Speech recognition processing method, system, device and medium based on conference scenario

Abstract

The application relates to a conference scene-based speech recognition processing method, system, device and medium, and belongs to the technical field of speech processing. The speech recognition processing method comprises the following steps: acquiring original conference audio stream collected by a microphone array; performing signal preprocessing on the collected original conference audio stream in a main channel pair, and outputting a pure speech signal; generating a sound source direction heat map based on the original conference audio stream; extracting a multi-dimensional voiceprint feature vector from the pure speech signal and dynamically grouping the multi-dimensional voiceprint feature vector, outputting a voice segment set labeled with a voiceprint ID, and generating an initial transcription text; dynamically correcting the initial transcription text, and outputting a transcription text stream with an industry term label; performing periodic memory enhancement processing on the transcription text stream, outputting a long text, and performing analysis to generate structured conference minutes data. The application can improve the automation level and accuracy of conference speech processing.

Description

Technical Field

[0001] This application relates to the field of speech processing technology, and in particular to a speech recognition processing method, system, device and medium based on a conference scenario. Background Technology

[0002] With the rapid development of artificial intelligence and intelligent office systems, meeting rooms, as important venues for corporate collaboration and decision-making, are experiencing a growing demand for high-quality voice capture and automated meeting recording. Traditional meeting recording and transcription methods often rely on single microphone devices, which struggle to cope with issues such as voice attenuation, reverberation interference, and noise pollution caused by long-distance sound pickup, leading to a significant decrease in recognition accuracy. Furthermore, in complex meeting environments with multiple speakers, effectively distinguishing between different speakers, accurately capturing their content, and achieving structured organization of meeting information have become critical issues that current intelligent meeting systems urgently need to address.

[0003] Especially in specialized fields such as transportation and infrastructure construction, industry conferences are characterized by highly specialized language, dense terminology, and complex contexts. Specialized terms in transportation planning meetings, such as road classification, traffic flow analysis, and intelligent transportation system deployment, as well as technical terms in infrastructure project meetings, such as engineering standards, construction specifications, and material specifications, place higher demands on the professional recognition capabilities of speech recognition systems. Traditional speech recognition systems often misrecognize or misunderstand specialized terms such as "interchange," "pavement bearing capacity," "prefabricated assembly structure," and "BIM modeling," seriously affecting the accuracy of meeting minutes and the progress of subsequent work.

[0004] While some microphone array-based speech enhancement and sound source localization methods have been applied to conference systems, traditional speech recognition systems are highly susceptible to interference in far-field sound pickup due to the prevalence of spatial reverberation, environmental noise, and simultaneous or alternating speech by multiple people in real-world conference room environments. This leads to a decline in speech signal quality and consequently affects subsequent recognition accuracy. Even simple multi-microphone acquisition schemes often lack effective spatial perception mechanisms and acoustic feature modeling methods, making it impossible to effectively separate and track multiple sound sources. Furthermore, existing speech recognition systems perform poorly in recognizing specialized terminology, easily leading to semantic ambiguity or recognition errors, which affects the professionalism and accuracy of the transcription results. This problem is particularly prominent in highly specialized industry conferences such as transportation and infrastructure. Summary of the Invention

[0005] To improve the automation level and accuracy of conference speech processing, this application provides a speech recognition processing method, system, device and medium based on conference scenarios.

[0006] Firstly, this application provides a speech recognition processing method based on a meeting scenario, employing the following technical solution:

[0007] A speech recognition processing method based on a meeting scenario, the recognition processing method includes,

[0008] Acquire the raw conference audio stream captured by a microphone array deployed in the conference room; wherein the microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones;

[0009] The main channel performs signal preprocessing on the acquired raw conference audio stream and outputs a clean voice signal after noise reduction.

[0010] Based on the original conference audio stream acquired by the microphone array, sound source localization is performed to generate a sound source location heatmap;

[0011] Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories are extracted from the pure speech signal to construct a multi-dimensional voiceprint feature vector;

[0012] Based on the heat map of the sound source location, a spatial constraint clustering algorithm is used to dynamically group the multi-dimensional voiceprint feature vectors. Combined with semantic coherence, conflicting segments in the clustering are arbitrated, and a set of speech segments labeled with voiceprint IDs is output.

[0013] The set of speech segments is input into a general speech recognition model to generate initial transcribed text;

[0014] The pre-configured industry terminology library is loaded to dynamically correct the initial transcribed text, and the transcribed text stream with industry terminology tags is output.

[0015] The transcribed text stream is subjected to periodic memory enhancement processing to output a semantically coherent long text.

[0016] The long text is parsed, and the core issues are identified by combining the industry terminology tags. The voiceprint ID is associated with the person responsible for the task to be completed, and structured meeting minutes data is generated.

[0017] By adopting the above technical solutions, a complete technical solution from physical layer signal acquisition to cognitive layer semantic understanding is constructed. Physical layer signal purification and spatial perception are achieved through microphone noise reduction and array localization technology; refined voiceprint expression at the perception layer is achieved through biometric modeling such as Mel frequency cepstral coefficients, fundamental frequency jitter, and formant trajectories; high-precision speaker segmentation at the cognitive layer is achieved through spatial constraint clustering and semantic conflict arbitration mechanisms; professional semantic adaptation at the cognitive layer is achieved through a collaborative mechanism between a general model and a domain terminology database, as well as meta-learning optimization; long-text semantic consistency at the decision layer is maintained through a periodic memory-enhanced LSTM architecture; and finally, automated output of meeting minutes is achieved through multimodal semantic parsing and structured generation mechanisms.

[0018] Optionally, the step of preprocessing the acquired raw conference audio stream on the main channel to output a noise-reduced clean speech signal includes:

[0019] The original conference audio stream acquired by the main channel is collected to obtain the first channel audio signal and the second channel audio signal;

[0020] Dynamic time warping is performed on the first and second channel audio signals. By calculating the peak-aligned phase difference of the cross-correlation function of the two signals, a phase-synchronized dual-channel audio signal is output.

[0021] A noise mask matrix is ​​generated based on the phase-synchronized dual-channel audio signal. Steady-state noise regions are identified by analyzing the energy distribution in the time-frequency domain, and a binary mask matrix is ​​constructed.

[0022] The phase-synchronized dual-channel audio signal is input into an adaptive filter bank, and nonlinear echo cancellation is performed using the binary mask matrix as a reference signal to output a primary noise reduction signal.

[0023] Subband spectral entropy analysis is performed on the primary noise-reduced signal to detect and filter out transient noise pulses, and a clean speech signal after noise reduction is output.

[0024] By adopting the above technical solutions, deep purification of voice signals in conference scenarios is achieved, significantly improving the accuracy of speech recognition and the efficiency of subsequent intelligent processing. The dual-channel design ensures spatial sound field modeling, laying the foundation for noise separation; cross-correlation phase synchronization eliminates propagation time difference and prevents signal distortion; the noise mask matrix combined with STFT accurately locates steady-state noise, guiding the adaptive filter bank to achieve directional noise reduction (such as NLMS to eliminate echo); spectral entropy analysis captures and filters out residual transient pulses, and the final output high-purity voice has a higher signal-to-noise ratio and better voice fidelity, suitable for applications such as high-quality voice communication, intelligent conference recording, and speech recognition preprocessing.

[0025] Optionally, the step of locating the sound source and generating a sound source location heatmap based on the original conference audio stream acquired by the microphone array includes:

[0026] The original conference audio streams collected by all microphones in the microphone array are processed by frame segmentation, and each frame signal is separated into multiple sub-band signals by a bandpass filter;

[0027] The generalized cross-correlation function between all microphone pairs is calculated based on the signal of each sub-band, and the time delay of searching for the peak of the generalized cross-correlation function is used as the time delay estimate of the sub-band.

[0028] Based on the estimated time delay, the spatial coordinates of the sound source in the current frame are calculated using a least squares optimization algorithm.

[0029] The physical space of the conference room is divided into a uniform grid. The frequency of the spatial coordinates of all sound sources falling into each grid within a preset time window is counted to generate a probability density distribution matrix.

[0030] Gaussian smoothing filtering is applied to the probability density distribution matrix to output a matrix-form heatmap of sound source location; where each matrix element represents the probability of the sound source presence in the corresponding grid.

[0031] By adopting the above technical solutions, the accuracy of time delay estimation is improved based on sub-band analysis, the robustness of positioning is enhanced by the time delay information of multiple microphone pairs, the accuracy of coordinate calculation is improved by least squares optimization, and a continuous and smooth heat map representation is obtained by probability statistics and Gaussian smoothing. This achieves high-precision and high-robust sound source localization function, providing an accurate spatial information foundation for conference voice processing systems.

[0032] Optionally, the steps of dynamically grouping the multi-dimensional voiceprint feature vectors using a spatial constraint clustering algorithm based on the sound source location heatmap, arbitrating conflicting cluster segments based on semantic coherence, and outputting a set of speech segments labeled with voiceprint IDs include:

[0033] Based on the directional information in the sound source location heatmap, the multi-dimensional voiceprint feature vectors are clustered according to spatial location similarity and voiceprint feature similarity to generate a preliminary speaker group set.

[0034] Detect clustered conflict segments in the preliminary speaker grouping set and output a list of conflict segments;

[0035] Extract the contextual semantic information of the clustered conflict fragments and calculate the semantic coherence score based on a natural language processing model;

[0036] Arbitrate the attribution group labels of the clustering conflict segments based on the semantic coherence score to obtain the arbitration decision result;

[0037] The preliminary speaker grouping set is updated based on the arbitration decision result, the grouping labels of the clustered conflict segments are adjusted, and voiceprint IDs are assigned to each speech segment. The set of speech segments labeled with voiceprint IDs is then output.

[0038] By employing the aforementioned technical solutions, the voiceprint clustering method improves the accuracy of initial clustering through spatial constraint clustering, identifies uncertain regions through conflict detection, provides content-level discrimination criteria through semantic coherence analysis, and achieves multi-dimensional information fusion decision-making through an intelligent arbitration mechanism, ultimately outputting high-precision speaker annotation results. This multi-dimensional clustering method, which integrates spatial, voiceprint, and semantic information, effectively solves the limitations of traditional single-feature clustering in complex meeting scenarios, significantly improving the accuracy and robustness of speaker separation and recognition.

[0039] Optionally, the step of loading a pre-configured industry terminology library to dynamically correct the initial transcribed text and outputting a transcribed text stream with industry terminology tags includes:

[0040] Load the industry terminology database pre-stored in the local database; the industry terminology database includes an industry-specific glossary, a mapping relationship between the full and abbreviations of terms, and a priority replacement weight value for terms;

[0041] The initial transcribed text is segmented to extract the segmented text units and timestamp intervals;

[0042] Based on the term priority replacement weight value, candidate terms that match the text unit are retrieved from the industry terminology database. If there are multiple candidate terms, the term with the highest weight value is selected as the target term.

[0043] When the target term and the text unit have a non-full name matching relationship, the text unit is replaced with the standard full name of the target term, and the timestamp interval of the term replacement position is recorded;

[0044] Based on the timestamp interval of the term replacement location, locate the acoustic confidence score of the corresponding speech segment;

[0045] When the acoustic confidence score is lower than a preset threshold, the semantic analysis module is invoked to analyze the semantic coherence of the context and verify the rationality of the term substitution.

[0046] The verified target terms are marked as industry term tags, and the transcribed text stream with timestamps, voiceprint IDs, and industry term tags is output.

[0047] By adopting the above technical solutions, a complete industry terminology knowledge base is constructed. Combined with a multi-layered verification mechanism (acoustic confidence verification and semantic coherence verification), intelligent and accurate correction of industry terms in speech-to-text text is achieved. The technical effect of this solution is to significantly improve the accuracy of speech recognition in professional fields, especially in professional scenarios where terminology is frequently used and abbreviations are numerous. It can effectively solve the shortcomings of general speech recognition systems in professional terminology recognition, and provide reliable technical support for speech data processing in professional fields such as finance, healthcare, and law.

[0048] Optionally, the step of performing periodic memory enhancement processing on the transcribed text stream to output semantically coherent long text includes:

[0049] The transcribed text stream is obtained as input data; the transcribed text stream contains text segments arranged in chronological order and their corresponding timestamps and voiceprint IDs;

[0050] The processing cycle is divided into fixed time intervals, and at the beginning of each cycle, the memory vector library of the previous cycle is loaded from the historical processing cache.

[0051] The transcribed text fragments within the current period are input into the pre-trained language model to generate corresponding dynamic semantic vectors, and then similarity matching is performed with the memory vector library.

[0052] Based on the similarity matching results, the associated memory vectors with similarity exceeding a preset threshold in the memory vector library are extracted and concatenated with the dynamic semantic vector to form an enhanced semantic vector;

[0053] The enhanced semantic vector is input into the text rewriting module to obtain a long text paragraph with semantic coherence;

[0054] Perform syntax tree alignment validation on the long text paragraph and output the final long text.

[0055] By adopting the above technical solutions, a periodic memory enhancement mechanism is constructed. Combined with the semantic representation capabilities of pre-trained language models and text rewriting technology, the common problems of semantic breaks and contextual incoherence in speech-to-text are effectively solved. This improves the semantic coherence and naturalness of expression in long text processing. In particular, when processing speech-to-text content in professional fields or complex contexts, the memory enhancement mechanism can maintain long-distance semantic consistency, providing a high-quality data foundation for subsequent applications such as text analysis, information retrieval, and knowledge extraction.

[0056] Optionally, after parsing the long text, identifying core topics by combining the industry terminology tags, associating the voiceprint ID with the person responsible for the to-do items, and generating structured meeting minutes data, the method further includes:

[0057] Extract the responsible persons and associated industry term tags for the to-do items from the structured meeting minutes data;

[0058] Construct a co-occurrence relationship graph of responsible persons and terms, and generate a feature vector of term usage preferences;

[0059] The preferred replacement weight values ​​in the industry terminology library are dynamically updated based on the terminology using the preference feature vector.

[0060] By adopting the above technical solutions, a dynamic optimization mechanism based on the terminology usage preferences of responsible personnel is constructed, realizing intelligent management of the industry terminology database and improving the personalized adaptability of terminology processing. By analyzing the actual terminology usage patterns of responsible personnel, the system can automatically optimize terminology selection strategies, making the generated text content more in line with the language habits and professional expression preferences of specific user groups. This adaptive terminology management mechanism not only improves the accuracy and naturalness of text processing, but also provides terminology services that are more tailored to the actual needs of users in different industries and positions, effectively solving the terminology usage mismatch problem caused by the static management of traditional terminology databases.

[0061] Secondly, this application provides a speech recognition processing system based on a meeting scenario, employing the following technical solution:

[0062] A speech recognition and processing system for a meeting scenario, the system comprising:

[0063] The acquisition module is used to acquire the raw conference audio stream collected by a microphone array deployed in the conference room; wherein, the microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones;

[0064] The preprocessing module is used to preprocess the original conference audio stream acquired by the main channel and output a clean voice signal after noise reduction.

[0065] The sound source localization module is used to locate the sound source based on the original conference audio stream collected by the microphone array and generate a sound source location heatmap.

[0066] The feature extraction module is used to extract Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories from the pure speech signal to construct a multi-dimensional voiceprint feature vector;

[0067] The voiceprint grouping module is used to dynamically group the multi-dimensional voiceprint feature vectors based on the sound source location heatmap using a spatial constraint clustering algorithm, arbitrate conflicting cluster segments by combining semantic coherence, and output a set of speech segments labeled with voiceprint IDs.

[0068] The speech recognition module is used to input the set of speech segments into a general speech recognition model to generate initial transcribed text;

[0069] The industry terminology correction module is used to load a pre-configured industry terminology library to dynamically correct the initial transcribed text and output a transcribed text stream with industry terminology tags.

[0070] An enhancement processing module is used to perform periodic memory enhancement processing on the transcribed text stream and output a semantically coherent long text.

[0071] The meeting minutes generation module is used to parse the long text, identify core topics by combining the industry terminology tags, associate the voiceprint ID with the person responsible for the task, and generate structured meeting minutes data.

[0072] Thirdly, this application provides a computer device, which adopts the following technical solution:

[0073] A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to perform the steps of the method as described in the first aspect.

[0074] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0075] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0076] In summary, this application includes at least one of the following beneficial technical effects: By constructing a complete conference voice processing system, it achieves fully automated processing from raw audio acquisition to the generation of structured meeting minutes. The system uses microphone array technology combined with spatial constraint clustering algorithms to accurately separate and identify the voice signals of different speakers, effectively solving the problem of voiceprint separation in multi-person dialogues in conference scenarios; through multi-dimensional voiceprint feature extraction and dynamic grouping mechanisms, it improves the accuracy and robustness of speech recognition; by introducing dynamic correction of industry terminology databases and semantic coherence enhancement processing, it significantly improves the quality and readability of transcribed text; finally, it automatically generates structured meeting minutes and associates them with responsible person information, greatly improving the efficiency and accuracy of meeting recording, providing an intelligent solution for meeting management and knowledge accumulation in enterprises and institutions, and has important practical value and promotional significance. Attached Figure Description

[0077] Figure 1 This is a first flowchart illustrating a speech recognition processing method for a meeting scenario according to one embodiment of this application.

[0078] Figure 2This is a second flowchart illustrating a speech recognition processing method based on a meeting scenario, according to one embodiment of this application.

[0079] Figure 3 This is a schematic diagram of the third process of a speech recognition processing method based on a meeting scenario, according to one embodiment of this application.

[0080] Figure 4 This is a schematic diagram of the fourth process of a speech recognition processing method based on a meeting scenario, according to one embodiment of this application.

[0081] Figure 5 This is a fifth flowchart of a speech recognition processing method based on a meeting scenario, according to one embodiment of this application.

[0082] Figure 6 This is a schematic diagram of the sixth process of a speech recognition processing method based on a meeting scenario, according to one embodiment of this application.

[0083] Figure 7 This is a schematic diagram of the seventh process of a speech recognition processing method based on a meeting scenario, according to one embodiment of this application. Detailed Implementation

[0084] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1 - Appendix Figure 7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0085] This application discloses a speech recognition processing method based on a meeting scenario.

[0086] Reference Figure 1 A speech recognition processing method based on a meeting scenario, the recognition processing method includes,

[0087] Step S101: Obtain the raw conference audio stream captured by the microphone array deployed in the conference room;

[0088] The microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones.

[0089] Specifically, the microphone array employs a specific topology design, where the main channel pair, consisting of the first and second microphones, serves as the core acquisition unit, responsible for acquiring the primary speech signal. Other microphones form auxiliary channels, providing spatial information and redundant acquisition. This architecture, where main and auxiliary channels work together, is based on the principle of spatial diversity. Multiple microphones acquire signals from the same sound source from different positions and angles, providing supplementary information and spatial constraints to the main channel signal. The spatial geometry of the microphone array directly affects the accuracy of sound source localization and beamforming. The baseline distance of the main channel pair needs to be balanced between spatial resolution and coherence, typically optimized based on the acoustic characteristics of the conference room and the expected speaker distribution range.

[0090] In one embodiment of this application, the main channel pair (first and second microphones) adopts a close-range symmetrical layout (typically 10-20cm apart), dedicated to capturing the direct sound waves of the core speaker. Their physical positioning must satisfy the near-field sound wave propagation model to ensure that the direct sound energy accounts for >60%. The auxiliary channel is... Two microphones (such as a third and fourth microphone) are used to expand the array aperture (total width 1-3m) to cover multi-angle sampling of reflected sound and ambient noise in the conference room.

[0091] Understandably, by using the main channel to focus on the core sound source and the auxiliary channels to expand spatial coverage, the problem of a single microphone being sensitive to far-field sound wave attenuation can be solved. For example, in a 20-person conference room, the main channel microphone is placed in the center of the conference table, and the auxiliary microphones are deployed in the corners, forming a full-domain sound field sampling matrix.

[0092] Step S102: Perform signal preprocessing on the original conference audio stream acquired by the main channel and output a clean voice signal after noise reduction.

[0093] In the signal preprocessing stage, the raw conference audio stream acquired by the main channel often contains environmental noise, echo interference, and phase deviation caused by microphone spacing. To eliminate the interference of these factors on the subsequent recognition process, the system can use the Dynamic Time Warping (DTW) algorithm for delay compensation. This algorithm determines the optimal alignment point by calculating the peak value of the cross-correlation function between the two signals, thereby constructing an accurate phase synchronization model.

[0094] Furthermore, considering the nonlinear echo problem caused by speaker feedback in a conference room environment, the system can further introduce an adaptive filter bank based on the LMS (Least Mean Square) algorithm to target typical resonant frequency bands (such as...). Energy suppression is performed to effectively suppress device feedback and improve the purity of the voice signal. Simultaneously, combined with a noise masking matrix, low-frequency noise from the air conditioner and transient interference such as keyboard clicks are filtered out in the time-frequency domain, ensuring a high-quality voice signal with a signal-to-noise ratio of no less than 15dB, providing a reliable input basis for subsequent voiceprint segmentation.

[0095] Step S103: Based on the original conference audio stream acquired by the microphone array, the sound source is located and a sound source location heatmap is generated.

[0096] Sound source localization technology, based on physical characteristics such as time difference, phase difference, or intensity difference in sound wave propagation, estimates the spatial location of a sound source by analyzing the spatiotemporal correlation between signals received by multiple microphones. In this application, the system calculates the generalized cross-correlation function between the signals from each microphone, extracts the time difference between the arrival of the sound source at different microphones, and combines this with the geometric position information of the microphone array to calculate the three-dimensional coordinates of the sound source using a geometric localization algorithm. The generated sound source orientation heatmap is essentially a spatial probability distribution map, where each pixel represents the probability density of a sound source existing at that location. The resolution and accuracy of the heatmap directly affect the effectiveness of subsequent spatial constraint clustering. This heatmap representation method can effectively handle the situation of multiple sound sources emitting sound simultaneously and provides continuous spatial trajectory information for each sound source.

[0097] Step S104: Extract Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories from the clean speech signal to construct a multi-dimensional voiceprint feature vector;

[0098] Specifically, after obtaining the denoised, clean speech, the system further extracts biologically specific acoustic features from the signal. Among these, Mel-frequency cepstral coefficients (MFCCs) simulate the frequency perception characteristics of the human cochlea. By compressing the speech spectrum using a triangular filter bank and performing a discrete cosine transform, 20-dimensional cepstral coefficients reflecting changes in the vocal tract shape are extracted. This feature not only possesses good noise resistance but also effectively distinguishes the formant structure differences between different speakers. Furthermore, the system introduces a fundamental frequency jitter index to measure the fluctuations between speech cycles; a smaller value indicates more stable speech, suitable for assisting in the judgment of emotional or physiological states. Formant trajectories track the time-varying paths of key formants such as F1 and F2, showing a significant advantage, especially in distinguishing vowel categories.

[0099] Step S105: Based on the heat map of the sound source location, a spatial constraint clustering algorithm is used to dynamically group the multi-dimensional voiceprint feature vectors, and semantic coherence is combined to arbitrate conflicting segments in the clustering, and output a set of speech segments labeled with voiceprint IDs.

[0100] Based on the heatmap of sound source location, the system employs a spatially constrained clustering algorithm to dynamically group multi-dimensional voiceprint feature vectors. Traditional spatially constrained clustering algorithms rely on neighborhood density judgments, while this application introduces a spatial orientation constraint mechanism, which mandates that the azimuth difference between samples within the same cluster be less than a certain value. This makes it easier to group speech segments with similar acoustic features and close spatial locations into the same category, thereby effectively eliminating the risk of mis-grouping caused by similar voiceprints from a distance.

[0101] Specifically, the system employs a graph-based clustering method, representing each speech segment as a node in a graph. The similarity between nodes is determined by both acoustic feature similarity and spatial distance, and clustering is achieved by optimizing graph segmentation. When clustering conflicts arise due to multiple speakers being spatially close or having similar acoustic features, the system further introduces a semantic coherence analysis mechanism. This mechanism assists in decision-making by analyzing the semantic relevance of speech content and the consistency of speaking style. For example, after obtaining text content through speech recognition, the semantic coherence of adjacent speech segments is analyzed, or prosodic features are used to analyze the speaker's intonation, speaking speed, and other stylistic features. The final output set of speech segments labeled with voiceprint IDs enables automatic identification of different speakers in a conference and the structured organization of speech content.

[0102] Step S106: Input the set of speech segments into the general speech recognition model to generate the initial transcribed text;

[0103] Step S107: Load the pre-configured industry terminology library to dynamically correct the initial transcribed text, and output the transcribed text stream with industry terminology tags.

[0104] The system feeds voice segments labeled with voiceprint IDs into a general speech recognition model to generate initial transcribed text. However, due to semantic gaps in specialized terminology within the general model, the system further introduces an industry-specific terminology database as a plug-in module. This database covers specialized terms from multiple vertical fields such as architecture, transportation, finance, and healthcare (e.g., "shear wall," "LTV," "MRI," "interchange," "pavement bearing capacity," "prefabricated assembly structure," etc.). This terminology database not only provides static dictionary support but can also be directly injected into the decoding network via a hash mapping mechanism, forcing the model to prioritize matching specialized terms in specific contexts, thereby improving recognition recall.

[0105] For the transportation infrastructure industry, a multi-dimensional professional corpus has been systematically constructed, including: policy and regulation categories (such as the "Outline for Building a Transportation Power" and the "Code for Design of Urban Road Engineering"), industry reports (transportation yearbooks, infrastructure development white papers, etc.), standards and specifications (national standards, industry standards, local standards, etc.), construction processes and methods (bridge construction processes, tunnel excavation technology, etc.), outline templates (project feasibility study report templates, construction organization design outlines, etc.), professional books and textbooks (classic textbooks and question banks such as "Road Engineering" and "Bridge Engineering"), data from authoritative resource websites, new technologies (terminology related to new materials, new processes, new technologies, and new equipment), excellent technical solution cases, and standardized templates for technical solutions, among other rich content categories. This professional corpus provides a solid language model foundation for speech recognition in the transportation infrastructure industry.

[0106] In addition, the system can also support a user feedback-driven meta-learning mechanism, allowing users to provide a certain number of error samples (such as misidentifying "LTV" as "LTP") and perform gradient backpropagation on the GPU to dynamically update the model output layer weight matrix, thereby achieving rapid optimization of the recognition capability of specific terms and greatly improving the recognition accuracy of professional terms.

[0107] Step S108: Perform periodic memory enhancement processing on the transcribed text stream to output a semantically coherent long text.

[0108] Among these solutions, periodic memory vector management and semantic enhancement techniques can address common issues of semantic fragmentation and contextual incoherence in speech-to-text transcription. By constructing a time-window-based memory management mechanism, the continuous transcribed text stream is divided into processing cycles at fixed time intervals. This allows the system to not only focus on the semantic information of the current text segment within each processing cycle but also effectively utilize the contextual information accumulated in the memory vector library from previous processing cycles.

[0109] Specifically, the system first converts text fragments within the current period into dynamic semantic vectors using a pre-trained language model. Then, it retrieves semantically related associated memory vectors from the historical memory vector library through a similarity matching mechanism and merges the two to generate an enhanced semantic vector. This enhanced semantic vector, which incorporates historical contextual information, provides richer contextual clues for the subsequent text rewriting module. This ensures that the rewritten text not only corrects the semantic breaks in the original transcription but also maintains semantic consistency with the historical content. Finally, the system uses syntactic tree alignment verification to ensure the correctness of the output text in terms of grammatical structure, thereby achieving an effective conversion from time-series speech-to-text to semantically coherent long text.

[0110] Step S109: Parse the long text, identify core issues by combining industry terminology tags, bind the responsible persons of the to-do items with voiceprint IDs, and generate structured meeting minutes data.

[0111] The system performs semantic parsing on long texts that have undergone memory enhancement, identifies core topics by combining industry terminology tags, and improves the accuracy of topic discovery by using the TextRank algorithm to increase the weight of terminology nodes. Building on this, the system uses Semantic Role Labeling (SRL) technology to extract <responsible person, action, time> triples, and then combines these with pre-registered voiceprint ID mappings to clearly bind task responsibilities to specific participants. For example, when the system detects that the voiceprint ID corresponding to "Zhang Ming" is spk02, it can automatically assign relevant to-do items to that person.

[0112] Furthermore, to ensure data security and compliance, the system can integrate a pre-defined regulatory thesaurus to filter sensitive content in real time and trigger audit logs to prevent the leakage of confidential information. The final output of structured meeting minutes is presented in Markdown format, including an overview of the agenda, key decisions, and a to-do list, making it highly readable and actionable.

[0113] The above implementation constructs a complete technical solution from physical layer signal acquisition to cognitive layer semantic understanding. Physical layer signal purification and spatial perception are achieved through microphone noise reduction and array localization technology; refined voiceprint expression at the perception layer is achieved through biometric modeling of Mel frequency cepstral coefficients, fundamental frequency jitter, and formant trajectories; high-precision speaker segmentation at the cognitive layer is achieved through spatial constraint clustering and semantic conflict arbitration mechanisms; professional semantic adaptation at the cognitive layer is achieved through a collaborative mechanism between a general model and a domain terminology database, as well as meta-learning optimization; long-text semantic consistency at the decision layer is maintained through a periodic memory-enhanced LSTM architecture; and finally, automated output of meeting minutes is achieved through multimodal semantic parsing and structured generation mechanisms.

[0114] This application, while ensuring localized deployment and data security, can not only effectively improve the quality of voice signals, but also accurately identify and distinguish different speakers, providing reliable technical support for applications such as intelligent meeting recording, speech transcription, and meeting content analysis, and significantly improving the automation level and accuracy of meeting voice processing.

[0115] Reference Figure 2 As one implementation of step S102, the step of preprocessing the original conference audio stream acquired by the main channel and outputting a noise-reduced clean speech signal includes:

[0116] Step S201: Acquire the original conference audio stream obtained from the main channel pair to obtain the first channel audio signal and the second channel audio signal;

[0117] The dual-microphone array design based on the main channel pair aims to capture multi-source audio information in meeting scenarios through spatial sound field sampling, solving the signal acquisition distortion problem of general speech recognition systems in noisy environments (such as background noise in meeting rooms or multiple speakers). Since meeting scenarios often face challenges such as environmental noise interference (such as air conditioning, page turning sounds) and confusion from multiple speakers (refer to the section "Speech Recognition Adaptation Challenges in Complex Meeting Scenarios"), the dual-channel design forms a differential input source by deploying two physically isolated microphones in the meeting room: the first channel mainly targets the direct sound waves of near-field speakers (i.e., the pure speech output by the speaker), while the second channel focuses on collecting environmental reflected sound and noise (such as reverberation or low-frequency steady-state noise). This channel-specific strategy provides a multi-angle data foundation for subsequent processing and can effectively separate the target speech from background interference.

[0118] Specifically, a "microphone array" refers to a sensor system composed of multiple microphones used for spatial signal capture; "direct sound waves" are speech energy that arrives directly at the microphone without reflection; and "ambient reflected sound" is the delayed signal formed after sound waves are reflected by objects such as room walls, which is easily superimposed with noise. For example, in a 20-person meeting, a dual-microphone array can be deployed on the desktop (channel 1) and the ceiling (channel 2). The former captures the clear voice of the near-field speaker, while the latter covers the overall ambient noise, thereby constructing a spatial sound field model.

[0119] Step S202: Perform dynamic time warping on the first channel audio signal and the second channel audio signal, and output a phase-synchronized dual-channel audio signal by calculating the peak-aligned phase difference of the cross-correlation function of the two signals.

[0120] The core of dynamic time warping is to locate and compensate for the relative delay between signals by calculating the cross-correlation function. Specifically, the cross-correlation function... Used to measure two signals at different time shifts The similarity peaks below, among which, and These represent the time-domain signals of the first and second channels, respectively. This represents the time offset. The maximum value of this cross-correlation function is found. This allows us to determine the time delay between the two signals and, based on this, perform reverse time shift compensation on the second channel signal. This operation effectively eliminates the propagation time difference caused by the microphone spacing, enabling the two signals to synchronize in the time domain, thus providing accurate dual-channel differential input for subsequent noise mask construction. The significance of this stage is to improve the accuracy of subsequent time-frequency analysis and avoid spectral aliasing or incorrect energy distribution judgments caused by phase inconsistency.

[0121] For example, in a conference room, microphone spacing can cause the same speaker's voice to be recorded with microsecond-level deviations in different channels. Cross-correlation analysis can locate and compensate for these deviations, synchronizing the dual-channel signals. By eliminating propagation time difference, a stable foundation is provided for subsequent frequency domain analysis, solving the recognition distortion caused by phase misalignment in existing technologies.

[0122] Step S203: Generate a noise mask matrix based on the phase-synchronized dual-channel audio signal, identify steady-state noise regions by analyzing the time-frequency domain energy distribution, and construct a binary mask matrix.

[0123] The synchronized dual-channel signal is converted to the time-frequency domain. The steady-state noise region is accurately identified and marked by the energy distribution model, and a binary mask matrix is ​​generated to guide directional noise reduction, thereby solving the problem of continuous noise (such as low-frequency air conditioning noise) interfering with the voice signal in the conference environment.

[0124] Specifically, the short-time Fourier transform (STFT) is applied to decompose the time-domain signal into time-frequency units. Each unit corresponds to an energy distribution at a specific frequency f and time t. Based on this, energy patterns are analyzed to identify noise characteristics: for example, low-frequency continuous energy (<500Hz) indicates air conditioning noise, while broadband transient high energy corresponds to interference such as keyboard typing. Subsequently, a binary mask matrix Mask(f,t) is constructed, where time-frequency units identified as noise are assigned a value of 0, and the rest are assigned a value of 1. This mask matrix acts like a "filter guide graph," providing a clear reference weight for the subsequent adaptive echo cancellation module, ensuring that noise reduction operations can accurately target noisy regions without damaging speech components.

[0125] For example, in a conference setting, STFT analysis can convert a 10ms audio segment to the frequency domain and mark the energy high points below 500Hz as the air conditioning noise zone. This solves the problem that existing technologies cannot accurately locate noise in noisy environments (such as "accuracy fluctuations in multiple people + noisy environment") and provides reference weights for subsequent adaptive noise reduction.

[0126] Step S204: Input the phase-synchronized dual-channel audio signal into the adaptive filter bank, use the binary mask matrix as the reference signal to perform nonlinear echo cancellation, and output the primary noise reduction signal.

[0127] The adaptive filter bank achieves closed-loop noise suppression by combining weighted guidance with a mask matrix to eliminate linear echoes (such as speaker feedback) and nonlinear distortion, significantly improving the signal-to-noise ratio and solving the problem of mixed interference from environmental reflections and electronic device feedback in meetings. The adaptive filter bank employs a multi-stage structure. For example, the primary filter uses the Normalized Least Mean Square (NLMS) algorithm to handle linear echoes (direct sound wave reflection), while the secondary filter targets nonlinear distortion (harmonics introduced by device circuitry). The binary mask matrix serves as a reference signal, applying additional attenuation (e.g., 20 dB) to regions marked as noise (Mask(f,t)=0) while preserving the energy of the speech region (Mask(f,t)=1), forming a directional noise reduction mechanism. For instance, when using a gooseneck microphone in a meeting, the adaptive filter can reduce low-frequency noise based on the mask while avoiding damage to the speaker's frequency range.

[0128] Understandably, this closed-loop guidance mechanism significantly improves the robustness and accuracy of the noise reduction system, avoiding the speech distortion or semantic information loss problems caused by over-filtering in traditional methods. Furthermore, since the entire process relies on automated time-frequency analysis and filtering operations without manual intervention, it exhibits good scalability and deployment adaptability.

[0129] Step S205: Perform subband spectral entropy analysis on the primary noise reduction signal, detect and filter out transient noise pulses, and output the noise-reduced clean speech signal.

[0130] Among them, residual transient noise (such as sudden keyboard sounds or paper flipping) is detected by spectral entropy analysis, abnormal pulses are identified by entropy value mutations, and interpolation technology is used to filter them out, finally outputting a high-purity speech signal, reducing the disturbance to voiceprint analysis or speech recognition.

[0131] Specifically, the primary noise reduction signal is divided into 20ms subframes, and the spectral entropy value of each subframe is calculated. , ;in, Let X(k) be the spectral energy normalized probability, and X(k) be the amplitude spectrum of the k-th frequency component. When the entropy value suddenly increases beyond a threshold (e.g., >2.5 bits), it indicates high-frequency transient noise; subsequently, these subframes are replaced using a time-domain interpolation algorithm (e.g., linear interpolation or adjacent frame replacement) to preserve coherent speech.

[0132] For example, if the peak entropy value of keystrokes is detected during a meeting, interpolation can smooth out this interval without affecting the speaker's content, thus solving the problem of limited accuracy in complex environments.

[0133] The above implementation achieves deep purification of voice signals in conference scenarios, significantly improving speech recognition accuracy and subsequent intelligent processing efficiency. The dual-channel design ensures spatial sound field modeling, laying the foundation for noise separation; cross-correlation phase synchronization eliminates propagation time difference, preventing signal distortion; the noise mask matrix combined with STFT accurately locates steady-state noise, guiding the adaptive filter bank to achieve directional noise reduction (such as NLMS echo cancellation); spectral entropy analysis captures and filters out residual transient pulses, resulting in high-purity output speech with a higher signal-to-noise ratio and better speech fidelity, suitable for applications such as high-quality voice communication, intelligent conference recording, and speech recognition preprocessing.

[0134] Reference Figure 3 As one implementation of step S103, the step of locating the sound source based on the original conference audio stream acquired by the microphone array and generating a sound source location heatmap includes:

[0135] Step S301: Perform frame segmentation processing on the original conference audio streams collected by all microphones in the microphone array, and separate each frame signal into multiple sub-band signals through a bandpass filter;

[0136] Specifically, this framing process is based on the assumption of short-time stationarity of speech signals, dividing continuous non-stationary speech signals into several short time intervals. Within each short time interval, the signal is approximated as stationary, thus facilitating time-frequency analysis. The length of each frame is typically in the range of 20-30 milliseconds, and the frame shift is set to 1 / 4 to 1 / 2 of the frame length to ensure time resolution. The use of bandpass filter banks is based on the characteristics of human hearing and the energy distribution of speech signals, decomposing the full-band signal into multiple sub-bands, each covering a specific frequency range.

[0137] Understandably, the advantage of this sub-band decomposition lies in the differences in sound wave propagation characteristics across different frequency bands. Low-frequency bands have longer wavelengths and weaker directionality but propagate over longer distances, while high-frequency bands have strong directionality but are more susceptible to environmental interference. By processing each sub-band signal separately, the robustness of localization can be improved. Commonly used filter banks include Mel-scale filter banks, proportional filter banks, or multi-resolution analysis methods based on wavelet transform.

[0138] Step S302: Calculate the generalized cross-correlation function between all microphone pairs based on the signal of each sub-band, and use the time delay of the peak value of the generalized cross-correlation function as the time delay estimate of the sub-band.

[0139] Generalized cross-correlation (GCC) is a classic algorithm for sound source localization. Its basic principle is to estimate the direction of the sound source by using the time difference between the arrival times of the sound source signal at different microphones. For any two microphones i and j, their received signals can be expressed as... and When the sound source is located in a certain direction, there is a time delay in the signal reaching the two microphones. .

[0140] Specifically, the generalized cross-correlation function is defined as:

[0141] ;

[0142] in, and Let be the frequency domain representations of the i-th and j-th microphone signals, respectively. Let W(f) represent the complex conjugate of the frequency domain representation of the signal received by the j-th microphone, where W(f) is the frequency domain weighting function, f is the frequency variable, and j in the exponent represents the imaginary unit. The optimal time delay estimate can be obtained by searching for the peak position of the cross-correlation function. In a multi-microphone array, it is necessary to calculate all possible microphone pair combinations. For an array of N microphones, there are N(N-1) / 2 microphone pairs. Each microphone pair generates a time delay estimate in each sub-band, forming a rich set of time delay information.

[0143] Step S303: Based on the time delay estimate, calculate the spatial coordinates of the sound source in the current frame using the least squares optimization algorithm;

[0144] This process is based on the principle of geometric positioning, which states that the distance difference between the sound source and each microphone should equal the speed of sound multiplied by the corresponding time delay. Let the coordinates of the sound source be (x, y, z), and the coordinates of the i-th microphone be... If the speed of sound is c, then:

[0145] ;

[0146] in, Let be the time delay of the i-th microphone relative to the reference microphone (usually the first microphone). For N microphones, N-1 such constraint equations can be established, forming an overdetermined system of equations. Due to measurement noise and model errors in real-world environments, direct solutions often fail to yield accurate results. Therefore, the least squares optimization method is used to find the optimal solution. The least squares criterion is:

[0147] The optimization problem is solved by iterative optimization algorithms (such as Gauss-Newton method, Levenberg-Marquardt algorithm, etc.) to obtain the estimated three-dimensional coordinates of the sound source.

[0148] Step S304: Divide the physical space of the conference room into a uniform grid, count the frequency of all sound source spatial coordinates falling into each grid within a preset time window, and generate a probability density distribution matrix.

[0149] This spatial discretization process is based on probabilistic and statistical principles, mapping a continuous three-dimensional space onto a discrete mesh structure, facilitating numerical computation and storage. The mesh density needs to be balanced between computational complexity and positioning accuracy; an overly dense mesh increases the computational burden, while an overly sparse mesh reduces spatial resolution. The mesh size is typically determined based on the size of the meeting room and the expected positioning accuracy. For example, for... The meeting room can be divided into The grid.

[0150] It should be noted that the preset time window is set based on the continuous nature of speech activity, and is usually set to a range of several seconds to tens of seconds. By statistically analyzing the sound source localization results of all frames within this time window, the impact of instantaneous localization errors can be effectively suppressed. The frequency statistics of each grid are actually estimating the probability that a sound source exists at that location; the higher the frequency, the more likely that location is the actual sound source location.

[0151] Step S305: Perform Gaussian smoothing filtering on the probability density distribution matrix to output a matrix-form heatmap of the sound source location; where each matrix element represents the probability of the sound source's presence in the corresponding grid.

[0152] The purpose of Gaussian smoothing filtering is to eliminate statistical noise and discretization errors, making the probability distribution smoother and more continuous. Since the spatial distribution of sound sources has a certain degree of continuity, the probabilities of sound sources in adjacent grids should be similar. This smoothing effect can be achieved through convolution using a Gaussian kernel function. The variance parameter of the Gaussian kernel controls the degree of smoothing; a smaller variance maintains higher spatial resolution but has limited smoothing effect, while a larger variance provides better smoothing but may obscure the true location of the sound sources. The filtered probability density distribution matrix forms the final sound source location heatmap. This heatmap not only contains the location information of the sound sources but also the confidence information of the location estimation, providing reliable spatial constraints for subsequent voiceprint clustering.

[0153] In the above implementation, the accuracy of time delay estimation is improved based on sub-band analysis, the robustness of positioning is enhanced by the time delay information of multiple microphone pairs, the accuracy of coordinate calculation is improved by least squares optimization, and a continuous and smooth heat map representation is obtained by probability statistics and Gaussian smoothing. This achieves high-precision and high-robust sound source localization function, providing an accurate spatial information foundation for conference voice processing systems.

[0154] Reference Figure 4 As one implementation of step S105, the steps of dynamically grouping multi-dimensional voiceprint feature vectors based on a sound source location heatmap using a spatial constraint clustering algorithm, arbitrating conflicting cluster segments based on semantic coherence, and outputting a set of speech segments labeled with voiceprint IDs include:

[0155] Step S401: Based on the directional information in the sound source location heatmap, cluster the multi-dimensional voiceprint feature vectors according to spatial location similarity and voiceprint feature similarity to generate a preliminary speaker group set.

[0156] Each group contains speech segments with similar spatial distribution and voiceprint features;

[0157] Specifically, each grid point in the sound source location heatmap corresponds to a sound source existence probability value. By extracting this directional information and jointly modeling it with the voiceprint feature vector, a more robust similarity metric can be constructed. Specifically, for any two speech segments, their similarity depends not only on the Euclidean distance or cosine similarity of the voiceprint features but also on their spatial proximity. This spatially constrained clustering algorithm can employ an improved K-means algorithm, spectral clustering algorithm, or density-based clustering algorithm, where the similarity metric function can be defined as: Where sim_voice(i,j) represents the voiceprint feature similarity, and sim_spatial(i,j) represents the spatial location similarity. and These are the weighting coefficients.

[0158] It is understandable that even if two speech segments have similar voiceprint features, they will not be classified into the same category if their spatial locations are far apart. This effectively avoids the problem of mis-clustering due to similar voiceprint features but different actual speakers.

[0159] Step S402: Detect clustered conflict segments in the preliminary speaker group set and output a list of conflict segments;

[0160] Specifically, the detection of conflicting segments in clustering is based on probability theory and fuzzy set theory, primarily identifying regions that overlap or have ambiguous boundaries among multiple groups. In real-world conference scenarios, due to the complexity of speech signals, the existence of localization errors, and the dynamic changes in speaker positions, some speech segments may simultaneously possess multiple candidate speaker identities.

[0161] These conflicting segments typically exhibit the following characteristics: multiple high-probability peak regions in the sound source location heatmap; proximity to multiple cluster centers in the speaker signature space; or location within ambiguous regions at the boundaries of different clusters. Conflict detection can be implemented using the membership concept from fuzzy C-means clustering. The membership value of each speech segment to each cluster is calculated, and a segment is identified as conflicting when the difference between the maximum and second-largest membership values ​​is less than a preset threshold. Alternatively, cluster confidence can be assessed, such as calculating the ratio of the distance from each segment to the cluster center to the average distance. A segment is also considered conflicting when this ratio exceeds a certain threshold. This conflict detection mechanism effectively identifies uncertain regions in the clustering results, providing accurate input for subsequent arbitration processing.

[0162] Step S403: Extract the contextual semantic information of clustered conflict fragments and calculate the semantic coherence score based on the natural language processing model;

[0163] This process is based on principles of linguistics and cognitive science, namely that human language possesses a high degree of contextual dependence and semantic coherence, and that the same speaker typically expresses related semantic content in consecutive speech segments. Extracting contextual semantic information first requires speech recognition of conflicting segments, converting them into text form, and then extracting the text content within a certain time window before and after the segment as contextual information. Natural language processing models can employ pre-trained language models, such as BERT and GPT, which are capable of capturing deep semantic relationships and contextual dependencies in text.

[0164] Specifically, semantic coherence scores can be calculated in several ways: one is to calculate the semantic similarity between conflicting segments and adjacent segments, such as using cosine similarity or semantic embedding distance based on pre-trained models; another is to assess the semantic consistency of the entire discourse, measuring semantic fluency by calculating conditional probabilities or perplexity using language models. This semantic coherence-based assessment method can provide important evidence for attributing conflicting segments at the linguistic content level, especially when voiceprint features and spatial information cannot clearly distinguish between them, semantic information often provides decisive discriminative ability.

[0165] Step S404: Arbitrate the grouping labels of clustering conflicting segments based on semantic coherence scores to obtain arbitration decision results;

[0166] Among them, the arbitration decision prioritizes adjacent segments with high semantic coherence to eliminate conflicts. This arbitration mechanism is based on Bayesian decision theory and the maximum a posteriori probability criterion, and uses a unified probabilistic model to integrate voiceprint feature similarity, spatial location similarity and semantic coherence score.

[0167] Specifically, for each clustering conflict segment, the system calculates its posterior probability of belonging to each candidate group, which comprehensively considers information from multiple dimensions: Where P(cluster|fragment) is the posterior probability, P(cluster) is the prior probability, and the likelihood function is... It can be decomposed into a product of voiceprint likelihood, spatial likelihood, and semantic likelihood. The arbitration decision adopts the maximum a posteriori probability criterion, selecting the group with the highest a posteriori probability as the final assignment. In particular, the system introduces a semantic coherence priority strategy, that is, when adjacent segments have already been assigned to a certain group, the currently conflicting segment tends to choose the group that can maintain semantic coherence. This strategy conforms to the natural law of human language communication, that is, the same speaker usually expresses related topic content in a short period of time.

[0168] Step S405: Update the preliminary speaker grouping set according to the arbitration decision result, adjust the grouping labels of the clustered conflict segments, assign voiceprint IDs to each speech segment, and output the speech segment set labeled with voiceprint IDs.

[0169] The system employs arbitration decisions to revise and refine the original clustering results. Voiceprint ID allocation is based on the principles of cluster stability and consistency. For each ultimately determined speaker group, the system assigns a unique voiceprint identifier, which can be numbers, letters, or a combination thereof. During the allocation process, the system also considers the reliability and confidence level of the clustering results. Groups with low confidence levels may require further manual verification or be marked as unidentified speakers. The final output set of voice segments annotated with voiceprint IDs not only contains the speaker's identity information for each segment but also carries corresponding confidence scores and semantic context information, providing a high-quality data foundation for subsequent applications such as meeting record organization, voice retrieval, and content analysis.

[0170] In the above embodiments, the voiceprint clustering method improves the accuracy of initial clustering through spatial constraint clustering, identifies uncertain regions through conflict detection, provides content-level discrimination criteria through semantic coherence analysis, and achieves multi-dimensional information fusion decision-making through an intelligent arbitration mechanism, ultimately outputting high-precision speaker annotation results. This multi-dimensional clustering method, which integrates spatial, voiceprint, and semantic information, effectively solves the limitations of traditional single-feature clustering in complex meeting scenarios, significantly improving the accuracy and robustness of speaker separation and recognition.

[0171] Reference Figure 5 As one implementation of step S107, the step of loading a pre-configured industry terminology database to dynamically correct the initial transcribed text and outputting a transcribed text stream with industry terminology tags includes:

[0172] Step S501: Load the industry terminology library pre-stored in the local database;

[0173] The industry terminology database includes an industry-specific glossary, a mapping relationship between the full and abbreviations of terminology standards, and a weight value for terminology priority replacement.

[0174] Specifically, an industry-specific glossary defines the range of professional terms within a specific field, ensuring the system can recognize industry-specific expressions. The mapping relationship between full and abbreviation terms addresses the common abbreviation problem in actual speech recognition, as professionals often use abbreviations or shortened forms of terms in specialized fields, which may be incorrectly identified as other words in general speech recognition models. The term priority replacement weight value reflects the differences in importance of different terms in a specific context. This weighting mechanism enables the system to make the optimal choice when faced with multiple possible candidate terms, and its weight value can be set based on factors such as the term's usage frequency in the industry, its professional importance, or its historical replacement success rate.

[0175] Specifically designed for the transportation infrastructure industry, the industry terminology database has constructed a multi-dimensional professional vocabulary system: Policy and regulatory terms include professional expressions from documents such as the "Outline for Building a Transportation Powerhouse" and the "Code for Design of Urban Road Engineering"; Standard and specification terms cover technical indicators and professional definitions in national, industry, and local standards; Process and method terms include technical terms from professional processes such as bridge construction, tunnel excavation, and pavement paving; New technology terms include cutting-edge professional terms related to new materials, new processes, new technologies, and new equipment; and Standardized technical solution template terms involve professional expressions from standardized documents such as feasibility study reports and construction organization designs. The introduction of these professional terms enables the system to accurately identify industry-specific technical terms such as "interchange," "pavement bearing capacity," "prefabricated assembly structure," and "BIM modeling," significantly improving the professionalism and accuracy of speech recognition in transportation infrastructure industry meetings.

[0176] Step S502: Perform word segmentation on the initial transcribed text and extract the segmented text units and timestamp intervals;

[0177] The word segmentation process here differs from general Chinese word segmentation. It needs to consider the integrity of industry terminology to avoid incorrectly splitting a complete professional term into multiple lexical units. The extraction of timestamp intervals provides a precise time reference for subsequent spatiotemporal localization and acoustic verification, enabling the system to accurately align text-level terminology corrections with corresponding speech signal features.

[0178] Step S503: Based on the term priority replacement weight value, retrieve candidate terms that match the text unit in the industry terminology database. If there are multiple candidate terms, select the term with the highest weight value as the target term.

[0179] When the system detects that a text unit may correspond to an industry term, it searches for all possible candidate terms in the industry terminology database using a fuzzy matching algorithm. This matching includes not only exact matching but also prefix matching, suffix matching, edit distance matching, and other forms. When multiple candidate terms exist, the system sorts and selects them based on weight values. This weight-based decision-making mechanism ensures that the most likely correct term is chosen in uncertain situations, improving the accuracy of the correction.

[0180] Step S504: When there is a non-full name matching relationship between the target term and the text unit, replace the text unit with the standard full name of the target term and record the timestamp range of the term replacement position.

[0181] Among these features, the standard full-name replacement operation addresses the common problem of term abbreviation recognition errors in speech recognition. This replacement is not merely a simple word substitution; more importantly, it ensures accuracy through predefined mapping relationships in a terminology database. Simultaneously, it records the timestamp interval of the term replacement location, providing precise temporal positioning information for subsequent acoustic verification. This timestamp recording mechanism enables the system to quickly locate the specific segment requiring verification within massive amounts of speech data.

[0182] Step S505: Based on the timestamp interval of the term replacement location, locate the acoustic confidence score of the corresponding speech segment;

[0183] The acoustic confidence score reflects the reliability of the speech recognition engine's recognition result for the speech segment. It is typically calculated based on multiple dimensions, including the output probability of the acoustic model, the recognition confidence at the phoneme level, and the quality of the speech signal. When the acoustic confidence score is below a preset threshold, it indicates that the recognition result has significant uncertainty. In this case, a higher-level verification mechanism is needed to ensure the correctness of the term substitution.

[0184] Step S506: When the acoustic confidence score is lower than the preset threshold, the semantic analysis module is invoked to analyze the semantic coherence of the context and verify the rationality of the term substitution.

[0185] The semantic analysis module uses deep learning or language models to analyze the semantic coherence changes of the entire sentence or paragraph before and after term substitution. This analysis considers not only lexical-level semantic matching but also syntactic structure, semantic logical relationships, and consistency of domain knowledge. For example, in the financial field, although "interest rate" and "exchange rate" are both financial terms, "exchange rate" has stronger semantic coherence if the context discusses foreign exchange transactions. This context-based semantic verification mechanism greatly reduces the risk of incorrect term substitution caused by acoustic recognition errors.

[0186] Step S507: Mark the verified target terms as industry term tags and output the transcribed text stream with timestamp, voiceprint ID and industry term tags.

[0187] In the above implementation, a complete industry terminology knowledge base is constructed, and combined with a multi-layered verification mechanism (acoustic confidence verification and semantic coherence verification), intelligent and accurate correction of industry terms in speech-to-text text is achieved. The technical effect of this solution is to significantly improve the accuracy of speech recognition in professional fields, especially in professional scenarios where terminology is frequently used and abbreviations are numerous. It effectively addresses the shortcomings of general speech recognition systems in recognizing professional terms, providing reliable technical support for speech data processing in professional fields such as finance, healthcare, and law.

[0188] Reference Figure 6 As one implementation of step S108, the step of performing periodic memory enhancement processing on the transcribed text stream to output a semantically coherent long text includes:

[0189] Step S601: Obtain the transcribed text stream as input data; wherein, the transcribed text stream contains text segments arranged in chronological order and their corresponding timestamps and voiceprint IDs;

[0190] Step S602: Divide the processing cycle into fixed time intervals, and load the memory vector library of the previous cycle from the historical processing cache at the beginning of each cycle;

[0191] The fixed time interval division ensures the regularity and controllability of the processing, avoiding the complexity issues caused by uneven text length. The memory vector library, as a carrier of historical information, stores key semantic features and contextual information extracted from previous processing cycles. This cycle-based memory management mechanism not only improves processing efficiency but also ensures the effective transmission of long-distance dependencies, enabling the system to fully utilize historical contextual information when processing the current text.

[0192] Step S603: Input the transcribed text fragments in the current period into the pre-trained language model to generate the corresponding dynamic semantic vectors, and perform similarity matching with the memory vector library;

[0193] Pre-trained language models, typically trained on large-scale corpora, possess powerful semantic representation capabilities, mapping text fragments to vector representations in a high-dimensional semantic space. These dynamic semantic vectors not only contain lexical-level semantic information but also incorporate deeper linguistic features such as syntactic structure and semantic relationships. In this way, the system can transform discrete text fragments into continuous semantic representations, providing a mathematical foundation for subsequent similarity matching and semantic enhancement.

[0194] For the transportation infrastructure industry, the pre-trained language model incorporates specialized corpora from this field for fine-tuning optimization. These include: policy and regulatory corpora (such as official documents like the "Outline for Building a Strong Transportation Nation" and the "Code for Design of Urban Road Engineering"), industry report corpora (annual transportation yearbooks, infrastructure development white papers, etc.), standard and specification corpora (various national, industry, and local standards documents), process and method corpora (bridge construction process manuals, tunnel excavation technology guidelines, etc.), outline and template corpora (project feasibility study report templates, construction organization design outlines, etc.), professional books and textbooks corpora (classic textbooks such as "Road Engineering" and "Bridge Engineering" and their accompanying question banks), authoritative resource website corpora, new technology corpora (technical documents related to new materials, new processes, new technologies, and new equipment), excellent technical solution case studies, and standardized technical solution templates. Through deep learning training on these specialized corpora, the model can better understand the unique semantic structure and expression habits of the transportation infrastructure industry, thereby generating more accurate domain-specific semantic vectors and improving the accuracy of similarity matching and the effect of semantic enhancement.

[0195] Step S604: Based on the similarity matching results, extract the associated memory vectors in the memory vector library whose similarity exceeds a preset threshold, and concatenate them with the dynamic semantic vectors to form an enhanced semantic vector;

[0196] Similarity matching typically employs cosine similarity, Euclidean distance, or other vector similarity calculation methods to measure the semantic relevance between the current semantic vector and the historical memory vector. A preset threshold ensures that only truly relevant memory information is incorporated, avoiding interference from irrelevant information. By concatenating the associated memory vector with the current dynamic semantic vector, the system achieves an organic fusion of historical contextual information and current semantic information. This enhanced semantic vector contains richer contextual information, contributing to improved quality in subsequent text rewriting.

[0197] Step S605: Input the enhanced semantic vector into the text rewriting module to obtain a semantically coherent long text paragraph;

[0198] In this embodiment, the text rewriting module can employ a sequence-to-sequence neural network architecture, capable of generating more semantically coherent and naturally expressed text content based on the input enhanced semantic vector. This rewriting process not only corrects potential grammatical errors and semantic breaks in the original transcribed text but also enhances the contextual consistency of the text by introducing historical memory information.

[0199] For example, a technical term mentioned in the current text segment may have a detailed definition in the preceding text. Through memory enhancement mechanisms, the rewriting module can better understand the meaning of the term and generate a more accurate expression.

[0200] Step S606: Perform syntax tree alignment verification on the long text paragraph and output the final long text.

[0201] Specifically, by constructing a syntactic structure tree of the text and performing consistency checks, the system ensures that the rewritten text maintains grammatical rationality and coherence. This verification mechanism can detect and correct syntactic errors that may be introduced during the rewriting process, ensuring that the output text is not only semantically coherent but also grammatically compliant with language norms. Through this multi-layered verification mechanism, the system can output high-quality, semantically coherent long texts.

[0202] In the above implementation, a periodic memory enhancement mechanism is constructed. Combined with the semantic representation capabilities of the pre-trained language model and text rewriting technology, it effectively solves the common problems of semantic breaks and incoherent context in speech-to-text transcription. It improves the semantic coherence and naturalness of expression in long text processing. In particular, when processing speech-to-text content in professional fields or complex contexts, the memory enhancement mechanism can maintain long-distance semantic consistency, providing a high-quality data foundation for subsequent applications such as text analysis, information retrieval, and knowledge extraction.

[0203] Reference Figure 7 As one implementation method of speech recognition processing, after parsing long text, identifying core issues by combining industry terminology tags, associating voiceprint IDs with responsible persons for to-do items, and generating structured meeting minutes data, it also includes:

[0204] Step S701: Extract the responsible persons for the tasks and the associated industry term tags from the structured meeting minutes data;

[0205] Specifically, a mapping relationship is established between the person responsible for a task and specific industry terminology. This mapping relationship is based on the binding relationship between the voiceprint ID clearly marked in the meeting minutes and the person responsible, as well as the industry terminology tags identified in the meeting content. In this way, the system can accurately track the use of professional terminology by each person responsible in specific meeting scenarios, laying a data foundation for subsequent personalized terminology management.

[0206] Step S702: Construct a co-occurrence relationship graph of responsible persons and terms, and generate a term usage preference feature vector;

[0207] Specifically, a co-occurrence graph is a graph data structure where nodes represent responsible persons and industry terms, edges represent the co-occurrence relationships between them, and edge weights reflect the frequency or importance of co-occurrence. This graph structure can intuitively show the strength of association and usage patterns between different responsible persons and various industry terms. A terminology usage preference feature vector is a data structure that quantitatively represents the terminology usage habits of each responsible person. It typically takes the form of a high-dimensional vector, with each dimension corresponding to a specific industry term or term category. The vector values ​​reflect the responsible person's frequency of use of the term, the importance of the usage scenario, and other characteristics.

[0208] Step S703: Dynamically update the priority replacement weight values ​​in the industry terminology library based on the terminology usage preference feature vector.

[0209] The priority replacement weight value is an important parameter in the industry terminology database, used to guide the system's preference for selecting synonyms or near-synonyms during text processing. This weight value is dynamically updated based on the actual usage preferences of the responsible personnel, using machine learning algorithms or statistical analysis methods to transform individual usage habits into optimization parameters for the terminology database.

[0210] For example, if a responsible person prefers to use the term "artificial intelligence" rather than "machine intelligence," the system will adjust the priority weight of these two terms in the relevant handling scenarios of that person accordingly.

[0211] In the above implementation, a dynamic optimization mechanism based on the terminology usage preferences of responsible personnel is constructed, realizing intelligent management of the industry terminology database and improving the personalized adaptability of terminology processing. By analyzing the actual terminology usage patterns of responsible personnel, the system can automatically optimize the terminology selection strategy, making the generated text content more in line with the language habits and professional expression preferences of specific user groups. This adaptive terminology management mechanism not only improves the accuracy and naturalness of text processing, but also provides terminology services that are more tailored to the actual needs of users in different industries and positions, effectively solving the terminology usage mismatch problem caused by the static management of traditional terminology databases.

[0212] This application also discloses a speech recognition processing system based on a conference scenario.

[0213] A speech recognition and processing system for a meeting scenario, the system comprising:

[0214] The acquisition module is used to acquire the raw conference audio stream collected by a microphone array deployed in the conference room; wherein the microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones.

[0215] The preprocessing module is used to preprocess the raw conference audio stream acquired by the main channel and output a clean voice signal after noise reduction.

[0216] The sound source localization module is used to locate the sound source based on the raw conference audio stream collected by the microphone array and generate a sound source location heatmap.

[0217] The feature extraction module is used to extract Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories from clean speech signals to construct multi-dimensional voiceprint feature vectors;

[0218] The voiceprint grouping module is used to dynamically group multi-dimensional voiceprint feature vectors based on the sound source location heatmap and a spatial constraint clustering algorithm. It also arbitrates conflicting cluster segments by combining semantic coherence and outputs a set of speech segments labeled with voiceprint IDs.

[0219] The speech recognition module is used to input a set of speech segments into a general speech recognition model to generate initial transcribed text;

[0220] The industry terminology correction module is used to load a pre-configured industry terminology library to dynamically correct the initial transcribed text and output a transcribed text stream with industry terminology tags.

[0221] The enhancement processing module is used to perform periodic memory enhancement processing on the transcribed text stream and output semantically coherent long text.

[0222] The meeting minutes generation module is used to parse long texts, identify core topics by combining industry terminology tags, bind voiceprint IDs to the responsible persons of to-do items, and generate structured meeting minutes data.

[0223] The speech recognition processing system based on a meeting scenario according to this application embodiment can implement any of the above-mentioned speech recognition processing methods, and the specific working process of each module in the speech recognition processing system can refer to the corresponding process in the above-mentioned method embodiments.

[0224] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0225] This application also discloses a computer device.

[0226] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a speech recognition processing method based on a conference scenario as described above.

[0227] This application also discloses a computer-readable storage medium.

[0228] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the speech recognition processing methods for a conference scenario.

[0229] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0230] It should be noted that the computer device and storage medium in the embodiments of this application are respectively electronic devices and storage media that apply the above-described speech recognition processing method based on a conference scenario. Therefore, all embodiments of the above-described speech recognition processing method are applicable to the computer device and storage medium, and can achieve the same or similar beneficial effects. For the computer device / storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple; relevant details can be found in the descriptions of the method embodiments.

[0231] In this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0232] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0233] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A speech recognition processing method based on a meeting scenario, characterized in that, The identification processing method includes: Acquire the raw conference audio stream captured by a microphone array deployed in the conference room; wherein the microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones; The main channel performs signal preprocessing on the acquired raw conference audio stream and outputs a clean voice signal after noise reduction. Based on the original conference audio stream acquired by the microphone array, sound source localization is performed to generate a sound source location heatmap; Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories are extracted from the pure speech signal to construct a multi-dimensional voiceprint feature vector; Based on the heat map of the sound source location, a spatial constraint clustering algorithm is used to dynamically group the multi-dimensional voiceprint feature vectors. Semantic coherence is combined to arbitrate conflicting segments in the clustering and output a set of speech segments labeled with voiceprint IDs. The set of speech segments is input into a general speech recognition model to generate initial transcribed text; The pre-configured industry terminology library is loaded to dynamically correct the initial transcribed text, and the transcribed text stream with industry terminology tags is output. The transcribed text stream is subjected to periodic memory enhancement processing to output a semantically coherent long text. The long text is parsed, and the core issues are identified by combining the industry terminology tags. The voiceprint ID is associated with the person responsible for the task to be completed, and structured meeting minutes data is generated. After parsing the long text, identifying core topics by combining the industry terminology tags, associating the voiceprint ID with the person responsible for the to-do items, and generating structured meeting minutes data, the method further includes: Extract the responsible persons and associated industry term tags for the to-do items from the structured meeting minutes data; Construct a co-occurrence relationship graph of responsible persons and terms, and generate a feature vector of term usage preferences; The preferred replacement weight values ​​in the industry terminology library are dynamically updated based on the terminology using the preference feature vector.

2. The speech recognition processing method based on a conference scenario according to claim 1, characterized in that, The steps for preprocessing the acquired raw conference audio stream through the main channel to output a clean, denoised speech signal include: The original conference audio stream acquired by the main channel is collected to obtain the first channel audio signal and the second channel audio signal; Dynamic time warping is performed on the first and second channel audio signals. By calculating the peak-aligned phase difference of the cross-correlation function of the two signals, a phase-synchronized dual-channel audio signal is output. A noise mask matrix is ​​generated based on the phase-synchronized dual-channel audio signal. Steady-state noise regions are identified by analyzing the energy distribution in the time-frequency domain, and a binary mask matrix is ​​constructed. The phase-synchronized dual-channel audio signal is input into an adaptive filter bank, and nonlinear echo cancellation is performed using the binary mask matrix as a reference signal to output a primary noise reduction signal. Subband spectral entropy analysis is performed on the primary noise-reduced signal to detect and filter out transient noise pulses, and a clean speech signal after noise reduction is output.

3. The speech recognition processing method based on a conference scenario according to claim 1, characterized in that, The steps for locating sound sources and generating a sound source location heatmap based on the original conference audio stream acquired by the microphone array include: The original conference audio streams collected by all microphones in the microphone array are processed by frame segmentation, and each frame signal is separated into multiple sub-band signals by a bandpass filter; The generalized cross-correlation function between all microphone pairs is calculated based on the signal of each sub-band, and the time delay of searching for the peak of the generalized cross-correlation function is used as the time delay estimate of the sub-band. Based on the estimated time delay, the spatial coordinates of the sound source in the current frame are calculated using a least squares optimization algorithm. The physical space of the conference room is divided into a uniform grid. The frequency of the spatial coordinates of all sound sources falling into each grid within a preset time window is counted to generate a probability density distribution matrix. Gaussian smoothing filtering is applied to the probability density distribution matrix to output a matrix-form heatmap of sound source location; where each matrix element represents the probability of the sound source presence in the corresponding grid.

4. The speech recognition processing method based on a conference scenario according to claim 3, characterized in that, Based on the aforementioned sound source location heatmap, the steps of dynamically grouping the multi-dimensional voiceprint feature vectors using a spatially constrained clustering algorithm, arbitrating conflicting cluster segments based on semantic coherence, and outputting a set of speech segments labeled with voiceprint IDs include: Based on the directional information in the sound source location heatmap, the multi-dimensional voiceprint feature vectors are clustered according to spatial location similarity and voiceprint feature similarity to generate a preliminary speaker group set. Detect clustered conflict segments in the preliminary speaker grouping set and output a list of conflict segments; Extract the contextual semantic information of the clustered conflict fragments and calculate the semantic coherence score based on a natural language processing model; Arbitrate the attribution group labels of the clustering conflict segments based on the semantic coherence score to obtain the arbitration decision result; The preliminary speaker grouping set is updated based on the arbitration decision result, the grouping labels of the clustered conflict segments are adjusted, and voiceprint IDs are assigned to each speech segment. The set of speech segments labeled with voiceprint IDs is then output.

5. The speech recognition processing method based on a conference scenario according to claim 1, characterized in that, The steps of loading a pre-configured industry terminology library to dynamically correct the initial transcribed text and outputting a transcribed text stream with industry terminology tags include: Load the industry terminology database pre-stored in the local database; the industry terminology database includes an industry-specific glossary, a mapping relationship between the full and abbreviations of terms, and a priority replacement weight value for terms; The initial transcribed text is segmented to extract the segmented text units and timestamp intervals; Based on the term priority replacement weight value, candidate terms that match the text unit are retrieved from the industry terminology database. If there are multiple candidate terms, the term with the highest weight value is selected as the target term. When the target term and the text unit have a non-full name matching relationship, the text unit is replaced with the standard full name of the target term, and the timestamp interval of the term replacement position is recorded; Based on the timestamp interval of the term replacement location, locate the acoustic confidence score of the corresponding speech segment; When the acoustic confidence score is lower than a preset threshold, the semantic analysis module is invoked to analyze the semantic coherence of the context and verify the rationality of the term substitution. The verified target terms are marked as industry term tags, and the transcribed text stream with timestamps, voiceprint IDs, and industry term tags is output.

6. The speech recognition processing method based on a conference scenario according to claim 5, characterized in that, The steps of performing periodic memory enhancement processing on the transcribed text stream to output semantically coherent long text include: The transcribed text stream is obtained as input data; the transcribed text stream contains text segments arranged in chronological order and their corresponding timestamps and voiceprint IDs; The processing cycle is divided into fixed time intervals, and at the beginning of each cycle, the memory vector library of the previous cycle is loaded from the historical processing cache. The transcribed text fragments within the current period are input into the pre-trained language model to generate corresponding dynamic semantic vectors, and then similarity matching is performed with the memory vector library. Based on the similarity matching results, the associated memory vectors with similarity exceeding a preset threshold in the memory vector library are extracted and concatenated with the dynamic semantic vector to form an enhanced semantic vector; The enhanced semantic vector is input into the text rewriting module to obtain a long text paragraph with semantic coherence; Perform syntax tree alignment validation on the long text paragraph and output the final long text.

7. A speech recognition processing system based on a meeting scenario, characterized in that, A speech recognition processing system for performing a speech recognition processing method based on a conference scenario as described in any one of claims 1 to 6, the speech recognition processing system comprising, The acquisition module is used to acquire the raw conference audio stream collected by a microphone array deployed in the conference room; wherein, the microphone array includes a main channel pair consisting of a first microphone and a second microphone, and an auxiliary channel consisting of at least two other microphones; The preprocessing module is used to preprocess the original conference audio stream acquired by the main channel and output a clean voice signal after noise reduction. The sound source localization module is used to locate the sound source based on the original conference audio stream collected by the microphone array and generate a sound source location heatmap. The feature extraction module is used to extract Mel frequency cepstral coefficients, fundamental frequency jitter and formant trajectories from the pure speech signal to construct a multi-dimensional voiceprint feature vector; The voiceprint grouping module is used to dynamically group the multi-dimensional voiceprint feature vectors based on the sound source location heatmap using a spatial constraint clustering algorithm, arbitrate conflicting cluster segments by combining semantic coherence, and output a set of speech segments labeled with voiceprint IDs. The speech recognition module is used to input the set of speech segments into a general speech recognition model to generate initial transcribed text; The industry terminology correction module is used to load a pre-configured industry terminology library to dynamically correct the initial transcribed text and output a transcribed text stream with industry terminology tags. An enhancement processing module is used to perform periodic memory enhancement processing on the transcribed text stream and output a semantically coherent long text. The meeting minutes generation module is used to parse the long text, identify core topics by combining the industry terminology tags, associate the voiceprint ID with the person responsible for the task, and generate structured meeting minutes data.

8. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 6.

Images