Speech recognition method and device, storage medium and computer equipment
By combining multi-dimensional confidence features and interaction scenario information in the speech recognition system and dynamically selecting error correction routing strategies, the problems of high misrecognition rate and large resource consumption in professional scenarios are solved, and efficient and accurate speech recognition is achieved.
Patent Information
- Application Number
- CN202610276016.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-06
- Publication Date
- 2026-04-10
AI Technical Summary
Existing speech recognition systems have a high false recognition rate in professional scenarios, and the deep model error correction strategies used to improve accuracy result in huge computational resource consumption. They cannot achieve a dynamic balance between accuracy requirements and resource constraints, leading to low efficiency.
By acquiring the speech to be recognized and the interaction scene information, and combining multi-dimensional confidence features and interaction context information, a three-level error correction routing strategy is dynamically selected, including direct output, expert database error correction, and large language model reconstruction path, and resources are called in a targeted manner for error correction processing.
While improving the accuracy of recognition in professional scenarios, it effectively controls system latency and computational overhead, thereby improving speech recognition efficiency.
Smart Images

Figure CN121838773A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of speech recognition technology, and can be applied to the fields of digital healthcare and finance. In particular, it relates to a speech recognition method and device, storage medium, and computer equipment. Background Technology
[0002] Speech recognition technology, as one of the core entry points for human-computer interaction, aims to automatically and accurately convert human speech signals into corresponding text information. With breakthroughs in artificial intelligence and deep learning technologies, the general accuracy of speech recognition has been significantly improved. Its applications have expanded from simple voice input to complex industry-wide intelligent solutions, becoming a key technology for improving efficiency, optimizing services, and driving innovation. In the digital healthcare field, doctors can generate electronic medical records and issue prescriptions through dictation, greatly improving diagnostic and treatment efficiency. In the financial insurance industry, claims adjusters can quickly record accident details via voice, and customer service representatives can obtain key policy information from customer voice recordings in real time to provide precise services. In the financial sector, account managers can conduct compliance interviews via voice, and traders can query market data using voice commands. These applications all require speech recognition systems not only to hear clearly but also to understand, especially in highly specialized business scenarios with dense terminology. The accuracy and reliability of recognition directly affect the smoothness of business processes, customer satisfaction, and even the correctness of business decisions.
[0003] Existing speech recognition systems designed for professional scenarios have high misrecognition rates for domain-specific terms, business abbreviations, and homophones in their general models. To ensure accuracy, the commonly adopted strategy of performing deep model error correction or full term matching on all recognition results leads to a significant increase in single-processing latency and huge computational resource consumption. It is impossible to achieve a dynamic balance between accuracy requirements and resource constraints, resulting in low speech recognition efficiency. Summary of the Invention
[0004] In view of this, the present invention provides a speech recognition method and apparatus, a storage medium, and a computer device, the main purpose of which is to solve the problem of low efficiency in existing speech recognition.
[0005] According to one aspect of the present invention, a speech recognition method is provided, comprising: Obtain the speech to be recognized and the interaction scenario information of the speech to be recognized, wherein the interaction scenario information includes interaction context information and interaction category; The speech to be recognized is identified to obtain the initial converted text and the multidimensional confidence features of the initial converted text; The interaction context information is used as the basis for scene matching, and together with the multi-dimensional confidence feature, the text confidence parameter of the initial converted text is calculated. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level route. The three-level route includes the direct output path, the expert database error correction path, and the text reconstruction path based on the large language model. The initial converted text is post-processed based on the target path to generate the target converted text of the speech to be recognized.
[0006] Furthermore, the step of using the interaction context information as a basis for scene matching and calculating the text credibility parameter of the initial converted text together with the multi-dimensional confidence features includes: Calculate scene relevance based on the initial converted text and the interaction context information; Based on the output confidence, keyword matching degree, text fluency and scene relevance in the multidimensional confidence features, the text credibility parameter is calculated; The step of determining the target path for performing the text correction task from the three-level routes based on the text credibility parameter, the interaction category, and the error correction routing strategy includes: A correction coefficient is determined based on the interaction category in the interaction scenario information, wherein the interaction category includes offline interaction and online interaction; The routing parameters are calculated based on the text credibility parameters and the correction coefficients. Based on the routing parameters and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level routes. The target path includes the direct output path, the expert database error correction path, and the text reconstruction path.
[0007] Furthermore, based on the routing parameters and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level routes, including: If the routing parameters are greater than a first preset threshold, the direct output path will be determined as the target path. If the routing parameters are less than or equal to the first preset threshold and greater than the second preset threshold, the expert database error correction path is determined as the target path. If the routing parameters are less than the second preset threshold, the text reconstruction path will be determined as the target path.
[0008] Furthermore, if the routing parameter is less than the second preset threshold, the method further includes: Obtain the real-time load of the text reconstruction path; If the real-time load is greater than or equal to a preset load threshold, and the deviation between the routing parameters and the second preset threshold is less than a preset deviation threshold, then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, then the text reconstruction path is determined as the target path.
[0009] Furthermore, the interaction scenario information also includes the business category, and based on the initial converted text, text correction is performed on the speech to be recognized based on the expert database correction path, including: The expert knowledge base matching the business category is retrieved as the target matching library, wherein the expert knowledge base includes a business-related product name library, a business-specific terminology library, and an abbreviation entity library; Each candidate in the initial converted text is matched with the pronunciation in the target matching library, and the candidate words are corrected based on the replacement words corresponding to the matched target pronunciations to obtain corrected candidate words; The target converted text is generated based on the modified candidate words and the candidate words that did not match the target word.
[0010] Furthermore, the text reconstruction path includes an expert database error correction path and a text reconstruction model based on a large language model; based on the initial converted text, the target converted text of the speech to be recognized is generated based on the text reconstruction path, including: Based on the error correction path of the expert database, the candidate words in the initial converted text are corrected to obtain corrected candidate words; The context information from the interaction scenario, the initial converted text, and the corrected candidate words are all used as prompt words and input into a lightweight large language model to reconstruct the target converted text.
[0011] Furthermore, after generating the target converted text of the speech to be recognized, the method further includes: Based on the business categories in the interaction scenario information, determine the target business model that matches the business category; Based on the target business big model, the target converted text is subjected to business intent recognition to obtain the business processing intent, and the business execution sequence is determined based on the business processing intent; The business execution agent is driven to perform business processing based on the business execution sequence.
[0012] According to another aspect of the present invention, a voice recognition device is provided, comprising: The acquisition module is used to acquire the speech to be recognized and the interaction scenario information of the speech to be recognized, wherein the interaction scenario information includes interaction context information and interaction category; The speech recognition module is used to recognize the speech to be recognized and obtain the initial converted text and the multidimensional confidence features of the initial converted text; The routing module is used to use the interaction context information as the basis for scene matching, and together with the multi-dimensional confidence features, calculate the text confidence parameter of the initial converted text. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, it determines the target path for performing the text error correction task from the three-level routes. The three-level routes include the direct output path, the expert library error correction path, and the text reconstruction path based on the large language model. The generation module is used to post-process the initial converted text based on the target path to generate the target converted text of the speech to be recognized.
[0013] Furthermore, the routing module includes: The first calculation unit is used to calculate the scene relevance based on the context information in the initial converted text and the interaction scene information; The second calculation unit is used to calculate the text credibility parameter based on the output confidence, keyword matching degree, text fluency and scene relevance in the multidimensional confidence features; The first determining unit is used to determine a correction coefficient based on the interaction category in the interaction scenario information, wherein the interaction category includes offline interaction and online interaction; The second determining unit is used to calculate routing parameters based on the text credibility parameters and the correction coefficients, and to determine the target path for performing the text correction task from the three-level routes based on the routing parameters and the error correction routing strategy. The target path includes the direct output path, the expert database error correction path, and the text reconstruction path.
[0014] Furthermore, in a specific application scenario, the second determining unit is specifically used to determine the direct output path as the target path when the routing parameters are greater than the first preset threshold. If the routing parameters are less than or equal to the first preset threshold and greater than the second preset threshold, the expert database error correction path is determined as the target path. If the routing parameters are less than the second preset threshold, the text reconstruction path will be determined as the target path.
[0015] Furthermore, the second determining unit is specifically used to: obtain the real-time load of the text reconstruction path when the routing parameter is less than the second preset threshold; If the real-time load is greater than or equal to a preset load threshold, and the deviation between the routing parameters and the second preset threshold is less than a preset deviation threshold, then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, then the text reconstruction path is determined as the target path.
[0016] Furthermore, the generation module includes: The retrieval unit is used to retrieve an expert knowledge base that matches the business category as the target matching base, wherein the expert knowledge base includes a business-related product name base, a business-specific terminology base, and an abbreviation entity base; The correction unit is used to match each candidate in the initial converted text with the pronunciation in the target matching library, and correct the candidate words based on the replacement words corresponding to the matched target pronunciations to obtain corrected candidate words; The generation unit is used to generate target converted text based on the modified candidate words and candidate words that do not match the target word.
[0017] Furthermore, the text reconstruction path includes an expert database error correction path and a text reconstruction model; The generation module further includes: correcting candidate words in the initial converted text based on the error correction path of the expert database to obtain corrected candidate words; The context information from the interaction scenario, the initial converted text, and the corrected candidate words are all used as prompt words and input into a lightweight large language model to reconstruct the target converted text.
[0018] Furthermore, the device also includes: The determination module is used to determine the target business model that matches the business category based on the business category in the interaction scenario information. The intent recognition module is used to perform business intent recognition on the target converted text based on the target business big model, obtain the business processing intent, and determine the business execution sequence based on the business processing intent; The business processing module is used to drive the business execution agent to perform business processing based on the business execution sequence.
[0019] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform an operation corresponding to the above-described speech recognition method.
[0020] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described speech recognition method.
[0021] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a speech recognition method, apparatus, storage medium, and computer device. First, it acquires the speech to be recognized and its interaction scenario information, including interaction context information and interaction category. Then, it recognizes the speech to be recognized to obtain initial converted text and multidimensional confidence features of the initial converted text. Using the interaction context information as a scenario matching basis, it calculates the text confidence parameter of the initial converted text together with the multidimensional confidence features. Based on the text confidence parameter, the interaction category, and an error correction routing strategy, it determines a target path for performing text error correction from a three-level routing system. The three-level routing system includes a direct output path, an expert database error correction path, and a text reconstruction path based on a large language model. Finally, it performs post-processing on the initial converted text based on the target path to generate the target converted text of the speech to be recognized. Compared with existing technologies, this method achieves intelligent dynamic routing of error correction strategies by integrating multi-dimensional confidence features and interaction scenario information. At the same time, it efficiently corrects errors by selectively calling different resources such as domain expert databases and large language model reconstruction. This improves the accuracy of professional scenario recognition while effectively controlling system latency and computational overhead, thereby significantly improving the efficiency of speech recognition.
[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a speech recognition method provided by an embodiment of the present invention is shown; Figure 2 A flowchart of another speech recognition method provided by an embodiment of the present invention is shown; Figure 3 This diagram illustrates a block diagram of a speech recognition device provided in an embodiment of the present invention. Figure 4 This diagram illustrates another speech recognition device provided in an embodiment of the present invention. Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation
[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0025] This invention provides a speech recognition method, such as... Figure 1 As shown, the method includes: 101. Obtain the speech to be recognized and the interaction scenario information of the speech to be recognized.
[0026] In this embodiment of the invention, the speech to be recognized can be a real-time acquired audio stream or a stored audio file. The interaction scenario information is metadata describing the background of this voice interaction, and may include, but is not limited to: the business category in which the interaction occurred, such as banking transactions, insurance claims, insurance appeals, online consultations; the interaction type, such as online real-time interaction or offline recording processing; the context text of the current dialogue, such as the content of previous rounds of dialogue; device information; and the ambient noise level. This information can be collected by the client and uploaded along with the voice data, or extracted from logs by the server based on the session ID.
[0027] 102. Recognize the speech to be recognized to obtain the initial converted text and the multidimensional confidence features of the initial converted text.
[0028] In this embodiment of the invention, basic automatic speech recognition models, such as end-to-end Deep Speech 2 and Conformer, are used to recognize the input audio and output preliminary text results, i.e., initial converted text. Simultaneously, the model outputs, or the system calculates, a set of multi-dimensional confidence features related to the recognition result. These features evaluate the credibility of the recognition from different dimensions. For example, output confidence: the probability score provided by the ASR model for each recognized word or character; keyword matching degree: whether there are scene-related key entity words in the recognized text and their matching degree; text fluency: the perplexity or fluency score of the recognized text under the language model. These features collectively constitute the quantitative basis for subsequent path decisions.
[0029] 103. Using the interaction context information as the basis for scene matching, and together with the multi-dimensional confidence features, calculate the text confidence parameter of the initial converted text, and determine the target path for performing the text correction task from the three-level routes based on the text confidence parameter, the interaction category, and the error correction routing strategy.
[0030] In this embodiment of the invention, the multidimensional confidence features (such as vectors or scalar groups) obtained in the above steps are comprehensively analyzed with interaction scenario information (such as interaction category labels and contextual text). The three-level routing includes preset text correction paths representing post-processing strategies with different processing intensities and resource consumption. Examples include direct output, rule-based or expert-based fast matching correction, and deep semantic reconstruction using a large language model. That is, the three-level routing includes a direct output path, an expert-based error correction path, and a text reconstruction path based on a large language model. A lightweight classifier or rule engine can be used to calculate which path to select based on the current input's multidimensional confidence and scenario combination; this selected path is the target path. Specifically, the confidence of the initial converted text can be determined from the scenario dimension based on the matching degree between the interaction context information and the initial converted text. Then, the scenario confidence and the confidence of other dimensions are used together to calculate the text credibility parameter of the initial converted text. Then, based on the text credibility parameter and interaction category, the subsequent text correction processing operations of the initial converted text are routed to the target path according to the pre-built error correction routing strategy, so as to perform error correction processing on the initial converted text based on the target path.
[0031] It's important to note that by incorporating interactive scenario information as a basis for routing decisions, the system can understand the context in which the speech occurs. For example, in specialized financial business scenarios, even if the general recognition confidence level is high, the system will tend to choose an error correction path that includes verification of professional terminology due to the sensitivity of the scenario. This prevents serious errors caused by misrecognition of professional terms in critical scenarios, thus consolidating accuracy at the application level. Furthermore, by analyzing multi-dimensional confidence features and interactive scenario information, "high-risk, low-confidence" recognition results can be accurately identified. Only these texts that truly require error correction undergo in-depth processing (such as expert database correction or text reconstruction), while high-confidence results are processed using lightweight or direct output. This on-demand allocation of error correction resources fundamentally avoids the omission of corrections for low-quality recognized text.
[0032] 104. Perform text correction processing on the initial converted text based on the target path, and generate the target converted text of the speech to be recognized based on the corrected initial converted text.
[0033] In this embodiment of the invention, after determining the target path, the initial converted text is input into the corresponding target path for processing. For example, if the target path is a direct output path, the initial converted text is directly used as the final result; if it is an expert database correction path, a terminology database matching the current business category is called for matching and replacement; if it is a text reconstruction path, a sequence-to-sequence model or a large language model may be used for rewriting to output an optimized, more accurate, and more semantically relevant target converted text.
[0034] It should be noted that the above process achieves intelligent and refined post-processing of speech recognition. It avoids the waste of computational resources and response latency caused by a single strong error correction model (when the recognition result itself has high credibility), and also avoids the accumulation of errors that may result from simple output (when the recognition result has low credibility). Through joint decision-making based on scenario and confidence level, the system can achieve the optimal balance between accuracy, efficiency, and resource consumption, significantly improving the practicality of the speech recognition system in real-world complex environments and the user experience.
[0035] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, the steps involve using the interaction context information as a basis for scene matching, calculating the text credibility parameter of the initial converted text together with the multi-dimensional confidence features, and determining the target path for performing the text correction task from the three-level routes based on the text credibility parameter, the interaction category, and the error correction routing strategy, including: 201. Calculate the scene relevance based on the context information in the initial converted text and the interaction scene information.
[0036] 202. Calculate the text credibility parameter based on the output confidence, keyword matching degree, text fluency and scene relevance in the multidimensional confidence features.
[0037] 203. Determine the correction coefficient based on the interaction category in the interaction scenario information.
[0038] 204. Calculate the routing parameters based on the text credibility parameters and the correction coefficients, and determine the target path for performing the text correction task from the three-level routes based on the routing parameters and the error correction routing strategy.
[0039] In this embodiment of the invention, the semantic coherence and contextual relevance of the initial converted text are evaluated not only based on the recognition results of single-sentence speech but also by utilizing contextual information from the dialogue, such as the question-and-answer content of the 1-2 rounds preceding the current speech. For example, a scene relevance score is obtained by calculating the cosine similarity between the current text and the context in the vector space, or by using a small language model to determine whether it constitutes a reasonable continuation of the dialogue. This score reflects the reasonableness of the current recognition result within the dialogue flow; a higher score indicates a better match with the context and potentially higher overall credibility.
[0040] After determining the contextual relevance, the underlying acoustics from speech recognition, output confidence (usually an average or weighted sum of word-level probabilities), the degree to which the recognized text hits business keywords (keyword matching), the degree to which the text conforms to grammatical conventions (text fluency, such as N-gram language model scores), and the contextual relevance calculated in the previous step are combined using a pre-defined weighted formula. For example, a linear weighted sum can be used: Confidence Parameter = α × Output Confidence + β × Keyword Matching + γ × Text Fluency + δ × Contextual Relevance. The weight coefficients (α, β, γ, δ) can be pre-trained or configured according to different business scenarios. The aforementioned confidence parameter is a comprehensive scalar; a higher value indicates a more reliable overall initial recognition result.
[0041] Interaction categories include offline and online interactions. Since offline and online interactions have different sensitivities to latency, online scenarios such as real-time voice assistants and online customer service have stringent real-time requirements and must prioritize low latency; while scenarios such as audio file transcription and non-real-time voice command processing are less sensitive to latency, allowing for more complex and time-consuming error correction models, but may require extremely high final accuracy. Therefore, the credibility parameter can be modified based on the interaction category, using different computing resources for different tolerances of response latency. Thus, a lower correction coefficient (e.g., <1) can be set for offline interactions, resulting in a relatively lower final routing parameter value, thereby increasing the tendency to choose error correction paths that require longer processing time but are more accurate (e.g., text reconstruction paths). For online interactions, a higher correction coefficient (e.g., ≥1) can be set, resulting in a relatively higher final routing parameter value, thereby guiding the system to choose faster paths (e.g., direct output or fast expert database matching), unless the recognition result has extremely low credibility.
[0042] Finally, routing parameters (which can be the product of the text credibility parameter and the correction coefficient) are calculated. These parameters characterize the quality of the current recognition result and the real-time constraints of the current scene. The routing parameters are compared with a preset path selection threshold, resulting in a three-level routing system that uniquely identifies the target path from the preset text correction paths. The target paths include the direct output path, the expert database correction path, and the text reconstruction path. Each path corresponds to different processing logic and resource overhead, thus achieving adaptive routing based on quantitative evaluation.
[0043] It should be noted that by introducing scene relevance calculation, the text credibility determination takes into account the coherence of the dialogue. This effectively detects recognition errors that appear highly confident in isolation but contradict the context. For example, if the preceding text asks for "transfer amount," it might be recognized as "war amount" in the following text. Furthermore, by using correction coefficients to differentiate between online and offline scenarios, it ensures that under stringent real-time requirements, the pursuit of ultimate accuracy will not trigger timeouts in deep processing, thus achieving a balance between accuracy and system responsiveness.
[0044] In one embodiment of the present invention, for further explanation and limitation, based on the routing parameters and the error correction routing strategy, a target path for performing the text error correction task is determined from the three-level routes, including: If the routing parameters are greater than a first preset threshold, the direct output path will be determined as the target path. If the routing parameters are less than or equal to the first preset threshold and greater than the second preset threshold, the expert database error correction path is determined as the target path. If the routing parameters are less than the second preset threshold, the text reconstruction path will be determined as the target path.
[0045] In this embodiment of the invention, the first preset threshold is greater than the second preset threshold. These two thresholds can be optimized and determined by analyzing the relationship between the distribution of routing parameters and the improvement in accuracy after error correction on the validation set. When the routing parameters are greater than the first preset threshold, it indicates that the overall credibility of the initial recognition result is very high, and the current interaction scenario, such as online real-time interaction, also supports rapid response. In this case, the initial converted text itself already has high accuracy and usability, and any additional error correction processing brings little marginal benefit, but instead introduces unnecessary processing delays and computational resource consumption. Therefore, direct output is used to maximize efficiency.
[0046] When the routing parameters fall between the first and second preset thresholds, it indicates that the initial identification results have a certain degree of unreliability, or that the scenario requires a certain level of verification. This may be due to inaccurate identification of individual technical terms or minor fluency issues. The expert database error correction path is a rule-based and domain-knowledge-based error correction method with moderate overhead. It can quickly and accurately match and replace high-frequency errors in specific domains, such as product names, terms, and abbreviations, with fast processing speed, thus effectively improving accuracy with controllable overhead.
[0047] When the routing parameters are less than or equal to the second preset threshold, it indicates that the initial recognition result is of poor quality, possibly containing serious semantic errors, a large number of irrelevant words, or completely incoherent sentences. In this case, local word substitution expert database correction is insufficient to solve the problem; a deep semantic understanding and reconstruction of the entire sentence is required. Text reconstruction paths typically involve more complex models (such as sequence-to-sequence error correction models or large language models) capable of utilizing broader semantic information and context for global rewriting.
[0048] In one embodiment of the present invention, for further explanation and limitation, when the routing parameter is less than the second preset threshold, the method further includes: Obtain the real-time load of the text reconstruction path; If the real-time load is greater than or equal to a preset load threshold, and the deviation between the routing parameters and the second preset threshold is less than a preset deviation threshold, then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, then the text reconstruction path is determined as the target path.
[0049] In this embodiment of the invention, when system resources are scarce, the system can intelligently balance the requirements for error correction intensity with real-time processing capabilities, avoiding service unavailability or overall latency deterioration due to overload. When routing parameters trigger the selection of a text reconstruction path, i.e., when the routing parameters are less than or equal to a second preset threshold, the system does not immediately execute the path. Instead, it acquires the real-time load, i.e., load indicators such as CPU utilization, GPU memory usage, request queue length, or average response time of the service cluster calling large language models or complex sequence-to-sequence models. If the real-time load is greater than or equal to a preset load threshold (this threshold is set according to the system service capacity, such as 80% utilization), and the deviation between the calculated routing parameters and the second preset threshold T2 is less than a preset deviation threshold (e.g., |routing parameters - T2| < ε), then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, or the deviation between the routing parameters and T2 is greater than or equal to the preset deviation threshold, i.e., the recognition result quality is very poor, far below T2, then regardless of the load, the text reconstruction path is determined as the target path.
[0050] It should be noted that by automatically downgrading some boundary reconstruction requests to expert database matching requests when overloaded, the dynamic fluctuations of system load in the actual production environment are fully considered, achieving a balance between accuracy and system security.
[0051] In one embodiment of the present invention, for further explanation and limitation, text correction of the speech to be recognized based on the expert database correction path includes: Retrieve the expert knowledge base that matches the business category as the target matching base; Each candidate in the initial converted text is matched with the pronunciation in the target matching library, and the candidate words are corrected based on the replacement words corresponding to the matched target pronunciations to obtain corrected candidate words; The target converted text is generated based on the modified candidate words and the candidate words that did not match the target word.
[0052] In this embodiment of the invention, multiple expert knowledge bases closely associated with different business categories, such as "banking," "stock trading," "medical consultation," and "flight booking," are pre-built. Specifically, the expert knowledge base includes a business-related product name library, a business-specific terminology library, and an abbreviation entity library. The business-related product name library contains the standard names of all products and services within the business domain. The business-specific terminology library contains professional terms, fixed collocations, and jargon for the domain, such as "netting," "market maker," and "contraindications." The abbreviation entity library contains commonly used abbreviations, shortened forms, and their corresponding full names, such as "ICBC" for "Industrial and Commercial Bank of China" and "CT" for "Computed Tomography." When expert library error correction is required, one or more corresponding libraries are retrieved from all knowledge bases based on the business category tag of the current voice interaction, serving as the target matching library for this error correction operation. This ensures a high degree of focus on error correction resources and avoids interference and inefficiency caused by cross-domain matching.
[0053] In the embodiments of the present invention, instead of directly performing exact matching of text strings, a fuzzy matching strategy based on pronunciation similarity is adopted to correct the problem of "semantically correct but word - using errors" caused by homophones and near - homophones. For example, misidentifying "China Merchants Bank" as "Zhao Shang Bank". Specifically, it includes: converting each candidate word in the initial conversion text (usually the result after word segmentation) and the standard word of each entry in the target matching library into a unified pronunciation representation sequence, such as a pinyin sequence, a phoneme sequence, or an acoustic feature vector. Calculate the similarity between the pronunciation representation of each candidate word and the pronunciation representation of the entries in the knowledge base, such as the edit distance and cosine similarity. For each candidate word, if its pronunciation similarity with a certain entry in the knowledge base exceeds a preset matching threshold, it is considered that the candidate word is a misidentified variant of this entry. Replace the candidate word in the original text with the standard text (replacement word) of the matched entry in the knowledge base to obtain the corrected candidate word. For example, if the recognized text is "three - fold line" or "mountain blocking danger", which has the closest pronunciation to "third - party liability insurance" and is in the insurance field, then replace the current word with "third - party liability insurance". After completing all matching and replacement attempts for candidate words, recombine the corrected candidate words that have been successfully matched and replaced and those candidate words that have not found sufficiently similar pronunciation matches in the expert library (these words remain unchanged) in the original word order to generate the final target conversion text. By directionally invoking the expert library according to the business category, the error - correction scope is highly relevant, reducing the risk of incorrect correction. Adopting fuzzy matching based on pronunciation can effectively capture and correct the most common homophone / near - homophone errors in speech recognition, which is difficult to solve by error - correction methods based on glyphs or simple word frequencies.
[0054] In an embodiment of the present invention, for further illustration and limitation, generating the target conversion text of the to - be - recognized speech includes: Correcting the candidate words in the initial conversion text based on the error - correction path of the expert library to obtain corrected candidate words; Taking the context information in the interaction scenario information, the initial conversion text, and the corrected candidate words as prompt words and inputting them into a lightweight large - language model to reconstruct and obtain the target conversion text.
[0055] In this embodiment of the invention, the text reconstruction path includes an expert-based error correction path and a text reconstruction model. When the routing decision points to the text reconstruction path, the system first calls the expert-based error correction module to preprocess the initial converted text. The execution logic of this step is consistent with that of the expert-based error correction path. After obtaining the correction candidate words, a structured prompt is constructed based on the context information, the initial converted text, and the correction candidate words, and then input into a lightweight large language model. The task instruction can be "Please output the most accurate and fluent corrected version based on the following dialogue context and potentially erroneous recognition text." After receiving this prompt, the lightweight large language model will comprehensively utilize its powerful semantic understanding, logical reasoning, and text generation capabilities to deeply integrate the input information. Finally, the model output is the deeply reconstructed target converted text. The expert database provides accuracy assurance and domain knowledge, while the large language model provides semantic generalization capabilities and contextual understanding. The collaboration between the two ensures that the final text reconstruction result is both accurate and reliable, as well as fluent and natural.
[0056] After post-processing the initial converted text based on the target path to generate the target converted text of the speech to be recognized, the method further includes: Based on the business categories in the interaction scenario information, determine the target business model that matches the business category; Based on the target business big model, the target converted text is subjected to business intent recognition to obtain the business processing intent, and the business execution sequence is determined based on the business processing intent; The business execution agent is driven to perform business processing based on the business execution sequence.
[0057] In this embodiment of the invention, after generating the target converted text, the matching target business model is invoked based on the business category in the interaction scenario information. Deep semantic analysis is then performed on the text to identify the precise business processing intent and its parameters, thereby generating an executable business execution sequence. Ultimately, this drives the business execution agent to automatically complete the entire business process from instruction to completion. Because the speech-recognized text undergoes accurate correction and reconstruction, more accurate intent parsing results can be obtained in subsequent business intent analysis, thus transforming it into correct and reliable actual business operations. This achieves loosely coupled integration with downstream vertical business systems.
[0058] This invention provides a speech recognition method. First, it acquires the speech to be recognized and its interaction scenario information, including interaction context information and interaction category. Then, it recognizes the speech to be recognized to obtain initial converted text and multidimensional confidence features of the initial converted text. Using the interaction context information as a scenario matching basis, it calculates the text confidence parameter of the initial converted text together with the multidimensional confidence features. Based on the text confidence parameter, the interaction category, and an error correction routing strategy, it determines a target path for performing text error correction from a three-level routing system. The three-level routing system includes a direct output path, an expert database error correction path, and a text reconstruction path based on a large language model. Finally, it performs post-processing on the initial converted text based on the target path to generate the target converted text of the speech to be recognized. Compared with existing technologies, this method achieves intelligent dynamic routing of error correction strategies by integrating multi-dimensional confidence features and interaction scenario information. At the same time, it efficiently corrects errors by selectively calling different resources such as domain expert databases and large language model reconstruction. This improves the accuracy of professional scenario recognition while effectively controlling system latency and computational overhead, thereby significantly improving the efficiency of speech recognition.
[0059] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this invention provides a speech recognition device, such as... Figure 3 As shown, the device includes: The acquisition module 31 is used to acquire the speech to be recognized and the interaction scene information of the speech to be recognized, wherein the interaction scene information includes interaction context information and interaction category; The speech recognition module 32 is used to recognize the speech to be recognized and obtain the initial converted text and the multidimensional confidence features of the initial converted text; The routing module 33 is used to use the interaction context information as the basis for scene matching, and together with the multi-dimensional confidence feature, calculate the text confidence parameter of the initial converted text. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, it determines the target path for performing the text error correction task from the three-level routes. The three-level routes include the direct output path, the expert library error correction path, and the text reconstruction path based on the large language model. The generation module 34 is used to post-process the initial converted text based on the target path to generate the target converted text of the speech to be recognized.
[0060] Furthermore, the routing module 33 includes: The first calculation unit is used to calculate the scene relevance based on the context information in the initial converted text and the interaction scene information; The second calculation unit is used to calculate the text credibility parameter based on the output confidence, keyword matching degree, text fluency and scene relevance in the multidimensional confidence features; The first determining unit is used to determine a correction coefficient based on the interaction category in the interaction scenario information, wherein the interaction category includes offline interaction and online interaction; The second determining unit is used to calculate routing parameters based on the text credibility parameters and the correction coefficients, and to determine the target path for performing the text correction task from the three-level routes based on the routing parameters and the error correction routing strategy. The target path includes the direct output path, the expert database error correction path, and the text reconstruction path.
[0061] Furthermore, in a specific application scenario, the second determining unit is specifically used to determine the direct output path as the target path when the routing parameters are greater than the first preset threshold. If the routing parameters are less than or equal to the first preset threshold and greater than the second preset threshold, the expert database error correction path is determined as the target path. If the routing parameters are less than the second preset threshold, the text reconstruction path will be determined as the target path.
[0062] Furthermore, the second determining unit is specifically used to: obtain the real-time load of the text reconstruction path when the routing parameter is less than the second preset threshold; If the real-time load is greater than or equal to a preset load threshold, and the deviation between the routing parameters and the second preset threshold is less than a preset deviation threshold, then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, then the text reconstruction path is determined as the target path.
[0063] Furthermore, the generation module 34 includes: The retrieval unit is used to retrieve an expert knowledge base that matches the business category as the target matching base, wherein the expert knowledge base includes a business-related product name base, a business-specific terminology base, and an abbreviation entity base; The correction unit is used to match each candidate in the initial converted text with the pronunciation in the target matching library, and correct the candidate words based on the replacement words corresponding to the matched target pronunciations to obtain corrected candidate words; The generation unit is used to generate target converted text based on the modified candidate words and candidate words that do not match the target word.
[0064] Furthermore, the text reconstruction path includes an expert database error correction path and a text reconstruction model; The generation module 34 further includes: The expert database correction unit is used to correct the candidate words in the initial converted text based on the expert database error correction path to obtain corrected candidate words; The reconstruction unit is used to input the context information in the interaction scenario information, the initial converted text, and the corrected candidate words as prompt words into a lightweight large language model to reconstruct the target converted text.
[0065] Furthermore, such as Figure 4 As shown, the device further includes: The determining module 35 is used to determine the target business model that matches the business category based on the business category in the interaction scenario information. The intent recognition module 36 is used to perform business intent recognition on the target converted text based on the target business big model, obtain the business processing intent, and determine the business execution sequence based on the business processing intent; The business processing module 37 is used to drive the business execution agent to perform business processing according to the business execution sequence.
[0066] This invention provides a speech recognition device. First, an acquisition module acquires the speech to be recognized and its interaction scenario information, including interaction context information and interaction category. Then, a speech recognition module recognizes the speech to be recognized, obtaining initial converted text and its multidimensional confidence features. Next, a routing module uses the interaction context information as a scenario matching basis, along with the multidimensional confidence features, to calculate the text confidence parameter of the initial converted text. Based on the text confidence parameter, the interaction category, and an error correction routing strategy, a target path for performing text error correction is determined from a three-level routing system. This three-level routing system includes a direct output path, an expert database error correction path, and a text reconstruction path based on a large language model. Finally, a generation module performs post-processing on the initial converted text based on the target path to generate the target converted text of the speech to be recognized. Compared with existing technologies, this method achieves intelligent dynamic routing of error correction strategies by integrating multi-dimensional confidence features and interaction scenario information. At the same time, it efficiently corrects errors by selectively calling different resources such as domain expert databases and large language model reconstruction. This improves the accuracy of professional scenario recognition while effectively controlling system latency and computational overhead, thereby significantly improving the efficiency of speech recognition.
[0067] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, the computer executable instruction being capable of executing the speech recognition method in any of the above method embodiments.
[0068] Figure 5The diagram illustrates a structural schematic of a computer device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.
[0069] like Figure 5 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.
[0070] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.
[0071] Communication interface 404 is used to communicate with other network elements such as clients or other servers.
[0072] The processor 402 is used to execute program 410, which can specifically execute the relevant steps in the above-described speech recognition method embodiment.
[0073] Specifically, program 410 may include program code that includes computer operation instructions.
[0074] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.
[0075] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.
[0076] Specifically, program 410 can be used to cause processor 402 to perform the following operations: Obtain the speech to be recognized and the interaction scenario information of the speech to be recognized, wherein the interaction scenario information includes interaction context information and interaction category; The speech to be recognized is identified to obtain the initial converted text and the multidimensional confidence features of the initial converted text; The interaction context information is used as the basis for scene matching, and together with the multi-dimensional confidence feature, the text confidence parameter of the initial converted text is calculated. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level route. The three-level route includes the direct output path, the expert database error correction path, and the text reconstruction path based on the large language model. The initial converted text is post-processed based on the target path to generate the target converted text of the speech to be recognized.
[0077] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.
[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A speech recognition method, characterized in that, include: Obtain the speech to be recognized and the interaction scenario information of the speech to be recognized, wherein the interaction scenario information includes interaction context information and interaction category; The speech to be recognized is identified to obtain the initial converted text and the multidimensional confidence features of the initial converted text; The interaction context information is used as the basis for scene matching, and together with the multi-dimensional confidence feature, the text confidence parameter of the initial converted text is calculated. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level route. The three-level route includes the direct output path, the expert database error correction path, and the text reconstruction path based on the large language model. The initial converted text is subjected to text correction processing based on the target path, so as to generate the target converted text of the speech to be recognized based on the corrected initial converted text.
2. The method according to claim 1, characterized in that, The step of using the interaction context information as a basis for scene matching and calculating the text credibility parameter of the initial converted text together with the multi-dimensional confidence features includes: Calculate scene relevance based on the initial converted text and the interaction context information; Based on the output confidence, keyword matching degree, text fluency and scene relevance in the multidimensional confidence features, the text credibility parameter is calculated; The step of determining the target path for performing the text correction task from the three-level routes based on the text credibility parameter, the interaction category, and the error correction routing strategy includes: Based on the interaction category in the interaction scenario information, a correction coefficient is determined, wherein the interaction category includes offline interaction and online interaction; Based on the text credibility parameter and the correction coefficient, routing parameters are calculated, and based on the routing parameters and the error correction routing strategy, the target path for performing the text error correction task is determined from the three-level routes.
3. The method according to claim 2, characterized in that, Based on the routing parameters and error correction routing policy, the target path for performing the text error correction task is determined from the three-level routes, including: If the routing parameters are greater than a first preset threshold, the direct output path will be determined as the target path. If the routing parameters are less than or equal to the first preset threshold and greater than the second preset threshold, the expert database error correction path is determined as the target path. If the routing parameters are less than the second preset threshold, the text reconstruction path will be determined as the target path.
4. The method according to claim 3, characterized in that, If the routing parameter is less than the second preset threshold, the method further includes: Obtain the real-time load of the text reconstruction path; If the real-time load is greater than or equal to a preset load threshold, and the deviation between the routing parameters and the second preset threshold is less than a preset deviation threshold, then the expert database error correction path is determined as the target path. If the real-time load is less than the preset load threshold, then the text reconstruction path is determined as the target path.
5. The method according to claim 1, characterized in that, The interaction scenario information also includes the business category. Based on the initial converted text, text correction is performed on the speech to be recognized based on the expert database correction path, including: The expert knowledge base matching the business category is retrieved as the target matching library, wherein the expert knowledge base includes a business-related product name library, a business-specific terminology library, and an abbreviation entity library; Each candidate in the initial converted text is matched with the pronunciation in the target matching library, and the candidate words are corrected based on the replacement words corresponding to the matched target pronunciations to obtain corrected candidate words; The target converted text is generated based on the modified candidate words and the candidate words that did not match the target word.
6. The method according to claim 5, characterized in that, The text reconstruction path based on the large language model includes an expert database error correction path and a text reconstruction model based on the large language model. Based on the initial converted text, the target converted text of the speech to be recognized is generated according to the text reconstruction path, including: Based on the error correction path of the expert database, the candidate words in the initial converted text are corrected to obtain corrected candidate words; The context information from the interaction scenario, the initial converted text, and the corrected candidate words are all used as prompt words and input into a lightweight large language model to reconstruct the target converted text.
7. The method according to claim 1, characterized in that, After generating the target converted text of the speech to be recognized, the method further includes: Based on the business categories in the interaction scenario information, determine the target business model that matches the business category; Based on the target business big model, the target converted text is subjected to business intent recognition to obtain the business processing intent, and the business execution sequence is determined based on the business processing intent; The business execution agent is driven to perform business processing based on the business execution sequence.
8. A voice recognition device, characterized in that, include: The acquisition module is used to acquire the speech to be recognized and the interaction scenario information of the speech to be recognized, wherein the interaction scenario information includes interaction context information and interaction category; The speech recognition module is used to recognize the speech to be recognized and obtain the initial converted text and the multidimensional confidence features of the initial converted text; The routing module is used to use the interaction context information as the basis for scene matching, and together with the multi-dimensional confidence features, calculate the text confidence parameter of the initial converted text. Based on the text confidence parameter, the interaction category, and the error correction routing strategy, it determines the target path for performing the text error correction task from the three-level routes. The three-level routes include the direct output path, the expert library error correction path, and the text reconstruction path based on the large language model. The generation module is used to perform text correction processing on the initial converted text based on the target path, so as to generate the target converted text of the speech to be recognized based on the corrected initial converted text.
9. A storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the speech recognition method as described in any one of claims 1-7.
10. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the speech recognition method as described in any one of claims 1-7.
