Voice identification method and system based on large model and MCP architecture

Abstract

The invention provides a voice identification method and system based on a large model and an MCP architecture, and belongs to the field of voice recognition, and the method comprises the steps: calling the large model to analyze a user request according to an RAG cue word, and obtaining key elements and condition constraints; obtaining an MCP tool set, calling the large model to perform task planning on each MCP tool in the MCP tool set according to the key elements to obtain a calling sequence, calling the MCP tools according to the calling sequence to analyze the input voice, and obtaining an analysis result; the analysis result is compared with the condition constraint through the large model, and when the analysis result meets the condition constraint, the large model is controlled to output a structure document containing the analysis result and the analysis process. According to the method, an identification tool does not need to be manually searched, and based on the large model and the MCP tool set, task requirements can be automatically disassembled, the MCP tool can be dynamically called, and a structured report can be automatically output, so that the identification efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and specifically to a speech identification method and system based on a large model and MCP architecture. Background Technology

[0002] In today's era of rapid digital information development, the demand for identity authentication and security verification is increasing daily. Among numerous identity recognition technologies, voiceprint identification stands out due to its unique advantages. Each person's vocal organ structure and vocal habits differ, making each person's voiceprint unique, just like a fingerprint. Utilizing this characteristic, voiceprint identification, through the analysis and processing of sound signals, can accurately determine the individual to whom the voice belongs, playing an irreplaceable role in multiple fields.

[0003] However, traditional voiceprint identification methods currently face numerous challenges, hindering their further development and widespread application. When faced with each case, traditional voiceprint identification requires examiners to manually break down the task meticulously. From initial screening of sound signals and extraction of feature parameters to comparative analysis with sample voiceprints, each step requires examiners to rely on their professional knowledge and experience. Because the entire process is cumbersome and reliant on manual labor, each case takes an extremely long time, typically 3-5 days to complete. Traditional voiceprint identification relies on expert rule engines, which operate based on pre-defined rules and fixed workflows. It's like a machine operating according to a predetermined program, only able to handle cases that conform to its preset rules. However, in actual voiceprint identification work, the identification needs are constantly changing.

[0004] In summary, traditional voiceprint identification faces two major technical challenges. First, the identification process is highly reliant on manual labor. From task breakdown to tool usage, each step requires manual completion by the identification personnel. This not only leads to low efficiency and excessively long processing times per case, making it difficult to meet the timeliness requirements of practical work, but also results in a lack of flexibility and adaptability in the fixed workflow of the expert rule engine used. This makes it unable to cope with increasingly complex and changing identification needs, severely impacting the accuracy and reliability of the identification results. Summary of the Invention

[0005] In view of this, it is necessary to provide a speech identification method and system based on a large model and MCP architecture to solve the technical problems of existing technologies that require manual task breakdown, manual tool calling, and fixed engine in voiceprint identification.

[0006] To address the aforementioned technical problems, in a first aspect, this invention provides a speech identification method based on a large model and MCP architecture, comprising: The large model is invoked to parse the user request based on the RAG prompt words, and key elements and conditional constraints are obtained. Obtain the MCP toolset, call the large model to perform task planning on each MCP tool in the MCP toolset according to the key elements to obtain the calling order, and call the MCP tools according to the calling order to parse the input speech and obtain the parsing result; The large model compares the analytical results with the conditional constraints. When the analytical results satisfy all the conditional constraints, the large model outputs a structured document containing the analytical results and the analysis process.

[0007] In one possible implementation, the invocation of the large model, before parsing the user request based on the RAG prompts to obtain key elements and conditional constraints, includes: Accept user input, and convert the input into text information when it is voice; The text information is vectorized to obtain the user request.

[0008] In one possible implementation, the large-scale invocation model parses the user request based on RAG prompts to obtain key elements and conditional constraints, including: Based on the user request, query the RAG knowledge base and construct the large model for inputting RAG prompt words; The large model is controlled by the RAG prompt words to parse the user request and extract key elements; The large model is invoked to parse the RAG prompt words and obtain conditional constraints.

[0009] In one possible implementation, the process of acquiring the MCP toolset, calling the large model to perform task planning on each MCP tool in the MCP toolset based on the key elements to obtain the calling order, calling the MCP tools to parse the input speech according to the calling order, and obtaining the parsing result includes: Based on the parsing results, determine whether the input speech meets the condition constraints; When the input speech does not meet the condition constraints, the input speech is marked and the remaining input speech is replaced.

[0010] In one possible implementation, comparing the analytical results with the conditional constraints using the large model includes: The large model is invoked to construct judgment prompts based on the user request, the parsing result, the MCP toolset, the conditional constraints, and the invocation result of the MCP tools. Based on the judgment prompts, it is determined whether the parsing result satisfies the conditional constraints.

[0011] In one possible implementation, after comparing the analytical result with the conditional constraints using the large model, the method further includes: When the analysis result does not meet the condition constraints, the large model outputs the MCP tool that needs to be called. After calling the output MCP tool to parse the speech and obtaining the parsing result, the step of comparing the parsing result with the conditional constraints through the large model is executed.

[0012] In one possible implementation, the MCP toolset includes, but is not limited to, tools for obtaining a list of evidence, voice quality analysis tools, voice speaker identification tools, voice separation tools, voice transcription and recognition tools, voice syllable comparison tools, statistical comparison tools, document generation tools, file attribute tools, case list tools, automatic comparison tools, text transcription tools, and tools for marking statistical information.

[0013] On the other hand, the present invention also provides a speech identification device based on a large model and MCP architecture, comprising: The condition acquisition module calls the large model to parse the user request based on the RAG prompt words and obtain key elements and condition constraints. The tool scheduling module obtains the MCP tool set, calls the large model to perform task planning on each MCP tool in the MCP tool set according to the key elements to obtain the calling order, and calls the MCP tools according to the calling order to parse the input speech and obtain the parsing result; The output judgment module compares the analysis results with the conditional constraints using the large model. When the analysis results satisfy all the conditional constraints, the module controls the large model to output a structured document containing the analysis results and the analysis process. In one possible implementation, the speech identification device based on the large model and MCP architecture further includes: The voice replacement module determines whether the input voice meets the condition constraints based on the parsing result. If the input voice does not meet the condition constraints, the input voice is marked and the remaining input voice is replaced.

[0014] In one possible implementation, the speech identification device based on the large model and MCP architecture further includes: The module is repeatedly invoked. When the parsing result does not meet the condition constraints, the large model is controlled to output the MCP tool that needs to be invoked. After the output MCP tool is invoked to parse the speech and obtain the parsing result, the step of comparing the parsing result with the condition constraints through the large model is executed.

[0015] The beneficial effects of the present invention are: The speech identification method based on the large model and MCP architecture provided by the present invention first calls the large model to parse the user request according to the RAG prompt words to obtain key elements and conditional constraints; The invention acquires the MCP toolset, calls a large model to plan the task order of each MCP tool in the toolset based on key elements, and calls the MCP tools to parse the input speech according to the calling order, obtaining the parsing results. The large model compares the parsing results with the conditional constraints. When the parsing results meet the conditional constraints, the large model outputs a structured document containing the parsing results and analysis process. This eliminates the need for manual identification tool searching; the large model automatically plans and calls the MCP tools, directly outputting the parsing results. Furthermore, because this invention is based on a large model and MCP framework, it can automatically decompose task requirements and dynamically call MCP tools, finally providing a structured report to improve identification efficiency. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic flowchart of an embodiment of the speech identification method based on a large model and MCP architecture provided by the present invention. Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of S101; Figure 3 For the present invention Figure 1 A schematic diagram of an embodiment of S102; Figure 4 For the present invention Figure 1 A schematic diagram of an embodiment of S103; Figure 5 This is a schematic diagram of an embodiment of the voice identity verification device provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] Before demonstrating the embodiments, the following terms will be explained.

[0021] MCP (Model Context Protocol) tool: It is the "universal adapter" in the field of AI. Through standardized communication protocols, it enables AI models to seamlessly call external resources and services, just like connecting a USB device. This breaks through the limitation that traditional AI can only process text and enables real-time interaction with databases, browsers, design tools and even the entire Internet.

[0022] RAG (Retrieval-Augmented Generation) knowledge base: It is an intelligent system that combines information retrieval technology with generative artificial intelligence models (such as large language models LLM). It combines external knowledge retrieval with the generation capabilities of large language models, and provides contextual support for generation through efficient retrieval, thereby improving the accuracy and timeliness of answers.

[0023] Large Model: An artificial intelligence model built on deep learning technology, with a huge number of parameters (usually in the billions to trillions) and a massive training data scale (covering massive amounts of text, images, speech and other multimodal data). By learning from large-scale data, it captures complex data patterns and semantic relationships, thereby having stronger generalization ability and general task processing ability.

[0024] This invention provides a speech identification method, device, electronic device, and storage medium based on a large model and MCP architecture, which are described below.

[0025] Figure 1 This is a schematic diagram of an embodiment of the speech identification method based on a large model and MCP architecture provided by the present invention, as shown below. Figure 1 As shown, speech identification methods based on large models and MCP architecture include: S101. Call the large model to parse the user request based on the RAG prompt words and obtain key elements and conditional constraints; S102. Obtain the MCP toolset, call the large model to perform task planning on each MCP tool in the MCP toolset according to the key elements to obtain the calling order, call the MCP tools according to the calling order to parse the input speech and obtain the parsing results; S103. By comparing the analytical results with the conditional constraints through the large model, when the analytical results meet all the conditional constraints, control the large model to output a structured document containing the analytical results and the analysis process.

[0026] It should be noted that: The RAG knowledge base, including forensic identification standards and voiceprint identification books, is retrieved based on user requests. RAG prompts are then constructed after querying the knowledge base, which stores a large amount of relevant knowledge. By querying this knowledge base, information relevant to the user's request is obtained, and combined with the user input, RAG prompts are constructed, providing richer context for the large model to parse the semantics. The prompts, constructed from the user input and RAG results, are then fed into the large model. The large model parses the user input and RAG prompts to obtain the conditional constraints and key elements.

[0027] In some embodiments of the present invention, the large model parses the core theme or main object of interest in the user request as key elements. For example, in the request "Find the two audio recordings most similar to the speaker in this case," the key elements might be "this case," "speaker," "most similar," "audio," and "two." Conditional constraints are the limitations, specific requirements, or preferences in the user request. For example, in the above request, conditional constraints might be "It is necessary to determine whether the evidence meets the criteria for identification," "It is necessary to determine whether the evidence and sample are comparable," "It is necessary to obtain results from more than 10 sets of different syllables for comparison," or "It is necessary to perform statistical analysis on the segments and record the results." By understanding natural language, the large model decomposes the user request into structured key information, enabling it to handle complex linguistic phenomena (such as implicit meanings and polysemous words), thus improving the accuracy of parsing.

[0028] It should be noted that the MCP toolset includes, but is not limited to, tools for obtaining evidence lists, voice quality analysis, voice speaker identification, voice separation, voice transcription and recognition, voice syllable comparison, statistical comparison, document generation, file attribute tools, case list tools, automatic comparison tools, text transcription tools, and tools for marking statistical information.

[0029] The following tools are included: **Exhibition List Acquisition Tool:** Input is empty; output is a list of audio files for the current case [{audio path, audio name, audio type}]. **Speech Quality Analysis Tool:** Input the audio path; return speech quality information {reverberation, cutoff, signal-to-noise ratio, result information}. **Speech Audience Determination Tool:** Input the audio path; return the number of people the current audio may contain. **Speech Separation Tool:** Input the audio path and the number of speakers; return a new audio file composed of the corresponding speaker's speech segments. **Speech Transcription and Recognition Tool:** Input the audio path; return the corresponding syllables, phonemes, phrases, and corresponding time slice information. **Speech Syllable Comparison Tool:** Input the time slice information of two corresponding audio files; return the similarity of the corresponding time slice information. **Statistical Comparison Tool:** Input the time slice information of two corresponding audio files; statistically analyze the frequency of identical syllables and formants, providing data such as deviation, mean, skewness, and kurtosis. **Document Generation Tool:** Input the document content information; generate the corresponding Word document. **File Attribute Tool:** Input the file path; return the file's corresponding hash value, etc. **Case List Tool:** Input the query conditions; output the corresponding case list. Automatic comparison: Input two audio paths and return corresponding likelihood similarity and other information. Transcription tool: Input an audio path and return the audio's text information. Statistical information tool: Input an audio path and obtain statistical information such as syllables, phonemes, and phrases. It should be noted that the selection of the MCP toolset can be adjusted according to actual needs, and will not be elaborated upon here.

[0030] The system obtains a list of available tools (MCP toolsets) through the MCP service to complete the task. Based on the results of the large model's analysis and the task requirements, it constructs prompts to invoke these tools. The large model then plans the task execution flow based on the constructed prompts and the MCP toolset, determining the invocation order and parameter settings for each tool. The corresponding MCP tools are invoked sequentially according to the task plan; for example, the voice quality analysis tool might be used to assess audio quality first, followed by the speech speaker identification tool to determine the number of speakers in the audio. Each tool returns its corresponding analysis results upon completion.

[0031] The model constructs judgment prompts based on user input, parsing results, and the MCP toolset. Based on these prompts, it determines whether the parsing results and input speech meet the user's needs and constraints. If the parsing results meet the user's requests and constraints, the large model can use it as the final output and may further generate structured documents or reports. If the parsing results do not meet the user's requests or constraints, the large model needs to take corresponding measures, which may include re-planning the tool call order, adjusting the parsing strategy, supplementing external knowledge or data, or interacting with the user to obtain more information. In some embodiments of the present invention, the input speech may include one or more components.

[0032] Based on the final task results, a structured document is generated. In the example of voice identification in a telecommunications fraud case, the output is "The two audios with the most similar speakers are: speaker 1 in audio 1 and speaker 1 in audio 2", and a document containing the analysis process and charts is automatically generated to provide users with a comprehensive and clear display of the task results.

[0033] The speech identification method based on a large model and MCP architecture provided in this invention first calls the large model to parse the user request based on RAG prompts to obtain key elements, and then calls the large model to parse the RAG prompts to obtain conditional constraints. Next, the large model performs task planning on the MCP toolset based on the key elements to obtain the calling order, and calls the MCP tools to parse the input speech according to the calling order to obtain the parsing results. When the input speech meets the user request and conditional constraints, a structured document containing the parsing results and analysis process is output. This eliminates the need for manual identification tool searching; the large model automatically plans and calls the MCP tools, directly outputting the parsing results. Furthermore, since this invention is based on a large model and MCP architecture, it can automatically decompose task requirements and dynamically call the MCP tools, finally outputting a structured report to improve identification efficiency.

[0034] In some embodiments of the present invention, before step S101, the following is included: It accepts user input and converts voice input into text information. The user request is obtained by vectorizing the text information.

[0035] It should be noted that the user input is: "Find the two audio recordings that most closely resemble the speakers in this case." This input may be provided directly in text form; if it is audio, it needs to be converted to text first. Converting the user-input text into vector form facilitates computer understanding and processing, and is beneficial for subsequent knowledge base queries and semantic analysis.

[0036] The large model can also incorporate error-tolerant mechanisms to post-process ASR results (such as spell correction) and improve text quality. Multi-turn dialogue support can also be added; if user input is incomplete, the large model can guide the user to supplement information (e.g., "You have not yet entered any speech.").

[0037] In some embodiments of the present invention, such as Figure 2 As shown, step S101 includes: S201. Based on the user's request, query the RAG knowledge base and construct a large-scale model for RAG prompt word input; S202. Based on the RAG prompt words, the large control model parses user requests and extracts key elements; S203. Call the large model to parse the RAG prompt words and obtain the conditional constraints.

[0038] RAG cue words are crucial for guiding large models to parse user requests. A good cue word should clearly state the task objective (parse key elements and constraints), provide clear examples to help the large model understand the output format, remain concise, and avoid introducing irrelevant information.

[0039] It should be noted that: when querying the RAG knowledge base according to user requests, the RAG knowledge base outputs RAG prompts, including "the sample needs to meet the identification conditions", "the sample and the evidence need to be comparable", "the test results need to be recorded, analyzed and statistically compared", and "the criteria for determining the identification opinion".

[0040] The system invokes a large-scale model to parse semantics. Based on the constructed RAG prompts, the large model performs in-depth analysis of user requests, extracting key elements. Simultaneously, the large model understands and parses relevant legal norms and constraints. It should be noted that the selection of constraints can be adjusted according to actual needs and is not limited to the four mentioned above; these will not be elaborated upon here.

[0041] This embodiment retrieves information related to the user request from the RAG knowledge base and constructs a large model for RAG prompt input. The large model parses the key elements in the user request and the conditional constraints in the RAG prompts, and automatically decomposes the task requirements through the large model.

[0042] In some embodiments of the present invention, such as Figure 3 As shown, step S104 includes: S301. Determine whether the input speech meets the condition constraints based on the parsing results; S302. When the input voice does not meet the condition constraints, mark the input voice and replace the remaining input voice.

[0043] In some embodiments of the present invention, the input speech to be processed and a list of input speech samples are obtained. For each input speech (1 to n), a speech quality analysis tool is scheduled for analysis. Based on the analysis results, the large model determines whether the input speech is suitable for identification. If not, the large model marks it and replaces the next file with the input speech. If suitable, the large model schedules a speech separation tool. For each input speech (1 to m), the speech separation tool is scheduled for separation processing. For each separated input speech (1 to m), a speech transcription and recognition tool is scheduled for transcription and recognition. Based on the transcription and recognition results, the large model determines whether there is comparability. If the input speech is not comparable, the large model marks it and replaces the next file with the input speech. If there is comparability, the large model schedules a syllable comparison tool. For each comparable input speech (1 to x), the syllable comparison tool is scheduled for comparison. For the comparison results of each input speech (1 to x), a statistical comparison tool is scheduled for statistical analysis. Finally, a document generation tool is scheduled to generate analysis results.

[0044] In some embodiments of the present invention, such as Figure 4 As shown, step S103 includes: S401. The large model is invoked to construct judgment prompts based on user requests, parsing results, MCP toolset, conditional constraints, and MCP tool invocation results. The judgment prompts are then used to determine whether the parsing results meet the conditional constraints. S402. When the parsing result does not meet the condition constraints, output the MCP tool that needs to be called. S403. After calling the output MCP tool to parse the speech and obtaining the parsing results, perform the step of comparing the parsing results with the conditional constraints through a large model.

[0045] It should be noted that: a judgment prompt word is constructed based on the user request, parsing results, MCP toolset, conditional constraints, and MCP tool call results. This prompt word is used by the large model to determine whether the task process has ended. If the large model analyzes whether the task process has ended, it will proceed to the next step. If the task has not ended, it will re-obtain available MCP tools, construct prompt words, perform task planning, and call MCP tools based on the current results.

[0046] If the parsing result does not meet the constraints, the large model identifies the violating fields or logic, outputs the specific MCP tool, automatically calls the MCP tool and iteratively verifies it, and inputs the parsing result returned by the MCP tool back into the S501 process, repeating until the constraints are met.

[0047] This embodiment automatically marks substandard speech by comparing the parsing results of the MCP tool with the conditional constraints, thus avoiding invalid calculations. When the parsing result does not meet the conditional constraints, the large model recommends the optimal tool combination calling order based on historical call data and current speech features, and continues to parse the input speech until the parsing result of the input speech meets all conditional constraints. Furthermore, it constructs judgment prompts based on the user request, parsing result, MCP tool set, conditional constraints, and MCP tool calling results, and uses these judgment prompts to determine whether the parsing result meets all conditional constraints.

[0048] Adaptive fault tolerance: Triggering backup data replacement or toolchain restart for tagged audio ensures the process is not interrupted by a single audio segment. For example, in a drug trafficking case, the system automatically detected a 0.8-second silent segment in a key recording (possibly due to editing) and immediately prompted for the complete recording to be supplemented, avoiding flawed evidence.

[0049] Compared to traditional manual task breakdown, this invention can automatically break down task requirements. Traditional voiceprint identification requires experts to manually complete more than 10 steps, including sample selection, speech separation, syllable annotation, and comparison calculations, with a single case processing cycle of 3-5 working days. Compared to traditional expert rule engines, this invention can automatically plan and dynamically load MCP tool combinations to adapt to changing identification needs. The table below shows a comparison between this invention and traditional methods.

[0050]

[0051] In summary, this solution integrates RAG knowledge enhancement, large-scale model conditional reasoning, and MCP toolchain scheduling to construct an automated voiceprint analysis method. This invention transforms judicial appraisal standards into executable constraints, ensuring that each step of the operation complies with legal standards. The large-scale model automatically schedules the MCP toolchain, compressing the process, and simultaneously records the calling parameters, parsing results, and large-scale model decision logic of all MCP tools, forming a complete logical analysis.

[0052] To better implement the speech identification method based on a large model and MCP architecture in this invention embodiment, based on the speech identification method based on a large model and MCP architecture, correspondingly, as follows: Figure 5 As shown, this embodiment of the invention also provides a speech identification device based on a large model and MCP architecture. The speech identification device 500 based on the large model and MCP architecture includes: The condition acquisition module 501 calls the large model to parse the user request based on the RAG prompt words and obtain key elements and condition constraints. The tool scheduling module 502 retrieves the MCP toolset, calls the large model to perform task planning on each MCP tool in the MCP toolset based on key elements to obtain the calling order, and calls the MCP tools to parse the input speech according to the calling order to obtain the parsing results; The output judgment module 503 compares the analysis results with the condition constraints through the large model. When the analysis results meet all the condition constraints, it controls the large model to output a structured document containing the analysis results and analysis process.

[0053] In some embodiments of the present invention, the speech identification device based on the large model and MCP architecture further includes: a speech replacement module, which determines whether the input speech meets the condition constraints based on the parsing results, and marks the input speech and replaces the remaining input speech when the input speech does not meet the condition constraints; The module is repeatedly called. When the parsing result does not meet the condition constraints, the large model is controlled to output the MCP tool that needs to be called. After the output MCP tool is called to parse the speech and obtain the parsing result, the large model is executed to compare the parsing result with the condition constraints.

[0054] The speech identity verification device 500 provided in the above embodiments can realize the technical solutions described in the above embodiments of the speech verification method based on the large model and MCP architecture. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the speech verification method based on the large model and MCP architecture, and will not be repeated here.

Claims

1. A speech identification method based on a large model and MCP architecture, characterized in that, include: The large model is invoked to parse the user request based on the RAG prompt words, and key elements and conditional constraints are obtained. Obtain the MCP toolset, call the large model to perform task planning on each MCP tool in the MCP toolset according to the key elements to obtain the calling order, and call the MCP tools according to the calling order to parse the input speech and obtain the parsing result; The large model compares the analytical results with the conditional constraints. When the analytical results satisfy all the conditional constraints, the large model outputs a structured document containing the analytical results and the analysis process.

2. The speech identification method based on large model and MCP architecture according to claim 1, characterized in that, Before the large model calls the user request based on the RAG prompts to obtain key elements and constraints, it includes: Accept user input, and convert the input into text information when it is voice; The text information is vectorized to obtain the user request.

3. The speech identification method based on large model and MCP architecture according to claim 1, characterized in that, The large-scale model parses the user request based on RAG prompts to obtain key elements and constraints, including: Based on the user request, query the RAG knowledge base and construct the large model for inputting RAG prompt words; The large model is controlled by the RAG prompt words to parse the user request and extract key elements; The large model is invoked to parse the RAG prompt words and obtain conditional constraints.

4. The speech identification method based on large model and MCP architecture according to claim 1, characterized in that, The process of acquiring the MCP toolset involves calling the large model to perform task planning on each MCP tool in the toolset based on the key elements to obtain the calling order, and then calling the MCP tools to parse the input speech according to the calling order. After obtaining the parsing results, the process includes: Based on the parsing results, determine whether the input speech meets the condition constraints; When the input speech does not meet the condition constraints, the input speech is marked and the remaining input speech is replaced.

5. The speech identification method based on a large model and MCP architecture according to claim 1, characterized in that, The step of comparing the analytical results with the conditional constraints using the large model includes: The large model is invoked to construct judgment prompts based on the user request, the parsing result, the MCP toolset, the conditional constraints, and the invocation result of the MCP tools. Based on the judgment prompts, it is determined whether the parsing result satisfies the conditional constraints.

6. The speech identification method based on large model and MCP architecture according to claim 5, characterized in that, After comparing the analytical results with the conditional constraints using the large model, the process further includes: When the analysis result does not meet the condition constraints, the large model outputs the MCP tool that needs to be called. After calling the output MCP tool to parse the speech and obtaining the parsing result, the step of comparing the parsing result with the conditional constraints through the large model is executed.

7. The speech identification method based on large model and MCP architecture according to claim 1, characterized in that, The MCP toolset includes, but is not limited to, tools for obtaining evidence lists, voice quality analysis, voice speaker identification, voice separation, voice transcription and recognition, voice syllable comparison, statistical comparison, document generation, file attribute tools, case list tools, automatic comparison tools, text transcription tools, and tools for marking statistical information.

8. A speech identification system based on a large model and MCP architecture, characterized in that, include: The condition acquisition module calls the large model to parse the user request based on the RAG prompt words and obtain key elements and condition constraints. The tool scheduling module obtains the MCP tool set, calls the large model to perform task planning on each MCP tool in the MCP tool set according to the key elements to obtain the calling order, and calls the MCP tools according to the calling order to parse the input speech and obtain the parsing result; The output judgment module compares the analysis result with the condition constraints through the large model. When the analysis result satisfies all the condition constraints, it controls the large model to output a structured document containing the analysis result and the analysis process.

9. The speech identification system based on a large model and MCP architecture according to claim 8, characterized in that, Also includes: The voice replacement module determines whether the input voice meets the condition constraints based on the parsing result. If the input voice does not meet the condition constraints, the input voice is marked and the remaining input voice is replaced.

10. The speech identification system based on a large model and MCP architecture according to claim 8, characterized in that, Also includes: The module is repeatedly invoked. When the parsing result does not meet the condition constraints, the large model is controlled to output the MCP tool that needs to be invoked. After the output MCP tool is invoked to parse the speech and obtain the parsing result, the step of comparing the parsing result with the condition constraints through the large model is executed.

Images