Voice transfer text error correction method and device, storage medium and computer equipment

Abstract

According to the voice transliteration text error correction method and device, the storage medium and the computer equipment, after the voice transliteration text is received in real time, the text is subjected to primary error detection, and the first error confidence coefficient is obtained; when the first error confidence coefficient is smaller than a fast error correction threshold value, the voice transcription text is directly output, and unnecessary error correction is avoided; otherwise, performing word-level rapid error correction on the voice transcription text to obtain a rapid error correction text, and performing secondary error detection on the rapid error correction text to obtain a second error confidence coefficient. Judging whether the second error confidence is smaller than a depth error correction threshold value or not; if yes, the rapid error correction text is directly output, and if not, semantic-level deep error correction is conducted on the rapid error correction text through a deep error correction model, and a deep error correction text is obtained and then output. Through a hierarchical error correction processing strategy, the method can be adapted to different service scenes, and meanwhile, dynamic balance between quality and efficiency of different service scenes can be realized by adjusting two error correction thresholds.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech recognition, and in particular to a speech transcription text error correction method and device, a storage medium and a computer device. BACKGROUND

[0002] With the deep expansion of human-computer interaction and human-human interaction scenarios, automatic speech recognition technology has become the core infrastructure in the fields of intelligent terminals, online meetings, social entertainment, etc. However, in complex acoustic scenarios such as multi-person online voice rooms, background noise interference, multi-person voice superposition, dialect accent differences, different speech speeds, and colloquial expressions will all lead to problems such as homonyms misuse, semantic rupture, and misspelled words in speech transcription text, directly affecting user communication efficiency and experience.

[0003] In order to improve the accuracy of speech transcription, post-processing error correction technology is generally used, which mainly includes rule-driven and model-driven solutions. Among them, rule correction quickly corrects specific error patterns through pre-set grammar, vocabulary library and other hard rules, such as directly replacing common misspelled words. Its advantage is fast processing speed and low resource consumption; model correction relies on deep learning models to achieve more flexible error correction through context semantic analysis, such as identifying and correcting homonym confusion caused by accent.

[0004] In summary, the rule correction method has limited coverage and is difficult to cope with unforeseen error types, and has insufficient generalization ability; although the model correction has significant effect, it has high reasoning delay and large computing power consumption, which is easy to cause lag in real-time speech stream scenarios, and the large-scale deployment cost is high. Therefore, the existing speech transcription text error correction method has great limitations, and it is difficult to balance between quality and efficiency, making it difficult for a single solution to adapt to different demand application scenarios, such as the strict requirement of low latency in entertainment scenarios or the absolute demand for high accuracy in conference scenarios. SUMMARY

[0005] The present application aims to at least solve one of the above technical defects, in particular the technical defect that the existing speech transcription text error correction method has great limitations and it is difficult to balance between quality and efficiency, making it difficult for a single solution to adapt to different demand application scenarios.

[0006] The present application provides a speech transcription text error correction method, which comprises:

[0007] Real-time receiving of speech transcription text, and adopting a multi-modal detection strategy to perform initial error detection on the speech transcription text to obtain a first error confidence;

[0008] if the first error confidence is not less than a fast correction threshold, performing word-level fast correction on the speech transcription text based on an AC automaton algorithm to obtain a fast correction text;

[0009] performing secondary error detection on the fast correction text by using the multi-modal detection strategy to obtain a second error confidence;

[0010] if the second error confidence is not less than a deep correction threshold, performing semantic-level deep correction on the fast correction text by using a preset deep correction model to obtain a deep correction text;

[0011] when the first error confidence is less than the fast correction threshold, the second error confidence is less than the deep correction threshold, or the deep correction model outputs the deep correction text, outputting the corresponding speech transcription text, fast correction text or deep correction text as a final transcription text.

[0012] Optionally, the first error confidence is obtained by performing initial error detection on the speech transcription text by using a multi-modal detection strategy, and the multi-modal detection strategy comprises:

[0013] performing error analysis on the speech transcription text based on word-level features and pinyin features to obtain an error probability value, and determining a fluency perplexity of the speech transcription text by using a preset statistical language model;

[0014] obtaining speech intelligibility and environmental noise features corresponding to the speech transcription text from a speech conversion engine, and determining a noise compensation factor of the speech transcription text according to the environmental noise features;

[0015] performing weighted fusion on the error probability value, the fluency perplexity, the speech intelligibility and the noise compensation factor to obtain the first error confidence.

[0016] Optionally, the fast correction text is obtained by performing word-level fast correction on the speech transcription text based on the AC automaton algorithm, and the AC automaton algorithm comprises:

[0017] obtaining a character index tree of a business type corresponding to the speech transcription text; each node in the character index tree represents an error word pattern and is provided with a failure pointer;

[0018] inputting each character in the speech transcription text into the character index tree in sequence for node-by-node matching, and marking the corresponding character in the speech transcription text as a candidate word group when a node representing an error word pattern is completely matched;

[0019] determining a corrected word group corresponding to the candidate word group according to a candidate correction rule corresponding to the character index tree, and determining a correction score of the corrected word group in the speech transcription text.

[0020] replace the corresponding candidate word group in the speech transcription text with the correction word group whose correction score is greater than the preset correction threshold, to obtain the fast correction text.

[0021] Optionally, the determining of the correction word group corresponding to the candidate word group according to the candidate correction rule corresponding to the character index tree comprises:

[0022] performing word group matching on the candidate word group from the candidate correction rule corresponding to the character index tree, to obtain a matching result;

[0023] determining whether the number of correction word groups in the matching result is one;

[0024] if yes, taking the correction word group in the matching result as the final correction word group of the candidate word group;

[0025] if no, calculating an overlapping interval score of each correction word group in the matching result, and taking the correction word group with the maximum overlapping interval score as the final correction word group of the candidate word group.

[0026] Optionally, the determining of the correction score of the correction word group in the speech transcription text comprises:

[0027] obtaining a global trigger frequency of the correction word group in a historical correction record, and a priority rule of a business type corresponding to the speech transcription text, and determining a priority of a word group type corresponding to the correction word group in the priority rule;

[0028] calculating a pinyin similarity between the correction word group and the candidate word group, and determining a fluency improvement value of the correction word group in the speech transcription text;

[0029] performing weighted fusion on the global trigger frequency, the priority, the pinyin similarity, and the fluency improvement value, to obtain the correction score.

[0030] Optionally, the performing of the semantic-level deep correction on the fast correction text by using the preset deep correction model to obtain a deep correction text comprises:

[0031] determining a deep correction model; the deep correction model is obtained by training a preset pre-training model using a Logit Bias constraint;

[0032] inputting the fast correction text into the deep correction model, so that the deep correction model performs context analysis and error word group correction on the fast correction text, and outputs a deep correction text.

[0033] Optionally, the method further comprises:

[0034] obtaining a correction word pair generated by the deep correction model in a semantic-level deep correction process and historical correction data of the correction word pair; the correction word pair is composed of word groups before and after correction;

[0035] based on the historical correction data, performing context analysis on the correction word pair by using Embedding clustering to obtain an analysis result, and generating a rule item of the correction word pair when the analysis result is strongly related;

[0036] pushing the rule item to a manual review queue for manual review, and performing node update on the character index tree based on the rule item when the manual review result is passed.

[0037] The application also provides a transcription text correction device, comprising:

[0038] a first detection module configured to receive a voice transcription text in real time and perform initial error detection on the voice transcription text by using a multi-modal detection strategy to obtain a first error confidence;

[0039] a fast correction module configured to perform word-level fast correction on the voice transcription text based on an AC automatic machine algorithm if the first error confidence is not less than a fast correction threshold to obtain a fast correction text;

[0040] a second detection module configured to perform secondary error detection on the fast correction text by using the multi-modal detection strategy to obtain a second error confidence;

[0041] a deep correction module configured to perform semantic-level deep correction on the fast correction text by using a preset deep correction model if the second error confidence is not less than a deep correction threshold to obtain a deep correction text;

[0042] a text output module configured to output corresponding voice transcription text, fast correction text or deep correction text as a final transcription text when the first error confidence is less than the fast correction threshold, the second error confidence is less than the deep correction threshold or the deep correction model outputs the deep correction text.

[0043] The application also provides a storage medium having computer readable instructions stored therein, wherein the computer readable instructions, when executed by one or more processors, cause the one or more processors to perform the steps of the voice transcription text correction method according to any one of the above embodiments.

[0044] The application also provides a computer device, comprising one or more processors and a memory.

[0045] The memory stores computer readable instructions that, when executed by the one or more processors, perform the steps of the voice-to-transcribed text error correction method as described in any of the above embodiments.

[0046] From the above technical solutions, the embodiments of the present application have the following advantages:

[0047] The voice-to-transcribed text error correction method, device, storage medium and computer device provided by the present application can detect errors in the transcribed text by using a multi-modal detection strategy after receiving the transcribed text in real time, obtain a first error confidence, and provide a dynamic basis for the selection of subsequent error correction strategies. For example, when the first error confidence is less than a fast error correction threshold, the transcribed text can be directly output as the final transcribed text, avoiding unnecessary error correction. Otherwise, the transcribed text is subjected to word-level fast error correction based on an AC automaton algorithm to obtain a fast error correction text, and the fast error correction text is subjected to error detection again to obtain a second error confidence. Then, it can be determined whether the second error confidence is less than a deep error correction threshold. If yes, the fast error correction text is output as the final transcribed text, and if no, the fast error correction text is subjected to semantic-level deep error correction by a pre-set deep error correction model to obtain a deep error correction text, which is output as the final transcribed text. By using the hierarchical error correction processing strategy, the present application can adapt to different business scenarios, and by adjusting the fast error correction threshold and the deep error correction threshold, a dynamic balance between quality and efficiency can be achieved in different business scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1 A flowchart of a voice-to-transcribed text error correction method provided by an embodiment of the present application;

[0050] Figure 2 A flowchart of a text error detection process provided by an embodiment of the present application;

[0051] Figure 3 A structural diagram of a voice-to-transcribed text error correction device provided by an embodiment of the present application;

[0052] Figure 4 An internal structure diagram of a computer device provided by an embodiment of the present application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0054] Rule-based error correction methods have limited coverage, struggle to handle unforeseen error types, and lack generalization ability. While model-based error correction is effective, it suffers from high inference latency and computational costs, easily causing stuttering in real-time audio streaming scenarios, and is prohibitively expensive for large-scale deployment. Therefore, existing speech-to-text error correction methods have significant limitations, making it difficult to strike a balance between quality and efficiency. This makes it difficult for a single solution to adapt to different application scenarios, such as the stringent low-latency requirements of entertainment scenarios or the absolute need for high accuracy in conferencing scenarios.

[0055] Based on this, this application proposes the following technical solution, as detailed below:

[0056] In one embodiment, such as Figure 1 As shown, Figure 1 This application provides a flowchart illustrating a speech-to-text error correction method according to an embodiment of the present application; the present application also provides a speech-to-text error correction method, specifically including the following:

[0057] S110: Receives speech-to-text in real time and uses a multimodal detection strategy to perform initial error detection on the speech-to-text to obtain the first error confidence level.

[0058] In this step, after receiving the speech-to-text in real time, the computer device can use a multimodal detection strategy to detect errors in the speech-to-text, thereby quantifying the amount of error in the transcribed text, obtaining the first error confidence level, and providing a dynamic basis for the selection of subsequent error correction strategies.

[0059] Among them, the multimodal detection strategy refers to a detection mechanism that integrates data from different information modalities or feature spaces to perform joint error analysis on speech-to-text results; for example, combining character-level features and pinyin features to detect errors in speech-to-text. Error confidence refers to a quantitative evaluation index generated after error detection of speech-to-text, used to characterize the probability and severity of errors in the current text.

[0060] Understandably, by employing a multimodal detection strategy, this application can cover multiple detection types for comprehensive error detection, and calibrate the error detection results through cross-modal feature consistency analysis, thereby generating an error confidence level to characterize the reliability of the current detected text, avoiding misjudgments caused by using a single rule or fixed threshold, and providing a precise basis for the dynamic selection and adaptive adjustment of subsequent error correction strategies.

[0061] S120: If the confidence level of the first error is not less than the fast error correction threshold, then the speech-to-text is fast corrected at the word level based on the AC automaton algorithm to obtain the fast error-corrected text.

[0062] In this step, after determining the first error confidence level of the speech-to-text in step S110, the computer device can perform word-level fast error correction on the speech-to-text based on the AC automaton algorithm when the first error confidence level is not less than the fast error correction threshold, thereby obtaining fast error-corrected text and improving text correction efficiency.

[0063] Among them, the AC automaton algorithm refers to a multi-pattern string matching algorithm based on a finite state machine. This algorithm constructs a Trie tree from multiple words or strings to be matched, and introduces failure pointers to form an automaton structure. Therefore, it can support high-speed scanning and replacement of transcribed text at the word level, realize the rapid identification and correction of common misspellings and fixed error patterns, and complete basic error correction tasks while ensuring extremely low processing latency, thus meeting the requirements of high-efficiency processing for real-time audio streams.

[0064] Specifically, after detecting the first error confidence level of the speech-to-text, the computer device can compare this first error confidence level with a pre-set fast error correction threshold. When the first error confidence level is not less than the fast error correction threshold, it indicates that the erroneous data in the speech-to-text exceeds the acceptable range for the current business scenario. Therefore, the computer device needs to trigger a word-level fast error correction process based on the AC automaton algorithm. This process efficiently matches the speech-to-text with a pre-built Trie tree, simultaneously completing multi-mode error recognition and replacement operations during a single linear scan, and then outputting the fast-corrected text.

[0065] Understandably, in order to meet the need for low latency, computer devices can prioritize a low-computational-cost word-level fast error correction process when speech-to-text needs error correction, which can significantly shorten the error correction processing path and response time, and avoid unnecessary complex model inference.

[0066] S130: A multimodal detection strategy is used to perform secondary error detection on the fast error-corrected text to obtain a second error confidence level.

[0067] In this step, after obtaining the fast error correction text through fast error correction in step S120, the computer device can perform secondary error detection on the fast error correction text to obtain a second error confidence, so as to evaluate the quality and verify the effect of the fast error correction result.

[0068] It can be understood that, through secondary error detection, the application can quantify the error correction benefit after completing one low-cost error correction, thereby providing an objective basis for whether to continue triggering a higher complexity error correction strategy, avoiding blind entry into a deep error correction phase with high algorithm consumption, thereby effectively controlling the overall processing time delay and resource consumption while ensuring continuous improvement of the transcription quality.

[0069] It should be noted that the secondary error detection process of the fast error correction text is completely consistent with the primary error detection process of the speech transcription text, and will not be described here.

[0070] S140: If the second error confidence is not less than a deep error correction threshold, performing semantic-level deep error correction on the fast error correction text through a preset deep error correction model to obtain a deep error correction text.

[0071] In this step, after determining the second error confidence of the fast error correction text in step S130, the computer device can perform semantic-level deep error correction on the fast error correction text through a preset deep error correction model when the second error confidence is not less than a fast error correction threshold, thereby obtaining a deep error correction text and improving the error correction quality of the text.

[0072] Specifically, after detecting the second error confidence corresponding to the fast error correction text, the computer device can compare the second error confidence with a pre-set deep error correction threshold. When the second error confidence is not less than the deep error correction threshold, it indicates that the error correction quality of the fast error correction text still does not meet the business requirements of the current business scenario. Therefore, the computer device needs to obtain a deep error correction model trained in advance to perform semantic-level deep error correction processing on the fast error correction text. The deep error correction model can model the overall context semantic relationship, syntactic structure, and cross-sentence dependency information, identify and correct complex errors introduced by homonym confusion, abnormal syntax, semantic missing, or expression ambiguity, thereby generating a corresponding deep error correction text.

[0073] It can be understood that, under the premise that the pre-set fast error correction has eliminated a large number of basic errors, the application uses a high-precision error correction model, which can fully utilize the advantages of the deep error correction model in complex semantic understanding, and avoid redundant calculation overhead in a low-error scenario, thereby effectively controlling the time delay and algorithm consumption of the overall error correction process while significantly improving the error correction quality, and achieving a more reasonable dynamic balance between efficiency and quality in the speech transcription text error correction process.

[0074] S150: When the confidence level of the first error is less than the fast error correction threshold, the confidence level of the second error is less than the deep error correction threshold, or the deep error correction model outputs deep error correction text, the corresponding speech-to-text, fast error correction text, or deep error correction text will be output as the final transcribed text.

[0075] In this step, during the hierarchical processing of the above steps, if the first error confidence is less than the fast error correction threshold, the second error confidence is less than the deep error correction threshold, or the deep error correction model outputs deep error-corrected text, then the computer device can output the corresponding speech-to-text, fast error-corrected text, or deep error-corrected text as the final transcribed text.

[0076] Specifically, when the confidence level of the first error is less than the fast error correction threshold, it indicates that there are few or no errors in the speech-to-text, which does not affect the application in the current business scenario. Therefore, the computer device can directly use the speech-to-text as the final transcribed text output, thereby avoiding unnecessary error correction calculations and improving processing efficiency. Similarly, when the confidence level of the second error is less than the deep error correction threshold, it indicates that the text after fast error correction has basically eliminated regularizable and high-frequency errors, meeting the application requirements of the current business scenario. Therefore, the computer device can use the fast-corrected text as the final transcribed text output, achieving a balance between low latency and high accuracy. In addition, if the deep error correction model successfully generates deep-corrected text, the computer device can use the deep-corrected text as the final transcribed text output, thereby further improving the semantic accuracy and overall readability of the transcribed text.

[0077] In the above embodiments, after receiving the speech-to-text in real time, in order to quantify the error rate of the transcribed text, this application can employ a multimodal detection strategy to detect errors in the speech-to-text and obtain a first error confidence level, providing a dynamic basis for the selection of subsequent error correction strategies. For example, when the first error confidence level is less than the fast error correction threshold, the speech-to-text can be directly output as the final transcribed text, avoiding unnecessary error correction; otherwise, word-level fast error correction is performed on the speech-to-text based on the AC automata algorithm to obtain fast-corrected text, and error detection is performed again on the fast-corrected text to obtain a second error confidence level. Then, it can be determined whether the second error confidence level is less than the deep error correction threshold; if so, the fast-corrected text is output as the final transcribed text; if not, semantic-level deep error correction is performed on the fast-corrected text using a preset deep error correction model to obtain the deep-corrected text, which is then output as the final transcribed text. Through the hierarchical error correction processing strategy, this application can adapt to different business scenarios, and can also achieve a dynamic balance between quality and efficiency in different business scenarios by adjusting the fast error correction threshold and the deep error correction threshold.

[0078] In one embodiment, such as Figure 2 As shown,Figure 2 A flowchart of a text error detection process provided for an embodiment of the present application is shown in FIG. 1. Figure 2 In the embodiment, the process of performing the initial error detection on the speech transcription text in step S110 using the multi-modal detection strategy can include:

[0079] S111: performing error analysis on the speech transcription text based on the word-level features and the pinyin features to obtain error probability values, and determining the fluency perplexity of the speech transcription text through a preset statistical language model.

[0080] S112: obtaining the speech intelligibility and the environmental noise features corresponding to the speech transcription text from the speech conversion engine, and determining the noise compensation factor of the speech transcription text according to the environmental noise features.

[0081] S113: performing weighted fusion on the error probability values, the fluency perplexity, the speech intelligibility, and the noise compensation factor to obtain the first error confidence.

[0082] In the embodiment, when detecting text errors, the computer device can perform error analysis on the speech transcription text based on the word-level features and the pinyin features to obtain error probability values, and determine the fluency perplexity of the speech transcription text through a preset statistical language model. Meanwhile, the computer device can obtain the speech intelligibility and the environmental noise features corresponding to the speech transcription text from the speech conversion engine, and determine the noise compensation factor of the speech transcription text according to the environmental noise features. Finally, the computer device can perform weighted fusion on the error probability values, the fluency perplexity, the speech intelligibility, and the noise compensation factor to obtain the first error confidence.

[0083] The error probability value refers to a numerical value quantifying the possible error degree of each word or phrase in the text; the higher the value, the greater the possibility of error for the word. The fluency perplexity is an index calculated by a statistical language model, which is used to measure the natural degree of the text in terms of semantics and syntax; the higher the perplexity, the more problems the text may have, such as semantic breaks, sequence errors, or incoherent context. The speech intelligibility refers to the degree of intelligibility of the original speech signal itself, which is usually obtained by analyzing the volume, spectral features, speech energy distribution, and signal-to-noise ratio of the speech signal; the higher the intelligibility, the better the quality of the speech signal, and the more reliable the transcription text generated by the speech recognition engine. The noise compensation factor refers to an adjustment parameter calculated according to the environmental noise level in the speech signal, which is used to correct the uncertainty of speech recognition caused by noise interference; when the environmental noise is high, the compensation factor will increase the weight of possible errors.

[0084] Specifically, the computer device can extract character-level features and pinyin features using FastText, and simultaneously capture the character form collocation errors and homophone errors in the speech transcription text based on the two features, and then calculate the error probability value of each word. At the same time, the computer device can use a preset statistical language model to analyze the overall language fluency of the speech transcription text to obtain the fluency perplexity of the text. In addition, the computer device can also obtain the speech signal characteristics corresponding to the speech transcription text from the speech conversion engine, including speech intelligibility and environmental noise features, wherein the environmental noise features can be used to calculate a corresponding noise compensation factor to adjust the recognition uncertainty caused by noise interference, for example, when the signal-to-noise ratio SNR <-5dB, the noise compensation factor is 0.05, otherwise it is 0. Finally, the computer device fuses the error probability value, fluency perplexity, speech intelligibility, and noise compensation factor through a preset weighting strategy to comprehensively form a first error confidence for representing the overall error probability and quality level of the speech transcription text.

[0085] More specifically, the weighted average formula of the error confidence can be as follows:

[0086]

[0087] In the formula, represents the error confidence; represents the error probability value; represents the fluency perplexity; represents the speech intelligibility; represents the noise compensation factor; , , , and respectively represent the adjustment weights of each parameter, which can be obtained by grid search on an offline validation set. In the present application, the priority values of each adjustment weight can be , , , .

[0088] Of course, in the secondary error detection process, the computer device can calculate the error probability value and fluency perplexity of the fast correction text, and use the error probability value and fluency perplexity to combine the speech intelligibility and noise compensation factor obtained in the first error detection process to calculate the error confidence, thereby improving the detection efficiency.

[0089] In one embodiment, the process of performing word-level fast correction on the speech transcription text based on the AC automaton algorithm in step S120 to obtain the fast correction text can include:

[0090] S121: Obtain a character index tree of a service type corresponding to the speech transcription text; each node in the character index tree represents an error word pattern and is provided with a failure pointer.

[0091] S122: Input each character in the speech transcription text into the character index tree in turn for node-by-node matching, and mark the corresponding character in the speech transcription text as a candidate word group when a node representing an error word pattern is completely matched.

[0092] S123: Determine a corrected word group corresponding to the candidate word group according to a candidate correction rule corresponding to the character index tree, and determine a correction score of the corrected word group in the speech transcription text.

[0093] S124: Replace the corresponding candidate word group in the speech transcription text with the corrected word group whose correction score is greater than a preset correction threshold, to obtain a fast correction text.

[0094] In the text fast correction process, the computer device can first obtain a character index tree of a service type corresponding to the speech transcription text, each node in the character index tree representing an error word pattern and being provided with a failure pointer; based on this, the computer device can input each character in the speech transcription text into the character index tree in turn for node-by-node matching, and mark the corresponding character in the speech transcription text as a candidate word group when a node representing an error word pattern is completely matched, then can determine a corrected word group corresponding to the candidate word group according to a candidate correction rule corresponding to the character index tree, and determine a correction score of the corrected word group in the speech transcription text, and finally can replace the corresponding candidate word group in the speech transcription text with the corrected word group whose correction score is greater than a preset correction threshold, to obtain a fast correction text.

[0095] The character index tree is a Trie tree constructed by an AC automatic machine algorithm, which organizes tens of thousands of error word patterns to be corrected in the form of nodes, each node corresponding to a possible error word pattern and being provided with a pinyin reverse index, and the tree structure is provided with a failure pointer to support efficient state jumping and multi-mode matching.

[0096] Specifically, the computer device can input each character in the speech transcription text into the character index tree in sequence, detect possible error patterns in the text in real time through the method of node-by-node matching, and when a node representing a specific error word pattern is completely matched, the computer device can mark the corresponding character in the speech transcription text as a candidate word group for subsequent correction processing. Subsequently, the computer device can determine the correction word group corresponding to each candidate word group according to the preset candidate correction rule in the character index tree, and calculate the correction score of the correction word group in the speech transcription text, which comprehensively considers factors such as historical global trigger frequency, pinyin similarity, and fluency improvement value, so as to quantify the reliability of the correction operation. Finally, the computer device can only replace the corresponding candidate word group in the speech transcription text with the correction word group whose correction score is greater than the preset correction threshold, thereby improving the basic accuracy and readability of the generated quick correction text.

[0097] In one embodiment, the process of determining the correction word group corresponding to the candidate word group according to the candidate correction rule corresponding to the character index tree in step S123 can include:

[0098] S1231: Perform word group matching on the candidate word group from the candidate correction rule corresponding to the character index tree to obtain a matching result.

[0099] S1232: Determine whether the number of correction word groups in the matching result is one.

[0100] S1233: If yes, the correction word group in the matching result is taken as the final correction word group of the candidate word group.

[0101] S1234: If no, calculate the overlap interval score of each correction word group in the matching result, and take the correction word group with the maximum overlap interval score as the final correction word group of the candidate word group.

[0102] In this embodiment, when confirming the correction word group, the computer device can first perform word group matching on the candidate word group from the candidate correction rule corresponding to the character index tree to obtain a matching result. If there is only one correction word group in the matching result, the computer device can directly take the correction word group as the final correction word group of the candidate word group; if there are multiple correction word groups in the matching result, the computer device can calculate the overlap interval score of each correction word group in the matching result, and take the correction word group with the maximum overlap interval score as the final correction word group of the candidate word group.

[0103] Specifically, in the matching result, if there is only one correction phrase, it indicates that the correction of the candidate phrase has uniqueness and certainty, and the computer device can directly take the correction phrase as the final correction phrase of the candidate phrase, thereby completing the correction operation of the candidate phrase. If there are multiple possible correction phrases in the matching result, it indicates that the candidate phrase may correspond to multiple correction rules in different semantic contexts. In order to select the optimal scheme from the multiple correction rules, the computer device can calculate the overlap interval score of the correction phrase in each correction rule, which reflects the compatibility of the correction phrase with other candidate correction phrases in the text in the coverage character interval and the overall text quality improvement potential. Subsequently, the computer device can select the correction phrase with the highest overlap interval score as the final correction phrase of the candidate phrase, thereby ensuring that in the case of conflict or multiple correction possibilities, the selected correction phrase can maximize the text correction effect without causing character coverage conflict.

[0104] For example, assuming that there is a candidate phrase "1 number microphone" in the speech transcription text, the candidate phrase corresponds to two correction rules, wherein the correction phrase of rule A is "1 number", and the correction phrase of rule B is "microphone". Therefore, the computer device can calculate the overlap interval score of each correction phrase, for example, "1 number" covers text positions 1 to 3, and the score is 5, and "microphone" covers text positions 3 to 5, and the score is 8. Finally, the system selects the correction phrase "microphone" with the highest overlap interval score as the final correction phrase of the candidate phrase "1 number microphone", thereby realizing accurate correction of the original text while ensuring the rationality of the correction result in the word level and the overall text semantics.

[0105] In one embodiment, the process of determining the correction score of the correction phrase in the speech transcription text in step S123 can include:

[0106] S1235: Obtain the global trigger frequency of the correction phrase in the historical correction record, and the priority rule of the business type corresponding to the speech transcription text, and determine the priority of the word group type corresponding to the correction phrase in the priority rule.

[0107] S1236: Calculate the pinyin similarity between the correction phrase and the candidate phrase, and determine the fluency improvement value of the correction phrase in the speech transcription text.

[0108] S1237: Weighted fusion of global trigger frequency, priority, pinyin similarity and fluency improvement value to obtain the correction score.

[0109] In this embodiment, when calculating the correction score of the correction word group, the computer device can first obtain the global trigger frequency of the correction word group in the historical correction record, and the priority rule of the business type corresponding to the voice transcription text, and determine the priority of the word group type corresponding to the correction word group in the priority rule; at the same time, the computer device can also calculate the pinyin similarity between the correction word group and the candidate word group, and determine the smoothness improvement value of the correction word group in the voice transcription text. After obtaining these parameters, the computer device can weight and fuse them to obtain the correction score of the correction word group.

[0110] Wherein, the correction frequency refers to the number of times or the frequency of occurrence that the correction word group is matched and finally adopted in the historical voice transcription error correction process; the higher the global trigger frequency, the higher the probability of the error mode corresponding to the correction word group. The priority refers to the importance level of the word group type to which the correction word group belongs under the current business type corresponding to the voice transcription text, which is determined by the preset business priority rule; for example, in different business scenarios such as meetings, customer service or live broadcast, the attention degree to word group types such as proper nouns, device names and role appellations is different. The pinyin similarity refers to an index for measuring the closeness of the pronunciation of the correction word group and the candidate word group, which is usually obtained by aligning and comparing the pinyin sequences of the two and calculating the similarity score; the higher the pinyin similarity, the more likely it is that the candidate word group is due to similar pronunciation, accent difference or recognition bias. The smoothness improvement degree refers to the degree of improvement in the smoothness of the voice transcription text under the language model evaluation after replacing the candidate word group with the correction word group; the greater the smoothness improvement value, the more effectively the correction word group can improve the text quality and reduce semantic ambiguity or sentence break problems.

[0111] Specifically, the computer device can first obtain the global trigger frequency of the correction word group in the historical correction record to reflect its stability and reliability in actual application; at the same time, in combination with the business type corresponding to the voice transcription text, the computer device can read the preset business priority rule and determine the priority of the word group type to which the correction word group belongs in the priority rule, so as to reflect the preference degree of specific vocabulary or expression form in different business scenarios. In addition, the computer device can also calculate the pinyin similarity between the correction word group and the candidate word group to determine whether the correction conforms to the objective law of voice recognition error; in addition, the computer device can also determine the improvement amplitude of the overall smoothness of the voice transcription text after replacing the candidate word group with the correction word group, and obtain the corresponding smoothness improvement value. Subsequently, the computer device can weight and fuse the global trigger frequency, the priority, the pinyin similarity and the smoothness improvement value according to the preset weight configuration, so as to calculate the correction score for representing the reliability and effectiveness of the correction word group.

[0112] More specifically, the weight fusion formula of the correction score can be as follows:

[0113]

[0114] In the formula, Represents the i-th modified phrase Corrected score; Indicates priority; Indicates the similarity of pinyin; Indicates the fluency boost value; Indicates the global trigger frequency; , , and These represent the adjustment weights for each parameter. In this application, each adjustment weight can be... , , , .

[0115] In one embodiment, step S140, which involves performing semantic-level deep error correction on the fast error-corrected text using a preset deep error correction model to obtain the deep error-corrected text, may include:

[0116] S141: Determine the deep error correction model; the deep error correction model is obtained by training a pre-trained model using Logit Bias constraints.

[0117] S142: Input the fast error-corrected text into the deep error-correction model so that the deep error-correction model can perform contextual analysis and error phrase correction on the fast error-corrected text and output it as deep error-corrected text.

[0118] In this embodiment, during deep text correction, the computer device can first obtain a deep correction model that has been pre-trained using Logit Bias constraints. Then, the fast-corrected text can be input into the model so that the deep correction model can perform contextual analysis and error phrase correction on the fast-corrected text and output it as deep-corrected text, thereby achieving high-quality text correction.

[0119] Understandably, Logit Bias constraints can controllably adjust the output bias of deep error correction models to reduce the risk of miscorrection and improve the model's error correction stability and reliability in specific business scenarios. For example, in this application, Logit Bias constraints can be used to suppress key business terms that should not be arbitrarily replaced, such as personal names, equipment names, numbered phrases, or industry-specific terms, preventing the model from incorrectly modifying them into seemingly more fluent but semantically inaccurate expressions during semantic inference. Furthermore, Logit Bias constraints can also apply a positive bias to high-confidence correct words, increasing their retention probability in the output results, thereby reducing overcorrection.

[0120] Therefore, after the computer device obtains the deep error correction model, the fast error correction text can be input into the deep error correction model as input, so that the model performs overall semantic analysis on the text based on context semantic relationship, syntax structure and cross-sentence dependency information, and accurately corrects the homonym confusion, semantic inconsistency, abnormal word order or implicit error word group that still exists under the premise of fully understanding the context meaning, and finally outputs the corresponding deep error correction text.

[0121] In an embodiment, the method can further include:

[0122] S160: obtaining error correction word group pairs generated by the deep error correction model in the semantic level deep error correction process and historical correction data of the error correction word group pairs; the error correction word group pair is composed of word groups before and after correction.

[0123] S170: based on the historical correction data, performing context analysis on the error correction word group pair by using Embedding clustering to obtain an analysis result, and generating a rule item of the error correction word group pair when the analysis result is strongly related.

[0124] S180: pushing the rule item to an artificial review queue for artificial review, and updating a node of a character index tree based on the rule item when the artificial review result is passed.

[0125] In this embodiment, whenever the deep error correction file completes a deep text error correction, the computer device can obtain error correction word group pairs generated by the model in the error correction process and historical correction data of the error correction word group pairs; then based on the historical correction data, the context analysis is performed on the error correction word group pair by using Embedding clustering to obtain an analysis result, and a rule item of the error correction word group pair is generated when the analysis result is strongly related; subsequently, the computer device can push the rule item to an artificial review queue for artificial review, and update a node of a character index tree based on the rule item when the artificial review result is passed.

[0126] It can be understood that the error correction word group pair is mainly used to represent the mapping relationship between the word groups before and after correction, and the historical correction data reflects the repeated occurrence of this type of error correction in different texts and contexts. Based on this, the computer device can map the error correction word group pair to a vector space and cluster its distribution in different contexts to judge the consistency of the error correction relationship in semantics and use scenarios, and when the analysis result shows that the error correction word group pair has high context similarity and stable correction direction in multiple occurrences, it can be determined that the error correction word group pair is a strongly related relationship.

[0127] Specifically, when the error correction phrase pair has a strong correlation in different text contexts, the computer device can generate a corresponding error correction rule item therefor, push it to a manual review queue, and manually check the rationality of the rule, business adaptability, and potential mis-correction risk. When the manual review result is passed, the computer device can update the node of the character index tree corresponding to its business scenario based on the rule item, so that it becomes a new rule that can be directly hit by the fast error correction process, thereby realizing the closed-loop cooperation between deep error correction and fast error correction, reducing the processing delay and computing power consumption of subsequent similar errors.

[0128] The voice-to-transcribed text error correction device provided by the embodiments of the present application is described below. The voice-to-transcribed text error correction device described below can be correspondingly referred to the voice-to-transcribed text error correction method described above.

[0129] In one embodiment, as shown in Figure 3 , Figure 3 is a structural schematic diagram of a voice-to-transcribed text error correction device provided by an embodiment of the present application. The present application also provides a voice-to-transcribed text error correction device, which comprises a first detection module 210, a fast correction module 220, a second detection module 230, a deep error correction module 240 and a text output module 250, and specifically comprises the following:

[0130] The first detection module 210 is configured to receive a voice-to-transcribed text in real time, and perform initial error detection on the voice-to-transcribed text by using a multi-modal detection strategy to obtain a first error confidence.

[0131] The fast correction module 220 is configured to perform word-level fast error correction on the voice-to-transcribed text based on an AC automatic machine algorithm if the first error confidence is not less than a fast error correction threshold, to obtain a fast error correction text.

[0132] The second detection module 230 is configured to perform secondary error detection on the fast error correction text by using a multi-modal detection strategy to obtain a second error confidence.

[0133] The deep error correction module 240 is configured to perform semantic-level deep error correction on the fast error correction text by using a pre-set deep error correction model if the second error confidence is not less than a deep error correction threshold, to obtain a deep error correction text.

[0134] The text output module 250 is configured to output the corresponding voice-to-transcribed text, fast error correction text or deep error correction text as a final transcribed text when the first error confidence is less than the fast error correction threshold, the second error confidence is less than the deep error correction threshold, or the deep error correction model outputs the deep error correction text.

[0135] In the above embodiments, after receiving the speech transcription text in real time, in order to quantify the error amount of the transcription text, the application can adopt a multi-modal detection strategy to detect errors in the speech transcription text to obtain a first error confidence, which provides a dynamic basis for the selection of subsequent error correction strategies. For example, when the first error confidence is less than a fast error correction threshold, the speech transcription text can be directly output as a final transcription text to avoid unnecessary error correction; otherwise, the speech transcription text is subjected to word-level fast error correction based on an AC automaton algorithm to obtain fast error correction text, and the fast error correction text is subjected to error detection again to obtain a second error confidence. Then, it can be judged whether the second error confidence is less than a deep error correction threshold; if yes, the fast error correction text is output as the final transcription text, and if no, the fast error correction text is subjected to semantic-level deep error correction through a preset deep error correction model to obtain deep error correction text, which is output as the final transcription text. Through the hierarchical error correction processing strategy, the application can adapt to different business scenarios, and by adjusting the fast error correction threshold and the deep error correction threshold, the application can achieve dynamic balance between quality and efficiency in different business scenarios.

[0136] In one embodiment, the one-time detection module 210 can include:

[0137] The first parameter determination sub-module is configured to analyze errors in the speech transcription text based on the word-level features and the pinyin features to obtain an error probability value, and determine a fluency perplexity of the speech transcription text through a preset statistical language model.

[0138] The second parameter determination sub-module is configured to obtain the speech intelligibility and the environmental noise features of the speech transcription text from the speech conversion engine, and determine a noise compensation factor of the speech transcription text according to the environmental noise features.

[0139] The weighted fusion sub-module is configured to perform weighted fusion on the error probability value, the fluency perplexity, the speech intelligibility, and the noise compensation factor to obtain the first error confidence.

[0140] In one embodiment, the fast correction module 220 can include:

[0141] The index tree acquisition sub-module is configured to acquire a character index tree of the business type corresponding to the speech transcription text; each node in the character index tree represents an error word pattern and is provided with a failure pointer.

[0142] The candidate marking sub-module is configured to input each character in the speech transcription text into the character index tree in sequence for node-by-node matching, and mark the corresponding character in the speech transcription text as a candidate word group when the node representing the error word pattern is completely matched.

[0143] The score calculation sub-module is configured to determine a corrected word group corresponding to the candidate word group according to the candidate correction rule corresponding to the character index tree, and determine a correction score of the corrected word group in the speech transcription text.

[0144] The text correction sub-module is configured to replace the candidate word group in the speech transcription text with the corrected word group having a correction score greater than a preset correction threshold, to obtain a fast correction text.

[0145] In an embodiment, the score calculation sub-module can include:

[0146] The word group matching unit is configured to perform word group matching on the candidate word group from the candidate correction rule corresponding to the character index tree, to obtain a matching result.

[0147] The number judgment unit is configured to judge whether the number of the corrected word groups in the matching result is one.

[0148] The first confirmation unit is configured to, if yes, take the corrected word group in the matching result as the final corrected word group of the candidate word group.

[0149] The second confirmation unit is configured to, if no, calculate an overlapping interval score of each corrected word group in the matching result, and take the corrected word group with the maximum overlapping interval score as the final corrected word group of the candidate word group.

[0150] In an embodiment, the score calculation sub-module can further include:

[0151] The first data acquisition unit is configured to acquire a global trigger frequency of the corrected word group in a historical correction record, and a priority rule of a business type corresponding to the speech transcription text, and determine a priority of a word group type corresponding to the corrected word group in the priority rule.

[0152] The second data acquisition unit is configured to calculate a pinyin similarity between the corrected word group and the candidate word group, and determine a fluency improvement value of the corrected word group in the speech transcription text.

[0153] The weighted fusion unit is configured to perform weighted fusion on the global trigger frequency, the priority, the pinyin similarity, and the fluency improvement value, to obtain the correction score.

[0154] In an embodiment, the deep correction module 240 can include:

[0155] The model determination sub-module is configured to determine a deep correction model; the deep correction model is obtained by training a preset pre-training model using a Logit Bias constraint.

[0156] The model correction sub-module is configured to input the fast correction text into the deep correction model, so that the deep correction model performs context analysis and error word group correction on the fast correction text, and outputs a deep correction text.

[0157] In an embodiment, the apparatus can further include:

[0158] a module configured to acquire a corrected word pair generated by the deep correction model in the semantic-level deep correction process and historical correction data of the corrected word pair, the corrected word pair being composed of a word group before correction and a word group after correction.

[0159] a module configured to perform context analysis on the corrected word pair based on the historical correction data using Embedding clustering to obtain an analysis result, and generate a rule item of the corrected word pair when the analysis result is strong correlation.

[0160] a module configured to push the rule item to a manual review queue for manual review, and perform node update on the character index tree based on the rule item when the manual review result is passed.

[0161] In an embodiment, the present application further provides a storage medium having computer readable instructions stored therein, the computer readable instructions being executed by one or more processors to cause the one or more processors to perform the steps of the voice transcription text correction method according to any one of the above embodiments.

[0162] In an embodiment, the present application further provides a computer device having computer readable instructions stored therein, the computer readable instructions being executed by one or more processors to cause the one or more processors to perform the steps of the voice transcription text correction method according to any one of the above embodiments.

[0163] As shown in Figure 4 , Figure 4 Fig. 1 is a schematic diagram of an internal structure of a computer device according to an embodiment of the present application. The computer device 300 can be provided as a server. As shown in Figure 4 , the computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by a memory 301, for storing instructions executable by the processing component 302, such as an application program. The application program stored in the memory 301 can include one or more than one module each corresponding to a set of instructions. In addition, the processing component 302 is configured to execute the instructions to perform the voice transcription text correction method according to any one of the above embodiments.

[0164] The computer device 300 can further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 can operate based on an operating system stored in the memory 301, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM, or the like.

[0165] Those skilled in the art can understand that Figure 4 The structure shown in FIG. 3 is merely a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. A specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0166] Finally, it should be noted that the terms such as first and second, etc., are merely used to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Also, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the phrase "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0167] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The various embodiments can be combined as needed, and the same or similar parts refer to each other.

[0168] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A speech-to-text error correction method, characterized in that, The method includes: The speech-to-text is received in real time, and a multimodal detection strategy is used to perform initial error detection on the speech-to-text to obtain a first error confidence level. If the first error confidence is not less than the fast error correction threshold, then the speech-to-text is subjected to word-level fast error correction based on the AC automata algorithm to obtain the fast error-corrected text. The multimodal detection strategy is used to perform secondary error detection on the fast error-corrected text to obtain a second error confidence level; If the second error confidence is not less than the deep error correction threshold, then the fast error correction text is subjected to semantic-level deep error correction through a preset deep error correction model to obtain the deep error correction text. When the first error confidence is less than the fast error correction threshold, the second error confidence is less than the deep error correction threshold, or the deep error correction model outputs the deep error correction text, the corresponding speech-to-text, fast error correction text, or deep error correction text is output as the final transcribed text.

2. The speech-to-text error correction method according to claim 1, characterized in that, The initial error detection of the speech-to-text using a multimodal detection strategy to obtain a first error confidence level includes: Error analysis is performed on the speech-to-text based on character-level features and pinyin features to obtain error probability values, and the fluency and perplexity of the speech-to-text are determined through a preset statistical language model. The speech clarity and environmental noise features corresponding to the speech-to-text are obtained from the speech conversion engine, and the noise compensation factor of the speech-to-text is determined based on the environmental noise features. The error probability value, the fluency confusion level, the speech clarity, and the noise compensation factor are weighted and fused to obtain the first error confidence level.

3. The speech-to-text error correction method according to claim 1, characterized in that, The AC automaton algorithm is used to perform word-level fast error correction on the speech-to-text, resulting in fast-corrected text, including: Obtain the character index tree corresponding to the business type of the speech-to-text text; each node in the character index tree represents an error word pattern and is set with a failure pointer; Each character in the speech-to-text is sequentially input into the character index tree for node-by-node matching. When a node representing an error word pattern is completely matched, the corresponding character in the speech-to-text is marked as a candidate word group. The correction word group corresponding to the candidate word group is determined according to the candidate correction rules corresponding to the character index tree, and the correction score of the correction word group in the speech-to-text is determined. The candidate word groups in the speech-to-text are replaced by correction words whose correction scores are greater than a preset correction threshold, resulting in a quickly corrected text.

4. The speech-to-text error correction method according to claim 3, characterized in that, The step of determining the corrected word group corresponding to the candidate word group according to the candidate correction rule corresponding to the character index tree includes: The candidate word groups are matched from the candidate correction rules corresponding to the character index tree to obtain the matching results; Determine whether the number of corrected phrases in the matching result is one; If so, the modified word group in the matching result will be taken as the final modified word group of the candidate word group; If not, calculate the overlap interval score of each corrected word group in the matching results, and take the corrected word group with the largest overlap interval score as the final corrected word group of the candidate word group.

5. The speech-to-text error correction method according to claim 3, characterized in that, Determining the correction score of the corrected phrase in the speech-to-text includes: Obtain the global trigger frequency of the corrected phrase in the historical correction record, as well as the priority rule of the business type corresponding to the speech-to-text, and determine the priority of the phrase type corresponding to the corrected phrase in the priority rule; Calculate the pinyin similarity between the corrected phrase and the candidate phrase, and determine the fluency improvement value of the corrected phrase in the speech-to-text. The global trigger frequency, priority, pinyin similarity, and fluency improvement value are weighted and fused to obtain the corrected score.

6. The speech-to-text error correction method according to claim 1, characterized in that, The step of performing semantic-level deep error correction on the fast error-corrected text using a preset deep error correction model to obtain deep error-corrected text includes: A deep error correction model is determined; the deep error correction model is obtained by training a pre-trained model using Logit Bias constraints. The fast error-corrected text is input into the deep error-correction model, so that the deep error-correction model performs contextual analysis and corrects erroneous phrases on the fast error-corrected text, and outputs it as deep error-corrected text.

7. The speech-to-text error correction method according to claim 3, characterized in that, The method further includes: Obtain the error-correcting word pairs generated by the deep error correction model during the semantic-level deep error correction process, as well as the historical correction data of the error-correcting word pairs; the error-correcting word pairs consist of word pairs before and after error correction; Based on the historical correction data, Embedding clustering is used to perform contextual analysis on the error correction word pair to obtain the analysis results. When the analysis results are strongly correlated, rule entries for the error correction word pair are generated. The rule entries are pushed to the manual review queue for manual review, and when the manual review result is passed, the nodes of the character index tree are updated based on the rule entries.

8. A transcription text error correction device, characterized in that, include: A primary detection module is used to receive speech-to-text in real time and perform initial error detection on the speech-to-text using a multimodal detection strategy to obtain a first error confidence level. The fast correction module is used to perform word-level fast error correction on the speech-to-text based on the AC automata algorithm if the first error confidence is not less than the fast error correction threshold, so as to obtain the fast error-corrected text. The secondary detection module is used to perform secondary error detection on the fast error-corrected text using the multimodal detection strategy to obtain a second error confidence level; The deep error correction module is used to perform semantic-level deep error correction on the fast error correction text through a preset deep error correction model if the second error confidence is not less than the deep error correction threshold, so as to obtain the deep error correction text. The text output module is used to output the corresponding speech-to-text, fast error correction text, or deep error correction text as the final transcribed text when the first error confidence is less than the fast error correction threshold, the second error confidence is less than the deep error correction threshold, or the deep error correction model outputs the deep error correction text.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the speech-to-text error correction method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the speech-to-text error correction method as described in any one of claims 1 to 7.

Images