Method, system and equipment for automatically generating cadre conversation research material based on multi-agent collaboration and speech enhancement recognition, and storage medium

Abstract

The invention discloses a cadre conversation investigation material automatic generation method, system and device based on multi-agent collaboration and speech enhancement recognition and a storage medium, and belongs to the technical field of artificial intelligence and natural language processing.The method comprises the following steps that cadre conversation investigation audio data are collected and preprocessed to obtain enhanced audio data, and the enhanced audio data are sent to a server; performing frame-level feature extraction, inputting the frame-level features into the large voice basic model, performing voiceprint recognition to obtain a structured voice transfer text, performing sensitive information recognition and semantic error correction optimization, performing policy term standardization processing to obtain a text conforming to language specifications and policy requirements, and matching a corresponding template to obtain a structured voice transfer text; and extracting core elements, filling the core elements into the template, generating a cadre reality presentation first draft, and performing semantic optimization and verification to obtain a cadre reality presentation material manuscript conforming to organization use specifications. According to the invention, the problem of low speech recognition accuracy can be solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and natural language processing, and particularly relates to a method, system and device for automatically generating cadre conversation investigation materials based on multi-agent collaboration and voice enhancement recognition, and a storage medium. BACKGROUND

[0002] Cadre conversation investigation is an important means for obtaining cadre performance information, assisting in cadre selection and performance evaluation in personnel work, and its results have important reference value in cadre evaluation, assessment and decision-making. The traditional cadre conversation process mainly relies on manual recording and subjective induction, which is not only low in efficiency and non-uniform in standard, but also easily interfered by human factors, and is difficult to meet the current needs of large-scale cadre management and efficient decision-making.

[0003] In recent years, with the rapid development of artificial intelligence technology, voice recognition and natural language processing have been gradually applied to the government affairs field. The voice recognition system based on deep learning has been widely used in tasks such as meeting recording and judicial record in the mandarin scenario. However, in the actual cadre investigation voice scenario, there are still many challenges, such as multiple people speaking in turn, different speech speeds, frequent appearance of local dialects, and serious background noise interference, which makes it difficult for existing voice recognition systems to ensure the accuracy and stability of recognition. At the same time, the text materials after voice recognition lack logical organization and language standardization, and the manual writing of cadre performance materials still relies on experience-driven, which has problems such as low writing efficiency, non-uniform expression style and non-standard policy terminology.

[0004] In addition, current text generation methods for organization management mainly focus on content completion driven by general language models, and lack understanding and constraints on organizational rules, cadre evaluation policies and personalized structures. Especially in the background of the increasing requirements of organizations on material quality and compliance, traditional technical means have been difficult to meet the generation needs of "compliance, safety, precision and efficiency" of government materials.

[0005] Therefore, it is urgent to build an intelligent cadre conversation material generation system that integrates voice enhancement recognition, semantic optimization, structure generation and compliance review, break through the key bottlenecks of complex voice processing and high-quality automatic text writing, and promote the standardization and intelligent upgrading of cadre management process. SUMMARY

[0006] The purpose of the present application is to provide a method, system, device and storage medium for automatically generating cadre conversation investigation materials based on multi-agent collaboration and voice enhancement recognition, which solves the problem of low voice recognition accuracy.

[0007] To achieve the above-mentioned purpose, the following technical solutions are adopted: In a first aspect, a method for automatically generating cadres' conversation research materials based on multi-agent collaboration and voice enhancement recognition includes the following steps: Collecting cadres' conversation research audio data for preprocessing to obtain enhanced audio data; Frame-level feature extraction is performed on the enhanced audio data, which is then input into a voice base large model, followed by voiceprint recognition to obtain structured voice transcription text; After sensitive information recognition and semantic error correction optimization of the structured voice transcription text, policy terminology standardization processing is performed to obtain text that meets language specifications and policy requirements; Matching the templates corresponding to the text that meets language specifications and policy requirements, extracting the core elements in the text that meets language specifications and policy requirements, and filling them into the templates to generate a preliminary draft of cadres' performance; After semantic optimization and verification of the preliminary draft of cadres' performance, a cadres' performance material manuscript that meets organizational usage specifications is obtained.

[0008] In some embodiments, the step of collecting cadres' conversation research audio data for preprocessing to obtain enhanced audio data specifically includes: After format unification and down-sampling of the cadres' conversation research audio, a mel-spectrogram is generated through short-time Fourier transform and mel filtering, and Gaussian noise disturbance is added; A voice enhancement model is used to process the mel-spectrogram with Gaussian noise added to obtain a repaired mel-spectrogram, which is converted into a time-domain audio signal as enhanced audio data; The voice enhancement model includes an encoder, a decoder, and an output layer. The encoder uses a multi-layer cascaded ResUNet structure, each layer of which consists of a convolution kernel, batch normalization, and a LeakyReLU activation function, and uses residual connection; The decoder is composed of multiple layers of transposed convolution, each layer of which fuses the skip-connection features of the corresponding layer of the encoder; The output layer uses a Sigmoid function.

[0009] In some embodiments, the step of frame-level feature extraction on the enhanced audio data and inputting it into a voice base large model for voiceprint recognition to obtain structured voice transcription text specifically includes: Mel-frequency spectrum features of the enhanced audio data are extracted and input into a voice base large model, followed by voiceprint recognition using a Cam++ model to obtain structured voice transcription text, which includes timestamps, speaker labels, speech content, and dialect types; The total loss function of the voice base model is composed of a voice recognition CTC loss, a cross-entropy loss and an InfoNCE loss.

[0010] In some embodiments, the step of performing sensitive information identification and semantic error correction optimization on the structured voice transcription text, and then performing policy terminology standardization processing to obtain a text that meets language specifications and policy requirements, specifically includes: The similarity between the to-be-detected text segment in the structured voice transcription text and the preset sensitive word is calculated using a BERT semantic similarity model. When the similarity is greater than or equal to a preset similarity threshold, the to-be-detected text is marked as ambiguous sensitive information and replaced with a preset candidate word to obtain a structured voice transcription text after sensitive information identification. Otherwise, no marking or replacement is performed. The perplexity and semantic coherence score of the to-be-detected text segment in the structured voice transcription text after sensitive information identification are calculated. When the perplexity is greater than or equal to a preset perplexity threshold, the to-be-detected text segment is a semantically abnormal sentence and is corrected. Otherwise, no correction is performed. When the semantic coherence score is less than a preset coherence score threshold, a logical link sentence for the to-be-detected text is supplemented. Otherwise, no supplementation is performed. Finally, a structured voice transcription text after semantic error correction optimization is obtained. The matching degree between the terminology in the structured voice transcription text after semantic error correction optimization and the standard terminology is calculated. When the matching degree is greater than or equal to a preset matching degree threshold, the terminology is replaced with the standard terminology. Otherwise, no replacement is performed. The compliance score of the replaced standard terminology in the structured voice transcription text after semantic error correction optimization is calculated. When the compliance score is greater than or equal to a preset compliance score threshold, a text that meets language specifications and policy requirements is obtained. Otherwise, the replaced standard terminology is marked as an unqualified terminology, and the step is performed again on the unqualified terminology.

[0011] In some embodiments, the step of matching the text that meets language specifications and policy requirements to a template, extracting core elements from the text that meets language specifications and policy requirements and filling them into the template to generate a preliminary draft of the cadre performance, specifically includes: The text that meets language specifications and policy requirements is converted into a semantic vector using a Sentence-BERT model, and the cosine similarity between the semantic vector and the semantic vector of each template in a preset template library is calculated. The template with the highest cosine similarity is selected. The template with the highest cosine similarity is converted into a slot-filling structure, and a core element slot is marked. The core elements in the text that meet the language norms and policy requirements are extracted and filled into the core element slot. Then, a paragraph-level content is generated by combining the prompt chaining strategy, and a perplexity check is performed to obtain a preliminary draft of cadre performance.

[0012] In some embodiments, the step of performing semantic optimization and checking on the preliminary draft of cadre performance to obtain cadre performance material manuscripts that meet the organization's usage specifications specifically includes: After deleting repetitive descriptions in the preliminary draft of cadre performance by a large language model and rewriting, the structure consistency verification, key information coverage review, and policy term compliance matching are performed in sequence based on the preset policy knowledge graph and organization material specification standards to obtain cadre performance material manuscripts that meet the organization's usage specifications.

[0013] In a second aspect, a cadre conversation research material automatic generation system based on multi-agent collaboration and voice enhancement recognition includes: A voice enhancement module is configured to collect cadre conversation research audio data for preprocessing to obtain enhanced audio data. A voice recognition module is configured to perform frame-level feature extraction on the enhanced audio data and input the enhanced audio data into a voice base large model to perform voiceprint recognition and obtain structured voice transcription text. A semantic optimization module is configured to perform sensitive information recognition and semantic error correction optimization on the structured voice transcription text, and then perform policy term standardization processing to obtain text that meets language norms and policy requirements. A multi-agent writing module is configured to match templates corresponding to the text that meets language norms and policy requirements, extract core elements in the text that meets language norms and policy requirements, and fill the core elements into the templates to generate a preliminary draft of cadre performance. A material proofreading module is configured to perform semantic optimization and checking on the preliminary draft of cadre performance to obtain cadre performance material manuscripts that meet the organization's usage specifications.

[0014] In a third aspect, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor. When the processor executes the computer program, the steps of the method for automatically generating a cadre conversation research material based on multi-agent collaboration and voice enhancement recognition are implemented.

[0015] In a fourth aspect, a computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition.

[0016] In a fifth aspect, a computer program product includes a computer program, and the computer program, when executed by a processor, implements the steps of the method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition.

[0017] Compared with the prior art, the present application has the following beneficial effects: The present application provides a method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition. By performing frame-level feature extraction on the enhanced audio data and inputting it into a voice base model, voiceprint recognition is performed to obtain structured voice transcription text. This can accurately process local accents and avoid semantic deviations caused by accents, significantly improving voice transcription accuracy and ensuring text input quality from the source. After sensitive information recognition and semantic error correction optimization of the structured voice transcription text, policy terminology standardization processing is performed to obtain text that meets language specifications and policy requirements. The text is matched with a corresponding template, core elements in the text are extracted and filled into the template, and a cadre performance draft is generated. Semantic optimization and verification are performed on the cadre performance draft to obtain cadre performance material manuscripts that meet organizational usage specifications. This can avoid format confusion and inconsistent expression, ensure consistent structure and standardized style of cadre performance materials, quickly screen high-risk content, automatically correct logical conflicts, supplement key information, significantly shorten compliance review time, and improve overall material generation efficiency.

[0018] Further, the Gaussian noise disturbance enhancement model is robust to noise, and the encoder of the ResUNet structure can accurately repair noise, reverberation and other distortions through residual connection and skip connection. The decoder is composed of multiple transpose convolutions, which can restore high-frequency details. The output enhanced audio can maximize the preservation of semantic information and avoid recognition errors caused by poor audio quality.

[0019] Further, the Mel spectrum feature as the model input can highlight the key information of speech acoustics, adapt to dialect and multi-role speech recognition requirements, realize voiceprint recognition combined with the Cam++ model, accurately distinguish the speaking roles of multiple people, and generate structured text containing corresponding speaker labels, solving the problem of role attribution confusion when multiple people speak in turns. The total loss function of the speech base large model is composed of speech recognition CTC loss, cross-entropy loss and InfoNCE loss, the speech recognition CTC loss optimizes the continuous speech transcription accuracy, the cross-entropy loss realizes accurate classification of dialect types, and the InfoNCE loss can strengthen the semantic consistency across dialects, effectively solving the recognition errors caused by local accents, and ensuring the semantic integrity and non-bias of the transcribed text.

[0020] Further, compared with traditional rule matching, the BERT semantic similarity model combined with a preset threshold can identify sensitive word variants, reduce missed and false judgments, replace sensitive information with candidate words, remove risk content while ensuring semantic equivalence, avoid information loss, and adapt to the processing requirements of confidential and private content involved in cadre conversations. Through perplexity judgment of semantic abnormal sentences and coherence score supplement of logical connection sentences, the problems of colloquial expression and logical break can be automatically corrected, and the readability of the text can be improved; finally, through the double mechanisms of matching degree threshold screening and compliance score verification, the non-standard terms can be accurately replaced by standard expressions, and the problem of mixed terms after replacement can be avoided, which meets the strict expression requirements of the organization policy.

[0021] Further, the Sentence-BERT model is used to generate semantic vectors and calculate cosine similarity, which can filter the most suitable templates for the text theme and scene from the semantic level, and avoid the deviation of material structure caused by template mismatch. The template is converted into a slot-filling structure, which can ensure that the extracted information is complete, and combined with the prompt chaining strategy to generate paragraph-level content, which can ensure the logical coherence and standard expression of each paragraph. Then, the perplexity verification is used to screen the best draft, avoiding the logical confusion and content redundancy problems in manual writing, and improving the quality of the draft.

[0022] Further, through the large language model, repeated expressions are deleted and sentence patterns are optimized to solve the language redundancy and grammatical errors in the draft, and improve the professionalism of the text; combined with the policy knowledge graph and organizational norms for structure consistency verification, key information coverage review and policy term compliance matching, to ensure that the manuscript structure meets the cadre material standards, key information is complete, and the term expression is completely consistent with the policy requirements. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flowchart of the cadre conversation research material automatic generation method based on multi-agent collaboration and speech enhancement recognition provided by the embodiment of the application; Figure 2 A structure diagram of a cadre conversation research material automatic generation system based on multi-agent collaboration and voice enhancement recognition is provided for an embodiment of the present application. Figure 3 A network structure diagram of a dialect recognition module is provided for an embodiment of the present application. Figure 4 A technical principle diagram of a cadre conversation research material automatic generation method based on multi-agent collaboration and voice enhancement recognition is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order for those skilled in the art to better understand the present application, the technical solutions of the present application will be further described in detail below with reference to the accompanying drawings, which are explanatory rather than limiting of the present application.

[0025] It should be noted that the terms "comprise" and "have" and any variations thereof in the specification and claims of the present application are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device containing a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, systems, products or devices.

[0026] As shown in Figure 4 To solve the problems of low voice recognition accuracy, serious local accent and noise interference, low material writing standardization degree, and low compliance review efficiency in existing cadre conversation research work, a cadre conversation research material automatic generation method based on multi-agent collaboration and voice enhancement recognition is provided, and the specific ideas are as follows: S1: Collecting cadre conversation research audio data, using a multi-task voice enhancement module containing a ResUNet structure to complete voice de-noising, de-reverberation and super-sampling processing in parallel, improving the voice restoration ability in low-quality and noisy environments, and providing high-fidelity input for voice recognition.

[0027] S2: Input the enhanced audio data into a voice basic large model based on dialect adaptation, strengthen the cross-dialect semantic consistency through multi-task fine-tuning (including voice recognition CTC loss, dialect classification cross-entropy loss and cross-dialect semantic comparison learning InfoNCE loss), complete the end-to-end structured transcription of voice, and output the transcription text containing timestamp, speaker role label, speech content and dialect type.

[0028] S3: Perform semantic-level compliance optimization on the structured text, including sensitive information identification and replacement, term standardization processing and context semantic error correction, and output the text result conforming to the language specification and policy requirements.

[0029] S4: input the compliance text into the multi-agent cadre material generation system, the system sequentially calls the template matching, structure extraction and draft writing module, completes the paragraph-level text writing of post responsibilities, performance of duties, personality characteristics and other contents, and generates the initial draft of cadre performance.

[0030] S5: input the initial draft material into the polishing and proofreading module, optimize the language expression, check the logical consistency and examine the policy terms, and finally output the cadre performance material manuscript that meets the organization's use specifications.

[0031] The step S1 comprises the following steps: after the input audio data is uniformly formatted and down-sampled, a multi-task voice enhancement model is used to perform spectrum masking, channel recovery and small sample voice reconstruction tasks in parallel, and a high-fidelity audio is output after reconstruction through spectrogram reconstruction and neural vocoder; Specifically, for the input cadre conversation audio data, first, the original audio is sampled and standardized to 16kHz, single-channel WAV format; then the audio data is input into a multi-task voice enhancement model containing a ResUNet structure, and voice denoising, dereverberation and up-sampling tasks are performed in parallel, wherein: the denoising module uses a time-frequency domain hybrid masking mechanism to adaptively filter out background noise (such as conference room noise, overlapping voices, and equipment running sound); the dereverberation module uses multi-scale residual connection and inverse convolution network to convolve and invert the reverberation residue; the up-sampling module based on the non-autoregressive reconstruction mechanism restores the low sampling rate audio to high-fidelity voice signal, improving the voice spectrum detail restoration degree; The finally output audio signal retains high semantic clarity and low distortion characteristics, providing enhanced usable audio input for the subsequent voice recognition module.

[0032] The step S2 comprises the following steps: the voice recognition model is constructed based on a dialect-adapted voice basic large model, the model's adaptation ability to mixed dialect contexts is strengthened through a multi-task fine-tuning strategy, voice recognition CTC loss is used to optimize the transcription accuracy, dialect classification cross-entropy loss is introduced to realize accurate recognition of dialect types such as Shanbei dialect, Shannan dialect and Guanzhong dialect, and InfoNCE loss is used for cross-dialect semantic comparison learning to enhance the semantic consistency capture between different dialects; the model outputs structured speech text, including timestamp, role label, speech content and corresponding dialect type, providing basic data support with accuracy and semantic coherence for subsequent text processing; Specifically, the enhanced audio data after denoising, dereverberation and oversampling processing in step S1 is input into a dialect-adapted voice base model. The model takes mixed languages (Mandarin and Shaanxi dialects, including northern Shaanxi dialect, southern Shaanxi dialect, and central Shaanxi dialect) and multi-role interactive context in the context of official conversation as the core design goal, and constructs an end-to-end structured voice transcription link. In the transcription process, first, frame-level feature extraction is performed on the enhanced audio to convert the audio signal into a mel-spectrum graph reflecting the acoustic characteristics of the voice, providing accurate acoustic feature input for the model. Then, the model is optimized and trained through a multi-task fine-tuning mechanism, in which: the voice recognition CTC (Connectionist Temporal Classification) loss is used to optimize the transcription accuracy of continuous speech sequences and reduce the text recognition rate in ambiguous pronunciation and fluctuating speech speed scenarios; the dialect classification cross-entropy loss is used to classify the dialect types in the audio, accurately distinguish the sub-types of Shaanxi dialect and Mandarin in the mixed context, and generate corresponding dialect type labels; the cross-dialect semantic comparison learning InfoNCE loss is used to construct semantic association sample pairs between different dialects (such as northern Shaanxi dialect and Mandarin) to strengthen the model's consistency in recognizing the same semantics in different dialects and avoid semantic deviation caused by dialect differences. At the same time, the model integrates multi-role recognition capability, analyzes the speaker's voiceprint features and speaking time sequence in the audio, and generates accurate speaker identity labels using the Cam++ model to clearly distinguish the roles in the multi-person speaking scenario. The final output is a structured voice transcription text containing four core pieces of information: the timestamp of the audio corresponding paragraph (used to trace the speaking time sequence), the speaker label (clearly attributing the content to the subject), the original speech content text (preserving semantic integrity), and the dialect type label (labeling the language type used in the content), which can directly provide semantic clear, role clear, and traceable basic text data support for subsequent sensitive information filtering, semantic correction, and material generation modules.

[0033] The step S3 includes the following steps: the sensitive information identification module uses a rule + semantic dual-channel mechanism to automatically screen and replace high-risk words by combining a pre-set sensitive word library with a BERT semantic vector model; the semantic correction module calls a large language model to optimize sentences at the syntax level and context logic level, improving the accuracy and readability of the text; Specifically, the structured transcription text output in step S2 is subjected to semantic-level optimization processing. First, the text is screened sentence by sentence by a sensitive information identification module. A BERT-based semantic similarity model is combined with a rule matching strategy to identify sensitive expressions and their variants in the annotated word library and automatically replace them with semantically equivalent compliant expressions. Then, the screened text is input into a semantic correction module driven by a large language model. Through a context window mechanism, colloquial expressions, semantic ambiguity, logical conflicts, and grammatical errors are identified and corrected. At the same time, the use of internal railway system terminology is aligned and standardized. Finally, the text content that meets the policy expression specification, semantic coherence, and grammatical accuracy is output, providing compliant and high-quality input text corpus for the cadre material generation module.

[0034] The step S4 includes the following steps: the template matching module uses Sentence-BERT to generate a semantic vector representation of the input text, and retrieves the top-K structure templates with the highest matching degree in the template library based on the Faiss index mechanism; the structure extraction module converts the matching template into a slot structure, and labels the paragraph composition and expression elements; the draft writing module calls a large language model, and completes the generation of each segment of content through prompt chaining and context caching mechanism, while retaining personalized expression labels; Specifically, the optimized text in step S3 is input into the cadre performance material multi-agent generation system, and the material writing process is completed by the template matching agent, the structure extraction agent, and the draft writing agent in sequence: the template matching agent uses Sentence-BERT to generate a semantic embedding representation of the input text, and retrieves the top-K candidate templates with the highest matching degree in the cadre material template library based on the vector similarity retrieval mechanism; the structure extraction agent converts the selected template into a slot-filling structure, identifies and labels the paragraph structure and key element slots required by the material, including job responsibilities, performance, ability characteristics, and individual highlights; the draft writing agent calls a large language model, fills the content according to the element information extracted from the input text and the template structure, and realizes segmented generation by combining the prompt chaining strategy, ensuring that the material is logically clear, expressionally standardized, and personalized. The generated preliminary draft material content is consistent with the input semantics and the format meets the organizational specification requirements, preparing for subsequent proofreading and output.

[0035] The step S5 includes the following steps: the polishing module calls the style conversion and automatic correction components in the language model to adjust the fluency and professionalism of the expression; the proofreading module combines the policy knowledge graph and the format rules to conduct compliance review and key element coverage verification on the draft content, and outputs the final material with a proofreading report; Specifically, the first draft material generated in step S4 is input into the polishing and proofreading intelligent agent module, and language optimization, logical verification and compliance review are performed in turn; the polishing intelligent agent uses the semantic optimization and sentence conversion mechanism built-in the large language model to intelligently identify and rewrite problems such as incorrect language, repetitive expression, inappropriate word choice and inconsistent language style in the material, improving the accuracy, fluency and professionalism of language expression; the proofreading intelligent agent combines the pre-set policy knowledge graph and organizational material standard to verify the structural consistency of the draft content, review the coverage of key information and match the compliance of policy terms, ensuring that the material meets the requirements of cadre assessment in terms of logical structure, policy expression and content elements; the system synchronously retains the original draft version and the version after proofreading and optimization, and provides a manual review interface to realize a controllable material quality management process combining "automatic generation + manual intervention". The final output is a high-quality cadre performance material manuscript after multi-dimensional intelligent review and polishing, meeting the formal archiving and organizational use standards.

[0036] A cadre conversation content automatic generation system based on voice enhancement and multi-agent collaboration, comprising: a voice enhancement module, a voice recognition module, a semantic optimization module, a multi-agent writing module and a material proofreading module; wherein the voice enhancement module is used to improve the audio quality in a noisy environment; the voice recognition module is used to transcribe multi-role voice into timestamped text; the semantic optimization module is used to complete compliance adjustment; the multi-agent writing module is used to match templates and generate cadre performance content; the material proofreading module is used to output formal manuscripts that meet organizational standards.

[0037] Based on the above technical ideas, as shown in Figure 1 and Figure 3 The embodiment provides a cadre conversation research material automatic generation method based on multi-agent collaboration and voice enhancement recognition, comprising the following steps: S1, collecting cadre conversation research audio data for preprocessing to obtain enhanced audio data; In actual cadre conversation scenarios, audio data is often disturbed by background noise, overlapping speech of multiple people, insufficient sampling rate and reverberation residue, which seriously affects the accuracy of subsequent recognition and material generation. To solve the above problems, the embodiment first collects and formats the audio data, and completes parallel repair of denoising, dereverberation and oversampling through a multi-task voice enhancement module containing a ResUNet structure.

[0038] First, the collected original voice is resampled to 16 kHz single-channel format. To reduce information loss caused by interpolation, a linear interpolation formula Smooth the signal to improve the sampling rate while preserving as much voice detail as possible.

[0039] In the speech enhancement preprocessing stage, the speech signal is converted into a mel-spectrogram using short-time Fourier transform (STFT) (80-dimensional mel-filter bank, T is the time frame), where the filter bank dimension is 80 and the number of time frames is T. To improve the robustness of the model, a Gaussian noise data augmentation strategy is introduced to perturb the spectrogram during the training phase, thereby improving the noise resistance of the model in noisy environments.

[0040] The speech enhancement module based on ResUNet is used as a pre-processing step for the speech recognition large model to solve the problem of reduced speech intelligibility caused by noise interference and reverberation. Specifically, the encoder part of the model uses a 5-layer cascaded ResUNet structure, where each layer consists of convolution (step size 2), batch normalization (BN), and LeakyReLU activation function, and is connected through residual connection to extract deep features of the noisy spectrogram and gradually compress the spatial dimension to In the decoder part, a non-autoregressive reconstruction mechanism is used, which gradually restores the spectrogram size through 4 layers of transposed convolution (step size 2), and each layer simultaneously fuses the skip-connection features from the encoder to alleviate gradient vanishing and enhance the recovery ability of local details. The final output layer uses a Sigmoid activation function to generate the repaired mel-spectrogram.

[0041] During training and optimization, a hybrid loss is used:

[0042] where is the pure spectrogram, (short-time objective intelligibility) is used to ensure the integrity of the speech semantics.

[0043] Through the above method, the output enhanced audio not only significantly reduces noise and reverberation interference, but also restores the high-frequency details of the low-sampling audio, providing high-fidelity input for the subsequent speech recognition module.

[0044] S2, frame-level feature extraction is performed on the enhanced audio data, and then input into the speech base large model, and voiceprint recognition is performed to obtain structured speech transcription text. In the scene of dry official talks, the speakers often mix standard Chinese with multiple local dialects (such as Shaanbei dialect, Shaannan dialect, Guanzhong dialect, etc.), and the speaking speed, tone and accent of different speakers differ greatly, which leads to the traditional speech recognition model to easily appear word recognition error and semantic deviation when transcribing. Therefore, the embodiment proposes a multi-dialect recognition optimization method based on cross-dialect joint learning, taking a speech base large model as the core, and through multi-task fine-tuning, the model has the ability of speech recognition and preliminary semantic understanding in the mixed dialect speaking scene, focuses on strengthening the semantic consistency across dialects, and solves the problem of data scarcity of low-resource dialects.

[0045] First, input feature construction is performed. The enhanced speech output by S1 is unified into a 16 kHz single-channel audio format, and 80-dimensional Mel-spectral features are extracted as model inputs to obtain a feature matrix , where represents the number of time frames. To avoid model overfitting to speaker features, the role label encoding related to the voiceprint is removed in the input stage, so that the model focuses more on the dialect speech features themselves.

[0046] In the multi-task fine-tuning process, a three-level optimization strategy is designed: (1) Core task loss: speech recognition CTC loss The speech recognition part uses Connectionist Temporal Classification (CTC) as the objective function to optimize the mapping from audio to character sequence. Given the target character at time , the prediction probability distribution is , and the CTC loss is defined as:

[0047] where is the target character at time , and is the model prediction probability. The dialect classification task is extended to a multi-classification task including Shaanbei dialect, Shaannan dialect, Guanzhong dialect, and standard Chinese, and the cross-entropy loss is used for optimization.

[0048] (2) Dialect classification task loss: cross-entropy loss To identify different dialect categories (Shaanbei dialect, Shaannan dialect, Guanzhong dialect, and standard Chinese), the model introduces a dialect classification sub-task. Given the dialect label to which the sample belongs, the prediction probability is , and its loss function is:

[0049] where, wherein, the total number of dialect categories, one-hot encoding of the true category.

[0050] (3) Cross-dialect semantic consistency: InfoNCE loss To strengthen the semantic consistency across dialects, an optimization objective based on contrastive learning is introduced. First, a positive sample pair is constructed: select voice clips of the same semantic content in different dialects, and generate feature vectors after model encoding and . Second, a negative sample pair is constructed: randomly select voice clips with different semantics or the same semantics but with large dialect differences, and generate feature vectors . The InfoNCE loss is used for optimization:

[0051] wherein, denotes the cosine similarity, is the temperature coefficient, is the number of negative samples. This loss forces the model to pull the synonymous dialect expressions closer in the semantic space and distinguish different semantic contents, thereby capturing cross-dialect common features.

[0052] (4) Total loss function The training objective of the final model is a weighted combination of the above three types of task losses:

[0053] wherein, is a dynamically adjusted weight coefficient used to balance the importance of different tasks during training.

[0054] Through the above method, the model not only achieves high-precision speech transcription in a dialect mixed speech environment, but also effectively ensures the consistency of cross-dialect semantics, providing structured and stable input for subsequent text compliance optimization and material generation.

[0055] S3, after sensitive information recognition and semantic error correction optimization on the structured speech transcription text, policy terminology standardization processing is performed to obtain a text that meets language specifications and policy requirements; The structured text generated by cross-dialect recognition of cadre conversation speech often has problems such as colloquialism, sensitive words, semantic ambiguity, and non-standard use of policy terminology. If not further processed, directly used for cadre performance material writing, it is easy to lead to imprecise expression, insufficient compliance, and even policy risks. Therefore, based on the transcription result, the semantic level compliance optimization module is designed to ensure that the text content meets the organizational specifications through sensitive information recognition, semantic error correction, and terminology standardization processing.

[0056] In the semantic level compliance optimization of structured transcription text, first, sensitive information in the text is identified and replaced, combining rule matching and semantic vector matching mechanisms for accurate screening. Direct marking of completely matched sensitive content is completed by traversing the preset sensitive word table. The BERT semantic similarity model is used to calculate the semantic similarity between the text segment to be detected and the sensitive words in the sensitive word library. The similarity calculation formula is: wherein is the BERT semantic vector (dimension 768) of the word or phrase to be detected, is the BERT semantic vector of a sensitive word in the sensitive word library, represents the L2 norm, and when the calculated similarity value is not less than 0.8, it is determined as fuzzy sensitive information and marked. Then, suitable replacement items are selected from the candidate word database, and the adaptation score formula is used to calculate the adaptation score: wherein takes 0.6 (weight coefficient, balancing semantic similarity and context adaptability), c is a replacement word in the candidate word library, w' is the context core keyword vector of the sensitive word, is the adaptation score of the candidate word and the context (based on the 3-gram language model to calculate the probability, with a value range of [0, 1]), and the candidate word with the highest score is selected for replacement. If the highest score is less than 0.5, manual review is triggered to ensure that the replacement is semantically equivalent and meets the context requirements.

[0057] After completing the sensitive information processing, further semantic error correction and context consistency optimization are carried out. The pre-trained large language model is used to calculate the sentence perplexity to quantify the semantic fluency. The perplexity calculation formula is: wherein is the sentence to be detected, and N is the length of the sentence, is the conditional probability prediction of the model M for the i-th word, and when the perplexity value is not less than 100, it is determined as a semantically abnormal sentence, which needs to be analyzed for error type (ambiguity / logical conflict / syntax error) and corrected. At the same time, the semantic coherence of adjacent sentences in the paragraph is calculated to avoid logical discontinuity problems: wherein P is the text paragraph, and N is the number of sentences contained in the paragraph, is the average semantic vector (BERT word vector mean) of the i-th sentence, is the cosine similarity (with a value range of [0, 1]), and when the coherence score is less than 0.3, a large language model is called to supplement the logical connection sentence to ensure the logical coherence of the context.

[0058] Finally, the policy terminology standardization process is performed to ensure that the text conforms to the organization's personnel terminology specifications. The matching degree of the terms in the text to be processed and the standard terminology library is calculated in two dimensions: string edit distance and semantic similarity. The matching degree calculation formula is: wherein 0.4 (weight coefficient, balancing string similarity and semantic similarity), t is the term to be standardized (e.g., "performance score"), is the term in the standard terminology library, is the normalized edit distance, and the calculation formula is: wherein is the original edit distance, is the term length, and when the matching degree is not less than 0.75, the term replacement is performed. After the replacement is completed, the term consistency is checked by the paragraph-level term compliance score formula to avoid the problem of "mixing standard and non-standard terms". The compliance score formula is: wherein is the set of all terms in the paragraph, K is the number of terms, is the corresponding standard term, and if the compliance score is not less than 0.8, the term standardization is determined to be qualified, otherwise the unqualified terms are marked and re-matched. Finally, the text content that conforms to the policy expression specifications, semantic coherence, and grammatical accuracy is output, providing high-quality input corpus for the subsequent cadre material generation module.

[0059] S4, matching the template corresponding to the text that conforms to the language specifications and policy requirements, extracting the core elements from the text that conforms to the language specifications and policy requirements and filling them into the template to generate the initial draft of the cadre performance. The text optimized at the semantic level in S3 is input into the cadre performance material multi-agent generation system. Through the division of labor and cooperation of multiple agents and the connection of the process, the automatic conversion from the compliant text to the structured initial draft is realized. First, the template matching agent starts the processing flow. This agent uses the Sentence-BERT model to perform semantic embedding processing on the compliant text, converting the text into a fixed-dimensional semantic vector. Then, based on the Faiss indexing mechanism, it searches in the pre-set cadre material template library. By calculating the cosine similarity between the semantic vector of the compliant text and the semantic vector of each template in the template library, the top-K candidate templates with the highest matching degree are selected (K value is pre-set to 3 according to the material type), ensuring that the selected template is highly adaptable to the theme and scene of the cadre conversation content, laying a structural foundation for subsequent content generation.

[0060] After the template is determined, the structure extraction agent takes over the processing task, converts the selected candidate template into a slot-filling structure, and clearly marks the core element slots required by the material, including the post responsibility description, work performance, core competence characteristics, personality highlights, and improvement direction of the cadre. At the same time, the agent analyzes the compliance text in depth through semantic extraction algorithm, extracts the information content corresponding to each slot from the text, such as the specific measures, effectiveness data, and colleague evaluation of the cadre in a certain key work from the conversation record, as the original material for slot filling; To ensure the accuracy of the extracted information, the system will check the extraction results, only when the element extraction accuracy (the ratio of the number of correctly extracted elements to the total number of elements required by the template) is not less than 90%, it will enter the next step, if it is lower than the threshold, it will trigger secondary extraction or manual intervention prompt.

[0061] Finally, the draft writing agent completes the first draft writing, which calls a large language model and realizes the ordered generation of paragraph-level content by combining the prompt chaining strategy: first, generate the introduction part of the material with "cadre basic information + core topic of the conversation" as the initial prompt, and clearly state the purpose of the material and the core evaluation direction; then generate the performance paragraph with the introduction content as the context cache, organize the extracted performance information into sentences according to the logic of "work task - implementation measures - effectiveness highlights"; subsequently, complete the writing of the ability characteristics, personality highlights and other modules, ensuring logical coherence and content focus between paragraphs. After the first draft is generated, the system automatically checks the perplexity of the text, calculates the conditional probability prediction value of the word sequence in the text to quantify the language fluency of the text, and only when the perplexity is lower than the preset threshold (usually 80), the first draft is determined to be fluent, and the cadre performance material first draft is output; if the perplexity exceeds the threshold, the agent will regenerate and optimize the sentence until the fluency requirement is met.

[0062] S5, semantic optimization and verification of the cadre performance first draft, to get the cadre performance material manuscript that meets the organization's usage specifications.

[0063] The initial draft of the S4 generated dry department performance material is input into the polishing and proofreading intelligent agent module. This module ensures that the manuscript fully meets the requirements of the dry department assessment in terms of language expression, logical structure, and policy compliance through multi-dimensional intelligent processing and verification. First, the polishing intelligent agent conducts language optimization. This intelligent agent relies on the semantic optimization and sentence conversion mechanisms built into the large language model to accurately identify and rewrite language problems in the initial draft. It addresses issues such as colloquial expressions, sentence errors, and repetitive statements by deleting redundant descriptions. At the same time, it unifies the language style of the material to ensure that the overall manuscript presents a formal and official style of official documents, avoiding mixed styles. To ensure that the rewritten semantics do not deviate from the original meaning, the polishing intelligent agent also calculates the semantic similarity between the initial draft and the polished text. Only when the similarity is not less than 95% is the polishing result considered valid, otherwise the rewriting strategy is adjusted again.

[0064] After polishing, the proofreading intelligent agent starts a multi-dimensional verification process. Based on the pre-set policy knowledge graph and organizational material specification standards, the intelligent agent conducts three layers of verification: the first layer is structure consistency verification, which checks whether the paragraph structure of the manuscript fully matches the standard structure of the dry department material (introduction - performance - ability characteristics - highlights and deficiencies - summary), and checks whether there are problems such as paragraph deletion or sequence disorder; the second layer is key information coverage review, which confirms whether the manuscript completely contains the key information required for dry department assessment, such as key work during the tenure, quantitative results, and democratic evaluation results, to avoid missing core information; the third layer is policy terminology compliance matching, which compares the consistency of the policy terminology used in the manuscript with the standard terminology library, corrects non-standard expressions, and ensures that all terminology meets the official expression requirements of organizational personnel work. Each layer of verification generates corresponding verification results, and only when all three layers of verification pass (structure matching degree, information coverage rate, and terminology compliance rate are not less than 95%) is the manuscript determined to be proofread and qualified.

[0065] After the proofreading is passed, the system automatically retains the original draft version and the optimized version after proofreading, forming a version comparison record for subsequent tracing and checking; at the same time, it provides a manual review interface to support organizational personnel to conduct secondary inspection on the intelligent processed manuscript. The staff can view the modification traces of intelligent polishing and proofreading through the interface, adjust the controversial modification contents, or supplement the missed personalized information. After the manual review confirms that there is no error, the system outputs the final dry department performance material manuscript, which fully meets the organizational management standards and dry department assessment requirements, and can be directly used in dry department evaluation, selection and appointment, and other formal work scenarios.

[0066] The embodiment can effectively cope with complex speech scenes such as local dialect, multi-person interaction and environmental noise interference, and improve the speech recognition accuracy; meanwhile, through the construction of a multi-agent collaborative mechanism and a content generation method driven by a structural template, an efficient, compliant and personalized cadre performance material writing process is realized. The speech enhancement module introduces a multi-task parallel repair mechanism, enabling the system to have end-to-end composite distortion repair capability; the speech recognition module adapts to multi-role and multi-language scenes, and the recognition result is time-stamped and identity-tagged to enhance the subsequent traceability. The writing module integrates semantic embedding matching, template structure reasoning and large model generation capability, and has the functions of logical layering, semantic focusing and style adaptation; the review module ensures the compliance, accuracy and readability of the material through the rule base and knowledge graph. The system as a whole has high automation and controllability, significantly reduces the burden of manual writing, and improves the informatization and intelligent level of cadre assessment work.

[0067] As shown in Figure 2 The embodiment also provides a cadre conversation research material automatic generation system based on multi-agent collaboration and speech enhancement recognition, comprising: A speech enhancement module is configured to collect cadre conversation research audio data for preprocessing to obtain enhanced audio data. A speech recognition module is configured to perform frame-level feature extraction on the enhanced audio data, and after inputting the enhanced audio data into a speech base large model, perform voiceprint recognition to obtain a structured speech transcription text. A semantic optimization module is configured to perform sensitive information recognition and semantic error correction optimization on the structured speech transcription text, and then perform policy terminology standardization processing to obtain a text meeting language specifications and policy requirements. A multi-agent writing module is configured to match a template corresponding to the text meeting language specifications and policy requirements, extract core elements in the text meeting language specifications and policy requirements, and fill the core elements into the template to generate a cadre performance draft. A material review module is configured to perform semantic optimization and verification on the cadre performance draft to obtain a cadre performance material manuscript meeting organizational use specifications.

[0068] The division of modules in the embodiment of the application is illustrative, and is only a logical functional division. In actual implementation, another division mode can be used. In addition, the functional modules in each embodiment of the application can be integrated in one processor, or can be physically separated, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module.

[0069] The embodiment also provides a computer device, which comprises a processor and a memory for storing a computer program (the computer program in the embodiment comprises a computing component and an iteration component, and can perform model computing and model updating), the computer program comprises program instructions, and the processor is used for executing the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method flow or a corresponding function; the processor in the embodiment can be used for the operation of the method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition.

[0070] The embodiment also provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in the computer device, and is used for storing programs and data. It can be understood that the computer readable storage medium herein can include an internal storage medium in the computer device, and of course can also include an expansion storage medium supported by the computer device. The computer readable storage medium provides a storage space, and the storage space stores an operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition in the above embodiment.

[0071] The embodiment also provides a computer program product, which comprises a computer program, and when the computer program is executed by the processor, the corresponding steps of the method for automatically generating cadre conversation research materials based on multi-agent collaboration and voice enhancement recognition in the above embodiment are implemented.

[0072] Those skilled in the art will appreciate that embodiments of the application can be readily used as a method, apparatus, or computer program product. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0073] The application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for (an apparatus including one or more of the flowchart illustrations and / or block diagrams and their associated elements as described herein).

[0074] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for (an apparatus including one or more of the flowchart illustrations and / or block diagrams and their associated elements as described herein).

[0075] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for (an apparatus including one or more of the flowchart illustrations and / or block diagrams and their associated elements as described herein).

[0076] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing and illustrating, not limiting, the technical solutions of the present application. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and any modifications or replacements of the specific embodiments of the present application without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition, characterized in that, Includes the following steps: Audio data from interviews and surveys with cadres is collected and preprocessed to obtain enhanced audio data. The enhanced audio data is subjected to frame-level feature extraction and then input into the basic speech model for voiceprint recognition to obtain structured speech-to-text. After performing sensitive information identification and semantic error correction optimization on the structured speech-to-text, policy terminology standardization processing is carried out to obtain text that conforms to language norms and policy requirements; Match the template corresponding to the text that conforms to language norms and policy requirements, extract the core elements from the text that conforms to language norms and policy requirements, and fill them into the template to generate a draft of the cadre's actual performance. The initial draft of the cadre's actual performance was semantically optimized and verified to obtain a draft of the cadre's actual performance material that conforms to the organization's usage standards.

2. The method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition as described in claim 1, characterized in that, The step of preprocessing the collected audio data from interviews with cadres to obtain enhanced audio data specifically includes: After standardizing the format and downsampling the audio of the cadre interviews, a Mel spectrogram was generated by short-time Fourier transform and Mel filtering, and Gaussian noise perturbation was added. The Mel speech spectrogram with added Gaussian noise was processed using a speech enhancement model to obtain the repaired Mel speech spectrogram, which was then converted into a time-domain audio signal as the enhanced audio data. The speech enhancement model includes an encoder, a decoder, and an output layer. The encoder adopts a multi-layer cascaded ResUNet structure, with each layer consisting of a convolutional kernel, batch normalization, and LeakyReLU activation function, and residual connections are used. The decoder consists of multiple layers of transposed convolutions, with each layer fusing the skip connection features of the corresponding layer of the encoder; The output layer uses the Sigmoid function.

3. The method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition as described in claim 1, characterized in that, The steps of extracting frame-level features from the enhanced audio data, inputting them into a large-scale speech model, and then performing speaker recognition to obtain structured speech-to-text transcription specifically include: The Mel spectrum features of the enhanced audio data are extracted and input into the basic speech model. Then, the Cam++ model is used for speaker recognition to obtain structured speech-to-text. The structured speech-to-text includes: timestamp, speaker tag, speech content and dialect type. The overall loss function of the speech-based large model is composed of a weighted average of speech recognition CTC loss, cross-entropy loss, and InfoNCE loss.

4. The method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition according to claim 1, characterized in that, The steps of performing sensitive information identification and semantic error correction optimization on the structured speech-to-text, followed by policy terminology standardization processing to obtain text that conforms to language norms and policy requirements, specifically include: The BERT semantic similarity model is used to calculate the similarity between the text segment to be detected in the structured speech-to-text and the preset sensitive words. When the similarity is greater than or equal to the preset similarity threshold, the corresponding text to be detected is marked as ambiguous sensitive information and replaced with preset candidate words to obtain the structured speech-to-text after sensitive information recognition; otherwise, no marking or replacement is performed. The perplexity and semantic coherence score of the text segment to be detected in the structured speech-to-text after the sensitive information is identified are calculated. When the perplexity is greater than or equal to a preset perplexity threshold, the text segment to be detected is a semantically abnormal sentence and is corrected; otherwise, no correction is performed. When the semantic coherence score is less than a preset coherence score threshold, logical connecting sentences are added to the text segment to be detected; otherwise, no addition is performed. Finally, the structured speech-to-text optimized by semantic correction is obtained. Calculate the matching degree between the terms in the semantically corrected and optimized structured speech-to-text and the standard terms. When the matching degree is greater than or equal to a preset matching degree threshold, the term is replaced with the standard term; otherwise, no replacement is performed. Calculate the compliance score of the replaced standard terms in the semantically corrected and optimized structured speech-to-text. When the compliance score is greater than or equal to a preset compliance score threshold, a text that conforms to language norms and policy requirements is obtained; otherwise, the replaced standard terms are marked as unqualified terms, and this step is repeated for the unqualified terms.

5. The method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition according to claim 1, characterized in that, The steps of matching the template corresponding to the text that conforms to language norms and policy requirements, extracting the core elements from the text that conforms to language norms and policy requirements, and filling them into the template to generate a draft of the cadre's actual performance include: The Sentence-BERT model is used to convert the text that conforms to language norms and policy requirements into semantic vectors, and the cosine similarity between the semantic vectors and the semantic vectors of each template in the preset template library is calculated to select the template with the highest cosine similarity. The template with the highest cosine similarity is transformed into a slot-filling structure, and the core element slots are marked. The core elements in the text that conforms to language norms and policy requirements are extracted and filled into the core element slots. Then, the prompt chaining strategy is combined to generate paragraph-level content and perform perplexity verification to obtain the first draft of the cadre's actual performance.

6. The method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition according to claim 1, characterized in that, The steps involved in semantically optimizing and validating the initial draft of the cadre's performance data to obtain a document conforming to organizational usage standards include: By removing repetitive descriptions from the initial draft of the cadre's actual performance using a large language model and rewriting it, and then combining it with a preset policy knowledge graph and organizational material standardization standards, the structural consistency verification, key information coverage review, and policy terminology compliance matching are carried out in sequence to obtain a cadre's actual performance material draft that conforms to the organization's usage standards.

7. A system for automatically generating interview and research materials for cadres based on multi-agent collaboration and enhanced speech recognition, characterized in that, include: The speech enhancement module is used to collect audio data from cadre interviews and surveys, preprocess it, and obtain enhanced audio data. The speech recognition module is used to extract frame-level features from the enhanced audio data, input them into the speech basic model, and then perform voiceprint recognition to obtain structured speech-to-text. The semantic optimization module is used to perform sensitive information identification and semantic error correction optimization on the structured speech-to-text, and then perform policy terminology standardization processing to obtain text that conforms to language norms and policy requirements. The multi-agent writing module is used to match the template corresponding to the text that conforms to language norms and policy requirements, extract the core elements from the text that conforms to language norms and policy requirements and fill them into the template to generate a draft of the cadre's actual performance. The document review module is used to semantically optimize and verify the initial draft of the cadre's actual performance, so as to obtain a document of the cadre's actual performance that conforms to the organization's usage standards.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable in the processor. When the processor executes the computer program, it implements the steps of the method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition as described in any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for automatically generating cadre interview survey materials based on multi-agent collaboration and speech enhancement recognition as described in any one of claims 1 to 6.

Images