, the target object indicated by the confirmation result is determined as the person responsible for the pending task. For example, when the target object referred to by the target word does not exist in the local contact database, the system outputs information requesting the user to manually input the actual identity of the target object. For example, when multiple contacts with the same name exist in the local contact database, the system outputs relevant information about the multiple contacts with the same name for the user to select and confirm. Thus, when the system cannot directly determine the actual identity of the target object, the human-machine collaborative confirmation mechanism avoids misjudging the person responsible for the pending task and improves the robustness of determining responsibility in complex call environments.

[0064] Optionally, if the target word for the task executor extracted in step S130 is a specific name or a third-person pronoun, the system also determines, based on the context of the text, whether the speaker has designated a temporary contact. A temporary contact refers to a contact provided by the call participants, containing their name or contact information, but not included in the local contact database. If the target object referred to by the target word matches the name of the temporary contact, the system temporarily adds the temporary contact to the contact database and designates the temporary contact as the responsible party. Furthermore, the system can also output a message asking whether to retain the temporary contact in the contact database for user confirmation.

[0065] It is understood that the above steps can be executed based on the inference results of the local model deployed on the terminal device, or based on the inference results of the large language model in the cloud server, without restriction.

[0066] In some embodiments, the method further includes: replacing privacy information in the text with corresponding de-identified placeholders locally on the terminal device to obtain de-identified text; sending the de-identified text to a cloud server so that a large language model on the cloud server can perform reasoning on the de-identified text to obtain a structured reasoning result expressed by the de-identified placeholders; receiving the structured reasoning result locally on the terminal device and backfilling the reasoning result according to the mapping relationship between the de-identified placeholders and privacy information to obtain execution suggestions for the pending task.

[0067] In this embodiment of the application, when event reasoning requires stronger large model capabilities, the system will anonymize the text in the call data and then upload the anonymized text to the cloud processor for event reasoning. For example... Figure 4 As shown, the system uses a local sensitive information recognizer to scan the text using a dual-channel approach based on rules and a NER model to automatically extract sensitive fields that belong to Personally Identifiable Information (PII). The rules channel identifies highly deterministic patterns such as phone numbers, email addresses, ID card numbers, and detailed addresses. The NER channel identifies context-sensitive entities such as names of people, organizations, and projects. After recognition, for each extracted PII item, the system generates a "original text-placeholder" mapping table locally. Subsequently, the system physically replaces all sensitive fields in the transcribed text with corresponding placeholders, generating desensitized text with all privacy entities removed. Only the placeholder text, structured fields, and necessary non-sensitive context are uploaded to the cloud. For example, the original call statement "Xiao Li, send the quote to Manager Zhang before 3 PM, his phone number is 13800000000" is desensitized and replaced with "[PERSON_1], send the quote to [PERSON_2] before 3 PM, his phone number is [PHONE_1]".

[0068] It is understandable that a Large Language Model (LLM) deployed on a cloud server has stronger reasoning capabilities than a local model. In this embodiment, the LLM receives anonymized context and cannot directly obtain the real name, phone number, or address. Furthermore, the cloud-based LLM matches the optimal execution flow in the knowledge base based on the identified task intent and automatically generates a structured reasoning result expressed by anonymized placeholders.

[0069] After receiving the structured reasoning result containing abstract placeholders from the cloud, the terminal device locally calls the "original text-placeholder" mapping table to replace the anonymized placeholders such as [PERSON_1], [PERSON_2], and [PHONE_1] in the structured reasoning result with their corresponding real personal identification information, thereby obtaining a complete planning result. Furthermore, the system can generate execution suggestions for to-do tasks based on the complete planning result and present them in card-style prompts. These execution suggestions can include the general flow and complete execution steps of the task, derived from the large language model reasoning. The system can output the complete planning result directly as the execution suggestion for the to-do task, or it can process the complete planning result to obtain the execution suggestion for the to-do task; there is no restriction.

[0070] Optionally, if a field is detected to be prohibited from being uploaded to the cloud during text scanning, such as a complete ID number or a precise home address, the system can directly choose to perform local small model inference or only output simplified suggestions, thus forming a hierarchical processing strategy of "allowed to upload to the cloud - restricted to upload to the cloud - prohibited to upload to the cloud".

[0071] Optionally, after the terminal device receives the structured reasoning results returned from the cloud, the system will also write "uploaded field type, call purpose, call time, and return summary hash" into the audit log.

[0072] According to the embodiments of this application, the terminal device can perform inference processing in conjunction with the model deployed in the cloud, and perform text desensitization processing before uploading data. It can utilize the stronger inference capabilities of the cloud model while ensuring user privacy, and realize intelligent services that combine privacy and security protection with an automated closed-loop experience.

[0073] S150, generate a prompt message, which includes the task to be done and the person responsible for the task.

[0074] This application does not limit the specific presentation method of the prompt information in its embodiments. In one example, the system can generate a single text notification, concisely presenting the task to be done and the person responsible for it. In another example, the system can write the prompt information to the system calendar of the terminal device, or to an application on the terminal device that records to-do items, such as a memo app. Figure 5 As shown, in another example, the system presents prompts in the user interface (UI) in the form of multimodal interactive cards. First, the card-based prompts may include a task title and a task content label. This section displays the core actions and objects of the task, such as structured action labels extracted from the call text through natural language understanding, like "organize contracts," "send quotes," or "buy bottled water," allowing users to intuitively understand the specific task.

[0075] Secondly, the card-style prompt information can include the responsible person's identification and physical association node, thus clearly indicating the person responsible for the task. When the person responsible for the task is included in the contact database, the system can integrate the contact's avatar, contact number, or instant messaging account entry in the local database, so that users can assign, follow up, or expedite tasks with one click.

[0076] Optionally, the card-style prompt message may also include links to trace the chain of evidence and play back the original audio segment. To facilitate user verification and prevent memory bias, the prompt message retains the original audio segment extracted from the event or its corresponding transcribed text sentence. Thus, by clicking the traceability button on the prompt card, the user can directly locate and play the original audio segment from the call recording, achieving highly reliable tracing.

[0077] Furthermore, after generating the reminder message, the system can combine the time and location of the task to be done with other conditions, and actively output the reminder message by popping up an interactive card, controlling the terminal device to vibrate, or controlling the terminal device to ring, thereby achieving more accurate reminders.

[0078] Optionally, after generating a reminder message, the terminal device can also synchronize the task information identified based on call data by adding the reminder message to the team calendar or enterprise management system, or by sending it to the person responsible for the task at the time via email or instant messaging software, thereby further improving the intelligence and humanization of the task reminder.

[0079] Continue to refer to Figure 5 When the system identifies the urgency of tasks, the prompt message can also include information indicating the urgency level of the task. Taking card-based prompts as an example, the system can directly display the identified urgency level as text, or indicate the urgency level by adjusting the font size or color of the text on the card, or by adjusting the order of the tasks on the card, thus presenting the urgency level of each task.

[0080] Furthermore, the system adjusts the prompt parameters of the prompt message based on the urgency of the task to be done. These parameters include at least one of the following: reminder priority, reminder frequency, and the vibration frequency or ringtone volume of the terminal device. In other words, the system can adjust the specific way the terminal device displays the prompt message in step S150 to the user based on the urgency of the task to be done.

[0081] For example, when the urgency level is determined to be normal, the system will set the reminder priority to low priority, generate a regular service card and display it statically and silently on the UI of the terminal device, or make a silent one-time pop-up in the notification bar without triggering a ringtone or physical vibration.

[0082] For example, when the urgency level is determined to be high urgency, the system sets the reminder priority to high priority and adjusts the prompt parameters mentioned above to ensure that the user is notified of the task as soon as possible, achieving closed-loop processing of the task. For instance, the system upgrades a single reminder to a progressively higher frequency reminder. Furthermore, if the user does not confirm the notification within a preset time after receiving it, the system will repeatedly trigger the push notification at preset time intervals (e.g., every 20 minutes) before the deadline. Another example is that the system actively sends a high-frequency drive command to the vibration motor of the terminal device while outputting the notification to increase the physical vibration frequency and intensity of the device. Yet another example is that the system actively increases the volume of the terminal device's speaker to a preset safe volume while outputting the notification, reducing the probability that the user will miss the task in noisy, active, or other adverse environments.

[0083] According to the embodiments of this application, by integrating acoustic physical features such as volume and speech rate with text semantics, accurate identification of the urgency of tasks to be performed is achieved. Furthermore, by dynamically adjusting prompt parameters such as vibration frequency and volume according to the urgency level, the probability of users missing highly urgent tasks due to insufficient prompts is effectively reduced, thus improving the intelligence level of the task closure loop.

[0084] In some embodiments, when the distance between the user's current location and the location of the task to be completed is less than a distance threshold, a prompt message is actively output.

[0085] like Figure 6 As shown in this embodiment, after identifying the target word representing the location of the task to be completed, the system can call the API to initiate a search to the local map service. The map service performs a geocoding inverse lookup on the target word to convert it into a Point of Interest (POI) with precise latitude and longitude coordinates. Subsequently, the system establishes a virtual geofence in the local background, centered on the latitude and longitude of the POI and using a preset distance value (e.g., 200 meters) as a distance threshold. During the user's movement, the system periodically calculates the distance between the user's current location coordinates and the center coordinates of the virtual geofence. When the calculated distance is less than the set distance threshold, the system confirms that the user's physical location has entered the virtual geofence range, and thus determines that the location trigger condition is met. At this time, the processor immediately sends a wake-up command to the device notification manager, and actively outputs the task to be completed and the responsible person's prompt information in the terminal device's UI in a multimodal form, such as a floating card, lock screen reminder, or top pop-up.

[0086] Let's take an example: During a phone call, Li Si tells Zhang San, "Please pick up the printed contract from the print shop next to the company on your way home." First, the NLU module in the system extracts the location target term as "print shop next to the company." Then, the system identifies the print shop as a POI type rather than a single place name, and uses the local map / POI service to retrieve the physical latitude and longitude coordinates of the "print shop" next to the company. Using these coordinates as the center and a distance threshold of 300 meters, a virtual geofence is constructed in the device's backend. Later, when Zhang San moves to a distance of 280 meters from the print shop on his way home from work, the device's positioning module detects that the current distance is less than the set 300-meter distance threshold. At this point, the system triggers a reminder mechanism, and an interactive card pops up on Zhang San's phone screen, displaying the message: "You have arrived near the print shop. Please pick up the contract Li Si asked you to pick up on your way home," accompanied by physical haptic feedback, thus accurately assisting Zhang San in performing the correct task at the correct physical location. In this way, by proactively constructing virtual geofences based on physical coordinates using the target word recognition results, the system intelligently binds to-do tasks to their real-world physical locations. Furthermore, by proactively triggering reminders when the user approaches the location of the task, the system further reduces the probability of the user missing tasks, contributing to improved efficiency and intelligence in the task completion process.

[0087] Optionally, during the creation of a virtual geofence, the system can also generate geofence parameters by integrating the user's commuting trajectory, current location, direction of travel, historical stop points, and effective task time period. Exemplary geofence parameters include one or more of the following: center latitude (center_lat), center longitude (center_lng), radius (radius), valid time window (valid_time_window), direction constraint (direction_constraint), and dwell time threshold (dwell_threshold). The direction constraint (direction_constraint) limits the triggering of alerts to only when the user is in a specific direction of travel (e.g., the "homeward direction"). The dwell time threshold (dwell_threshold) sets the minimum static or slow dwell time required to trigger an alert, preventing repeated alerts when the user is passing through the area at high speed or briefly. Thus, by adding constraints based on user habits when creating a geofence, the probability of blindly triggering distractions in inappropriate situations or when the user is unable to perform their task is reduced, improving the accuracy of location alerts.

[0088] Optionally, after confirming that a user has entered the virtual geofence, the system does not immediately issue an indiscriminate alert. Instead, it performs a secondary filtering and verification of the alert event based on at least one of the following criteria: whether the task is of high urgency, whether the current time falls within the task's valid window, whether the task is not yet completed, whether a similar alert has already been triggered within the set time, and whether the user is in a state unsuitable for interaction, such as driving. Only after passing the filtering will the system proactively trigger the alert. This effectively reduces unnecessary disturbances to users caused by invalid alerts, further improving the intelligence and user-friendliness of task reminders.

[0089] Optionally, before generating the prompt message, method 100 further includes the following steps: S145, compare the time of the pending task with the existing schedule in the system calendar. When the time of the pending task conflicts with the existing schedule, output suggested information based on the available time slots in the system calendar. The suggested information includes solutions for adjusting the time of the pending task and / or the time of the existing schedule.

[0090] like Figure 3 As shown, the system can also perform calendar conflict detection before generating a prompt message. In actual implementation, after the system identifies the time of the task to be done in step S130, the system reads the system calendar database through the Application Programming Interface (API). The system retrieves existing scheduled events that overlap with the time interval of the task to be done, and extracts the start and end timestamps of the existing events. The system determines whether there is a conflict by calculating the intersection of the two time intervals. If the intersection is empty, it is determined that there is no conflict, and a regular prompt message is generated directly. If the intersection is not empty, it is determined that the current task to be done has a time conflict with an existing event, and the system will automatically mark the conflicting event and trigger an automatic rescheduling process.

[0091] Specifically, the system retrieves the user's scheduling preferences, defines the daily allowed time window for inserting tasks, and searches the system calendar for available time slots to adjust the timing of tasks. For example, if a quote is sent at 3 PM, and a schedule comparison reveals a meeting at 2:30 PM, the system outputs a reordering result based on three constraints: "available time slot length before deadline," "estimated task duration," and "minimum reminder interval." If the estimated quote duration is 20 minutes, a strategy can be generated to send the first reminder at 1:30 PM, a second reminder at 2:00 PM, and a final reminder at 2:20 PM.

[0092] Optionally, the system compares the priority levels of the new task with existing schedules and generates a plan to adjust the timing of the task and / or the existing schedule. For example, if the existing schedule has a higher priority than the new task, the system selects the optimal empty time slot during idle periods that meets the duration requirements of the new task and generates a first adjustment plan to move the new task to that empty time slot. Alternatively, if the new task has a higher priority than the existing schedule, the system finds an alternative time slot for the existing schedule during idle periods and generates a second adjustment plan to shift the existing schedule while keeping the new task running at its original time. Furthermore, if the priorities are the same, the system calculates a third adjustment plan involving two-way fine-tuning of the schedule, such as shortening or slightly advancing the existing meeting schedule by 30 minutes while delaying the new task by 30 minutes to accommodate both events. Further, the system integrates one or more of the generated adjustment plans and outputs suggested information.

[0093] After generating the suggested information, the system outputs the suggested information for the user to select and confirm. After the user confirms, the system generates a reminder message based on the updated time of the to-do task and adjusts the schedule in the system calendar, thereby achieving dynamic adjustment of the schedule through human-computer collaboration.

[0094] Optionally, the system can also dynamically adjust the prompt parameters based on the current time and the time of the task to be completed, thereby achieving time-based differentiated reminders. For example, if the difference between the current time and the time of the task to be completed is greater than a time threshold, the system sets the reminder priority to low priority and adjusts the reminder parameters to those corresponding to low priority. If the difference between the current time and the time of the task to be completed is less than or equal to the time threshold, the system sets the reminder priority to high priority and adjusts the reminder parameters to those corresponding to high priority.

[0095] According to the embodiments of this application, an intelligent time conflict retrieval and adaptive scheduling mechanism is established by comparing and dynamically adjusting the time of pending tasks with existing schedules, thereby improving the intelligence level of the task closure loop.

[0096] Optionally, after generating the prompt message, method 100 further includes the following steps: S155 assigns a task identifier to the pending task and dynamically updates the execution status corresponding to the task identifier based on the data related to the task identifier. The data related to the task identifier includes email data and instant messaging data.

[0097] like Figure 7As shown in this embodiment, after the system generates a prompt message based on the to-do task locally, it assigns a globally unique task identifier (Task Identifier, Task ID) to each to-do task. Simultaneously, the task-level aggregator in the system also acquires data related to the to-do task from the terminal device's local storage or other application services, integrates and processes the data, and generates an indication graph based on the task identifier. In subsequent processing, the system uses this task identifier as an index to manage information related to the to-do task.

[0098] In practice, after obtaining user authorization, the system can detect changes in the drafts and sent items databases of local email clients, as well as updates to the local message databases of commonly used instant messaging software. When new email sending data or instant messaging chat logs are detected, the system determines whether the recipient's address matches the email address of the person responsible for the task, or whether the account of the instant messaging chat partner matches the instant messaging account of the person responsible for the task. If the determination is successful, the system automatically adds these newly appearing email bodies, email attachments, and instant messaging text chat fragments as evidence of the subsequent execution of the task to the database node corresponding to the task identifier, forming a complete chain of evidence for the evolution of events across telephone, email, and instant messaging.

[0099] Specifically, the system can establish evidence nodes from data from different sources, forming a structure of "main task table + multi-source evidence table". For example, telephone evidence can record call time, evidence fragments, and speaker tags. Email evidence can record subject, sender, attachment name, and sending time. Instant messaging software evidence can record group / one-on-one chat source, message body, and last update time. Thus, the aggregator can combine topic similarity, contact overlap, time proximity window, and attachment naming rules to determine whether different sources correspond to the same task. That is, a user receives a "send quote" task over the phone, then receives a quote template in an email, and then receives supplementary requests from the client in instant messaging software. The system aggregates these three under the same task identifier. When the user calls again, the system can use the task identifier to trace back the email attachment name, the latest status of the instant messaging chat, and the conclusion of the last call.

[0100] Furthermore, the system analyzes the database corresponding to the task identifier and dynamically updates the execution status of the task identifier. Taking the execution status as including Created, In Progress, Completed, and Overdue as an example: The system initially sets the execution status of the task identifier to Created. When it detects that the user has started writing an email draft strongly related to the task identifier, or has sent a preliminary communication message containing project keywords to the responsible person in instant messaging software, the system determines that the task has started and automatically updates the execution status of the task identifier to "In Progress." When the system detects that the user has successfully sent an email containing the target file to the responsible person's email address, or has sent a message confirming completion to the responsible person in instant messaging software, the system determines that the task has been completed and automatically changes the execution status of the task identifier to "Completed." Optionally, the system can simultaneously mark the item as completed in the UI interface and send a completion notification to the user.

[0101] Optionally, when processing related calls again, such as calls from the person responsible for the task or the recipient, the system does not cram all historical content into the context. Instead, it performs targeted retrieval based on the task identifier, extracting four types of summary information—"latest status," "key attachments," "incomplete subtasks," and "most recent conclusion"—and injecting them into the current inference. This maintains cross-cycle memory while avoiding increased inference costs due to excessively long contexts.

[0102] According to the embodiments of this application, by assigning a unique task identifier to the to-do tasks extracted based on call data and establishing a multimodal task fusion mechanism, fragmented information from multiple channels can be automatically aggregated into the same task chain, reducing the information retrieval and maintenance costs for users in multi-application collaborative scenarios and improving the intelligence level of task closed-loop processing.

[0103] S160: Based on the type of the task to be done, a corresponding execution step chain is generated from a preset step library. Based on the execution step chain, the local application software required to execute the task is automatically opened on the terminal device.

[0104] like Figure 8 As shown, after extracting the target words for the tasks to be done, the system can further classify the tasks using multiple tags to predict their corresponding task types. For example, if the extracted action and object are "sending a quote" or "issuing an invoice," the system predicts and classifies it as a first-type task. If the extracted action and object are "organizing contracts" or "drafting agreements," the system predicts and classifies it as a second-type task. If the extracted action and object are "booking flights" or "confirming schedules," the system predicts and classifies it as a third-type task.

[0105] After determining the task type, the terminal device's backend service calls the local preset step database. This database pre-stores mappings between various common task types and standard execution flows. The system retrieves and matches the corresponding standard execution step chain. In actual implementation, the system first calculates a short-term priority score based on the candidate item set. Scoring factors may include: deadline proximity, item value, dependency on other items, suitability of the current location for execution, length of the user's current idle time slot, and historical completion probability. The prediction module outputs a ranking of A / B / C candidate items for the next half hour, while the step planner further selects the corresponding execution chain from the local preset step database. For example, the first type of task can be mapped to opening a template - completing a document - sending an email - callback confirmation.

[0106] Furthermore, the system can sequentially open target application software based on the matched execution step chain and send specific startup parameters to the application software. For example, when launching a local email application, it automatically passes the email address of the person responsible for the task identified in the call text to the sender or recipient parameter slot.

[0107] Optionally, the system can also utilize accessibility services or secure APIs between applications to automatically pre-populate the structured slot data extracted by NLU into the populated application form, thereby achieving automatic data filling across applications.

[0108] Continue to refer to Figure 8 In this embodiment, the system also supports further analysis of pending tasks, providing users with operational suggestions for executing them. For example, the system prioritizes tasks based on their history, deadline, current location, and available time slots, thus providing suggestions such as "A / B / C should be completed first in the next half hour." For instance, if a quotation task has only 90 minutes left before its deadline and the user is still in the office, the system prompts: 1) Open the quotation template; 2) Complete the price; 3) Send an email; 4) Automatically call back confirmation in 30 minutes. Furthermore, after the step chain is generated, the system can either only display suggestions or automatically trigger pre-actions based on permissions, such as opening the corresponding template document, popping up an email draft, or setting a callback reminder in 30 minutes. If the confidence level of the highest-rated task is below a threshold, the system can only display a "recommended candidate task list" without directly executing it automatically.

[0109] Optionally, if a user repeatedly rejects the execution suggestion for a certain type of task, the feedback can be written into the user preference model to reduce the frequency of similar triggers in the future.

[0110] It's understandable that when multiple tasks exist in the system calendar, the system can integrate the execution steps of these tasks and provide cross-reminders for each step in chronological order. For example, if the system calendar contains Task 1 and Task 2, where Task 1 includes steps A1 and A2, and Task 2 includes steps B1 and B2, the system can output that Task 1 and Task 2 should be executed in the following order (A1, B1, B2, A2) within the next half hour.

[0111] According to the proposed solution, combined with a pre-defined step library, the system can automatically generate an execution step chain based on the task type identified from call data, and automatically launch local application software based on the execution step chain. This simplifies the user's operation process for performing tasks through the device, and improves the device's operational efficiency and intelligent experience in complex office and call scenarios.

[0112] It is understood that the embodiments of this application do not restrict the order of system execution steps S155 and S160.

[0113] Figure 9 This document illustrates a flowchart of a process for providing notifications based on call data, as provided in an embodiment of this application. Table 2 shows a schematic diagram of the inputs, intermediate states, and outputs in the process. Figure 9 As shown, the process includes the following steps: S201. Collect call audio and verify whether the user has authorized recording, transcription, event extraction, and cross-application linkage. This step is performed by... Figure 10 The input layer is executed.

[0114] S202. Perform automatic speech recognition and speaker separation on the call audio to obtain transcribed text with speaker tags.

[0115] S203. Perform time recognition, location recognition, action phrase recognition, object recognition, and constraint condition recognition on the transcribed text to obtain a structured semantic frame.

[0116] S204. The speech acoustic features and text emotion features are fused to output the urgency level and reminder priority.

[0117] S205. Determine the responsible person based on voiceprint features and contact mapping. Steps S202 to S205 above are... Figure 10 The perception and understanding layer in the process is used to implement functions such as speech recognition, natural language understanding, tone / emotion recognition, and voiceprint recognition.

[0118] S206. Generate an event object, which includes at least the event name, responsible person, source call, deadline, location, urgency, confidence level, and evidence fragments. Table 1 shows an example of an event object.

[0119] S207. Perform conflict detection between the task object and the existing calendar / to-do list to obtain recommended time slots or rearrangement suggestions.

[0120] S208. If the policy allows, automatically create calendar events, to-do items, alarms, and notifications, and generate a reminder policy. Table 1 shows an example of a reminder policy. Steps S206 to S208 above are... Figure 10 The decision orchestration layer is executed in the middle, and it is used to realize functions such as task extraction, conflict detection and scheduling, reminder strategy generation, and action suggestions and step planning.

[0121] S209. Aggregate call summaries, emails, IMs, and calendar events related to the matters at the event level to form cross-round contextual memory. This step is performed by... Figure 10 The memory and security layer is executed within the framework, and it is used to implement functions such as cross-round context, privacy anonymization, policies and auditing, and knowledge base and rule base. The knowledge base and rule base contain user preferences and enterprise rules stored locally or in the cloud. Table 1 shows an example of a context memory entry.

[0122] S210. When time, location, or state trigger conditions are met, proactively output a reminder, step suggestions, and subsequent predicted actions. This step is initiated by... Figure 10 The output execution layer is executed in the process. The output execution layer is used to present prompt information through calendar time, to-do list, alarm / notification, operation suggestions and location-triggered reminders.

[0123] Table 1

[0124] Table 2

[0125] It is understood that the specific implementation methods of the above steps S201 to S210 can be referred to the relevant content above, and will not be repeated here.

[0126] This application provides a system for providing event reminders based on call data. The system includes a call interface module, a speaker separation module, a natural language understanding module, a responsible person determination module, and an output execution module.

[0127] The call interface module is used to acquire call data, including call audio and corresponding text. Specifically, the call interface module uses sound acquisition hardware, such as a microphone and related amplification circuit in the terminal device, to acquire the call audio of the local user, and uses a baseband processor and a wireless radio frequency chip to acquire call audio from other speakers through a communication network. Furthermore, when the initial input of the system only includes call audio, the call interface module also calls an ASR model to transcribe the call audio into corresponding text. This ASR model is stored in the terminal device's memory and runs in the Central Processing Unit (CPU) and / or a dedicated Neural Processing Unit (NPU).

[0128] The speaker separation module is used to determine the speaker corresponding to multiple sentences in the text based on call audio and voiceprint features. Specifically, at the hardware level, the speaker separation module calls the NPU to provide computational support for the algorithms and models used, and caches the extracted voiceprint feature vectors in the terminal device's memory, i.e., Random Access Memory (RAM). At the software level, the speaker separation module uses the VAD algorithm, OSD algorithm, and a local model for voiceprint extraction to determine the speaker corresponding to each sentence in the text, and optionally performs contact identity matching for the speakers involved in the call through a vector retrieval database. Detailed explanations of the aforementioned algorithms and models can be found in the relevant content above, and will not be repeated here.

[0129] The Natural Language Understanding (NLP) module analyzes text and extracts target words representing tasks and executors. The models used in this module can be deployed on either terminal devices or cloud servers. When deployed on a terminal device, at the hardware level, the NLP module utilizes the CPU or NPU for computation and loads the NLU model into memory. At the software level, it extracts target words representing tasks and executors from the text using NLU, NER, and SRL models. When deployed on a cloud server, at the hardware level, the NLP module uses the CPU to schedule algorithms for text anonymization and backfilling, transmits the anonymized text to the cloud server via baseband processors and RF chips, and stores a mapping table of personal information and anonymized placeholders in memory. At the software level, it anonymizes the text using NER models and anonymization algorithms. Specific explanations of the aforementioned models can be found in the relevant sections above and will not be repeated here.

[0130] The responsible party determination module is used to identify the target object referred to by the task executor based on the speaker and contact databases corresponding to the multiple statements, and to determine the target object as the responsible party for the pending task. The target object includes at least one of the one or more speakers and other contacts corresponding to the multiple statements. Specifically, at the hardware level, the responsible party determination module calls the CPU to execute the required algorithms and allocates memory to store the contact database. At the software level, the responsible party determination module employs relevant algorithms to achieve cross-modal data alignment and target object identification.

[0131] The output execution module generates prompts, including the task to be done and the person responsible for it. Specifically, at the hardware level, the output execution module displays the prompts via a UI on a screen, plays ringtones via a speaker circuit, provides tactile vibration cues via a vibration motor, and collects the terminal device's current location via a positioning chip such as a Global Positioning System (GPS) chip. At the software level, the output execution module uses API calls to applications such as memos, smart notification centers, email, and instant messaging software to achieve cross-modal application collaboration and status synchronization of the prompts, and optionally calls local map applications to establish virtual geofences.

[0132] The call interface module may be included in Figure 10 The input layer, speaker separation module, natural language understanding module, and responsible person determination module can be included in... Figure 10 The perception and understanding layer in the middle, the output execution module can be included in Figure 10 The decision-making and scheduling layer in the process.

[0133] This application also provides an electronic device for providing reminders based on call data. The electronic device includes: a memory for storing a computer program; and a processor for executing the computer program stored in the memory, so that the device performs the above-described method 100 for providing reminders based on call data.

[0134] This application also provides a computer-readable storage medium storing program code that, when executed on a computer, causes the computer to perform the above-described method 100 for providing event reminders based on call data.

[0135] This application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute the above-described method 100 for providing event reminders based on call data.

[0136] This application embodiment also provides a device for providing event reminders based on call data, the device including a unit for the above-described method 100 for providing event reminders based on call data.

[0137] Although this application has been described with reference to preferred embodiments, various modifications can be made thereto and components or steps can be replaced with equivalents without departing from the scope of this application. In particular, the technical features mentioned in the various embodiments can be combined in any manner, provided there is no conflict in structure or method steps. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for providing event reminders based on call data, characterized in that, include: Acquire call data, which includes call audio and corresponding text; Based on the call audio and voiceprint features, the speaker corresponding to each of the multiple sentences in the text is determined; Analyze the text and extract target words that represent pending tasks and task performers. Based on the spokesperson and contact databases corresponding to the multiple statements, the target object referred to by the task executor is determined, and the target object is identified as the person responsible for the task to be done. The target object includes at least one of the spokespersons corresponding to the multiple statements and other contacts in the contact database. Generate a prompt message, which includes the task to be done and the person responsible for the task.

2. The method according to claim 1, characterized in that, The target object also includes potential contacts, and the method further includes: When it fails to determine the target object based on the speaker and contact databases corresponding to the multiple statements, the target object is marked as a pending contact. Output information used to request the user to confirm the identity of the pending contact person; Receive the user's confirmation of the identity of the pending contact person, and determine the person responsible for the pending task based on the confirmation result.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Extract the volume features and / or speech rate features of the sentences containing the target word in the call audio; The urgency of the task expressed by the target word is determined based on the volume characteristics and / or speech rate characteristics. The prompt parameters of the prompt information are adjusted according to the urgency of the task to be done. The prompt parameters include at least one of the following: reminder priority, reminder frequency, and vibration frequency or ringtone volume of the terminal device.

4. The method according to any one of claims 1 to 3, characterized in that, The target word also includes the time of the task to be done. Before generating the prompt message, the method further includes: Compare the time of the task to be done with the existing schedule in the system calendar; When the time of the task to be done conflicts with the time of the existing schedule, the system outputs suggested information based on the available time period in the system calendar. The suggested information includes a plan to adjust the time of the task to be done and / or the time of the existing schedule.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Analyze the text and extract target words that express the location of the task to be done; When the distance between the user's current location and the location of the task to be completed is less than a distance threshold, the prompt message is actively output.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Based on the type of the task to be done, a corresponding execution step chain is generated by matching from the preset step library; According to the execution step chain, the local application software or document template required for the task to be done is automatically opened on the terminal device.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Replace the privacy information in the text with the corresponding desensitized placeholders to obtain the desensitized text; The de-identified text is sent to the cloud server; The system receives the structured reasoning result expressed by the de-identified placeholders returned by the cloud server, and fills in the reasoning result according to the mapping relationship between the de-identified placeholders and the privacy information to obtain the execution suggestion for the task to be done.

8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: Assign a task identifier to the pending task; The execution status corresponding to the task identifier is dynamically updated based on the data associated with the task identifier, which includes email data and instant messaging data.

9. A system for providing event reminders based on call data, characterized in that, include: The call interface module is used to acquire call data, which includes call audio and corresponding text. The speaker separation module is used to determine the speaker corresponding to each of the multiple sentences in the text based on the call audio and voiceprint features. The natural language understanding module is used to analyze the text and extract target words from the text that represent tasks to be done and task executors. The responsible person determination module is used to determine the target object referred to by the task executor based on the spokesperson and contact databases corresponding to the multiple statements, and to determine the target object as the responsible person for the task to be done. The target object includes at least one of one or more spokes and other contacts corresponding to the multiple statements. The output execution module is used to generate prompt information, which includes the task to be done and the person responsible for the task.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 8, or to implement the system as described in claim 9.