Speech separation and paragraph attribution method and system for teleconference scenarios

Abstract

This invention discloses a method and system for speech separation and segment attribution in remote conferencing scenarios, relating to the field of speech recognition technology. The method includes: triggering speech recording in a target meeting window via a conferencing platform to generate meeting audio; performing two-way speech separation processing on the meeting audio using a first engine to construct a speech logic graph; matching a target meeting template using a second engine and performing speech segment attribution based on automated meeting record layout according to the speech logic graph to generate a target meeting record; and storing the target meeting record in a platform database for retrieval and management using access settings. This invention solves the technical problem in existing multi-party conferencing systems where speech interleaving makes it difficult to effectively distinguish speakers and the logical structure of their statements, leading to unclear meeting record attribution and low editing efficiency. It achieves two-way semantic logic separation and structured attribution of remote meeting speech content, improving the accuracy of meeting records and the efficiency of automatic layout.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and more specifically to a method and system for speech separation and segmentation in remote conferencing scenarios. Background Technology

[0002] With the increasing prevalence of modern remote work and collaboration, remote conferencing has become an important form of team communication, project reporting, and business negotiation. However, in meeting scenarios with multiple participants, due to overlapping, intersecting, and frequent switching of speakers, traditional speech recognition and transcription systems often struggle to accurately distinguish the speaker's identity and corresponding semantic structure. This is especially true in complex contexts such as controversial discussions and clashes of viewpoints, where the attribution and logical division of audio content become even more confusing. This not only affects the completeness and readability of meeting minutes but also significantly increases the burden of manual organization and archiving afterward. Summary of the Invention

[0003] This application provides a method and system for speech separation and paragraph attribution in remote conferencing scenarios, which solves the technical problem in existing technologies where speech interleaving in multi-party conferences makes it difficult to effectively distinguish speakers and the logical structure of their speech content, resulting in unclear attribution of meeting records and low editing efficiency.

[0004] The first aspect of this application provides a method for speech separation and segment attribution in remote conferencing scenarios, the method comprising:

[0005] As multiple users join the meeting, the meeting platform triggers voice recording of the target meeting window, generating meeting audio. The voice processing engine matches and temporarily connects to the target meeting window. Based on the temporarily connected first engine, two-way voice separation processing is performed on the meeting audio to construct a voice logic graph. This two-way voice separation processing includes constructing a first logic graph based on the claim-argument relationship and a second logic graph based on the main line-point of contention. Based on the temporarily connected second engine, the target meeting template is matched, and voice segment attribution based on automated meeting record layout is performed according to the voice logic graph to generate the target meeting record. The target meeting record is stored in the platform database and managed for retrieval using access permissions.

[0006] A second aspect of this application provides a speech separation and segmentation system for remote conferencing scenarios, the system comprising:

[0007] Audio Generation Module: As multiple users join the meeting, the meeting platform triggers voice recording of the target meeting window, generating meeting audio. Voice Processing Module: Performs voice processing engine matching and temporary access to the target meeting window. Based on the temporarily accessed first engine, it performs two-way voice separation processing on the meeting audio, constructing a voice logic graph. The two-way voice separation processing includes constructing a first logic graph based on the claim-argument relationship and a second logic graph based on the main line-point of contention. Record Generation Module: Based on the temporarily accessed second engine, it matches the target meeting template and, based on the voice logic graph, performs voice segment attribution based on automated meeting record layout, generating the target meeting record. Management Module: Stores the target meeting record in the platform database and manages its retrieval through permission settings.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] As multiple users join the meeting, the conferencing platform triggers voice recording of the target meeting window, generating meeting audio. Next, the voice processing engine matches and temporarily accesses the target meeting window. Based on the temporarily accessed first engine, two-way voice separation processing is performed on the meeting audio to construct a voice logic graph. This two-way voice separation processing includes constructing a first logic graph based on the claim-argument relationship and a second logic graph based on the main line-point of contention. Then, based on the temporarily accessed second engine, the target meeting template is matched, and voice segment attribution based on automated meeting record formatting is performed according to the voice logic graph, generating the target meeting record. Finally, the target meeting record is stored in the platform database and managed for retrieval through access control. This solves the technical problem in existing technologies where voice interweaving in multi-party meetings makes it difficult to effectively distinguish speakers and the logical structure of their statements, leading to unclear meeting record attribution and low editing efficiency. It achieves two-way semantic logic separation and structured attribution of remote meeting voice content, improving the accuracy of meeting records and the efficiency of automatic formatting. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic flowchart of a speech separation and segment attribution method for remote conferencing scenarios provided in an embodiment of this application;

[0012] Figure 2This is a schematic diagram of the speech separation and segment attribution system for remote conferencing scenarios provided in an embodiment of this application.

[0013] Explanation of reference numerals in the attached diagram: Audio generation module 11, Speech processing module 12, Recording generation module 13, Management module 14. Detailed Implementation

[0014] This application provides a method and system for speech separation and paragraph attribution in remote conferencing scenarios, which solves the technical problem in existing technologies where speech interleaving in multi-party conferences makes it difficult to effectively distinguish speakers and the logical structure of their speech content, resulting in unclear attribution of meeting records and low editing efficiency.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] Example 1, as Figure 1 As shown, this application provides a method for speech separation and segment attribution in remote conferencing scenarios, wherein the method includes:

[0018] As multiple users join the meeting, the meeting platform triggers voice recording in the target meeting window and generates meeting audio.

[0019] Before a remote meeting begins, the conferencing platform establishes a meeting session instance and generates a unique target meeting window identifier. As multiple user clients join the meeting through this target meeting window, the platform detects the access status of these clients in real time and triggers the initialization process of the meeting audio acquisition module. Specifically, the platform uses its built-in audio acquisition engine to receive, synchronize, and preprocess the audio stream data uploaded by participating user clients, and timestamps and channels all incoming audio signals. After preprocessing the audio data stream, the platform merges the various audio streams into a unified meeting audio data stream and continuously records this stream throughout the meeting to form a complete original meeting audio.

[0020] The speech processing engine is matched and temporarily accessed by the target conference window. Based on the temporarily accessed first engine, two-way speech separation processing is performed on the conference audio to construct a speech logic graph. The two-way speech separation processing includes the construction of a first logic graph based on the claim-argument relationship and the construction of a second logic graph based on the main line-point of contention.

[0021] After the conference audio is generated, the conference platform, based on the access status and conference window identifier, dynamically matches idle speech processing engine groups by traversing the deployed speech processing engine array to determine the target engine group adapted to the target conference window. The conference platform establishes a temporary logical connection between the target speech processing engine group and the target conference window through system protocols, enabling the engine group to access and process the conference audio stream in real time. After establishing the connection, the first engine in the matched engine group is invoked to perform two-way speech separation processing. Specifically, the first engine performs semantic reconstruction and deep separation based on the pre-separated speech segments annotated in the conference audio (such as speech streams initially separated based on voiceprints or speaker channels).

[0022] Two-way speech separation processing includes:

[0023] Claim-Argument Relationship Extraction: By using natural language understanding algorithms to identify structural elements such as arguments, reasons, and supplementary explanations in the speech content, a first logical graph is constructed with "claim nodes" as the center and "argument edges" as the connection, which is used to express the argument structure and logical chain of each speaker.

[0024] Main thread-point structure analysis: Identify the main thread of the topic and points of conflict or disagreement between speakers, and construct a second logical graph with "main thread nodes" as the core and "points of disagreement" as the links to reflect the discussion direction and disagreement structure throughout the meeting.

[0025] Furthermore, the construction of the speech processing engine array before performing matching by the speech processing engine includes:

[0026] A first-class engine is deployed based on speech separation and reconstruction, and a second-class engine is deployed based on speech typesetting. The first-class engine and the second-class engine are combined to form a speech processing engine group. The speech processing engine group is subjected to logical supervised training to determine a first engine subarray. The first engine subarray is multi-bit mirrored twinned to form a speech processing engine array, and the speech processing engine array is embedded and deployed in the conference platform.

[0027] Before performing speech processing engine matching, the speech processing engine array is pre-constructed and deployed. Specifically, two types of dedicated speech processing engines are constructed according to their functions: one type of engine is dedicated to performing the separation and reconstruction of conference speech, possessing capabilities such as speaker recognition, semantic segmentation, and speech structure extraction; the other type of engine is used for attribution inference and layout mapping of processed speech segments, possessing capabilities such as semantic attribution judgment, template matching, and layout generation. The first and second types of engines are combined in a one-to-one pairing manner to form multiple speech processing engine groups, each containing a complete set of pre-processing (separation) and post-processing (attribution) capabilities. Multiple speech processing engine groups undergo unified logical supervised training. The training process is based on historical conference datasets, using labeled claim / dispute data and the attribution relationship between speech segments and conference records to iteratively optimize model parameters. Through this training process, several engine groups with the best performance are selected as the first engine subarray for subsequent dynamic task allocation. After the first engine subarray is built, the platform performs multiple mirrored twin replications of it. This involves building multiple mirrored engine instances with identical structure and function on different physical or virtual computing nodes, and uniformly managing their scheduling and load. Finally, all mirrored engines are combined into a speech processing engine array, which is then deployed in an embedded manner in the conferencing platform server environment to ensure high availability, low latency, and high parallelism speech processing capabilities in multi-concurrency, multi-user conferencing scenarios.

[0028] Furthermore, for multi-concurrent conferences, the platform's computing power is dynamically balanced and allocated to the voice processing engine array.

[0029] To address scenarios where the platform simultaneously hosts multiple concurrent conference instances, and to ensure stable execution of voice processing tasks under high concurrency, the conference platform employs a dynamic power balancing strategy for the voice processing engine array. Specifically, the platform uses a resource monitoring module to obtain real-time operational status data for each voice processing engine node, including current CPU utilization, memory usage, engine load metrics, and processing queue length. Simultaneously, it monitors the number of concurrent conferences, the amount of voice data in each conference, and the processing priority.

[0030] When multiple concurrent meetings are detected, the platform invokes the resource scheduling unit to allocate instances of each sub-engine in the engine array based on scheduling strategy models (such as weighted minimum load, resource affinity scheduling, or priority-based task distribution strategies). When an engine group reaches a preset load threshold, the platform automatically allocates new meeting instances to engine groups with lower loads or in standby mode. It also supports dynamic scaling or task migration of mirror twin nodes to achieve a balanced distribution of overall platform computing resources. Furthermore, the platform can flexibly adjust the allocation strategy of the voice processing engines based on meeting priority, user permissions, or real-time performance requirements. For example, it can prioritize reserving high-performance node resources for key users or high-level meetings to ensure the stability and timeliness of voice processing.

[0031] Furthermore, the matching process performed by the speech processing engine includes:

[0032] As multiple users join the meeting, the speech processing engine array is traversed, and idle engine subarrays are matched to determine the target speech processing engine. According to the platform interaction protocol, a temporary connection is established between the target speech processing engine and the target meeting window, wherein the connection is disconnected when the meeting ends.

[0033] After multiple user terminals access the target conference window, the platform automatically triggers the voice processing resource scheduling process, executing a voice processing engine matching operation. Specifically, it traverses the currently deployed voice processing engine array, identifying idle or accessible engine subarrays based on the operating status indicators of each engine subarray (such as CPU load, memory usage, processing thread ratio, and number of idle channels). Combining this with parameters such as the current conference's voice data scale and processing priority, it performs an adaptation score for candidate engine groups to determine the most suitable target voice processing engine group. According to the preset platform interaction protocol, the selected target voice processing engine group is temporarily logically bound to the target conference window, establishing a task processing session channel. This task processing session channel can achieve bidirectional communication and status synchronization based on communication protocols such as RESTful API, WebSocket, or gRPC, and supports automatic authentication and task status feedback mechanisms. After the temporary connection is established, the real-time audio stream captured by the conference window will be automatically forwarded to the bound engine for processing.

[0034] During the meeting, the target speech processing engine continuously receives and processes the audio stream from the target meeting window, and synchronizes the structured output results to the platform database after each round of speech analysis and processing. When a meeting end event is detected (such as all users leaving the meeting, the host closing the window, etc.), the platform will automatically trigger an engine unbinding operation, disconnect the temporary connection between the target speech processing engine and the meeting window, and reset the engine resource status to "idle" for subsequent task scheduling.

[0035] Furthermore, before performing two-way speech separation processing on the conference audio, the following steps are included:

[0036] According to the voiceprint recognition plugin, the conference audio is subjected to voiceprint recognition and separation to determine the voiceprint separation speech, wherein the separation standard is based on the multiple speaking parties; for the multiple user terminals, multiple party identification codes are generated to annotate the voiceprint separation speech and determine the pre-separated speech.

[0037] Before performing two-way speech separation processing on the conference audio, the platform needs to perform preliminary voiceprint recognition and speech annotation to achieve basic differentiation of the speech content of multiple users. Specifically, the conference platform calls the built-in voiceprint recognition plugin, which uses a deep neural network (such as a CNN-BLSTM structure or an ECAPA-TDNN voiceprint encoder) to extract and recognize voiceprints from the real-time acquired conference audio. The conference platform performs temporal segmentation of the entire conference audio, extracts the feature vector of each speech segment, and compares it with the user voiceprint feature library pre-stored in the platform's database to identify the user identity of the speaker.

[0038] After identifying the speakers, the platform uses the speakers from multiple users as the separation criterion. Based on interval clustering of voiceprint features and time segmentation rules, it divides the original conference audio into several voiceprint-separated speech segments, each clearly corresponding to a specific speaker. Simultaneously, the conference platform generates a unique multi-party identifier (e.g., a combination of conference ID and user ID) for all participants and embeds this identifier into the metadata of each voiceprint-separated speech segment. Finally, after annotating all the voiceprint-separated speech segments, the pre-separated speech is obtained.

[0039] Furthermore, performing two-way speech separation processing on the conference audio includes:

[0040] The first engine is triggered, and the pre-separated speech is reconstructed by taking the argument items as the identification target and the claim-evidence as the construction elements to determine the first logical graph; the pre-separated speech is reconstructed by taking the main line-point of contention as the construction elements to determine the second logical graph; the first logical graph and the second logical graph are added to the speech logical graph.

[0041] By triggering the first engine, which identifies argument entries as the target, the semantic structure of each segment of speech in the pre-separated speech is analyzed. Based on a natural language understanding model (such as BERT+BiLSTM-CRF or a Transformer-based structured semantic recognition network), the first engine identifies elements such as assertion sentences, supporting arguments, rebuttal information, and data citations in each speech segment, and constructs a first logical graph according to the semantic dependencies between assertions and arguments. In the first logical graph, each node represents an independent argument or semantic unit, and edges represent the connection paths between assertions and supporting information.

[0042] Using the main theme and points of contention as building blocks, the first engine drives the re-semantic re-separation and reconstruction of the same batch of pre-separated speech. The first engine analyzes the development trajectory of the main theme reflected in the conference speeches, as well as the related disagreements, objections, or points of contention, identifying the aggregation and opposition of different user viewpoints under a certain topic, and then constructing a second logical graph. In this graph, the main theme serves as the structural backbone, and the points of contention serve as branch nodes, forming a multi-faceted viewpoint intersection structure around a specific topic.

[0043] The first and second logic graphs are added to the speech logic graph, which serves as a deep semantic structure representation of the conference speech. It will then be used to drive core processes such as semantic attribution of conference record content, automated typesetting, and storage mapping.

[0044] Based on the temporarily accessed second engine, the target meeting template is matched, and the voice segment attribution is performed based on the voice logic graph to generate the target meeting record.

[0045] The meeting platform invokes a temporarily connected second engine to execute the meeting template matching and audio segment assignment layout generation process. First, the meeting platform's embedded meeting template library has preset structure templates for various typical meeting scenarios, including but not limited to: roundtable discussion, agenda advancement, report and defense, and brainstorming. Each template includes corresponding paragraph structure, role format, content hierarchy tags, and logical assignment constraints.

[0046] The second engine, based on structural features and semantic tags extracted from the speech logic graph (such as the main line of speech, the distribution of disputed nodes, and the density of claims and arguments), performs template similarity calculation and matching decision-making mechanisms to determine the most suitable target conference template. During the matching process, a joint comparison strategy based on semantic embedding (such as Sentence-BERT) and structural vectorization (as shown in the figure embedding) can be adopted to improve matching accuracy.

[0047] Based on the matched target meeting template, the system triggers the attribution mapping and automatic formatting process for audio segments. The second engine, according to the attribution constraints of the meeting template, maps each speaking node in the audio logic graph to a designated location unit in the template, and performs attribution alignment by combining information such as the speaker's identifier code, timestamp, and argument number. Specifically, this includes: dividing the audio segments into paragraphs based on the main topic grouping determined by the second logic graph; determining the internal structure of the speech based on the claim-argument relationship of the first logic graph; and completing the semantic attribution of the audio segments. Finally, the meeting platform uses a structured formatting engine to fill the attributed audio content into the template, automatically generating a target meeting record with a unified format, clear logic, and well-defined content.

[0048] Furthermore, matching the target meeting template includes:

[0049] For diverse meeting scenarios, a meeting template library is constructed, which is embedded in the platform database; for a target meeting scenario, a meeting template library matching based on scenario elements is performed to determine the target meeting template.

[0050] Based on diverse meeting scenarios, a meeting template library is constructed and embedded in the platform database as a structured knowledge component. The meeting template library contains a variety of standardized template styles. Each template is predefined based on meeting attributes (such as participant roles, agenda structure, speaking rounds, topic progression logic, etc.), covering but not limited to: project reporting, expert review, brainstorming, task scheduling, technical discussion and other meeting scenarios.

[0051] During the actual remote meeting, the platform uses a second engine to parse the initial meeting data, including user role information, meeting title / topic keywords, and historical template references, forming a set of meeting scenario element features. Based on these scenario elements, the platform performs a fast matching operation from the template library. This involves using scenario vector embedding and template index comparison, and invoking a strategy model based on semantic similarity calculation and template structure matching evaluation to select the target meeting template that best fits the current meeting context.

[0052] Furthermore, performing voice segment attribution based on automated meeting record formatting according to the aforementioned voice logic graph includes:

[0053] The second engine generates attribution constraints by scanning the target meeting template; using the attribution constraints, it performs voice attribution determination based on the second logic graph to determine first attribution data for the pre-separated speech; based on the first logic graph, it fine-tunes the first attribution data to determine the voice attribution result; based on the voice attribution result, it automatically typesets the target meeting template to determine the target meeting record.

[0054] The second engine performs structural scanning and field parsing on the selected target meeting template, extracting format elements such as paragraph types, role nodes, topic order, and affiliation tags. Based on this, it dynamically generates affiliation constraints. Affiliation constraints are used to define the classification rules of audio content in the structural template. For example, speeches on specific topics are preferentially affixed to topic paragraphs, controversial points are classified into discussion units, and different speakers correspond to different role sections.

[0055] Based on the attribution constraints and combined with the obtained pre-separated speech (tagged with voiceprint information and time series), the speech segments are initially attributed under the guidance of the second logical graph (main line-dispute point relationship), generating the first attribution data. The first attribution data represents the preliminary corresponding position of each speech segment in the conference structure template.

[0056] The first logical graph (claim-argument logical relationship) is used to perform semantic fine-tuning on the first attribution data in order to improve the logical integrity and expression coherence of speech attribution, such as aligning argument fragments to their claim speech paragraphs and aggregating speech blocks with the same stance, and finally outputting the speech attribution results.

[0057] Based on the audio attribution results, the meeting platform automatically fills in and structures the target meeting template, automatically generating target meeting records with standard format and clear semantic classification.

[0058] The target meeting records are stored in the platform database and access is managed through permission settings.

[0059] After generating the target meeting record, the conferencing platform implements a unified data archiving and access control strategy. Specifically, it associates and binds the target meeting record with its corresponding voice logic graph (including the first logic graph and the second logic graph) and stores it in the platform's database.

[0060] During database storage, the system sets access permission levels based on user identity information, meeting organizational structure, and speaker role hierarchy, forming a multi-level access control policy. For example, the meeting host has editing and approval permissions for all audio segments and structure records, ordinary participants can only view their own speeches or authorized topic segments, and administrators have archiving, auditing, and global access permissions.

[0061] Furthermore, an index mapping table is constructed using node elements in the speech logic graph (such as speakers, topic tags, and locations of points of contention), supporting targeted retrieval based on multi-dimensional conditions such as role, content, topic, and time. When a retrieval request is triggered, the index fields are matched according to the current user's permission tags, returning only meeting segments and structured content that conform to the access policy, effectively ensuring data security and controllability.

[0062] Furthermore, the target meeting records are stored in the platform database and accessed and managed using permission settings, including:

[0063] Using the first and second logical graphs as intermediate layers, a mutual mapping is established with the target meeting records, and permission levels are set and stored in the platform database; wherein, the target meeting records are retrieved in a targeted manner through the graph node elements in the first and second logical graphs.

[0064] After the target meeting minutes are generated, the first and second logical graphs are used as semantic intermediate layers to establish one-to-one or many-to-many mapping relationships with the corresponding meeting minutes content, forming a mutual mapping structure. This mutual mapping structure is implemented in the form of a mapping table or index table, including but not limited to the correspondence between speaker identifiers and segment numbers, logical nodes and speech content paragraphs, and the temporal relationship between graph paths and meeting topics.

[0065] Based on this, access permission levels are set for each audio segment in the target meeting record according to parameters such as user role (e.g., host, recorder, participant), semantic sensitivity (e.g., content of dispute, decision items), and meeting level.

[0066] By writing the target meeting minutes along with their logical graph index information into the platform database in a composite form of structured data and graph structure data, it supports mixed text and graph retrieval and graph-level relationship parsing.

[0067] When retrieving or accessing meeting content, the platform can perform targeted retrieval of target meeting records based on the graph node elements (including node type, path structure, key semantic identifiers, etc.) in the first and second logical graphs. For example, when retrieving content related to a specific topic, it can quickly aggregate information through the "topic main line" node in the graph; when retrieving a speaker's viewpoint, it can perform backtracking extraction through the mapping relationship between the "claim node" and its speech, thereby achieving precise content access driven by logical structure and controllable retrieval within permissions.

[0068] In summary, the embodiments of this application have at least the following technical effects:

[0069] As multiple users join the meeting, the conferencing platform triggers voice recording of the target meeting window, generating meeting audio. Next, the voice processing engine matches and temporarily accesses the target meeting window. Based on the temporarily accessed first engine, two-way voice separation processing is performed on the meeting audio to construct a voice logic graph. This two-way voice separation processing includes constructing a first logic graph based on the claim-argument relationship and a second logic graph based on the main line-point of contention. Then, based on the temporarily accessed second engine, the target meeting template is matched, and voice segment attribution based on automated meeting record formatting is performed according to the voice logic graph, generating the target meeting record. Finally, the target meeting record is stored in the platform database and managed for retrieval through access control. This solves the technical problem in existing technologies where voice interweaving in multi-party meetings makes it difficult to effectively distinguish speakers and the logical structure of their statements, leading to unclear meeting record attribution and low editing efficiency. It achieves two-way semantic logic separation and structured attribution of remote meeting voice content, improving the accuracy of meeting records and the efficiency of automatic formatting.

[0070] Example 2, based on the same inventive concept as the speech separation and segment attribution method for remote conferencing scenarios in the foregoing examples, such as... Figure 2 As shown, this application provides a speech separation and segmentation system for remote conferencing scenarios, wherein the system includes:

[0071] Audio generation module 11: As multiple user terminals join the meeting, the meeting platform triggers voice recording of the target meeting window, generating meeting audio; Voice processing module 12: Performs voice processing engine matching and temporary access to the target meeting window, and performs two-way voice separation processing on the meeting audio according to the temporary access first engine, constructing a voice logic graph, wherein the two-way voice separation processing includes the construction of a first logic graph based on the claim-argument relationship and the construction of a second logic graph based on the main line-point of contention; Record generation module 13: Matches the target meeting template according to the temporary access second engine, and performs voice segment attribution based on the voice logic graph based on automated meeting record layout, generating the target meeting record; Management module 14: Stores the target meeting record in the platform database and manages its retrieval according to permission settings.

[0072] Furthermore, the voice processing module 12 is used to perform the following methods:

[0073] A first-class engine is deployed based on speech separation and reconstruction, and a second-class engine is deployed based on speech typesetting. The first-class engine and the second-class engine are combined to form a speech processing engine group. The speech processing engine group is subjected to logical supervised training to determine a first engine subarray. The first engine subarray is multi-bit mirrored twinned to form a speech processing engine array, and the speech processing engine array is embedded and deployed in the conference platform.

[0074] Furthermore, the voice processing module 12 is used to perform the following methods:

[0075] As multiple users join the meeting, the speech processing engine array is traversed, and idle engine subarrays are matched to determine the target speech processing engine. According to the platform interaction protocol, a temporary connection is established between the target speech processing engine and the target meeting window, wherein the connection is disconnected when the meeting ends.

[0076] Furthermore, the voice processing module 12 is used to perform the following methods:

[0077] According to the voiceprint recognition plugin, the conference audio is subjected to voiceprint recognition and separation to determine the voiceprint separation speech, wherein the separation standard is based on the multiple speaking parties; for the multiple user terminals, multiple party identification codes are generated to annotate the voiceprint separation speech and determine the pre-separated speech.

[0078] Furthermore, the voice processing module 12 is used to perform the following methods:

[0079] The first engine is triggered, and the pre-separated speech is reconstructed by taking the argument items as the identification target and the claim-evidence as the construction elements to determine the first logical graph; the pre-separated speech is reconstructed by taking the main line-point of contention as the construction elements to determine the second logical graph; the first logical graph and the second logical graph are added to the speech logical graph.

[0080] Furthermore, the record generation module 13 is used to perform the following method:

[0081] For diverse meeting scenarios, a meeting template library is constructed, which is embedded in the platform database; for a target meeting scenario, a meeting template library matching based on scenario elements is performed to determine the target meeting template.

[0082] Furthermore, the record generation module 13 is used to perform the following method:

[0083] The second engine generates attribution constraints by scanning the target meeting template; using the attribution constraints, it performs voice attribution determination based on the second logic graph to determine first attribution data for the pre-separated speech; based on the first logic graph, it fine-tunes the first attribution data to determine the voice attribution result; based on the voice attribution result, it automatically typesets the target meeting template to determine the target meeting record.

[0084] Furthermore, the voice processing module 12 is used to perform the following methods:

[0085] Using the first and second logical graphs as intermediate layers, a mutual mapping is established with the target meeting records, and permission levels are set and stored in the platform database; wherein, the target meeting records are retrieved in a targeted manner through the graph node elements in the first and second logical graphs.

[0086] Furthermore, the management module 14 is used to perform the following methods:

[0087] For multi-concurrent conferences, the platform's computing power is dynamically balanced and allocated to the speech processing engine array.

[0088] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0089] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0090] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method of speech separation and paragraph attribution for teleconference scenarios, characterized in that, The method comprises: With the entry of multi-party users, the conference platform triggers voice recording of the target conference window, and generates conference audio; Performing matching of the voice processing engine and temporary access of the target conference window, according to the first engine of temporary access, performing double-pass voice separation processing on the conference audio, and constructing a voice logical graph, wherein the double-pass voice separation processing includes first logical graph construction based on argument-persuasion relationship and second logical graph construction based on main line-disputed point; According to the second engine of temporary access, matching the target conference template, and performing voice segment attribution based on automatic conference record layout according to the voice logical graph, to generate a target conference record; The target conference record is stored in the platform database for access management according to the permission settings; Before performing matching of the voice processing engine, constructing a voice processing engine array, including: Deploying a class of engines for voice separation reconstruction, deploying a class of engines for voice layout attribution, combining the class of engines for voice separation reconstruction and the class of engines for voice layout attribution as a voice processing engine group; Logical supervision training is performed on the voice processing engine group to determine a first engine subarray; Multi-bit mirror twin is performed on the first engine subarray to form a voice processing engine array, and the voice processing engine array is embedded and deployed on the conference platform; Before performing double-pass voice separation processing on the conference audio, including: According to the voiceprint recognition plug-in, voiceprint recognition and separation are performed on the conference audio to determine voiceprint separated voice, wherein the multi-party speaking subjects are used as the separation standard; For multi-party user terminals, a multi-party identification code is generated, and the voiceprint separated voice is labeled to determine pre-separated voice; Performing double-pass voice separation processing on the conference audio, including: Triggering the first engine, taking argument items as the recognition target, and taking argument-persuasion as the construction element, performing complex separation reconstruction on the pre-separated voice to determine a first logical graph; Taking main line-disputed point as the construction element, performing complex separation reconstruction on the pre-separated voice to determine a second logical graph; The first logical graph and the second logical graph are added to the voice logical graph.

2. The method for speech separation and paragraph attribution for teleconference scenarios of claim 1, wherein, Performing matching of the voice processing engine, including: With the entry of multi-party users, the voice processing engine array is traversed to match an idle engine subarray, and a target voice processing engine is determined; According to the platform interaction protocol, temporary connection is established between the target voice processing engine and the target conference window, and the connection is disconnected when the conference ends.

3. The method for speech separation and paragraph attribution for teleconference scenarios of claim 1, wherein, Matching the target conference template, including: For multi-element conference scenarios, a conference template library is constructed, wherein the conference template library is embedded in the platform database; For target conference scenarios, conference template library matching based on scene elements is performed to determine the target conference template.

4. The method for speech separation and paragraph attribution for teleconference scenarios of claim 3, wherein, Performing voice segment attribution based on automatic conference record layout according to the voice logical graph, including: The second engine generates attribution constraint conditions by scanning the target conference template; Using the attribution constraint conditions, the second engine performs voice attribution determination on the pre-separated voice based on the second logical graph to generate first attribution data; According to the first logical graph, the first attribution data is fine-tuned to determine voice attribution results; According to the voice attribution result, the target conference template is automatically laid out to determine the target conference record.

5. The method for speech separation and paragraph attribution for teleconference scenarios of claim 1, wherein, The target conference record is stored in a platform database for access management according to a permission setting, including: The first logical graph and the second logical graph are used as intermediate layers to establish mutual mapping with the target conference record, and a permission level is set and stored in the platform database. The target conference record is accessed by a graph node element in the first logical graph and the second logical graph.

6. The method for speech separation and paragraph attribution for teleconference scenarios of claim 1, wherein, If it is a multi-concurrent conference, the voice processing engine array is dynamically balanced and distributed according to platform computing power.

7. A speech separation and segment attribution system for a teleconference scenario, characterized by, The system for implementing the voice separation and paragraph attribution method for the remote conference scene according to any one of claims 1-6, the system comprising: An audio generation module: when multiple user terminals join the conference, the conference platform triggers voice recording of a target conference window to generate conference audio; A voice processing module: performing matching of a voice processing engine and temporary access of a target conference window, performing double-pass voice separation processing on the conference audio according to a first engine for temporary access, and constructing a voice logical graph, wherein the double-pass voice separation processing includes construction of a first logical graph based on a claim-argument relationship and construction of a second logical graph based on a main line-dispute point; A record generation module: according to a second engine for temporary access, matching a target conference template, and performing voice segment attribution based on automatic conference record layout according to the voice logical graph to generate a target conference record; A management module: storing the target conference record in a platform database for access management according to a permission setting.

Images