Speaker separation method in intelligent conference system based on voiceprint recognition

Abstract

The invention relates to the technical field of computers, and discloses a voiceprint recognition-based speaker separation method in an intelligent conference system, and the method comprises the steps: obtaining a personnel sound signal of a conference participant through a microphone, carrying out the real-time preprocessing of the personnel sound signal, extracting a personnel voiceprint feature from the preprocessed sound signal, and carrying out the recognition of the voiceprint feature; the voiceprint features of the persons are coded, and a personalized voiceprint vector of each speaker is generated; the method comprises the following steps: when a conference is carried out, segmenting a voice stream in the conference into different speaking segments in real time, and sorting each speaking segment based on a time sequence; and identifying a corresponding speaker for each sorted speaking segment based on the personalized voiceprint vector, distributing a corresponding identity identifier for each speaker, and associating the identity identifier with the corresponding speaking content to output. According to the invention, the speaker separation accuracy in the conference can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of technology machines, and in particular to a method for separating speakers in an intelligent conference system based on voiceprint recognition. Background Art

[0002] With the increasing popularity of remote work and global communication, intelligent conference systems have become an indispensable tool for enterprises and organizations. However, existing conference recording technology has difficulty in automatically distinguishing the voices of different speakers, making post-conference collation and review complicated.

[0003] Although voiceprint recognition technology has made some progress in single speaker recognition, it still faces many challenges in actual meeting environments, such as environmental noise interference, difficulty in handling multiple people speaking at the same time, real-time requirements, cross-device consistency issues, and privacy protection.

[0004] From the above, we can see that how to improve the accuracy of speaker separation in meetings still needs to be solved. Summary of the invention

[0005] In order to improve the accuracy of speaker separation in a conference, the present application provides a method and device for speaker separation in an intelligent conference system based on voiceprint recognition.

[0006] In a first aspect, the present application provides a method for separating speakers in an intelligent conference system based on voiceprint recognition, which adopts the following technical solutions:

[0007] A method for separating speakers in an intelligent conference system based on voiceprint recognition comprises: obtaining voice signals of conference participants through a microphone, preprocessing the voice signals in real time, extracting voiceprint features from the preprocessed voice signals, encoding the voiceprint features, and generating a personalized voiceprint vector for each speaker, wherein the preprocessing comprises dynamic noise suppression and adaptive audio enhancement; when a conference is in progress, dividing the voice stream in the conference into different speech segments in real time, and sorting each speech segment based on a time series; identifying the corresponding speaker for each sorted speech segment based on the personalized voiceprint vector, assigning a corresponding identity to each speaker, and associating the identity with the corresponding speech content and outputting it.

[0008] Optionally, an identity authentication mechanism based on blockchain technology is included: a decentralized distributed ledger is configured to store the user's voiceprint data and its related identity information, each block contains the hash value of the previous block, forming an unalterable chain structure, each node in the blockchain network keeps a complete copy of the ledger, and all access and operations to the voiceprint data are recorded on the blockchain.

[0009] Optionally, in the case of multiple microphones in a conference room, the method includes: obtaining the acoustic characteristics of the conference room based on built-in sensors or external devices, wherein the acoustic characteristics include but are not limited to the reverberation time, background noise level, reflection and absorption characteristics of the room; and adjusting the microphone positions corresponding to the microphone array based on the acoustic characteristics.

[0010] Optionally, it includes: obtaining corresponding historical meeting data, which includes the speaker's historical speech records, voiceprint features, topics and discussion content in the meeting, meeting time, location and participant information, acoustic characteristics of the conference room, and system settings and user feedback; inputting a pre-trained prediction model based on the historical meeting data to obtain corresponding future speech patterns, which include predicting the speaking frequency of a specific speaker in the next meeting, the possible distribution of speech content, and the main topics that may appear in the meeting; and adjusting relevant settings of the intelligent conference system based on the speech pattern.

[0011] Optionally, it includes: calling the corresponding speech segment, obtaining the corresponding audio signal from the speech segment, converting the audio signal into text, preliminarily marking each speech segment, and generating an initial identity tag; calling the speaker's personal file, obtaining the context segment corresponding to the speech segment, the personal file including speech records in previous meetings, voiceprint features, common vocabulary and expression habits; calling the corresponding verification model, inputting the context segment and the personal file into the verification model, obtaining the corresponding verification result, and judging whether the identity tag needs to be adjusted based on the verification result.

[0012] Optionally, in the process of determining whether the identity tag needs to be adjusted based on the verification result, the method also includes: obtaining a verification score in the verification result, calling a preset verification threshold, and comparing the verification score with the verification threshold; generating a corresponding label correction report based on the comparison result, if the verification score is greater than the verification threshold, the identity tag is not adjusted, otherwise the identity tag is adjusted; after each meeting, the label correction report is displayed on the user visualization interface, allowing the user to view and confirm the label correction suggestions proposed by the system.

[0013] In a second aspect, the present application provides a speaker separation device in an intelligent conference system based on voiceprint recognition, which adopts the following technical solution:

[0014] A speaker separation device in an intelligent conference system based on voiceprint recognition, comprising:

[0015] A personalized voiceprint vector generation module, which obtains voice signals of conference participants through a microphone, performs real-time preprocessing on the voice signals, extracts voiceprint features from the preprocessed voice signals, and encodes the voiceprint features to generate a personalized voiceprint vector for each speaker, wherein the preprocessing includes dynamic noise suppression and adaptive audio enhancement;

[0016] A speech segment segmentation module, which is used to segment the speech stream in the conference into different speech segments in real time during the conference, and sort each speech segment based on the time series;

[0017] The identity identifier allocation module identifies the corresponding speaker of each of the sorted speech segments based on the individual voiceprint vector, is used to allocate a corresponding identity identifier to each speaker, and associates the identity identifier with the corresponding speech content and outputs it.

[0018] In a third aspect, the present application provides a method for separating speakers in an intelligent conference system based on voiceprint recognition, which adopts the following technical solutions:

[0019] A method for separating speakers in an intelligent conference system based on voiceprint recognition comprises a processor, wherein the program of any one of the above-mentioned methods for separating speakers in an intelligent conference system based on voiceprint recognition is run in the processor.

[0020] In a fourth aspect, the present application provides a storage medium, which adopts the following technical solution:

[0021] A storage medium stores a program of any one of the above-mentioned methods for separating speakers in an intelligent conference system based on voiceprint recognition.

[0022] In summary, the present application includes at least one of the following beneficial technical effects:

[0023] Accurately identify the speaker: Based on the individual voiceprint vector, each speech segment is identified to ensure that the speech content of each speaker is correctly marked.

[0024] Generate structured records: Assign an identity to each speaker and associate it with the corresponding speech content to generate structured meeting records for easy post-meeting organization and review.

[0025] Data security: Voiceprint data and its related identity information are stored in a decentralized distributed ledger to ensure data security and immutability.

[0026] Transparency and traceability: All access and operations to voiceprint data are recorded on the blockchain, providing a high degree of transparency and traceability.

[0027] Building trust: Through blockchain technology, users’ trust in the system is enhanced, especially in identity authentication scenarios involving sensitive information.

[0028] Predictive speaking patterns: Based on historical meeting data, use pre-trained prediction models to predict future speaking patterns, including speaking frequency, content distribution, and main topics.

[0029] Adjust system settings in advance: Based on the prediction results, adjust the relevant settings of the smart conference system (such as microphone array configuration, audio processing parameters, etc.) in advance to improve the conference quality and user experience.

[0030] Continuous optimization: By continuously accumulating and analyzing historical data, the system can continue to learn and optimize, further improving the accuracy of predictions and the performance of the system.

[0031] Preliminary labeling and verification: Each speech segment is preliminarily labeled and verified through personal profile and context information to generate an initial identity label.

[0032] Label correction: Generate a label correction report based on the verification results. If the verification score is lower than the threshold, adjust the identity label to ensure the accuracy of identity recognition.

[0033] User confirmation: After each meeting, the label correction report is displayed on the user visualization interface, allowing the user to view and confirm the label correction suggestions made by the system, thereby further improving the accuracy and reliability of the system through user feedback. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention is a flowchart showing a method for separating speakers in an intelligent conference system based on voiceprint recognition according to an exemplary embodiment.

[0035] Figure 2 The present invention is a structural block diagram of a speaker separation device in an intelligent conference system based on voiceprint recognition according to an exemplary embodiment. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0037] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0038] The present application embodiment discloses a method for separating speakers in an intelligent conference system based on voiceprint recognition. Figure 1 ,include:

[0039] S100, obtaining voice signals of conference participants through a microphone, preprocessing the voice signals in real time, extracting voiceprint features from the preprocessed voice signals, encoding the voiceprint features, and generating a personalized voiceprint vector for each speaker.

[0040] Among them, one or more microphones are used to capture all sound signals in the conference room. These microphones can be fixed or portable and distributed in different locations to ensure comprehensive coverage. In an embodiment of the present application, the conference room uses multiple microphones, and the number of microphones turned on and the array can be switched and controlled.

[0041] Preprocessing includes dynamic noise suppression and adaptive audio enhancement. Specifically, dynamic noise suppression: using adaptive filters or other noise reduction algorithms to detect and reduce the impact of background noise (such as air conditioning sound, keyboard tapping, etc.) in real time; adaptive audio enhancement: improving the quality and clarity of voice signals through technologies such as spectrum shaping and dynamic range compression. Spectrum shaping can highlight the key frequency components of voice and suppress noise frequencies, while dynamic range compression can reduce volume fluctuations, making the low-volume part clearer and more audible, and the high-volume part will not be overloaded and distorted.

[0042] After good preprocessing of the human voice signal, the quality of the voice signal can be improved, the interference of environmental noise can be reduced, and clean data can be provided for subsequent voiceprint feature extraction.

[0043] Then, advanced voiceprint feature extraction techniques, such as Mel-frequency cepstral coefficients (MFCC), linear predictive coding (LPC) or advanced features extracted by deep neural networks, are used to extract the voiceprint features of each speaker from the preprocessed speech signal; by extracting the voiceprint features that can uniquely identify each speaker, basic data is provided for subsequent identity recognition.

[0044] In addition, the extracted voiceprint features are input into a deep learning model (such as a convolutional neural network CNN, a recurrent neural network RNN, or a transformer) for encoding to generate a personalized voiceprint vector for each speaker. By generating a highly distinctive voiceprint vector, a reliable basis is provided for subsequent real-time recognition.

[0045] S110, when the meeting is in progress, the voice stream of the meeting is divided into different speech segments in real time, and each speech segment is sorted based on the time series.

[0046] Among them, the voice stream in the meeting is divided into different speech segments in real time, and endpoint detection can be used to determine the start and end time of the speech. The multi-speaker tracking algorithm can also be used to track the continuous speeches of different speakers.

[0047] By dividing the continuous speech stream into independent speech segments, it is convenient for subsequent speaker identification and content association, and each speech segment is sorted by time series to facilitate subsequent matching.

[0048] S120, identifying the corresponding speaker of each sorted speech segment based on the individual voiceprint vector, assigning a corresponding identity identifier to each speaker, and associating the identity identifier with the corresponding speech content and outputting the result.

[0049] Among them, the voiceprint features of each speech segment are compared with all known individual voiceprint vectors. Specifically, the similarity or distance between the feature vector of each speech segment and each individual voiceprint vector is calculated. Similarity calculation: Commonly used similarity calculation methods include cosine similarity, Euclidean distance, Manhattan distance, etc. For example, cosine similarity is used to measure the angle between two vectors, and the closer the value is to 1, the more similar they are. Threshold setting: In order to improve the accuracy of recognition, a similarity threshold can be set. Only when the similarity exceeds this threshold, the speech segment is considered to belong to a specific speaker. Speech segments below the threshold may be marked as unknown speakers or require further verification.

[0050] In addition, considering that speeches in meetings are usually continuous, time series information can be used to improve recognition accuracy. For example, if the previous speech segment has been correctly identified as a certain speaker, then the subsequent adjacent speech segments are likely to be the same person's continued speech. At this time, contextual information, such as the content of the conversation, the changing trend of voice features, etc., can be combined to help the system better understand the identity of the current speaker.

[0051] Once the speaker corresponding to each speech segment is identified, a corresponding identity identifier (such as ID number, name, etc.) is assigned to the speaker; the identity identifier is associated with the corresponding speech content to generate a structured meeting record.

[0052] By comparing the voiceprint features of each speech segment with the personalized voiceprint vector, the system can accurately identify the speaker corresponding to each speech segment. Accurate speaker identification helps to generate accurate meeting minutes, which can facilitate post-meeting organization and review, and improve the readability and practicality of meeting minutes.

[0053] In the process of assigning a corresponding identity to each speaker, it is necessary to ensure that the assigned identity is correct, so the following steps are also performed:

[0054] S1201, retrieve the corresponding speech segment, obtain the corresponding audio signal from the speech segment, convert the audio signal into text, preliminarily mark each speech segment, and generate an initial identity tag.

[0055] Among them, automatic speech recognition (ASR) technology is used to convert audio signals into text; the voice content is converted into analyzable text form to facilitate subsequent natural language processing and context analysis; and then each speech segment can be preliminarily marked by using preliminary voiceprint recognition technology to generate an initial identity tag.

[0056] S1202, retrieve the personal profile of the speaker and obtain the context segment corresponding to the speech segment.

[0057] Among them, the personal profile of each speaker is retrieved from the database, which includes speech records in previous meetings, voiceprint features, common vocabulary and expression habits, etc.; the context fragments related to the current speech fragment are extracted from the transcribed text of the meeting. The context fragments may include the previous and subsequent speech content, conversation topics, etc., which can help to more accurately understand the true identity of the current speaker through context information.

[0058] S1203, calling the corresponding verification model, inputting the context fragment and the personal profile into the verification model, obtaining the corresponding verification result, and judging whether the identity tag needs to be adjusted based on the verification result.

[0059] The context fragments and the speaker's personal profile are input into the verification model together. The verification model will comprehensively consider this information to verify the identity and determine whether the preliminary marked identity label needs to be adjusted based on the confidence score and other indicators of the verification result. If the verification result shows that the preliminary label is incorrect or the confidence is low, the identity label is adjusted.

[0060] In addition, in the process of determining whether the identity tag needs to be adjusted based on the verification result, the method further includes:

[0061] S1204, obtaining a verification score in the verification result, retrieving a preset verification threshold, and comparing the verification score with the verification threshold.

[0062] Among them, the verification score is usually a numerical value, which provides a quantitative indicator to evaluate the accuracy of the preliminary marked identity tag, which can indicate the degree to which the identity tag of the current speech segment matches the personal profile; the verification threshold is a pre-set numerical value used to determine whether the verification score is high enough to confirm the correctness of the identity tag.

[0063] In addition, the verification score obtained from the verification model is compared with the preset verification threshold. If the verification score is higher than or equal to the verification threshold, the identity label is considered correct; if the verification score is lower than the verification threshold, the identity label is considered to be incorrect.

[0064] S1205, generating a corresponding label correction report based on the comparison result, if the verification score is greater than the verification threshold, the identity label is not adjusted, otherwise the identity label is adjusted;

[0065] Among them, a label correction report is generated based on the comparison results. The report contains the initial identity label, verification score, verification threshold, and information on whether the identity label needs to be adjusted for each speech segment. The label correction report provides a detailed report that records the identity label verification status of each speech segment, which is convenient for subsequent processing and user review.

[0066] If the verification score is greater than or equal to the verification threshold, the system will keep the existing identity label unchanged. If the verification score is lower than the verification threshold, the system will adjust the identity label. The adjustment can include re-identifying the speaker, using contextual information to make inferences, or marking it as an unknown speaker.

[0067] S1206, after each meeting, the label correction report is displayed on the user visual interface, allowing the user to view and confirm the label correction suggestions proposed by the system.

[0068] After the meeting, the system will display the generated label correction report on the user visual interface. Users can view the identity label and verification status of each speech segment through this interface. Users can view the label correction suggestions made by the system on the visual interface and can choose to accept or reject these suggestions. If users reject certain suggestions, the system can record these feedbacks and use them for future optimization. In this way, the accuracy and reliability of the system can be further improved through user feedback, while providing a better user experience.

[0069] In an embodiment of the present application, in the case of multiple microphones in a conference room, the following are included:

[0070] S121, obtains the acoustic characteristics of the conference room based on built-in sensors or external devices.

[0071] Among them, built-in sensors in the conference room (such as microphones, acoustic probes, etc.) are used to measure and analyze the acoustic characteristics of the conference room. These sensors can be fixed or portable. External equipment (such as professional acoustic test equipment, sound level meters, reverberation time meters, etc.) is used to perform more detailed acoustic characteristic measurements. These devices usually have higher accuracy and more functions.

[0072] Acoustic properties include, but are not limited to, the room's reverberation time, background noise level, reflection and absorption characteristics. Specifically: Reverberation time: the time it takes for sound to gradually decay in a room, usually expressed as RT60 (i.e., the time it takes for the sound to decay by 60 decibels). Background noise level: the continuous noise level in the environment, such as air conditioning noise, traffic noise, etc. Reflection and absorption characteristics: the reflection and absorption characteristics of the surfaces in the room (such as walls, floors, ceilings) to sound. Different materials and structures will affect the propagation and quality of sound.

[0073] S122, adjusting microphone positions corresponding to the microphone array based on the acoustic characteristics.

[0074] The acquired acoustic characteristic data is input into the system for analysis and processing, and the system calculates the optimal microphone position and configuration based on the data.

[0075] Microphone Position Adjustment: The system can automatically adjust the position of the microphone through mechanical means or intelligent algorithms. For example, using a movable microphone stand or a motorized track system. The system generates microphone position adjustment suggestions and displays them to the user through the user interface. The user can manually adjust the microphone position based on the suggestions.

[0076] In an array of multiple microphones, the system takes into account the relative position and angle of each microphone to optimize the overall sound capture effect. For example, beamforming technology forms a directional beam to focus on a specific speaker while suppressing noise and interference from other directions.

[0077] First, through optimization based on acoustic characteristics, the system can better capture clear sound signals and reduce the impact of background noise and reverberation; by providing a high-quality audio experience, meeting participants can hear and understand the speech more clearly; the system can automatically adjust the microphone position under different environmental conditions to adapt to the acoustic characteristics of various conference rooms. Through automated and intelligent adjustments, the complexity and workload of manually setting the microphone position are reduced.

[0078] In addition, in the embodiment of the present application, the method further includes:

[0079] S123, obtaining corresponding historical conference data,

[0080] Among them, all historical meeting data related to the current meeting are retrieved from the database. These data include but are not limited to: Speaker's historical speech records: record the content, duration and frequency of each speaker's speech in past meetings. Voiceprint features: The voiceprint features of each speaker are used to identify and verify the identity of the speaker. Topics and discussion content in the meeting: record the main topics, topics of discussion and key points of each meeting. Time, location and participant information of the meeting: record the specific time, location and list of participants of each meeting. Acoustic characteristics of the conference room: including reverberation time, background noise level, reflection and absorption characteristics, etc. System settings and user feedback: record the system configuration of each meeting (such as microphone position, audio processing parameters, etc.) and user feedback.

[0081] S124, inputting a pre-trained prediction model based on historical conference data to obtain corresponding future speech patterns.

[0082] The acquired historical meeting data is input into a pre-trained prediction model. This model can be based on machine learning or deep learning technology, such as: Time series analysis: Use time series models such as ARIMA and LSTM to predict speech frequency and content. Cluster analysis: Use clustering algorithms such as K-means and DBSCAN to discover different types of speech patterns. Natural language processing (NLP): Use models such as BERT and Transformer to analyze and predict the main topics in the meeting. Multi-task learning: Combine multiple tasks (such as speech frequency prediction, content prediction, topic prediction) for joint training to improve overall performance.

[0083] In addition, the prediction model outputs future speech patterns, including but not limited to: Speech frequency: predict the number of times and duration a specific speaker will speak in the next meeting. Probability distribution of speech content: predict the topics and content that the speaker may discuss, as well as the probability of these contents appearing. Main topics that may appear in the meeting: predict the main topics and discussion points that may be involved in the meeting.

[0084] S125, adjusting relevant settings of the intelligent conference system based on the speaking mode.

[0085] Among them, the relevant settings include: microphone array configuration: automatically adjust the position and gain of the microphone according to the predicted speaker position and speaking frequency to ensure the best sound capture effect; audio processing parameters: adjust audio processing parameters such as noise suppression level and dynamic range compression to adapt to the predicted background noise and speech content; voiceprint recognition preloading: load the voiceprint model of a specific speaker in advance to speed up recognition; keyword and phrase preset: according to the predicted main topics, preset relevant keywords and phrases to generate meeting minutes more accurately; resource allocation: pre-allocate computing resources and storage space according to the predicted speaking frequency and content to ensure efficient operation of the system.

[0086] By executing steps S123 to S125, the audio quality and user experience of the meeting can be significantly improved by predicting future speaking patterns and adjusting system settings in advance; and the system can be flexibly responded to in different meeting scenarios, improving its adaptability and robustness; at the same time, through automated prediction and adjustment, the need for users to manually set system parameters is reduced, simplifying the operation process. By continuously accumulating and analyzing historical data, the system can continue to learn and optimize, further improving the accuracy of predictions and the performance of the system.

[0087] The method of the present application includes an identity authentication mechanism based on blockchain technology: a decentralized distributed ledger is configured to store the user's voiceprint data and its related identity information. Each block contains the hash value of the previous block to form an unalterable chain structure. Each node in the blockchain network stores a complete copy of the ledger, and all access and operations to the voiceprint data are recorded on the blockchain.

[0088] Among them, distributed ledger: Create a distributed ledger system to store the user's voiceprint data and its related identity information. This ledger is distributed on multiple nodes (such as servers, computers, etc.) instead of being stored in a centralized database. Decentralization: Each node saves a complete copy of the ledger, with no single control point or failure point, thereby improving the reliability and security of the system.

[0089] Through decentralization, the security and availability of data are ensured, and even if some nodes fail, the entire system can still operate normally.

[0090] In addition, block structure: Each block contains a set of transaction records (for example, the addition, modification or deletion of voiceprint data), as well as the hash value of the previous block. The hash value is a string of fixed length calculated by a hash function, which is unique and irreversible. Any slight change in data will cause a significant change in the hash value. Chain structure: Since each block contains the hash value of the previous block, a continuous chain structure is formed. New blocks are continuously added to the end of the chain, forming a growing blockchain.

[0091] The chain structure ensures that the data cannot be tampered with. Once the data of a block is tampered with, its hash value will also change, thus affecting the hash values ​​of all subsequent blocks, making the tampering behavior easy to detect. This ensures the integrity and credibility of the voiceprint data and its related identity information.

[0092] At the same time, full nodes: Each node keeps a complete copy of the ledger, which means that each node can independently verify all transaction records. Synchronization mechanism: The nodes maintain the consistency of the ledger through a synchronization mechanism. When a new block is added to the blockchain, all nodes will update their own copies of the ledger. Consensus mechanism: In order to ensure that the ledgers on all nodes are consistent, the blockchain network usually adopts a consensus mechanism (such as proof of work PoW, proof of stake PoS, etc.). Only when most nodes reach a consensus will the new block be added to the blockchain.

[0093] By having each node save a complete copy of the ledger, data transparency and verifiability are ensured. Even if some nodes are attacked or damaged, other nodes can still provide complete data backup. The consensus mechanism ensures the consistency of the ledger on all nodes and prevents malicious nodes from tampering with the data.

[0094] Whenever a user accesses or operates voiceprint data, these operations are recorded as a transaction and packaged into blocks and added to the blockchain. Timestamp: Each transaction contains a timestamp that records the time when the transaction occurred; Digital signature: Each transaction is digitally signed by the user to ensure the authenticity and integrity of the transaction; Transparency: All transaction records are publicly visible and can be viewed and verified by anyone.

[0095] By recording all access and operations, the transparency and traceability of data use are ensured. Any operation on voiceprint data can be traced and audited, enhancing the security and trust of the system. At the same time, digital signatures ensure the authenticity and integrity of transactions and prevent unauthorized access and operation.

[0096] The present application embodiment discloses a speaker separation device in an intelligent conference system based on voiceprint recognition, referring to Figure 2,include:

[0097] The personalized voiceprint vector generation module 200 obtains the voice signals of the participants in the conference through a microphone, performs real-time preprocessing on the voice signals, extracts the voiceprint features of the participants from the preprocessed voice signals, and encodes the voiceprint features of the participants to generate the personalized voiceprint vector of each speaker, wherein the preprocessing includes dynamic noise suppression and adaptive audio enhancement;

[0098] The speech segment segmentation module 210 is used to segment the speech stream in the conference into different speech segments in real time during the conference, and to sort each speech segment based on the time sequence;

[0099] The identity identifier assignment module 220 identifies the corresponding speaker of each sorted speech segment based on the individual voiceprint vector, is used to assign a corresponding identity identifier to each speaker, and associates the identity identifier with the corresponding speech content and outputs it.

[0100] Furthermore, the device includes but is not limited to: including an identity authentication mechanism based on blockchain technology:

[0101] A decentralized distributed ledger is configured to store the user's voiceprint data and its related identity information. Each block contains the hash value of the previous block, forming an unalterable chain structure. Each node in the blockchain network keeps a complete copy of the ledger, and all access and operations to the voiceprint data are recorded on the blockchain.

[0102] Furthermore, the device includes but is not limited to:

[0103] An acoustic characteristic acquisition module, which is used to acquire the acoustic characteristics of the conference room based on built-in sensors or external devices, wherein the acoustic characteristics include but are not limited to the reverberation time, background noise level, reflection and absorption characteristics of the room;

[0104] The microphone adjustment module is used to adjust the microphone positions corresponding to the microphone array based on the acoustic characteristics.

[0105] Furthermore, the device includes but is not limited to:

[0106] The historical meeting data acquisition module is used to acquire the corresponding historical meeting data, which includes the speaker's historical speech records, voiceprint features, the theme and discussion content of the meeting, the time, location and participant information of the meeting, the acoustic characteristics of the meeting room, and system settings and user feedback;

[0107] The speech pattern acquisition module uses a pre-trained prediction model based on historical meeting data to obtain the corresponding speech pattern in the future. The speech pattern includes predicting the speaking frequency of a specific speaker in the next meeting, the possible distribution of speech content, and the main topics that may appear in the meeting.

[0108] The setting adjustment module is used to adjust the relevant settings of the intelligent conference system based on the speaking mode.

[0109] Furthermore, the device includes but is not limited to:

[0110] A speech segment retrieval module is used to retrieve the corresponding speech segment, obtain the corresponding audio signal from the speech segment, convert the audio signal into text, preliminarily mark each speech segment, and generate an initial identity tag;

[0111] The personal profile adjustment module is used to retrieve the speaker's personal profile and obtain the context fragment corresponding to the speech fragment. The personal profile includes speech records in previous meetings, voiceprint features, common vocabulary and expression habits;

[0112] The verification result acquisition module calls the corresponding verification model, inputs the context fragment and the personal profile into the verification model to obtain the corresponding verification result, and determines whether the identity tag needs to be adjusted based on the verification result.

[0113] Furthermore, the device includes but is not limited to:

[0114] A verification score acquisition module is used to obtain the verification score in the verification result, call a preset verification threshold, and compare the verification score with the verification threshold;

[0115] A label correction report generation module is used to generate a corresponding label correction report based on the comparison result. If the verification score is greater than the verification threshold, the identity label is not adjusted, otherwise the identity label is adjusted;

[0116] After each meeting, the label correction report is displayed on the user visual interface, allowing the user to review and confirm the label correction suggestions made by the system.

[0117] An embodiment of the present application also discloses a method for separating speakers in an intelligent conference system based on voiceprint recognition, including a processor in which a program of any one of the above-mentioned methods for separating speakers in an intelligent conference system based on voiceprint recognition is run.

[0118] The embodiment of the present application further discloses a storage medium storing a program of any one of the above-mentioned methods for separating speakers in an intelligent conference system based on voiceprint recognition.

[0119] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A speaker separation method in an intelligent conference system based on voiceprint recognition, characterized in that: include: Acquire voice signals of conference participants through microphones, preprocess the voice signals in real time, extract voiceprint features from the preprocessed voice signals, encode the voiceprint features, and generate a personalized voiceprint vector for each speaker, wherein the preprocessing includes dynamic noise suppression and adaptive audio enhancement; When the meeting is in progress, the voice stream in the meeting is divided into different speech segments in real time, and each speech segment is sorted based on the time series; The speaker corresponding to each of the sorted speech segments is identified based on the individual voiceprint vector, a corresponding identity identifier is assigned to each speaker, and the identity identifier is associated with the corresponding speech content and outputted.

2. The speaker separation method based on voiceprint recognition in an intelligent conference system according to claim 1 is characterized in that: Including an identity verification mechanism based on blockchain technology: A decentralized distributed ledger is configured to store the user's voiceprint data and its related identity information. Each block contains the hash value of the previous block, forming an unalterable chain structure. Each node in the blockchain network keeps a complete copy of the ledger, and all access and operations to the voiceprint data are recorded on the blockchain.

3. The speaker separation method based on voiceprint recognition in an intelligent conference system according to claim 1 is characterized in that: In the case of multiple microphones in a conference room, these include: Acquiring acoustic characteristics of the conference room based on built-in sensors or external devices, wherein the acoustic characteristics include but are not limited to the reverberation time, background noise level, reflection and absorption characteristics of the room; The microphone positions corresponding to the microphone array are adjusted based on the acoustic characteristics.

4. The speaker separation method based on voiceprint recognition in an intelligent conference system according to claim 1 is characterized in that: include: Obtain the corresponding historical meeting data, which includes the speaker's historical speech records, voiceprint features, meeting topics and discussion content, meeting time, location and participant information, meeting room acoustic characteristics, system settings and user feedback; Based on the historical meeting data, a pre-trained prediction model is input to obtain the corresponding future speech pattern, which includes predicting the speaking frequency of a specific speaker in the next meeting, the possible distribution of the speech content, and the main topics that may appear in the meeting; Adjust relevant settings of the intelligent conference system based on the speaking mode.

5. The speaker separation method based on voiceprint recognition in an intelligent conference system according to claim 1 is characterized in that: include: Retrieving corresponding speech segments, obtaining corresponding audio signals from the speech segments, converting the audio signals into text, preliminarily marking each speech segment, and generating initial identity tags; Retrieve the speaker's personal profile and obtain the context fragment corresponding to the speech fragment. The personal profile includes speech records in previous meetings, voiceprint features, common vocabulary and expression habits; The corresponding verification model is retrieved, the context segment and the personal profile are input into the verification model, a corresponding verification result is obtained, and whether the identity tag needs to be adjusted is determined based on the verification result.

6. The speaker separation method based on voiceprint recognition in an intelligent conference system according to claim 5 is characterized in that: In the process of determining whether the identity tag needs to be adjusted based on the verification result, the method further includes: Obtaining a verification score in the verification result, retrieving a preset verification threshold, and comparing the verification score with the verification threshold; Generate a corresponding label correction report based on the comparison result, if the verification score is greater than the verification threshold, do not adjust the identity label, otherwise adjust the identity label; After each meeting, the label correction report is displayed on the user visual interface, allowing the user to review and confirm the label correction suggestions made by the system.

7. A speaker separation device in an intelligent conference system based on voiceprint recognition, characterized in that: include: A personalized voiceprint vector generation module, which obtains voice signals of conference participants through a microphone, performs real-time preprocessing on the voice signals, extracts voiceprint features from the preprocessed voice signals, and encodes the voiceprint features to generate a personalized voiceprint vector for each speaker, wherein the preprocessing includes dynamic noise suppression and adaptive audio enhancement; A speech segment segmentation module, which is used to segment the speech stream in the conference into different speech segments in real time during the conference, and sort each speech segment based on the time series; The identity identifier allocation module identifies the corresponding speaker of each of the sorted speech segments based on the individual voiceprint vector, is used to allocate a corresponding identity identifier to each speaker, and associates the identity identifier with the corresponding speech content and outputs it.

8. A method for speaker separation in an intelligent conference system based on voiceprint recognition, characterized in that: The method comprises a processor, wherein a program of the speaker separation method in an intelligent conference system based on voiceprint recognition as described in any one of claims 1 to 6 is run in the processor.

9. A storage medium, characterized in that: A program for the method for separating speakers in an intelligent conference system based on voiceprint recognition as described in any one of claims 1 to 6 is stored.

Images