TWI900067B - Method and system for generating meeting minutes - Google Patents

Abstract

A method for generating meeting minutes includes the follow steps. A video signal, an audio signal and source localization information of a video conference are obtained. Face recognition is performed on a plurality of image frames of the video signal to obtain a plurality of face recognition results. Voice print recognition is performed on a plurality of audio segments of the audio signal to obtain a plurality of voice recognition results at a plurality of timestamps. The voice recognition results are matched to the face recognition results according to the source localization information, in order to obtain a plurality of speaker identities. Speech to text is performed on the audio segments of the audio signal to obtain a transcript. The speaker identities are marked in the transcript to obtain a context according to the timestamps. Context understanding is performed on the context to obtain a meeting minutes report.

Description

Method and system for generating meeting minutes

This case relates to a method and system for generating meeting minutes, and more particularly to a method and system for generating meeting minutes suitable for multi-person meetings.

Current conferencing systems can automatically generate meeting transcripts through features like artificial intelligence speech-to-text recognition. Some systems then treat these transcripts as written text and use text summarization technology to comprehend the transcripts and generate meeting records.

However, a meeting involving multiple participants is different from a single article. A purely speech-to-text transcript fails to accurately capture the context of a multi-person meeting, limiting the practical application of this feature. A comprehensive meeting record must consider the entire context of the meeting. When summarizing a meeting, multiple factors must be considered to produce an appropriate transcript, such as the identities of the participants and the emotional dynamics of the interaction. These factors often cannot be captured by a single textual interpretation.

On the other hand, in real life, meetings can take many forms, including in-person, online, or hybrid. The interactive process in actual meetings is often multimodal. For example, participants may use cameras, microphones, text, and other devices. These audio and video devices may not all be used in a single meeting. Common scenarios include cameras not turning on, multiple participants sharing the same microphone, or accounts joining meetings.

Furthermore, meeting participants can change at any time, for example, if an invitee becomes unavailable due to an emergency, or someone not originally on the attendee list is added to the meeting. In scenarios where multiple people share a microphone, it can be difficult for attendees to identify the speaker. In multi-person meetings, participants often struggle to recognize everyone, leading to confusion and difficulty recording information.

Therefore, how to propose a method and system for generating meeting minutes to solve the above problems is an important issue in this field.

In some embodiments of this disclosure, a multimodal conferencing system architecture is proposed to enhance the multi-person conferencing experience and improve the recording of meeting context. The disclosed method and system simultaneously receive multimodal inputs, including video, audio, text input, and account numbers, to analyze participant identities and meeting content. This allows for real-time identification and tagging, and the delivery of meeting keynote recordings that take into account the complete meeting context.

This disclosure provides a method for generating a conference transcript. The method comprises the following steps: obtaining a video signal, a voice signal, and sound source localization information from a video conference; performing facial recognition on multiple image frames in the video signal to obtain multiple facial recognition results; performing voiceprint recognition on multiple voice segments in the voice signal to obtain multiple voiceprint recognition results at multiple time stamps; matching the voiceprint recognition results with the facial recognition results based on the sound source localization information to obtain multiple speaker identities at the time stamps; and performing speech-to-text recognition on the voice segments in the voice signal to obtain a text transcript. Based on these time stamps, the identities of these speakers are annotated to the transcript to obtain the text. Contextual understanding is performed on the text to obtain the meeting record.

In some embodiments, these voiceprint recognition results and these facial recognition results include multiple known identities and at least one unknown identity.

In some embodiments, the voiceprint recognition results include multiple unknown identities, and the facial recognition results include multiple known identities.

In some embodiments, the conference record generation method further includes the following steps: Determining the identities of the speakers among the known identities based on the sound source localization information and facial recognition results. Based on the time stamps, updating the unknown identities in the voiceprint recognition results with the speaker identities.

In some embodiments, the voiceprint recognition results include a plurality of known identities, and the facial recognition results include a plurality of unknown identities.

In some embodiments, the sound source localization information includes at least one of the angle and direction of each of the multiple sound sources.

In some embodiments, the facial recognition results include the locations of facial bounding boxes in these image frames and the identities corresponding to these facial bounding boxes.

In some embodiments, the meeting transcript generation method further includes the following steps: obtaining text input configured by at least one user; inserting the text input into the transcript based on a time sequence to generate an updated transcript; tagging the speakers' identities to the updated transcript based on the time stamps to obtain text; and performing contextual understanding on the text to obtain a meeting transcript.

In some embodiments, the meeting transcript generation method further includes the following steps: performing contextual understanding on the text to obtain multiple emotional meanings of multiple sentences in the text; removing portions of the text based on these emotional meanings to generate an updated text; and performing summary extraction on the updated text to obtain the meeting transcript.

This disclosure provides a conference record generation system. The conference record generation system includes a memory and a processor. The memory is used to store multiple instructions and data. The processor is electrically connected to the memory. The processor is used to access these instructions and data stored in the memory to execute the following steps. A video signal, a voice signal, and sound source localization information in a video conference are obtained. Facial recognition is performed on multiple image frames in the video signal to obtain multiple facial recognition results. Voiceprint recognition is performed on multiple voice segments in the voice signals to obtain multiple voiceprint recognition results at multiple time stamps. Based on the sound source localization information, the voiceprint recognition results are matched with the facial recognition results to obtain the identities of multiple speakers at these time stamps. Speech-to-text recognition is performed on the speech segments in the voice signal to obtain a text transcript. Based on the time stamps, the speaker identities are annotated to the text transcript to obtain a text. Contextual understanding is performed on the text to obtain a meeting transcript.

To make the above and other objects, features, advantages and embodiments of the present disclosure more clearly understood, the accompanying symbols are explained as follows:

100: Meeting Minutes Generation System

110: External device

111:Keyboard

112: Mouse

113: Microphone Array

114: Camera

120: Electronic devices

121: Processor

122: Memory device

123:(Touch) Display

124: Sound Sensor

125: Image sensor

126:Keyboard

130: Internet

140: Server

200:Meeting record generation method

410: Voice and facial features

420: Text Records

421~426: Sentences

430: Text

431~436: Emotional Semantics

440: Meeting Minutes

441: Meeting Title

442: Action Items

TRX: Text Record

S210, S220, S230, S240, S250, S260, S270: Steps

S281, S282, S290, S291, S292, S293, S294: Steps

S212, S214, S216: Steps

S222, S224, S226, S228, S229: Steps

S230, S242, S244, S246, S248, S249, S250: Steps

S272, S274, S276, S278, S281, S282: Steps

To make the above and other objects, features, advantages, and embodiments of the present disclosure more clearly understood, the accompanying drawings are described as follows: Figure 1 is a schematic diagram of a meeting record generation system according to some embodiments of the present disclosure.

Figure 2 is a schematic diagram of a method for generating meeting minutes according to some embodiments of the present disclosure.

Figure 3 is a schematic diagram of a method for generating meeting minutes according to some embodiments of the present disclosure.

Figure 4 is a schematic diagram of the process of generating a meeting record according to some embodiments of the present disclosure.

The following examples are illustrated in detail with accompanying diagrams. However, these examples are not intended to limit the scope of this disclosure, and the description of the structural operation is not intended to limit the order of execution. Any device with equivalent functionality produced by recombining these components is within the scope of this disclosure. Furthermore, the diagrams are for illustrative purposes only and are not drawn to scale. To facilitate understanding, identical or similar components will be labeled with the same reference numerals in the following description.

Unless otherwise specified, terms used throughout this specification and claims generally have the ordinary meanings they have in the art, within the context of this disclosure, and within the specific context. Furthermore, the terms "include," "including," "having," "having," and "containing" are intended to be open-ended, meaning "including, but not limited to." Furthermore, the term "and/or" as used herein includes any and all combinations of one or more of the listed items.

Please refer to Figure 1, which is a schematic diagram of a meeting record generation system 100 according to some embodiments of the present disclosure.

As shown in FIG1 , the conference transcript generation system 100 includes an external device 110 , an electronic device 120 , and a server 140 . In some embodiments, the external device 110 includes a keyboard 111 , a mouse 112 , a microphone array 113 , and a camera 114 . In some embodiments, the microphone array 113 functions as a sound source localization technology.

In some embodiments, the microphone array 113 uses sound localization algorithms such as Time Difference of Arrival (TDOA) and beamforming to locate the source of a sound and thereby obtain sound source localization information. In some embodiments, the sound source localization information includes at least one of an angle and a direction for each of multiple sound sources. In some embodiments, the sound source localization information includes the angle and/or direction of each speaker during a meeting in a conference venue.

In some embodiments, the electronic device 120 is electrically connected to an external device to receive audio, video, and/or text input from the external device 110. In some embodiments, the electronic device 120 includes a processor 121, a memory device 122, a (touch) display 123, an audio sensor 124, an image sensor 125, and a keyboard 126. In some embodiments, users capture images of the conference room using a camera 114 of an external device 110 or an image sensor 125 of an electronic device 120. Microphone array 113 or an audio sensor 124 receives sounds within the conference room and localizes the sound sources. These images are transmitted via a network 130 to a server 140, where they are streamed to generate video and audio signals for the video conference.

In some embodiments, the user inputs text via the keyboard 111, mouse 112 of the external device 110 and/or the touch display 123, keyboard 126 of the electronic device 120.

It should be noted that Figure 1 only depicts one electronic device, but it is understood that the video conference is conducted using electronic devices used at each conference venue. In some embodiments, the equipment used at each conference venue corresponds to electronic device 120 and/or external device 110 in Figure 1. In some embodiments, server 140 receives data such as images, audio, and/or sound source data from each conference venue via network 130. Based on the data such as images, audio, and/or sound source data from each conference venue, server 140 identifies the speakers in the audio recordings during the conference and generates a conference record based on text content tagged with the speakers. In some embodiments, the meeting minutes referred to in this disclosure refer to meeting minutes. In some embodiments, the meeting minutes referred to in this disclosure refer to meeting summaries, but this disclosure is not limited to these.

In some embodiments, server 140 includes a memory and a processor. The memory is used to store data and computer-executable instructions. The processor is electrically coupled to the memory and accesses data or instructions from the memory to execute the steps of the meeting record generation method 200 in Figures 2 and 3. In some embodiments, the memory may include dynamic memory, static memory, a hard drive, and/or flash memory. In some embodiments, the processor includes a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), an application-specific integrated circuit (ASIC), or any equivalent processing circuit. Therefore, the present disclosure is not limited thereto.

Please refer to Figure 1 and Figure 2. Figure 2 is a schematic diagram of a meeting record generation method 200 according to some embodiments of the present disclosure. As shown in Figure 2, the meeting record generation method 200 includes steps S210 to S294. In some embodiments, steps S210 to S294 can be performed by the meeting record generation system 100 in Figure 1. In some embodiments, steps S210 to S294 can be performed by the server 140 in Figure 1. In some embodiments, the server 140 receives images, sounds, text inputs, and accounts collected by devices/equipment at each conference venue, and uses the four modalities to perform weighted correlation and integration to analyze the meeting situation record. After the meeting concludes, the system will send a meeting record to participants based on the above information and simultaneously update the profile database. This database stores personnel data and historical data, including voiceprints, images, accounts, behavioral analysis, and other data, for continued use in subsequent meetings, continuously and automatically improving the system's integrity. In some embodiments, the four modalities include an image branch, an audio branch, a text branch, and an account branch. In some embodiments, step S220 corresponds to the image branch, and steps S230-S240 correspond to the audio branch. In some embodiments, steps S250-S260 correspond to the text branch, and step S270 corresponds to the account branch.

In step S210 , the video conference begins. In some embodiments, the video conference begins after participants in each conference venue join the video conference.

In step S220, facial recognition is performed. In some embodiments, the facial recognition task can be performed by a facial recognition neural network. In some embodiments, the facial recognition task can be performed by a facial recognition neural network whose main architecture is YOLOv5. In other embodiments, the facial recognition task can be performed by other neural networks that can recognize facial features, but the present disclosure is not limited thereto. Face detection is performed on multiple image frames in the video signal of the video conference to obtain all facial bounding boxes in each conference venue, and the identities of the participants in these facial bounding boxes are identified based on the data in the database to obtain facial recognition results. For example, if the facial bounding boxes of one or more participants are recognized as known identities, the identities of the corresponding participants may be labeled on those facial bounding boxes. On the other hand, if the facial bounding boxes of one or more participants are recognized as unknown identities, the identities of those participants may be labeled as unknown. In some embodiments, some of the aforementioned participants may be participating in a video conference at the same conference venue using the same account, while other participants may be participating in a video conference at different conference venues using different accounts.

In step S230 , sound source localization is performed. In some embodiments, multiple participants in a conference venue may use the same account to conduct a video conference with participants in other venues. The identities of these participants cannot be directly determined through the conference account. Therefore, in such a multi-person conference venue, sound source localization can be performed using microphone array 113 to obtain sound source localization information. This sound source localization information includes the angle and/or direction of each speaker within the conference venue during the meeting.

In step S240, voice recognition is performed. In some embodiments, voice recognition can be a voiceprint recognition task, which can be performed by a voiceprint recognition neural network. In some embodiments, the voiceprint recognition task can be performed by a voiceprint recognition neural network based on a machine learning architecture using PyTorch. In other embodiments, voiceprint recognition can be performed by a neural building block based on pyannote.audio or other neural networks capable of voiceprint recognition, but the present disclosure is not limited thereto. In some embodiments, voiceprint recognition is performed on the voice signal of a video conference based on database data to identify whether the speaker of each voice segment in the voice signal is known or unknown. Based on the time sequence/time stamp, the identified identity is tagged to the corresponding voice segment to obtain the voiceprint recognition result.

In some embodiments, the voiceprint recognition result obtained in step S240 and the facial recognition result obtained in step S210 include multiple known identities and at least one unknown identity.

In some embodiments, the voiceprint recognition result obtained in step S240 includes multiple unknown identities, and the facial recognition result obtained in step S210 includes multiple known identities. In some embodiments, the voiceprint recognition result obtained in step S240 includes multiple unknown identities, and the facial recognition result obtained in step S210 includes multiple known identities.

In some embodiments, the voiceprint recognition result obtained in step S240 includes multiple known identities, and the facial recognition result obtained in step S210 includes multiple unknown identities. In some embodiments, the voiceprint recognition result obtained in step S240 includes multiple known identities, and the facial recognition result obtained in step S210 includes multiple unknown identities.

In step S281, speaker matching is performed. In some embodiments, the sound source angle obtained in step S230 is used to match the voiceprint recognition results with the facial recognition results to obtain the speaker's identity at certain time stamps. In some embodiments, the position of the microphone array is known, and the voiceprint recognition results can be compared and linked with the facial recognition results based on the sound source angle. In some embodiments, the facial recognition results include the size and position of the facial bounding boxes in the video frame of the video conference, as well as the known or unknown identities corresponding to each facial bounding box.

In some embodiments, if the voiceprint recognition result at a time stamp indicates an unknown identity and the facial recognition result at the same time stamp indicates a known identity, the speaker's identity can be identified by comparing the position of the participant's facial bounding box with the sound source angle, and the speaker's voiceprint features can be added to the database. For example, if the database does not contain a participant's voiceprint features but does contain the participant's facial features and their corresponding identity, the sound source angle can be used to compare the position of the facial bounding box in the image. Based on the identity marked on the facial bounding box, the participant's voiceprint data/features can be added to the database to update the participant's personal profile. In other embodiments, if the voiceprint recognition result at a time stamp indicates a known identity and the facial recognition result at the same time stamp indicates an unknown identity, the speaker's facial bounding box can be identified by comparing the sound source angle with the speaker's position in the image, and the speaker's facial features can be added to the database. For example, if the database does not contain a participant's facial features but does contain the participant's voiceprint features and the corresponding identity, the sound source angle can be used to compare the facial bounding box position in the image. Based on the identity and time stamp tagged to the voice clip, the participant's facial data/features can be added to the database to update the participant's profile. In some further embodiments, if both the voiceprint recognition result and the facial recognition result at a time stamp are of a known identity, the voiceprint recognition result and the facial recognition result can be weighted and correlated based on the sound source angle, thereby improving the accuracy of speaker identification.

In step S250, speech-to-text recognition is performed. In some embodiments, speech-to-text recognition is performed on multiple speech segments in the speech signal to obtain a text transcript. In some embodiments, the speech-to-text recognition task can be performed by a convolutional neural network. In some embodiments, the speech-to-text recognition task can be performed by a neural network primarily based on Whisper. In other embodiments, the speech-to-text recognition task can be performed by Maestro or other neural networks capable of speech-to-text conversion, but this disclosure is not limited thereto.

In step S260, text input is obtained. In some embodiments, text input by the participant via keyboard 111, 126, mouse 112, or (touch) display 123 is obtained.

In step S270, user configuration is obtained. In some embodiments, text input configured by each user in the video conference may be obtained.

In step S282, context matching is performed. In some embodiments, the text input is inserted into the text record based on the time sequence to obtain an updated text record.

In step S290 , context understanding is performed. In some embodiments, the speaker's identity is tagged to the text record/updated text record to obtain text, and context understanding is performed on the text record to generate and obtain the meeting record in step S293 . In some embodiments, step S290 may be followed by step S291 or step S293 .

In step S291, the configuration database is updated. In some embodiments, the speaker matching results obtained in step S281 can be used to update the personal data in the database (e.g., facial features, voiceprint features, account data, and/or behavioral analysis). In some embodiments, the text obtained in step S290 can be used to update the meeting data in the database for continued use in subsequent meetings.

In step S292, the video conference ends.

In step S294, the meeting minutes are sent to all participants.

Please refer to Figures 1, 2, and 3. Figure 3 is a schematic diagram of a conference record generation method 200 according to some embodiments of the present disclosure. The conference record generation method 200 includes step S210. In some embodiments, all the steps in Figure 3 can be performed by the conference record generation system 100 of Figure 1. In some embodiments, all the steps in Figure 3 can be performed by the server 140 of Figure 1. In some embodiments, steps S212 and S222-S229 correspond to the image branch, and steps S214, S230, and S242-S249 correspond to the audio branch. In some embodiments, steps S250 and S262 correspond to the text branch, and steps S216 and S272-S278 correspond to the account branch. In some embodiments, facial recognition in step S220 of FIG. 2 includes steps S222-S229, and voice recognition in step S220 of FIG. 2 includes steps S242-S249. In some embodiments, step S260 of FIG. 2 includes step S262, and step S270 of FIG. 2 includes steps S272-S278.

In step S212 , it is determined whether the camera is turned on. In some embodiments, if the camera is turned on, it can capture images of the conference room and transmit them to the server for streaming, thereby obtaining the video conference image signal S _VID . If the camera is not turned on, the identity of the person at the conference room is determined to be unknown in the facial recognition branch.

In step S222 , face detection is performed. In some embodiments, face detection is performed on the image frame of the image signal S _VID to generate a face bounding box and facial features FT _FACE within the face bounding box.

In step S224, feature matching is performed. In some embodiments, the facial features FT _FACE in the face bounding box are matched with facial feature data in a database.

In step S226, a determination is made as to whether the facial features FT _FACE within the facial bounding box match the facial feature data in the database. If so, step S228 is executed to display the identity. If not, step S229 is executed in the face recognition branch, where the identity is determined to be unknown. This yields the facial recognition result R _EACE for the facial bounding box. The facial recognition result R _EACE represents the known identity and/or unknown identity for each facial bounding box.

In step S230, sound source localization is performed to obtain the sound source angle ANG _SD .

In step S214 , it is determined whether the microphone is on. In some embodiments, if the microphone is on, it can pick up audio from the conference room and transmit it to the server for streaming, thereby obtaining the video conference voice signal S _AUD . If the microphone is not on, the identity of the person at the conference room is determined to be unknown in the voiceprint recognition branch.

In step S242, a voiceprint is extracted. In some embodiments, a voiceprint feature FT _SD of a speech segment in the speech signal S _AUD is extracted.

In step S244, the speaker's identity is identified. In some embodiments, the voiceprint features _FTSD of the speech segment are matched with voiceprint features stored in a database.

In step S246, a determination is made as to whether the voiceprint features FT _SD of the speech segment match those stored in the database. If so, step S248 is executed to display the identity. If not, step S249 is executed in the voiceprint recognition branch, where the identity is determined to be unknown. This yields the voiceprint recognition result R _SD for the speech segment. The facial recognition result R _SD represents the known and/or unknown identity for each speech segment.

In step S250, speech-to-text recognition is performed. In some embodiments, speech-to-text recognition is performed on speech segments in the speech signal to obtain a text record TRX.

In step S216, it is determined whether the input account number is received. In some embodiments, the account number for joining the video conference is obtained.

In step S272, a check is performed to determine whether the account joining the video conference matches the account information stored in the database. If the user joins the conference using a known account, such as a video conferencing software account or a known email address, this branch compares the account information with the database data. If the account information exists in the database, the account identity is marked as the identity displayed in the database; otherwise, it is marked as unknown. The identified account identity is passed to the text processing branch for subsequent marking. If a match is found, step S274 is executed to display the identity. If not, step S276 is executed to determine that the identity is unknown.

In step S278, configuration identification is performed to identify the account information R _PROF (e.g., personal identity, company name, etc.) of the account (e.g., video conferencing software account or mailbox) that joins the video conference.

In step S262, determine whether text input has been received. If not, directly output the text record TRX and proceed to step S293. If text input has been received, execute step S282.

In step S282, context matching is performed. In some embodiments, the text input corresponding to the account data R _PROF is inserted into the text record TRX according to the time sequence to generate an updated text record, and the speaker identity is annotated to the updated text record according to the time stamp to obtain text.

In step S293, context understanding is performed. In some embodiments, context understanding is performed on the text to obtain a meeting transcript. In some embodiments, context understanding is performed on the text to obtain multiple emotional meanings of multiple sentences in the text, and based on these emotional meanings, some content in the text is removed to generate an updated text. In some embodiments, summary extraction is performed on the updated text to obtain a meeting transcript.

Please refer to Figure 4. Figure 4 is a schematic diagram illustrating the process of generating a meeting transcript according to some embodiments of the present disclosure. In some embodiments, the context understanding operation in step S290 may be implemented by a context processing module. The context processing module combines previously identified identities, emotional semantics, and context to create a comprehensive summary of the meeting's key points and identify the corresponding speakers. The generated meeting transcript can be automatically emailed to all identified attendees after the meeting. In some embodiments, the context processing module may be implemented by a summary extraction neural network model. In some embodiments, the context processing module may be implemented by a context understanding neural network model. In some embodiments, the context processing module may be implemented by a cross-modal neural network that detects emotions from images and sounds. In some embodiments, the context processing module can be implemented by a neural network model capable of performing multi-person conversation language parsing tasks. Therefore, the present disclosure is not limited thereto.

As shown in FIG. 4 , in step S290 , the speaker and emotional meaning 431 - 436 of sentences 421 - 426 in the text transcript 420 are identified using the participant's voice and facial features 410 to generate text 430 . In step S293 , the meeting transcript 440 includes a meeting title 441 and action items 442 corresponding to the speaker's identity. Furthermore, joking content is removed from the text transcript 420 , reducing unimportant content and improving the quality of the transcript 440 .

In summary, the disclosed meeting transcript generation system 100 and method 200 can improve the situation where voiceprint data is not available in the database by matching facial features with voiceprint features. This allows different speakers in the same meeting to be identified even when facial data is available but voiceprint data is not. Furthermore, by tagging speaker identities to the generated text, contextual understanding of multi-person conversations is facilitated, resulting in higher-quality meeting transcripts.

Although this disclosure has been described above in terms of implementation, it is not intended to limit the present disclosure. Anyone with ordinary skill in the art may make various modifications and enhancements without departing from the spirit and scope of this disclosure. Therefore, the scope of protection of this disclosure shall be determined by the scope of the attached patent application.

100: Meeting Minutes Generation System

110: External device

111:Keyboard

112: Mouse

113: Microphone Array

114: Camera

120: Electronic devices

121: Processor

122: Memory device

123:(Touch) Display

124: Sound Sensor

125: Image sensor

126:Keyboard

130: Internet

140: Server

Claims (9)

A method for generating a conference record comprises: obtaining a video signal, a voice signal, and sound source positioning information in a video conference; performing facial recognition on a plurality of image frames in the video signal to obtain a plurality of facial recognition results; performing voice print recognition on a plurality of voice segments in the voice signal to obtain a plurality of voice print recognition results at a plurality of time stamps; and matching the voice print recognition results with the facial recognition results according to the sound source positioning information to obtain a The invention relates to a method for obtaining a meeting record by identifying a plurality of speakers at the time stamps; performing speech-to-text recognition on the speech segments in the speech signal to obtain a text record; obtaining a text input configured by at least one user; inserting the text input into the text record according to a time sequence to generate an updated text record; tagging the speaker identities to the updated text record according to the time stamps to obtain a text; and performing context understanding on the text to obtain a meeting record. The method for generating meeting records as described in claim 1, wherein the voiceprint recognition results and the facial recognition results include multiple known identities and at least one unknown identity. The method for generating a meeting record as described in claim 1, wherein the voiceprint recognition results include a plurality of unknown identities, and wherein the facial recognition results include a plurality of known identities. The conference record generation method as described in claim 3 further includes: determining the identities of the speakers among the known identities based on the sound source positioning information and the facial recognition results; and updating the unknown identities of the voiceprint recognition results according to the speaker identities based on the time stamps. The method for generating a meeting record as described in claim 1, wherein the voiceprint recognition results include a plurality of known identities, and wherein the facial recognition results include a plurality of unknown identities. A method for generating conference records as described in claim 1, wherein the sound source localization information includes at least one of an angle and a direction of each of a plurality of sound sources. The method for generating a meeting record as described in claim 1, wherein the facial recognition result includes the position of the facial bounding boxes in the image frames and an identity corresponding to each of the facial bounding boxes. The meeting record generation method as described in claim 1 further includes: performing contextual understanding on the text to obtain multiple emotional semantics of multiple sentences in the text; removing part of the content in the text based on the emotional semantics to generate an updated text; and extracting a summary of the updated text to obtain the meeting record. A conference record generation system includes: a memory for storing a plurality of instructions and data; and a processor electrically connected to the memory, the processor for accessing the instructions and data stored in the memory to execute the following steps: obtaining a video signal, a voice signal, and sound source positioning information in a video conference; performing facial recognition on a plurality of image frames in the video signal to obtain a plurality of facial recognition results; performing voiceprint recognition on a plurality of voice segments in the voice signal to obtain a plurality of voiceprint recognition results at a plurality of time stamps; and Based on the sound source localization information, the voiceprint recognition results are matched with the facial recognition results to obtain a plurality of speaker identities at the time stamps; speech-to-text recognition is performed on the speech segments in the voice signal to obtain a text transcript; text input configured by at least one user is obtained; the text input is inserted into the text transcript based on a time sequence to generate an updated text transcript; the speaker identities are annotated to the updated text transcript based on the time stamps to obtain a text; and contextual understanding is performed on the text to obtain a meeting transcript.