A method for intelligently generating meeting minutes files

Abstract

The present invention discloses a method for intelligently generating meeting minutes files, comprising the following steps: step one, collecting and uploading the audio data of the meeting to a cloud server through a recording device; step two, using a voiceprint recognition large model to identify different speakers in the audio after receiving the audio data uploaded in step one; step three, converting the voice segments segmented in step two into text content, and improving the accuracy of the text through an error correction model; step four, based on vectorization technology and cosine similarity algorithm; step five, summarizing and extracting the topics, key points, and decision summaries corresponding to each set; step six, summarizing the meeting summary; step seven, forming meeting minutes; step eight, extracting various information of the meeting minutes from a database. The method for intelligently generating meeting minutes files of the present invention, by automatically processing the content of the meeting, the system can not only identify different speakers, but also automatically distinguish topics, summarize discussion points, and refine decision content.

Description

Technical Field

[0001] The present invention relates to a file generation system, and more specifically to an intelligent generation method for conference minutes files. Background Art

[0002] Traditional methods of generating meeting minutes rely on manual writing or voice recorder conversion, which is inefficient and prone to errors. Manual writing may miss key information, and voice recorders directly converting to text generates unstructured documents with complex information and a large number of typos, making it difficult to quickly generate high-quality meeting minutes. In addition, these methods cannot automatically extract and organize key information, affecting the completeness and accuracy of the information.

[0003] iFLYTEK Co., Ltd. has invented an audio recognition method, system and related device, which involves obtaining audio to be recognized, inputting the audio into a target recognition model, and obtaining target recognition text that matches the input audio. The target recognition model is obtained based on a trained initial recognition model, which includes an audio parsing network and a text parsing network that share at least some parameters. The audio parsing network is trained using multiple training audios, and the text parsing network is trained using multiple training texts, wherein the training audios are matched with text labels, and the training texts are matched with category labels. This method can improve the accuracy of audio recognition.

[0004] In its invention patent "Method for generating meeting minutes, electronic device and computer-readable storage medium", Huawei Technologies Co., Ltd. extracts key information from meeting recordings according to predefined extraction rules, determines the category of the extracted information, classifies the information according to the category, and generates meeting minutes based on the classification results. This invention cannot fully cover all the contents of the meeting, but can only extract key information. Although this can generate refined minutes, some important information may be omitted. In addition, information extraction rules need to be set in advance, such as keyword extraction, which requires certain manual intervention and adjustment and cannot be fully automated.

[0005] Industrial and Commercial Bank of China Limited describes a method for automatically generating meeting minutes using artificial intelligence in its invention patent "Method, device, equipment, medium and program product for generating meeting minutes". The method involves transcribing the audio / video of the meeting, identifying the type of speaker (leader or non-leader), compressing the statements of non-leaders using a language model, and then generating a meeting summary based on the compressed text. The method aims to reduce the cost and effort required to manually generate meeting minutes, especially for long and complex meetings. The method mainly performs semantic compression on the speech content of non-leaders, while the speech content of leaders is not specially processed. This may result in the meeting minutes not being able to fully highlight the key points of the meeting. In addition, the method does not mention how to handle other elements in the text of the meeting record, such as meeting time, location, participants and other information. This may result in the generated meeting minutes lacking completeness.

[0006] Some other invention patents have not realized the automatic generation of structured meeting minutes documents, and there are problems such as missing the core content of the meeting, unrefined content, informal description, incomplete structure, etc. The present invention aims to realize a system and device for intelligent generation of meeting minutes documents, which realizes intelligent generation of meeting minutes documents by combining software and hardware, based on large model technology, greedy thinking and thought chain mechanism. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for intelligently generating meeting minutes files by combining technologies such as voiceprint recognition, speech-to-text (ASR) and large language model (LLM) to automatically generate high-quality, structured and complete meeting minutes.

[0008] To achieve the above object, the present invention provides the following technical solution: a method for intelligently generating a meeting minutes file, characterized in that it comprises the following steps:

[0009] Step 1: Collect and upload the audio data of the meeting to the cloud server through the recording device;

[0010] Step 2: After receiving the audio data uploaded in step 1, use the voiceprint recognition model to identify different speakers in the audio, and cut the complete audio into segments of single-person speech audio arranged in chronological order, and mark the speaker and start and end time of each audio segment;

[0011] Step 3: Convert the speech segments segmented in step 2 into text content, and improve the accuracy of the text through an error correction model;

[0012] Step 4: Based on vectorization technology and cosine similarity algorithm, the speech texts in the text content of step 3 are grouped into several sets according to chronological order to identify and classify topics;

[0013] Step 5: Summarize and extract the topics, key points, and decision summaries corresponding to each set in order;

[0014] Step 6: Concatenate the text content of each topic into a long text containing all the topics, and then process the long text through a large language model to summarize the meeting summary.

[0015] Step 7: Store the meeting start and end time, participants, address information, meeting summary and agenda content in the corresponding meeting audio record in the database to form meeting minutes;

[0016] Step eight, initiated by the client, extracts various information of the meeting minutes from the database, combines them to generate a word file of the meeting minutes, and downloads them to the local computer to complete the generation and export of the meeting minutes.

[0017] As a further improvement of the present invention, the recording device in step 1 includes:

[0018] Processor and Wi-Fi connection module. The processor uses low-power microcontroller chips such as ESP32, which is responsible for core functions such as recording control, Wi-Fi management, and file transfer. It can access the Internet or local area network through the built-in Wi-Fi module;

[0019] A storage unit, used for storing recording files;

[0020] Microphone, used for recording, usually a built-in high-sensitivity condenser microphone, to collect audio data of speeches during meetings;

[0021] Audio processing chips enable recording devices to record high-quality, clear audio signals through analog-to-digital conversion, signal processing, noise reduction, and echo cancellation, and provide a basis for subsequent audio uploading and processing;

[0022] The GPS module is used to obtain the current longitude and latitude information when recording starts, and convert it into a specific address on the server through Baidu Map API, Amap API, and Tencent location service;

[0023] A power management module, used to provide power supply for the recording device, including a battery, a charging circuit and a power control module;

[0024] Physical or virtual buttons, USE interface, used to provide a physical interface or touch interface for controlling recording and Wi-Fi settings.

[0025] As a further improvement of the present invention, the cloud server in step one includes an audio library and a voiceprint library. The audio library is divided into an audio directory library and a meeting minutes database. The voiceprint library completes the training of SincNet through an open source data set. In the process of generating meeting minutes files, the voiceprint vector of the speaker is generated by SincNet and compared with the data in the vector library. If it already exists, the speaker information is proposed; otherwise, a speaker ID is created and stored in the vector library together with the voiceprint vector.

[0026] As a further improvement of the present invention, the specific steps of performing voiceprint recognition and audio segmentation in step 2 are as follows:

[0027] Step 21, preprocessing the audio;

[0028] Step 22, dividing the long audio data into shorter segments;

[0029] Step 2 and 3: Process each audio clip through the voiceprint feature extraction model SincNet to convert the voice data into a feature vector that can represent the speaker's identity. The model first extracts frame-level features from the audio, and then performs temporal aggregation through a global statistical pooling layer to obtain a vector representation of fixed dimensions.

[0030] Step 24: extract voiceprint features from the segmented audio clips and perform similarity matching with speaker vectors in the voiceprint library;

[0031] Step 25: After the speaker is identified, each continuous audio of the same speaker will be regarded as a speech in chronological order. The audio splicing belonging to the same speech content will be realized using Python libraries such as pydub or librosa, so as to merge discrete audio clips into audio clips with speech as the unit and reconstruct the complete dialogue scene.

[0032] As a further improvement of the present invention, the audio is preprocessed in step 21 as follows: first, noise reduction is performed by spectrum subtraction or Wiener filtering to remove background noise and make the speech signal clearer; second, the silent segments in the audio are removed, firstly, the silent area is detected by Librosa using short-time energy technology, and compared with the set energy threshold to identify the silent segment; then, the silent segment is removed by audio cutting technology, the non-silent segment is retained, and the audio stream is restored by audio splicing;

[0033] The sampling rate is converted to a standard value, such as 16kHz or 8kHz, and the speech is spectrally analyzed through short-time Fourier transform to enhance the speech features and highlight the pronunciation details.

[0034] As a further improvement of the present invention, the specific steps of converting the voice segment into text content in step 3 are as follows:

[0035] Step 31, converting the speech fragment into text through automatic speech recognition technology;

[0036] Step 32: Structuring the transcribed text;

[0037] Step 33: Verify and correct the text to improve its accuracy.

[0038] As a further improvement of the present invention, the specific steps of topic identification and classification in step 4 are as follows:

[0039] Step 41: Identify the boundaries of the topic;

[0040] Step 42: Merge the topics.

[0041] As a further improvement of the present invention, the specific steps of identifying the topic boundary in step 41 are as follows:

[0042] Step 4: Initialize the topic set topic T = {t1, t2, .., t m}, each of its elements is a set of speech texts on different topics, m is the total number of topics; the conference speech text set is S = {s1, s2, .., s n}, whose elements are arranged in the order of speech, where n represents the total number of speeches, assuming that the speech text is s i , the text set of the current topic is t j , it either contains only s i-1 , either with s i-1 A continuous subset of speech text, such as t j ={s2,s3,...,s i-2 ,s i-1};

[0043] Step 4.12: Calculate the current speech text s based on BERT i The embedding vector v i ; Set t j All speech texts are concatenated into a long text, and the topic vector is calculated

[0044] Step 413: Calculate the current speech vector s i and the issue vector , and determine the relationship between the similarity and the threshold:

[0045] When the similarity is lower than the threshold, it means that the current text belongs to a new topic. The text set of the new topic is initialized with the current speech text, which is recorded as t j+1={s i}, let i = i + 1; and set the new topic as the current topic, let j = j + 1, jump to e to execute;

[0046] When the similarity is higher than or equal to the threshold: then the next speech text s i+1 Transformed into a high-dimensional vector v i+1 and Topic Vector If the similarity is higher than the threshold, it means that the current text does not belong to the boundary, and the current speech text s i Summarize to the current topic text set t j , t j ={s2,s3,...,s i-2 ,s i-1 ,s i}, let i=i+1, jump to e to execute; if the similarity is lower than the threshold, it means that the current speech belongs to the boundary;

[0047] Step 414, for the current text s i It is a boundary problem. Assume that s i It contains two topics, i Topic segmentation is performed, starting from the beginning of the formal speech content, increasing the fixed step size d each time, and cutting the text into two sub-texts;

[0048] If s b If it is empty, it means s i Only belongs to t j Single topic, will s i Included to j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s i}, let i=i+1, and j=j+1, jump to e for execution;

[0049] Otherwise, convert the previous and next subtexts into vectors v respectively. a and v b , and respectively with the current topic vector Compare, if and only if s a With t j Belong to the same topic and s b When it is a new topic, stop cutting and a Add to t j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s a}, and s b After adding the speaker information, add it to t j+1 In, tj+1 ={s b}, let i=i+1, and j=j+1, jump to e for execution;

[0050] Step 4-15, start again from stage b until all speech texts are processed and classified.

[0051] As a further improvement of the present invention, the specific steps of merging topics in step 42 are as follows:

[0052] Step 421: Construct a new topic set T′={t′1,t′2,..,t′ k}, k is the final number of topics, which is the set of topics after merging;

[0053] Step 422, starting from the first topic in the original set, delete it from the original set and set it as the current topic to be merged, and compare it with the remaining topics in the original set in turn for similarity. If the similarity exceeds the threshold, merge them and delete the compared topic from the original set. Otherwise, continue to compare the next set until the comparison with all elements of the original set is completed, and add the newly merged set as the merged topic to the new topic set;

[0054] Step 423, continue to execute operation b until the original topic set is empty.

[0055] As a further improvement of the present invention, the specific steps of summarizing and extracting the topics, key points, and decision summaries corresponding to each set in step 5 are as follows:

[0056] Step 51: Use the pre-trained large language generation model combined with fine-tuning to complete the refinement of topic content, hot topics, and final conclusions or decisions;

[0057] Step 52: Process each topic in turn. Based on the semantic understanding ability of the large language model, analyze the speech content, extract the views of each speaker, identify the hot issues discussed, and regenerate the topics, arguments and decision summaries through the generative ability of the model.

[0058] The beneficial effects of the present invention are that the intelligent generation system and device of meeting minutes of the present invention significantly improve the efficiency and quality of meeting minutes generation by combining technologies such as voiceprint recognition, speech-to-text (ASR) and large language model (LLM).

[0059] First, voiceprint recognition technology can accurately identify different speakers, ensuring that the opinions of each speaker in the meeting minutes can be accurately attributed. This function effectively solves the problem of speaker identity confusion in traditional methods, allowing the generated minutes to clearly distinguish between speakers and their speech content.

[0060] Secondly, speech-to-text technology converts conference recordings into text, laying the foundation for subsequent text processing and analysis. Through the application of large language models, the system can not only automatically process text content, but also perform semantic analysis and summary generation to extract key discussion points and decision-making content in the meeting. Compared with traditional methods, LLM can generate more structured and coherent minutes, avoiding omissions and errors that may occur in traditional manual records.

[0061] In addition, the system of the present invention can automatically identify and classify topics to ensure that the content of the meeting minutes is comprehensive and well-structured. Unlike the previous information extraction rules that require manual adjustment, the system can reduce manual intervention through intelligent processing, thereby improving the accuracy and completeness of meeting records.

[0062] Ultimately, this technical solution improves work efficiency, reduces human errors, and is suitable for a variety of scenarios such as company meetings and project discussions, with broad practical application significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The structure diagram of the system used by the intelligent generation method of meeting minutes files of the present invention;

[0064] Figure 2 This is a schematic diagram of the internal modules of the recording device;

[0065] Figure 3 This is a schematic diagram of the method for collecting and uploading recording information;

[0066] Figure 4 This is a schematic diagram of the server system architecture;

[0067] Figure 5 This is a schematic diagram of the method for building a voiceprint library;

[0068] Figure 6 This is a schematic diagram of the speech-to-text algorithm method;

[0069] Figure 7 Schematic diagram of text cutting for topic boundaries;

[0070] Figure 8 Identify algorithm flow charts for topics;

[0071] Fig. 9 Merge algorithm flow chart for the topic;

[0072] Fig.10 A schematic diagram of the topic content generation method;

[0073] Fig.11 This is a schematic diagram of the complete meeting document;

[0074] Fig.12Automatically generate front-end system architecture diagram for meeting minutes. DETAILED DESCRIPTION

[0075] The present invention will be further described in detail below in conjunction with the embodiments given in the accompanying drawings. Traditional manual meeting minutes recording is inefficient and easy to miss key information. The present invention can automatically generate high-quality meeting minutes by using voiceprint recognition, speech-to-text (ASR) and large model technology to ensure the accuracy and completeness of the content.

[0076] Reference Figures 1 to 12 As shown, a method for intelligently generating a meeting minutes file in this embodiment includes the following steps:

[0077] (1) Collect and upload audio data through a recording device: Complete the conference audio collection through a recording device and upload it to the cloud server. Store the audio file in a specified directory, and store the audio address, audio start and end time, and address information in the database.

[0078] (2) Voiceprint recognition and audio segmentation: Use the voiceprint recognition model to identify different speakers in the audio, and cut the complete audio into segments of single-person speech audio arranged in chronological order, and mark the speaker and start and end time of each segment. The identification of the speaker is based on a pre-established voiceprint library. The voiceprint library is a vector database. When building the database, the audio data of relevant personnel is collected, and the feature vector is extracted through a deep learning algorithm, which is stored together with the person's real name, position and other information;

[0079] (3) Speech-to-text: Use speech-to-text technology (ASR, Automatic Speech Recognition) to convert the segmented speech segments into text content, and improve the accuracy of the text through the error correction model. In addition, based on the annotation information of the audio segment, add the prefix [speaker + position + start / end time] to each text segment, and connect them in sequence to form a long text, such as [Zhang San - R&D Director - 11:30-11:25: Today we will mainly discuss the XX project].

[0080] (4) Topic identification and classification: Based on vectorization technology and cosine similarity algorithm, speech texts are grouped into several sets according to chronological order. And through the idea of ​​merging, non-adjacent speeches on the same or similar topics are merged into the same set.

[0081] (5) Topics, key points, and decision summaries: Summarize and extract the topics, key points, decision summaries, and other contents corresponding to each set in order. The main approach is to concatenate the speech texts of the corresponding set into a long text, and generate the corresponding topics, key points, decision summaries, and other contents based on the pre-trained and fine-tuned large language model.

[0082] (6) Meeting summary generation: The text content of each topic (topic, key points, decision summary) is spliced ​​into a long text containing all topic contents, and the meeting summary is summarized using a large language model.

[0083] (7) Meeting minutes construction: The meeting start and end time, participants, address information, meeting summary and agenda content are stored in the corresponding meeting audio records in the database, and the agenda content is stored in JSON format.

[0084] (8) Meeting minutes generation and export: Initiated by the client, the meeting minutes information is extracted from the database, combined to generate a meeting minutes word file, and downloaded to the local computer.

[0085] The specific functional modules and their implementation are as follows:

[0086] Recording device

[0087] The recording device, including but not limited to mobile phones, tablet computers, recording equipment and other hardware devices, should include Figure 2 The structural modules are shown.

[0088] Core hardware components

[0089] Processor (MCU / SoC) and wifi connection: It uses low-power microcontroller chips such as ESP32, which are responsible for core functions such as recording control, Wi-Fi management, and file transfer. It can access the Internet or local area network through the built-in Wi-Fi module.

[0090] Storage unit (Flash / SD card): used to store recording files. The storage unit can be a built-in flash memory or an inserted SD card. The data will be stored here during recording and will be uploaded to the server via Wi-Fi later.

[0091] Microphone: Responsible for recording, usually a built-in high-sensitivity condenser microphone to collect audio data of speeches during meetings.

[0092] Audio processing chip: Through analog-to-digital conversion (ADC), signal processing, noise reduction, echo cancellation and other functions, the recording device can record high-quality, clear audio signals and provide a basis for subsequent audio uploading and processing.

[0093] GPS module: Get the current longitude and latitude information when recording starts, and convert it into a specific address on the server through Baidu Map API, Amap API, Tencent location service, etc.

[0094] Power management module: provides power supply for recording equipment, including battery, charging circuit and power control module.

[0095] Physical or virtual buttons, USE interface: Provide physical or touch interfaces for controlling recording and Wi-Fi settings, such as device on / off buttons, start / stop recording buttons, USB interfaces, etc.

[0096] The implementation process of audio recording and uploading

[0097] In order to realize audio recording and uploading to the specified path of the server, a script program needs to be built into the recording device to realize logic control. The written firmware script is compiled into machine code, and then the compiled firmware is burned into the Flash memory through a programming tool (such as a JTAG / SWD debugger or burner). When the device is powered on or restarted, the processor reads the firmware program from the Flash memory and loads it into the memory to start running. Based on the above firmware and the following steps, complete the audio recording and upload to the specified directory of the server:

[0098] ① Recording: The recording device collects sound through the built-in microphone. The sound collected by the microphone is converted into a digital signal through the ADC function of the audio processing chip. The processor is responsible for processing the signal and storing it as an audio file (common formats such as WAV or MP3) in the local storage unit.

[0099] ②Wi-Fi connection and server communication: The recording device enters Wi-Fi configuration mode when it starts up. Users can use their mobile phones or computers to connect to the recording device's Wi-Fi through the hotspot generated by the device. The recording device provides a local configuration page (implemented through a built-in web server) where users can enter the Wi-Fi network and server address.

[0100] ③ Audio data upload: After successfully connecting to the server via Wi-Fi, the recorded audio and related data can be uploaded to the pre-configured server through the interface and Http, Https and FTP protocols. The audio is uploaded to the specified directory of the server via FTP and the corresponding address is returned; the audio address is combined with the recording start and end time (i.e. the start and end time of the meeting), as well as the latitude and longitude obtained by GPS, and uploaded to the server. Add audio records to the audio database.

[0101] ④ Reconnection mechanism and power-off resume: If the connection is interrupted, the recording device will automatically reconnect via Wi-Fi and continue uploading audio files after the connection is restored. Figure 3 shown.

[0102] Server selection and system architecture

[0103] When choosing an operating system, you can choose Ubuntu Server or CentOS. Both operating systems are open source and stable, suitable for use as a web server. Build a server-side application to handle related processes. The overall functional architecture of the server-side application is as follows: Figure 4 shown.

[0104] Audio Library - Audio Catalog

[0105] Configure a Web server: Next, configure a Web server (such as

[0106] Apache or Nginx), which will process the audio files uploaded from the recording device.

[0107] Audio file directory: Set a file storage directory (for example

[0108] / var / www / html / uploads) and create an interface to handle file uploads. This interface will receive the file and save it to the specified directory, while updating the MySQL database to record the file's metadata.

[0109] Audio Library - Meeting Minutes Database

[0110] Choose MySQL as the database management system for audio files to record the metadata of the files, including the file name, path, and upload date. Install the MySQL database management system and create a database (such as audio_library) in which to create a table to store the information of audio files. The fields of the table should include file_name (file name), file_path (file path), and upload_date (upload date).

[0111] Voiceprint Library

[0112] SincNet is trained through open source data sets. When generating meeting minutes, the speaker's voiceprint vector is generated through SincNet and compared with the data in the vector library. If the speaker information already exists, the speaker information is proposed; otherwise, a speaker ID is created and stored in the vector library together with the voiceprint vector. The specific technical route is as follows:

[0113] ① Selection and preparation of open source audio datasets

[0114] Dataset: Select the VoxCeleb2 dataset. This is a large-scale open source audio dataset that contains recordings of more than 10,000 speakers and provides rich voiceprint samples. The VoxCeleb2 dataset covers multiple languages ​​and accents, allowing the model to learn diverse voiceprint features.

[0115] Preprocessing: Use tools such as Librosa to perform audio preprocessing, including denoising, normalization, and segmentation. Ensure that each audio sample is of high quality and perform standardization to improve training results.

[0116] ②Voiceprint feature extraction

[0117] Feature extraction: Mel Frequency Cepstral Coefficients (MFCC) and Log-Mel Spectrogram are used as the basis for feature extraction. MFCC is a classic acoustic feature used to capture the spectral characteristics of audio, while Log-Mel Spectrogram provides higher-dimensional spectral information. Librosa is used for feature extraction.

[0118] Feature vectorization: The DeepSBD toolkit is used to extract voiceprint feature vectors. These vectors have high discriminative power and are used for speaker identification.

[0119] ③Voiceprint model training

[0120] Model selection: Use SincNet, an advanced voiceprint recognition model that can process raw audio waveforms without feature extraction pre-processing. SincNet performs well in speaker identification tasks and can better capture voiceprint features.

[0121] Training: Use TensorFlow or PyTorch framework for training. Regularly evaluate model performance and adjust hyperparameters to optimize results.

[0122] ④ Voiceprint library creation

[0123] a. Select the vector database as the voiceprint database;

[0124] b. After the recordings of different people are collected and feature vectors are generated through SincNet, they are stored in the voiceprint database together with the name, position and other information of the associated speakers;

[0125] Based on the efficient approximate search algorithm of the vector database, voiceprint and personnel information matching can be achieved, such as Figure 5 shown.

[0126] Meeting minutes file intelligent generation system

[0127] 1) Voiceprint recognition and audio segmentation

[0128] The present invention ensures the integrity of the speaker information through the audio segmentation and overlapping strategy, uses the vector generated by the voiceprint feature extraction model SincNet to effectively characterize the speaker's voice features, and accurately matches the speaker through cosine similarity based on the voiceprint library data. On this basis, the audio clips of each speaker are merged in chronological order according to the timeline (one speech content is merged into one audio clip) to maintain a complete conversation scene. Tags and metadata are added to each merged audio file to mark the speaker information, conversation time, etc.

[0129] ① Audio preprocessing

[0130] First, noise reduction is performed through spectral subtraction or Wiener filtering to remove background noise and make the speech signal clearer. Secondly, remove the silent segments in the audio. First, the short-time energy (RMS) technology is used to detect the silent area through Librosa, and compared with the set energy threshold to identify the silent segment. Then, the silent segment is removed by audio cutting technology, the non-silent segment is retained, and the audio stream is restored by audio splicing. In order to ensure that the audio is consistent with the input standard during model training, the sampling rate is converted to a standard value, such as 16kHz or 8kHz. In addition, the speech is spectrally analyzed by short-time Fourier transform (STFT) to enhance the speech features and highlight the pronunciation details.

[0131] ② Audio Slicing and Overlap Strategy (Sliding Window with Overlap)

[0132] First, long audio data needs to be divided into shorter segments. The Sliding Window strategy is used to divide the audio into fixed lengths (such as 2 seconds or 3 seconds), and each window covers a content segment of audio. However, direct segmentation may cause the speech to be truncated at the boundary, affecting the accuracy of speaker recognition. Therefore, the Overlap technology is added to make adjacent windows partially overlap (such as overlapping 0.5 seconds) to ensure that the speech information in the audio transition area can be fully captured and avoid a single speaker being segmented at the window boundary. This method ensures the integrity of the audio content.

[0133] ③Speaker Embedding Extraction

[0134] Each audio clip is processed by the voiceprint feature extraction model SincNet, which converts the voice data into a feature vector that can represent the speaker's identity. The model first extracts frame-level features from the audio, and then performs temporal aggregation through the global statistical pooling layer to obtain a fixed-dimensional vector representation.

[0135] ④ Speaker recognition based on similarity calculation

[0136] After extracting the voiceprint features from the segmented audio clips, they are matched by similarity with the speaker vectors in the voiceprint library. Cosine similarity is used to quantify their similarity by calculating the angle between the two vectors, with a range of [-1,1]. The closer the cosine value is to 1, the higher the match between the current speaker and a speaker in the voiceprint library. Based on the matching results, the system can determine the identity of the speaker of each audio clip.

[0137] Since the audio clips are overlapped and segmented, there may be multiple clip recognition results covering the same time period. In order to enhance the robustness of recognition, the final speaker identity within a certain time period is determined by majority voting or confidence weighting. If no existing voiceprint template is matched, it is marked as a new speaker and stored in the voiceprint library.

[0138] ⑤ Audio segment merging and labeling

[0139] After the speaker is identified, each continuous audio of the same speaker is regarded as a speech in chronological order, and the audio splicing belonging to the same speech content is realized by programming using Python libraries such as pydub or librosa, so as to merge discrete audio clips into audio clips with speech as the unit and reconstruct the complete dialogue scene. The speaker information and timeline annotation are embedded in the merged audio file, which contains key information such as the speaker's name, position, and speech time.

[0140] 2) Speech to text

[0141] The process of converting the segmented speech segments into structured text includes three steps: speech-to-text conversion, text structuring, and text verification and correction. The conversion from speech to text is achieved through ASR technology. The speaker, position, and speaking time information are added to the text header through annotation information, and the text quality is improved through the error correction model, thereby generating accurate and easy-to-understand structured text. The details are as follows:

[0142] ①Convert the speech segments into text through automatic speech recognition (ASR) technology. ASR converts the speech signal into corresponding text by analyzing its frequency, amplitude and other characteristics. In this step, the system transcribes each speech segment into corresponding text.

[0143] ②Structural processing of the transcribed text. By annotating information, add speakers, positions and time tags to each text. For example, [Li Si, R&D manager, 12:01-12:05: Let’s discuss the deadline of the project today]”.

[0144] Verify and correct text to improve its accuracy. Use natural language processing (NLP) technology and the error correction model BERT to automatically correct text. These models can identify and correct spelling, grammar, and context errors to ensure that the final structured text is accurate and natural, such as Figure 6 shown.

[0145] Issue identification and merging

[0146] The present invention proposes a greedy-based topic identification and merging strategy, and uses the BERT model to identify and merge topics in speech texts.

[0147] ①Model selection and technical principles

[0148] The large language model BERT is used to generate context embedding through the bidirectional Transformer architecture, which can effectively capture the semantic information in the text. The pooler_output is used to represent the semantics of the entire sentence. The cosine similarity technology is used to calculate the similarity between the topic and the speech text. The similarity between two vectors is determined by calculating the cosine value of the angle between them. The value range is [-1,1], and the closer to 1, the higher the similarity.

[0149] ② Topic identification and merging algorithm process

[0150] Identification of topic boundaries:

[0151] a. Initialize the topic set topic T = {t1, t2, .., t m}, each of its elements is a set of speech texts on different topics, m is the total number of topics; the conference speech text set is S = {s1, s2, .., s n}, whose elements are arranged in the order of speech, where n represents the total number of speeches. Assume that the speech text is s i , the text set of the current topic is t j , it either contains only s i-1 , either with s i-1 A continuous subset of speech text, such as t j ={s2,s3,...,s i-2 ,s i-1};

[0152] b. Based on BERT, calculate the current speech text s i The embedding vector v i ; Set t j All speech texts are concatenated into a long text, and the topic vector is calculated

[0153] c. Calculate the current speech vector si and the issue vector The cosine similarity of:

[0154] If the similarity is lower than the threshold, it means that the current text belongs to a new topic. The text set of the new topic is initialized with the current speech text, which is recorded as t j+1 ={s i}, let i = i + 1; and set the new topic as the current topic, let j = j + 1. Jump to e to execute;

[0155] If the similarity is higher than or equal to the threshold, then the next statement s i+1 Transformed into a high-dimensional vector v i+1 and Topic Vector If the similarity is higher than the threshold, it means that the current text does not belong to the boundary, and the current speech text s i Summarize to the current topic text set t j , t j ={s2,s3,...,s i-2 ,s i-1 ,s i}, let i=i+1, jump to e to execute; if the similarity is lower than the threshold, it indicates that the current speech belongs to the boundary.

[0156] d. For the current text s i It is a boundary problem. Assume that s i It contains two topics, i Topic segmentation. Starting from the start position of the formal speech content, increase the fixed step size d each time, and cut the text into two sub-texts (the text from the start position to the cutting position is the former sub-text s a , the following content is the post-subtext s b ).

[0157] If s b If it is empty, it means s i Only belongs to t j Single topic, will s i Included to j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s i}, let i=i+1, and j=j+1, jump to e to execute, such as Figure 7 shown.

[0158] Otherwise, convert the previous and next subtexts into vectors v respectively. a and v b, and respectively with the current topic vector Compare, if and only if s a With t j Belong to the same topic and s b When it is a new topic, stop cutting. a Add to t j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s a}, and s b After adding the speaker information, add it to t j+1 In, t j+1 ={s b}, let i=i+1, and j=j+1, jump to e for execution;

[0159] e. Start again from stage b until all speeches are processed and classified, such as Figure 8 shown.

[0160] Merge of topics:

[0161] After obtaining the topic set T = {t1, t2, .., t m}, by calculating the similarity between topic sets, topics with high similarity are merged. The specific steps are as follows:

[0162] a. Construct a new topic set T′={t'1,t'2,..,t' k}, k is the final resolution

[0163] The number of topics, as a collection of merged topics;

[0164] b. Start with the first topic in the original set, delete it from the original set, set it as the current topic to be merged, and compare it with the remaining topics in the original set in turn. If the similarity exceeds the threshold, merge it and delete the topic being compared from the original set. Otherwise, continue to compare the next set until the comparison with all elements of the original set is completed. Add the newly merged set as the merged topic to the new topic set.

[0165] c. Continue to perform operation b until the original topic set is empty, such as Fig. 9 shown.

[0166] The specific process is shown in the following example: Assuming m=7, k=4, calculate T={t1,t2,t3,t4,t5,t6,t7} from t1 to t7 in sequence. Assuming that the current calculation is t1, let the corresponding merged topic be t'1, t'1 contains t1 by default, and calculate the cosine similarity with t1 from t2. Assuming that the similarity between t5 and t6 and t1 is greater than the threshold, the final t'1={t1,t5,t6} is obtained. Delete t1,t5,t6 from the original set T to obtain the remaining t={t2,t3,t4,t7}, and use the same method to calculate t'2={t2,t3}, t'3={t4}, t'4={t7}. Thus, the final speech topic content set T′={t'1,t'2,..,t'4} is obtained, that is, T′={{t1,t5,t6},{t2,t3},{t4},{t7}}. Based on this method, the meeting topics of very long meetings can be summarized and sorted.

[0167] Intelligent generation of topics, key points and decision summaries

[0168] The present invention proposes to extract the content of the discussion topic, hot topics, and the final conclusions or decisions based on the generative model to form a structured topic summary document. To achieve this goal, we can rely on the powerful natural language understanding and generation capabilities of the large language model, and through appropriate technologies and methods, enable the model to automatically identify the topic of discussion and extract key information. The details are as follows:

[0169] ① Selection and fine-tuning of large language models

[0170] The pre-trained large language generation model is combined with fine-tuning to complete the refinement of topic content, hot topics, and the final conclusions or decisions. The generation model selects T5 or BART, etc., and uses historical meeting records, meeting minutes and other data to fine-tune the model to make it more suitable for processing professional terms and language structures in meeting scenarios.

[0171] ② Generate topic content based on semantic understanding

[0172] Each topic is processed in turn. Based on the semantic understanding ability of the large language model, the speech content is analyzed to extract the views of each speaker, identify the hot issues of discussion, and regenerate the topics, arguments, decision summaries and other content through the generative ability of the model.

[0173] Semantic understanding: The generative model can process long sequences of text and capture the contextual relationships in the text through the pre-trained self-attention mechanism. When processing each speech, the model pays attention to the semantic direction of the speech through the self-attention mechanism and determines what the core argument is. Through the processing of word vectors and context vectors, the model can identify the deep semantics of the speech and determine the relevance of the speech to the previous and next speeches.

[0174] Hotspot identification: The model performs semantic analysis on speech content and uses word vectors and sentence vectors to calculate semantic similarity to identify frequently discussed keywords or concepts. Based on these hot topics, the model focuses on describing the topics discussed, core arguments, and related decisions, such as Fig.10 shown.

[0175] 5) Meeting summary generation

[0176] The text content of each issue that will be generated (issue, key points, decision summary)

[0177] The text is concatenated into a long text containing all the topics and input into the large language model as a prompt to generate a summary of the entire meeting.

[0178] 6) Meeting minutes construction

[0179] The overall structure of the meeting minutes includes: meeting time, meeting location, participants, meeting summary, agenda and resolutions, etc. The content of each department is constructed as follows:

[0180] ① Meeting time: obtained when the recording device starts recording and ends recording, and added to the detailed data table when uploading to the server;

[0181] ② Meeting location: When recording, the address is automatically obtained according to the locator, and the meeting time is uploaded to the server together;

[0182] ③ List of participants: obtained during voiceprint recognition. If the name is unknown, after creating a speaker database record based on the voiceprint, use the speaker ID as a placeholder, generate a word file, and manually adjust it. For example: Zhang San, speaker 00003, Li Si

[0183] ④ Conference summary generation: obtained according to the above generation algorithm;

[0184] ⑤Topics, key points, and decision summaries: obtained according to the above generation algorithm; and converted into JSON format for storage:

[0185]

[0186] },

[0187]

[0188] ⑥ Storage of meeting minutes content: The meeting start time, end time, list of participants, content summary, topic content summary, etc. are stored in the database. When the front end calls, they are encapsulated into complete file content and generated into a word file, such as Fig.11 shown.

[0189] (2) User Terminal

[0190] The terminal hardware can be a mobile phone, computer and other devices, and the conference recording minutes generation and download functions are implemented on the terminal. Use React, Vue.js or Angular as the main framework to build a dynamic and responsive user interface. These frameworks can efficiently manage component status and user interaction, display the recording list, handle recording selection and playback control. In order to improve the efficiency of interface design and development, integrate Bootstrap, Material-UI or Ant Design component library to provide beautiful UI elements and layout. The recording playback function is implemented through HTML5Audio API or React Audio Player component, supporting operations such as play, pause and fast forward. The front-end interacts with the server through Axios or Fetch API, which is responsible for submitting recording selection and minutes generation requests, and handling progress feedback and errors. The file download function is implemented using Anchor tags or JavaScript Blob, allowing users to download the generated Word document. Combining these technologies, the front-end realizes the full-process user experience from recording management to minutes generation and downloading. The specific functions and implementation methods are as follows:

[0191] ① Meeting recording list display

[0192] Function: The terminal interface needs to display a list of conference recordings available to users. Each recording needs to include the meeting name, time, and duration, so that users can quickly identify the required meeting.

[0193] Implementation: You can obtain the recording metadata from the server and display the list in the form of a table or card.

[0194] ②Recording selection and playback

[0195] Function: Users can select any recording and provide a simple playback function for listening to the recording, ensuring that users can preview the meeting content.

[0196] Implementation: The terminal calls the audio player component to support basic functions such as play, pause, fast forward, and rewind.

[0197] ③Generate meeting minutes

[0198] Function: After the user selects the recording and clicks the "Generate Meeting Minutes" button, the terminal will send a request to the server to generate a Word document of meeting minutes based on the meeting recording. After generation, the server returns a download link for the document.

[0199] Implementation: The terminal interacts with the server through the API, sends the unique identifier (ID) of the conference recording to request the server to generate minutes, and waits for the generation to be completed.

[0200] ④Get and download the download link

[0201] Function: Once the minutes are generated, the terminal will display a download link, and the user can click the link to download the minutes to the local device.

[0202] Implementation: After the server generates the Word document, it sends the download link to the terminal. The terminal interface displays a "click to download" button and calls the browser or the system's default download manager to download the file.

[0203] ⑤Status feedback and error handling

[0204] Function: During the minutes generation, the terminal should provide progress feedback, such as "generating", "downloading", etc., and handle abnormal situations such as network interruption and file generation failure, and display friendly error prompts to users, such as Fig.12 shown.

[0205] In summary, the intelligent generation method of meeting minutes file in this embodiment:

[0206] (1) A recording device-server-client-based overall architecture for generating meeting minutes is proposed. This architecture decouples the overall process of recording collection, generating meeting minutes, and converting them into word documents, so that the practicality of the recording device and the accuracy of the meeting minutes can be continuously improved by optimizing and iterating the meeting minutes generation process on the server side.

[0207] (2) Voiceprint and speech data fusion processing technology: This technology combines voiceprint recognition with speech-to-text technology to accurately identify different speakers and accurately record their speech content. This method improves the accuracy of the recording and the clarity of the speaker's attribution information.

[0208] (2) Intelligent topic identification and classification merging method: This method is based on greedy thinking, large models and vectorization technology. It analyzes and processes speech texts in order of speech. Through the cosine similarity algorithm, it automatically identifies whether adjacent speeches belong to the same topic. At the same time, it uses the adjacent calculation method to identify topic boundaries, and through the text segmentation and comparison method, it assigns boundary speeches to different topic sets. Finally, through the proposed merging algorithm, it merges non-adjacent speeches on the same topic, thereby obtaining a speech text set based on topic distinction. This method ensures the structure and comprehensiveness of the meeting minutes, so that the generated meeting minutes can clearly show the discussion of each topic.

[0209] (3) Intelligent summary method of conference content based on the idea of ​​whole-split-merge: Based on the identification and merging of topics, a large language model is used to perform semantic analysis on individual topics in turn to generate topics, key points, and decision summaries. Finally, the topics are merged to generate content and input into the large model to generate meeting summaries. This method can greatly improve the pertinence and accuracy of topic content extraction.

[0210] (4) Structured meeting minutes automatic generation process: Through a series of intelligent and automated processes from audio collection, speaker voiceprint recognition, speech text conversion and error correction, topic identification and classification merging, meeting summary generation to meeting minutes integration, a complete meeting minutes including meeting time, meeting location, participants, meeting summary, and topic summary is generated. This method effectively improves the efficiency of meeting minutes generation and reduces the errors and omissions that may occur in the traditional manual recording process.

[0211] The method of this embodiment aims to automatically generate high-quality, structured and complete meeting minutes by combining technologies such as voiceprint recognition, speech-to-text (ASR) and large language model (LLM), so as to solve the problem of low efficiency and easy omission of key information in traditional manual recording. Its main purpose is to improve the accuracy and completeness of meeting minutes, reduce errors in manual recording, and improve work efficiency. By automatically processing the content of the meeting, the system can not only identify different speakers, but also automatically distinguish topics, summarize discussion points and refine decision-making content. The invention is widely used in scenarios such as company meetings and project discussions, has important practical significance, and can be effectively applied to various scenarios that need to record and generate meeting minutes, such as:

[0212] Internal company meetings: including board meetings, project progress meetings, etc.

[0213] Government decision-making meetings: record the core topics discussed in the meetings and the decision-making process;

[0214] Other high-density information conference scenarios: such as product launches, customer meetings, etc.

[0215] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for intelligently generating meeting minutes files, characterized in that: The steps include: Step 1: Collect and upload the audio data of the meeting to the cloud server through the recording device; Step 2: After receiving the audio data uploaded in step 1, use the voiceprint recognition model to identify different speakers in the audio, and cut the complete audio into segments of single-person speech audio arranged in chronological order, and mark the speaker and start and end time of each audio segment; Step 3: Convert the speech segments segmented in step 2 into text content, and improve the accuracy of the text through an error correction model; Step 4: Based on vectorization technology and cosine similarity algorithm, the speech texts in the text content of step 3 are grouped into several sets according to chronological order to identify and classify topics; Step 5: Summarize and extract the topics, key points, and decision summaries corresponding to each set in order; Step 6: Concatenate the text content of each topic into a long text containing all the topics, and then process the long text through a large language model to summarize the meeting summary. Step 7: Store the meeting start and end time, participants, address information, meeting summary and agenda content in the corresponding meeting audio record in the database to form meeting minutes; Step eight, initiated by the client, extracts various information of the meeting minutes from the database, combines them to generate a word file of the meeting minutes, and downloads them to the local computer to complete the generation and export of the meeting minutes.

2. The method for intelligently generating meeting minutes files according to claim 1 is characterized in that: The recording device in step 1 includes: Processor and Wi-Fi connection module. The processor uses low-power microcontroller chips such as ESP32, which is responsible for core functions such as recording control, Wi-Fi management, and file transfer. It can access the Internet or local area network through the built-in Wi-Fi module; A storage unit, used for storing recording files; Microphone, used for recording, usually a built-in high-sensitivity condenser microphone, to collect audio data of speeches during meetings; Audio processing chips enable recording devices to record high-quality, clear audio signals through analog-to-digital conversion, signal processing, noise reduction, and echo cancellation, and provide a basis for subsequent audio uploading and processing; The GPS module is used to obtain the current longitude and latitude information when recording starts, and convert it into a specific address on the server through Baidu Map API, Amap API, and Tencent location service; A power management module, used to provide power supply for the recording device, including a battery, a charging circuit and a power control module; Physical or virtual buttons, USE interface, used to provide a physical interface or touch interface for controlling recording and Wi-Fi settings.

3. The method for intelligently generating meeting minutes files according to claim 2 is characterized in that: The cloud server in step one includes an audio library and a voiceprint library. The audio library is divided into an audio directory library and a meeting minutes database. The voiceprint library completes the training of SincNet through an open source data set. In the process of generating meeting minutes files, the speaker's voiceprint vector is generated by SincNet and compared with the data in the vector library. If it already exists, the speaker information is proposed; otherwise, a speaker ID is created and stored in the vector library together with the voiceprint vector.

4. The method for intelligently generating meeting minutes files according to claim 1, 2 or 3, characterized in that: The specific steps of performing voiceprint recognition and audio segmentation in step 2 are as follows: Step 21, preprocessing the audio; Step 22, dividing the long audio data into shorter segments; Step 2 and 3: Process each audio clip through the voiceprint feature extraction model SincNet to convert the voice data into a feature vector that can represent the speaker's identity. The model first extracts frame-level features from the audio, and then performs temporal aggregation through a global statistical pooling layer to obtain a vector representation of fixed dimensions. Step 24: extract voiceprint features from the segmented audio clips and perform similarity matching with speaker vectors in the voiceprint library; Step 25: After the speaker is identified, each continuous audio of the same speaker will be regarded as a speech in chronological order. The audio splicing belonging to the same speech content will be realized using Python libraries such as pydub or librosa, so as to merge discrete audio clips into audio clips with speech as the unit and reconstruct the complete dialogue scene.

5. The method for intelligently generating meeting minutes files according to claim 4 is characterized in that: The method of preprocessing the audio in step 21 is: first, performing noise reduction processing by spectrum subtraction or Wiener filtering to remove background noise and make the speech signal clearer; second, removing the silent segments in the audio, first using short-time energy technology to detect the silent area through Librosa, and comparing it with the set energy threshold to identify the silent segment; then, using audio cutting technology to remove the silent segment, retaining the non-silent segment, and restoring the audio stream by audio splicing; The sampling rate is converted to a standard value, such as 16kHz or 8kHz, and the speech is spectrally analyzed through short-time Fourier transform to enhance the speech features and highlight the pronunciation details.

6. The method for intelligently generating meeting minutes files according to claim 1, 2 or 3, characterized in that: The specific steps of converting the voice segment into text content in step 3 are as follows: Step 31, converting the speech fragment into text through automatic speech recognition technology; Step 32: Structuring the transcribed text; Step 33: Verify and correct the text to improve its accuracy.

7. The method for intelligently generating meeting minutes files according to claim 1, 2 or 3, characterized in that: The specific steps of topic identification and classification in step 4 are as follows: Step 41: Identify the boundaries of the topic; Step 42: Merge the topics.

8. The method for intelligently generating meeting minutes files according to claim 7 is characterized in that: The specific steps for identifying the topic boundaries in step 41 are as follows: Step 4: Initialize the topic set topic T = {t1, t2, .., t m }, each of its elements is a set of speech texts on different topics, m is the total number of topics; the conference speech text set is S = {s1, s2, .., s n }, whose elements are arranged in the order of speech, where n represents the total number of speeches, assuming that the speech text is s i , the text set of the current topic is t j , it either contains only s i-1 , either with s i-1 A continuous subset of speech text, such as t j ={s2,s3,...,s i-2 ,s i-1 }; Step 4.12: Calculate the current speech text s based on BERT i The embedding vector v i ; Set t j All speech texts are concatenated into a long text, and the topic vector is calculated Step 413: Calculate the current speech vector s i and the issue vector , and determine the relationship between the similarity and the threshold: When the similarity is lower than the threshold, it means that the current text belongs to a new topic. The text set of the new topic is initialized with the current speech text, which is recorded as t j+1 ={s i }, let i=i+1; Set the new topic as the current topic, set j = j + 1, and jump to e for execution; When the similarity is higher than or equal to the threshold: then the next speech text s i+1 Transformed into a high-dimensional vector v i+1 and Topic Vector If the similarity is higher than the threshold, it means that the current text does not belong to the boundary, and the current speech text s i Summarize to the current topic text set t j , t j ={s2,s3,...,s i-2 ,s i-1 ,s i }, let i=i+1, jump to e to execute; If the similarity is lower than the threshold, it indicates that the current speech belongs to the boundary; Step 414, for the current text s i It is a boundary problem. Assume that s i It contains two topics, i Topic segmentation is performed, starting from the beginning of the formal speech content, increasing the fixed step size d each time, and cutting the text into two sub-texts; If s b If it is empty, it means s i Only belongs to t j Single topic, will s i Included to j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s i }, let i=i+1, and j=j+1, jump to e for execution; Otherwise, convert the previous and next subtexts into vectors v respectively. a and v b , and respectively with the current topic vector Compare, if and only if s a With t j Belong to the same topic and s b When it is a new topic, stop cutting and a Add to t j In, t j ={s2,s3,...,s i-2 ,s i-1 ,s a }, and s b After adding the speaker information, add it to t j+1 In, t j+1 ={s b }, let i=i+1, and j=j+1, jump to e for execution; Step 4-15, start again from stage b until all speech texts are processed and classified.

9. The method for intelligently generating meeting minutes files according to claim 7 is characterized in that: The specific steps of topic merging in step 42 are as follows: Step 421: Construct a new topic set T′={t′1,t′2,..,t′ k }, k is the final number of topics, which is the set of topics after merging; Step 422, starting from the first topic in the original set, delete it from the original set and set it as the current topic to be merged, and compare it with the remaining topics in the original set in turn for similarity. If the similarity exceeds the threshold, merge them and delete the compared topic from the original set. Otherwise, continue to compare the next set until the comparison with all elements of the original set is completed, and add the newly merged set as the merged topic to the new topic set; Step 423, continue to execute operation b until the original topic set is empty.

10. The method for intelligently generating meeting minutes files according to claim 1, 2 or 3, characterized in that: The specific steps of summarizing and extracting the topics, key points, and decision summaries corresponding to each set in step 5 are as follows: Step 51: Use the pre-trained large language generation model combined with fine-tuning to complete the refinement of topic content, hot topics, and final conclusions or decisions; Step 52: Process each topic in turn. Based on the semantic understanding ability of the large language model, analyze the speech content, extract the views of each speaker, identify the hot issues discussed, and regenerate the topics, arguments and decision summaries through the generative ability of the model.

Images