Speech recognition error correction method, model training method and device

Abstract

The present disclosure provides a speech recognition error correction method, model training method, and device, which relate to the fields of artificial intelligence technology, and in particular to speech recognition and natural language processing technology. A specific implementation scheme is as follows: obtaining the original speech recognition result of speech information; identifying the presence of target text in the original speech recognition result; replacing the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result; and correcting the target text based on the original speech recognition result and the candidate speech recognition result. Thus, automatic error correction of speech recognition results can be achieved based on preset labels, which has good applicability and improves the error correction efficiency and accuracy of speech recognition.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a speech recognition error correction method, model training method, device, electronic device, storage medium, and computer program product. Background Art

[0002] Currently, with the development of technologies such as artificial intelligence and natural language processing, speech recognition technology has been widely used in scenarios such as simultaneous interpretation, smart home appliances, robot voice interaction, and in-car voice. For example, in the in-car voice scenario, an in-car voice assistant can recognize the speech content of the vehicle's passengers, understand the user's intention based on the speech recognition content, and automatically execute the corresponding instructions, without the need for manual user operation. The response speed is fast and it is beneficial to driving safety. However, the speech recognition results of related technologies may be erroneous due to external factors, non-standard user pronunciation, and excessive speaking speed, resulting in a low accuracy rate. Summary of the Invention

[0003] The present disclosure provides a speech recognition error correction method, a model training method, an apparatus, an electronic device, a storage medium, and a computer program product.

[0004] According to one aspect of the present disclosure, a method for correcting speech recognition errors is provided, comprising: obtaining an original speech recognition result of speech information; identifying the presence of target text in the original speech recognition result; replacing the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result; and correcting the target text based on the original speech recognition result and the candidate speech recognition results.

[0005] According to another aspect of the present disclosure, a model training method is provided, which obtains a training sample, wherein the training sample includes a sample original speech recognition result and a sample candidate speech recognition result, the sample original speech recognition result includes a sample target text, the training sample also includes a text label of the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label; a masked language model is trained based on the training sample, wherein the masked language model is used to obtain a first training score of the sample target text at a sample text position and a second training score of the preset label, the sample text position is the text position of the sample target text in the sample original speech recognition result, the first training score is positively correlated with the predicted probability that the sample target text is at the sample text position, and the second training score is positively correlated with the predicted probability that the preset label is at the sample text position; in response to not meeting a model training end condition, returning to use the next training sample to continue training the masked language model until the model training end condition is met, thereby generating a target masked language model.

[0006] According to another aspect of the present disclosure, a speech recognition error correction device is provided, comprising: an acquisition module for acquiring an original speech recognition result of speech information; a recognition module for identifying the presence of a target text in the original speech recognition result; a replacement module for replacing the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result; and an error correction module for correcting the target text based on the original speech recognition result and the candidate speech recognition result.

[0007] According to another aspect of the present disclosure, a model training device is provided, comprising: an acquisition module for acquiring training samples, wherein the training samples include a sample original speech recognition result and a sample candidate speech recognition result, the sample original speech recognition result includes a sample target text, the training sample also includes a text label of the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label; a training module for training a masked language model based on the training samples, wherein the masked language model is used to obtain a first training score of the sample target text at a sample text position and a second training score of the preset label, the sample text position being the text position of the sample target text in the sample original speech recognition result, the first training score being positively correlated with a predicted probability that the sample target text is at the sample text position, and the second training score being positively correlated with a predicted probability that the preset label is at the sample text position; the training module is further configured to, in response to a model training termination condition not being met, return to continue training the masked language model using the next training sample until the model training termination condition is met to generate a target masked language model.

[0008] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute an error correction method for speech recognition, or to execute a model training method.

[0009] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute an error correction method for speech recognition or a model training method.

[0010] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the steps of a speech recognition error correction method or the steps of a model training method.

[0011] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0013] Figure 1 is a flowchart of a method for correcting speech recognition errors according to the first embodiment of the present disclosure;

[0014] Figure 2 is a flow chart of a method for correcting speech recognition errors according to a second embodiment of the present disclosure;

[0015] Figure 3 is a flowchart of a method for correcting speech recognition errors according to a third embodiment of the present disclosure;

[0016] Figure 4 is a flowchart of a model training method according to the first embodiment of the present disclosure;

[0017] Figure 5 is a block diagram of a speech recognition error correction device according to a first embodiment of the present disclosure;

[0018] Figure 6 is a block diagram of a model training device according to a first embodiment of the present disclosure;

[0019] Figure 7 It is a block diagram of an electronic device used to implement the error correction method and / or model training method of speech recognition in the embodiments of the present disclosure. DETAILED DESCRIPTION

[0020] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0021] Artificial Intelligence (AI) is a discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. Currently, AI technology has been widely used due to its high degree of automation, high precision, and low cost.

[0022] Speech recognition is a technology that allows machines to convert speech signals into corresponding text or commands through the process of recognition and understanding. It mainly includes three aspects: feature extraction technology, pattern matching criteria and model training technology.

[0023] NLU (Natural Language Processing) is a science that studies computer systems that can effectively implement natural language communication, especially the software systems therein. It is an important direction in the fields of computer science and artificial intelligence.

[0024] Figure 1 It is a flowchart of the error correction method for speech recognition according to the first embodiment of the present disclosure.

[0025] like Figure 1 As shown, the speech recognition error correction method of the first embodiment of the present disclosure includes:

[0026] S101, obtaining original speech recognition results of speech information.

[0027] It should be noted that the execution entity of the speech recognition error correction method in the embodiments of the present disclosure is a hardware device with data information processing capabilities and / or the necessary software to drive the operation of the hardware device. Optionally, the execution entity may include a workstation, server, computer, user terminal, and other intelligent devices. Among them, user terminals include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, and in-vehicle terminals.

[0028] It should be noted that there are no specific limitations on the types of voice information. For example, voice information includes but is not limited to Chinese, English, Japanese, etc., and includes but is not limited to conversations, meetings, speeches, etc. There are also no specific limitations on the methods for acquiring voice information. For example, a voice collection device can be installed on the smart device or in the area surrounding the smart device to collect voice information. The voice collection device may include a microphone.

[0029] In the embodiments of the present disclosure, the original speech recognition result of the speech information may be obtained in an offline and / or online manner, which is not limited here.

[0030] In one embodiment, a speech recognition device may be pre-installed on the smart device, and the speech recognition device may output the original speech recognition result of the speech information, thereby achieving offline speech recognition. The speech recognition device may include a speech recognition model.

[0031] In one embodiment, a smart device can establish a network connection with a server, and the smart device can send voice information to the server. A voice recognition device in the server outputs a raw voice recognition result of the voice information, and the server can send the raw voice recognition result to the smart device, thereby achieving online voice recognition. The server can include a cloud server.

[0032] S102: Identify whether the target text exists in the original speech recognition result.

[0033] It should be noted that the target text refers to the text that may need to be corrected, which can be set according to the actual situation and is not excessively limited here.

[0034] In one embodiment, the target text includes but is not limited to a person's name, a company name, a place name, etc.

[0035] In one embodiment, identifying the presence of the target text in the original speech recognition result may include inputting the original speech recognition result into a NER (Named Entity Recognition) model, and having the NER model output the target text.

[0036] In one embodiment, the target text can be obtained based on the number of repetitions of the recognized text. It is understandable that when the voice recognition result is wrong, the user will generally repeat the voice information, and the voice recognition result may still be wrong, that is, the more times the voice recognition result is repeated, the greater the possibility of error. The target text can be obtained based on the number of repetitions of the recognized text. For example, if it is recognized that the original voice recognition results for a first set number of consecutive times have the same recognized text for a second set number of times, the above-mentioned recognized text can be determined as the target text. It should be noted that the first set number of times is greater than or equal to the second set number of times, and there are no excessive restrictions on the first set number and the second set number of times. For example, the first set number of times can be 10 times, and the second set number of times can be 5 times.

[0037] S103: Replace the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result.

[0038] It should be noted that there are no excessive restrictions on the preset tags. For example, the preset tags include but are not limited to <ne> 、 <tt>wait.

[0039] For example, the original speech recognition result is "Just now, Youping introduced the open platform to everyone", the target text is "Youping", and the preset label is " <ne>", then the candidate speech recognition result is "just now <ne>Introduced the open platform to everyone."

[0040] For example, the original speech recognition result is "Hello everyone, I am Zhang San, a Chinese teacher", the target text is "Zhang San", and the preset label is " <ne>", then the candidate speech recognition result is "Hello everyone, I am <ne>, is a Chinese teacher."

[0041] For example, the original speech recognition result is "Please navigate to Gugong", the target text is "Gugong", and the preset label is " <ne>", the candidate speech recognition result is "Please navigate to <ne>”.

[0042] For example, the original speech recognition result is "start washing machine", the target text is "washing machine", and the preset label is " <ne>", the candidate speech recognition result is "Start <ne>”.

[0043] S104: Correct the target text based on the original speech recognition result and the candidate speech recognition results.

[0044] In one embodiment, whether the target text needs error correction can be determined based on the original speech recognition result and the candidate speech recognition result. If the target text needs error correction, indicating that the target text recognition error occurred, the target text is corrected. Alternatively, if the target text does not need error correction, indicating that the target text recognition was correct, the target text does not need error correction.

[0045] For example, the original speech recognition result is "Just now, Youping introduced the open platform to everyone", the target text is "Youping", and the preset label is " <ne>", the candidate speech recognition result is "just now <ne>I introduced the open platform to you", based on "I just introduced the open platform to you" and "I just <ne>Just now, Youping introduced the open platform to everyone. If the recognition of "Youping" needs to be corrected, you can correct "Youping" and change "Youping" to "Youping". The target speech recognition result will be "Just now, Youping introduced the open platform to everyone."

[0046] For example, the original speech recognition result is "I introduced the open platform to everyone", the target text is "introduction", and the preset label is " <ne>", the candidate speech recognition result is "I tell everyone <ne>I introduced the open platform to you and <ne>The target speech recognition result is "I introduced the open platform to everyone". If "introduction" does not require error correction, the target speech recognition result is "I introduced the open platform to everyone".

[0047] For example, the original speech recognition result is "Please navigate to Gugong", the target text is "Gugong", and the preset label is " <ne>", the candidate speech recognition result is "Please navigate to <ne>", based on "Please navigate to Gugong" and "Please navigate to <ne>", if the recognition of "Gu Gong" needs correction, then "Gu Gong" can be corrected and "Gu Gong" can be corrected to "Gu Gong". Then the target speech recognition result is "Please navigate to the Forbidden City".

[0048] For example, the original speech recognition result is "start washing machine", the target text is "washing machine", and the preset label is " <ne>", the candidate speech recognition result is "Start <ne>", based on "Start the washing machine" and "Start <ne>", if the recognition of "washing machine" needs correction, then "washing machine" can be corrected and "washing machine" can be corrected to "dishwasher", and the target speech recognition result will be "start dishwasher".

[0049] In summary, according to the speech recognition error correction method of the disclosed embodiment, upon identifying the presence of target text in the original speech recognition result, the target text in the original speech recognition result can be replaced with a preset label to generate a candidate speech recognition result. The target text can then be corrected based on the original speech recognition result and the candidate speech recognition result. This method can thus automatically correct speech recognition results based on the preset labels, providing good applicability and improving the error correction efficiency and accuracy of speech recognition.

[0050] Figure 2 4 is a flow chart of a method for correcting speech recognition errors according to the second embodiment of the present disclosure.

[0051] like Figure 2 As shown, the speech recognition error correction method of the second embodiment of the present disclosure includes:

[0052] S201, obtaining original speech recognition results of speech information.

[0053] S202: Identify whether the target text exists in the original speech recognition result.

[0054] S203: Replace the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result.

[0055] The relevant contents of steps S201-S203 can be found in the above embodiment and will not be repeated here.

[0056] S204: Obtain the text position of the target text in the original speech recognition result.

[0057] S205, obtaining a first score of the target text at the text position and a second score of the preset label, wherein the first score is positively correlated with the predicted probability that the text position is the target text, and the second score is positively correlated with the predicted probability that the text position is the preset label.

[0058] In the embodiments of the present disclosure, the probabilities of the target text and the preset label at the text position can be predicted respectively. It can be understood that the higher the predicted probability, the greater the possibility that the text position is the corresponding text, that is, the higher the score, the greater the possibility that the text position is the corresponding text.

[0059] For example, if the first score of the target text at the text position is less than the second score of the preset label at the text position, then the possibility that the text position is the target text is less than the possibility that the text position is the preset label; or, if the first score of the target text at the text position is greater than or equal to the second score of the preset label at the text position, then the possibility that the text position is the target text is greater than or equal to the possibility that the text position is the preset label.

[0060] In one embodiment, obtaining a first score for a target text at a text location and a second score for a preset label may include inputting the original speech recognition result and the candidate speech recognition result into a target MLM (Masked Language Model), and having the target masked language model output the first score and the second score. It is understandable that the target masked language model may obtain the first score for the target text at the text location and the second score for the preset label based on the original speech recognition result and the candidate speech recognition result. It should be noted that the structure of the target masked language model is not excessively limited. Thus, the method can obtain the first score and the second score through the target masked language model, thereby achieving automatic acquisition of the scores.

[0061] S206: Correct the target text based on the first score and the second score.

[0062] In one embodiment, whether the target text needs error correction can be identified based on the first score and the second score. If the target text needs error correction, indicating that the target text recognition error occurred, the target text is corrected. Alternatively, if the target text does not need error correction, indicating that the target text recognition was correct, the target text does not need error correction.

[0063] In one embodiment, if the first recognition score is less than the second score, it indicates that the possibility of the target text at the text position is less than the possibility of the preset label at the text position, that is, the target text recognition error, then the target reference text corresponding to the target text can be obtained, and the target text in the original speech recognition result can be replaced with the target reference text to generate the target speech recognition result. It should be noted that the target reference text can be set according to the actual situation and is not too limited here. It is understandable that different target texts can correspond to different target reference texts. Therefore, the method can replace the target text with the target reference text when the first score is less than the second score, thereby achieving automatic error correction of the target text, which helps to improve the accuracy of speech recognition.

[0064] For example, the original speech recognition result is "Just now, Youping introduced the open platform to everyone", the target text is "Youping", and the preset label is " <ne>", the candidate speech recognition result is "just now <ne>I introduced the open platform to you", based on "I just introduced the open platform to you" and "I just <ne>Introduced the open platform to everyone", the first score of "Youping" at the text position is 20 points, the first score of "Youping" at the text position is 20 points, <ne>The second score of " is 88 points, the first score is less than the second score, and the target reference text is "Youping". The target speech recognition result is "Just now Youping introduced the open platform to everyone."

[0065] In one embodiment, obtaining a target reference text corresponding to a target text may include obtaining a reference text library, obtaining a first similarity between the target text and a reference text, and determining the reference text corresponding to the greatest first similarity as the target reference text. It should be noted that the reference text library may include a large number of reference texts, which may be pre-set based on actual circumstances. The categories of the first similarity are not particularly limited; for example, the first similarity includes, but is not limited to, phonemes, glyphs, etc. Thus, the method can obtain the target reference text based on the first similarity between the target text and the reference text, thereby enabling automatic acquisition of the target reference text.

[0066] In one embodiment, obtaining the first similarity between the target text and the reference text may include obtaining the phoneme similarity between the phoneme features of the target text and the phoneme features of the reference text, and / or obtaining the glyph similarity between the glyph features of the target text and the glyph features of the reference text, and obtaining the first similarity based on the phoneme similarity and / or glyph similarity.

[0067] In one embodiment, phoneme features of the target text and the reference text may be extracted respectively, and the extracted phoneme features of the target text and the phoneme features of the reference text may be input into a similarity algorithm, which then outputs the phoneme similarity.

[0068] In one embodiment, glyph features of the target text and the reference text may be extracted respectively, and the extracted glyph features of the target text and the glyph features of the reference text may be input into a similarity algorithm, which then outputs the glyph similarity.

[0069] In one embodiment, obtaining the first similarity based on the phoneme similarity and / or the glyph similarity may include obtaining a first weight for the phoneme similarity and a second weight for the glyph similarity, obtaining a first product of the phoneme similarity and the first weight, and a second product of the glyph similarity and the second weight, and using the sum of the first product and the second product as the first similarity. It should be noted that neither the first weight nor the second weight is particularly limited.

[0070] For example, the reference text library may include "Youping, Forbidden City, dishwasher, washing machine", the target text is "Youping", and the reference text corresponding to the maximum first similarity is "Youping", then "Youping" can be determined as the target reference text; or, the target text is "Gu Gong", and the reference text corresponding to the maximum first similarity is "Forbidden City", then "Forbidden City" can be determined as the target reference text; or, the target text is "washing machine", and the reference text corresponding to the maximum first similarity is "dishwasher", then "dishwasher" can be determined as the target reference text.

[0071] In one embodiment, if the first recognition score is greater than or equal to the second score, it indicates that the possibility of the text position being the target text is greater than or equal to the possibility of the text position being the preset label, that is, the target text is recognized correctly, and the original speech recognition result can be used as the target speech recognition result.

[0072] For example, the original speech recognition result is "I introduced the open platform to everyone", the target text is "introduction", and the preset label is " <ne>", the candidate speech recognition result is "I tell everyone <ne>I introduced the open platform to you and <ne>The first score of the text position "Introduction" is 88 points, and the first score of the text position " <ne>The second score of " is 20 points. The first score is greater than the second score. The target speech recognition result is "I introduced the open platform to everyone."

[0073] In summary, according to the speech recognition error correction method of the embodiment of the present disclosure, the target text can be corrected based on the first score of the target text at the text position and the second score of the preset label, and the automatic error correction of the speech recognition result can be realized based on the preset label. It has good applicability and improves the error correction efficiency and accuracy of speech recognition.

[0074] Figure 3 It is a flowchart of a method for correcting speech recognition errors according to the third embodiment of the present disclosure.

[0075] like Figure 3 As shown, the speech recognition error correction method of the third embodiment of the present disclosure includes:

[0076] S301, obtaining the original speech recognition result of the speech information.

[0077] S302: Acquire a reference text library.

[0078] The relevant contents of steps S301-S302 can be found in the above embodiment and will not be repeated here.

[0079] S303: Divide the original speech recognition result into at least one word segment.

[0080] It can be understood that the original speech recognition result may include at least one word segment, and the original speech recognition result may be divided into at least one word segment.

[0081] In one embodiment, the original speech recognition result may be divided into at least one segmented word according to a preset segmentation unit, wherein the segmentation unit includes but is not limited to a word, a character, a word, a phrase, etc.

[0082] For example, when the word segmentation unit is a word, the original speech recognition result "Just now, you have introduced the open platform to everyone" can be divided into "just now", "you have commented", "with", "everyone", "introduction", "released", "open", and "platform".

[0083] For example, when the word segmentation unit is a word, the original speech recognition result "Please navigate to Gu Gong" can be divided into "please", "navigate", "to", and "Gu Gong".

[0084] S304: Compare the segmented words with the reference text to identify candidate segmented words that are inconsistent with any reference text.

[0085] In the embodiment of the present disclosure, the candidate word segmentation is inconsistent with any reference text in the reference text library.

[0086] For example, the reference text library may include "Youping, Forbidden City, dishwasher, washing machine", and the original speech recognition result "Just now, Youping introduced the open platform to everyone" can be divided into "just now", "Youping", "with", "everyone", "introduction", "released", "open", and "platform", and "Youping", "introduction", and "open" can be identified as candidate word segmentations.

[0087] For example, the reference text library may include "Youping, Forbidden City, dishwasher, washing machine", and the original speech recognition result "Please navigate to Gugong" can be divided into "please", "navigate", "to", and "Gugong", so "navigate" and "Gugong" can be identified as candidate word segmentations.

[0088] S305 : Obtain a second similarity between the candidate segmentation word and the reference text, and determine the candidate segmentation word whose second similarity is greater than a first preset threshold as the target text.

[0089] It should be noted that there are no excessive restrictions on the first preset threshold. For example, the first preset threshold may be set to 60%.

[0090] In one embodiment, obtaining the second similarity between the candidate segmentation and the reference text may include obtaining a first phoneme feature of the candidate segmentation and a second phoneme feature of the reference text, obtaining a third similarity between the first phoneme feature and the second phoneme feature, and using the third similarity as the second similarity. Thus, the method can obtain the second similarity between the candidate segmentation and the reference text based on the similarity between the phoneme features of the candidate segmentation and the reference text.

[0091] It should be noted that, for the content related to obtaining the second similarity between the candidate segmentation word and the reference text, reference may be made to the content related to obtaining the first similarity between the target text and the reference text in the above embodiment, which will not be repeated here.

[0092] In one embodiment, a second similarity between the candidate segmentation and each reference text may be obtained. If the second similarity between the candidate segmentation and each reference text is greater than a first preset threshold, the candidate segmentation may be determined as a target text. Alternatively, if the second similarity between the candidate segmentation and each reference text is not greater than the first preset threshold, the candidate segmentation may be determined as a non-target text.

[0093] For example, the reference text library may include "Youping, Forbidden City, dishwasher, washing machine", and the candidate segmentation words include "Youping". The second similarities between "Youping" and "Youping", "Forbidden City", "dishwasher", and "washing machine" can be obtained as 80%, 20%, 10%, and 10% respectively. The first preset threshold is 60%. It can be identified that "Youping" has a second similarity (80%) greater than the first preset threshold, and "Youping" is determined as the target text.

[0094] Alternatively, if the candidate word segment includes "open", the second similarities between "open" and "Youping", "the Forbidden City", "dishwasher", and "washing machine" can be obtained as 20%, 20%, 10%, and 10% respectively. The first preset threshold is 60%. Then, it can be recognized that there is no second similarity greater than the first preset threshold for "open", and "open" is determined as non-target text.

[0095] S306. Replace the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result.

[0096] S307. Correct the target text based on the original speech recognition result and the candidate speech recognition result.

[0097] For the relevant content of steps S306 - S307, reference can be made to the above embodiments, which will not be elaborated here.

[0098] In summary, according to the speech recognition error correction method of the present disclosure embodiment, the original speech recognition result can be divided into at least one word segment, the word segment is compared with the reference text, the candidate word segment inconsistent with any reference text is recognized, and the target text is obtained according to the second similarity between the candidate word segment and the reference text, so as to achieve automatic recognition of the target text.

[0099] Figure 4 It is a flowchart of the model training method according to the first embodiment of the present disclosure.

[0100] As Figure 4 shown, the model training method of the first embodiment of the present disclosure includes:

[0101] S401. Obtain training samples, where the training samples include sample original speech recognition results and sample candidate speech recognition results. The sample original speech recognition results include sample target texts, and the training samples further include text labels of the sample target texts. The sample candidate speech recognition results are generated by replacing the sample target texts in the sample original speech recognition results with preset labels.

[0102] It should be noted that the execution subject of the model training method of the present disclosure embodiment is a hardware device with data information processing capabilities and / or the necessary software for driving the hardware device to work. Optionally, the execution subject may include workstations, servers, computers, user terminals, and other intelligent devices. Among them, the user terminal includes but is not limited to mobile phones, computers, intelligent voice interaction devices, intelligent home appliances, vehicle-mounted terminals, etc.

[0103] In the embodiments of the present disclosure, a large number of training samples can be obtained, each of which includes a sample original speech recognition result, a sample candidate speech recognition result, and a text label of a sample target text. The sample original speech recognition result includes the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label.

[0104] It should be noted that the text label is used to indicate whether the sample target text requires error correction and can be set according to actual circumstances. No further restrictions are imposed here. In one embodiment, the text label includes a target text label and a non-target text label, wherein the target text label is used to indicate whether the sample target text requires error correction, and the non-target text label is used to indicate that the sample target text does not require error correction.

[0105] It should be noted that there are no excessive restrictions on the preset tags. For example, the preset tags include but are not limited to <ne> 、 <tt>wait.

[0106] For example, the default label is " <ne>”, the target text label is “T” and the non-target text label is “N”.

[0107] Training sample 1 may include the sample original speech recognition result "Just now, you have introduced the open platform to everyone", the candidate speech recognition result "Just now, <ne>Introduced the open platform to everyone. The sample target text is "You Ping", and the text label of "You Ping" is "T".

[0108] The training sample 2 may include the original speech recognition result of the sample "Just now, You Ping introduced the open platform to everyone.", and the candidate speech recognition result "Just now, You Ping <ne>The sample target text is "Introduction", and the text label of "Introduction" is "N".

[0109] Training sample 3 may include the sample original speech recognition result "Please navigate to Gugong", the candidate speech recognition result "Please navigate to <ne>”, the sample target text is "Gugong", and the text label of "Gugong" is "T".

[0110] The training sample 4 may include the sample original speech recognition result "Please navigate to Gugong", and the candidate speech recognition result "Please <ne>To Gugong", the sample target text is "Navigation", and the text label of "Navigation" is "N".

[0111] S402, training the masked language model based on the training sample, wherein the masked language model is used to obtain a first training score of the sample target text at the sample text position and a second training score of the preset label, the sample text position is the text position of the sample target text in the sample original speech recognition result, the first training score is positively correlated with the predicted probability that the sample text position is the sample target text, and the second training score is positively correlated with the predicted probability that the sample text position is the preset label.

[0112] It should be noted that there are no excessive restrictions on the structure of the masked language model.

[0113] In an embodiment of the present disclosure, a masked language model is used to obtain a first training score for a sample target text at a sample text position and a second training score for a preset label based on a sample original speech recognition result and a sample candidate speech recognition result. The sample text position is the text position of the sample target text in the sample original speech recognition result. The first training score is positively correlated with the predicted probability that the sample text position is the sample target text, and the second training score is positively correlated with the predicted probability that the sample text position is the preset label.

[0114] In one embodiment, training a masked language model based on training samples may include inputting the sample original speech recognition results and the sample candidate speech recognition results into the masked language model, outputting a first training score and a second training score from the masked language model, determining a training target based on the text label, and adjusting the model parameters of the masked language model based on the training target.

[0115] In one embodiment, in response to the text label being the target text label, indicating that the sample target text requires error correction, the model parameters of the masked language model are adjusted with the training target being that the second training score is greater than the first training score. Thus, when the text label is the target text label, the method can ensure that the second training score output by the masked language model is greater than the first training score based on the training target being that the second training score is greater than the first training score.

[0116] For example, training sample 1 may include the sample original speech recognition result "Just now, you have introduced the open platform to everyone", the candidate speech recognition result "Just now, <ne>Introduced the open platform to everyone. The sample target text is "You Ping", and the text label of "You Ping" is "T".

[0117] It can be "Just now You Ping introduced the open platform to everyone", "Just now" <ne>I introduced the open platform "input to the masked language model, and the masked language model outputs the first training score and " <ne>The second training score is 88 points, and the first training score is 20 points. In response to the text label being the target text label, indicating that "there is a comment" needs to be corrected, the model parameters of the masked language model are adjusted with the second training score being greater than the first training score as the training objective.

[0118] In one implementation, in response to the text label being a non-target text label, indicating that the sample target text does not need to be corrected, the model parameters of the masked language model are adjusted with the second training score being less than the first training score as the training objective. Thus, when the text label is a non-target text label, based on the second training score being less than the first training score as the training objective, the masked language model can output a second training score less than the first training score.

[0119] For example, training sample 2 may include the original speech recognition result of the sample "Just now, You Ping introduced the open platform to everyone", and the candidate speech recognition result "Just now, You Ping <ne>The sample target text is "Introduction", and the text label of "Introduction" is "N".

[0120] You can change "Just now I introduced the open platform to you" and "Just now I introduced the open platform to you" <ne>The open platform is input to the masked language model, and the masked language model outputs the first training score and the <ne>"The second training score is 88 points, the first training score is 88 points, and the second training score is 20 points. In response to the text label being a non-target text label, indicating that "introduction" does not need error correction, the second training score is set as the training target to be less than the first training score, and the model parameters of the masked language model are adjusted.

[0121] S403: In response to the model training end condition not being met, returning to continue training the masked language model using the next training sample until the model training end condition is met, thereby generating a target masked language model.

[0122] In an embodiment of the present disclosure, after a model is trained using a training sample, it can be identified whether the model training end condition is met. In response to not meeting the model training end condition, the next training sample is returned to continue training the masked language model until the model training end condition is met. The masked language model obtained from the last training can be used as the target masked language model.

[0123] It should be noted that the model training termination conditions can be set according to actual conditions and are not limited here. For example, the model training termination conditions include but are not limited to the model training times reaching a set number threshold and the model accuracy reaching a set accuracy threshold.

[0124] In summary, according to the model training method of the embodiments of the present disclosure, a masked language model can be trained based on training samples. The training samples include the sample original speech recognition results, the sample candidate speech recognition results, and the text labels of the sample target text. The masked language model is used to obtain a first training score for the sample target text at the sample text location and a second training score for the preset labels. This allows the construction of training samples and the training of the masked language model based on the preset labels, making the construction of training samples and model training simpler and improving model training efficiency.

[0125] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0126] Figure 5 4 is a block diagram of a speech recognition error correction device according to the first embodiment of the present disclosure.

[0127] like Figure 5 As shown, the speech recognition error correction device 500 of the embodiment of the present disclosure includes: an acquisition module 501, a recognition module 502, a replacement module 503 and an error correction module 504.

[0128] The acquisition module 501 is used to obtain the original speech recognition result of the speech information;

[0129] The recognition module 502 is used to recognize that the target text exists in the original speech recognition result;

[0130] The replacement module 503 is used to replace the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result;

[0131] The error correction module 504 is configured to correct errors in the target text based on the original speech recognition result and the candidate speech recognition results.

[0132] In one embodiment of the present disclosure, the error correction module 504 includes: an acquisition unit for acquiring the text position of the target text in the original speech recognition result; an acquisition unit for acquiring a first score of the target text at the text position and a second score of the preset label, wherein the first score is positively correlated with the predicted probability that the text position is the target text, and the second score is positively correlated with the predicted probability that the text position is the preset label; an error correction unit for correcting the target text based on the first score and the second score.

[0133] In one embodiment of the present disclosure, the acquisition unit is further configured to: input the original speech recognition result and the candidate speech recognition result into a target masked language model MLM, and have the target masked language model output the first score and the second score.

[0134] In one embodiment of the present disclosure, the error correction unit is further used to: identify that the first score is less than the second score, obtain a target reference text corresponding to the target text; replace the target text in the original speech recognition result with the target reference text, and generate a target speech recognition result.

[0135] In one embodiment of the present disclosure, the error correction unit is further configured to: obtain a reference text library; obtain a first similarity between the target text and the reference text, and determine the reference text corresponding to the maximum first similarity as the target reference text.

[0136] In one embodiment of the present disclosure, the error correction module 504 is further configured to: identify that the first score is greater than or equal to the second score, and use the original speech recognition result as a target speech recognition result.

[0137] In one embodiment of the present disclosure, the recognition module 502 is further used to: obtain a reference text library; divide the original speech recognition result into at least one segmentation; compare the segmentation with the reference text to identify candidate segmentations that are inconsistent with any reference text; obtain a second similarity between the candidate segmentation and the reference text, and determine the candidate segmentation whose second similarity is greater than a first preset threshold as the target text.

[0138] In one embodiment of the present disclosure, the recognition module 502 is further used to: obtain a first phoneme feature of the candidate word segmentation and a second phoneme feature of the reference text; obtain a third similarity between the first phoneme feature and the second phoneme feature, and use the third similarity as the second similarity.

[0139] In summary, the speech recognition error correction device of the disclosed embodiment recognizes the presence of target text in the original speech recognition result, replaces the target text in the original speech recognition result with a preset label, generates a candidate speech recognition result, and corrects the target text based on the original speech recognition result and the candidate speech recognition result. This allows automatic error correction of speech recognition results based on the preset label, which has good applicability and improves the error correction efficiency and accuracy of speech recognition.

[0140] Figure 6 4 is a block diagram of a model training device according to the first embodiment of the present disclosure.

[0141] like Figure 6 As shown, the model training device 600 of the embodiment of the present disclosure includes: an acquisition module 601 and a training module 602.

[0142] The acquisition module 601 is used to acquire a training sample, wherein the training sample includes a sample original speech recognition result and a sample candidate speech recognition result, the sample original speech recognition result includes a sample target text, the training sample also includes a text label of the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label;

[0143] The training module 602 is used to train the masked language model based on the training sample, wherein the masked language model is used to obtain a first training score of the sample target text at a sample text position and a second training score of the preset label, wherein the sample text position is the text position of the sample target text in the sample original speech recognition result, the first training score is positively correlated with the predicted probability that the sample text position is the sample target text, and the second training score is positively correlated with the predicted probability that the sample text position is the preset label;

[0144] The training module 602 is further configured to, in response to a failure to meet a model training end condition, return to continue training the masked language model using the next training sample until the model training end condition is met, thereby generating a target masked language model.

[0145] In one embodiment of the present disclosure, the training module 602 is further used to: input the sample original speech recognition result and the sample candidate speech recognition result into the masked language model, and the masked language model outputs the first training score and the second training score; in response to the text label being the target text label, the model parameters of the masked language model are adjusted with the second training score being greater than the first training score as the training target; or, in response to the text label being a non-target text label, the model parameters of the masked language model are adjusted with the second training score being less than the first training score as the training target.

[0146] In summary, the model training device of the disclosed embodiments can train a masked language model based on training samples. The training samples include the sample original speech recognition results, the sample candidate speech recognition results, and the text labels of the sample target text. The masked language model is used to obtain a first training score for the sample target text at the sample text location and a second training score for the preset labels. This allows the construction of training samples and the training of the masked language model based on the preset labels, making the construction of training samples and model training simpler and improving model training efficiency.

[0147] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0148] Figure 7 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0149] like Figure 7 As shown, the device 700 includes a computing unit 701, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 702 or a computer program loaded from a storage unit 708 into a random access memory (RAM) 703. Various programs and data required for the operation of the device 700 can also be stored in the RAM 703. The computing unit 701, the ROM 702, and the RAM 703 are connected to each other via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0150] Various components in device 700 are connected to I / O interface 705, including an input unit 706, such as a keyboard, mouse, etc.; an output unit 707, such as various types of displays, speakers, etc.; a storage unit 708, such as a magnetic disk, optical disk, etc.; and a communication unit 709, such as a network card, modem, wireless communication transceiver, etc. The communication unit 709 allows device 700 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0151] The computing unit 701 can be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 701 performs the various methods and processes described above, such as the error correction method for speech recognition. For example, in some embodiments, the error correction method for speech recognition can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded into the RAM 703 and executed by the computing unit 701, one or more steps of the error correction method for speech recognition described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the error correction method for speech recognition by any other appropriate means (e.g., by means of firmware).

[0152] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0153] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0154] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0155] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0156] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0157] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0158] According to an embodiment of the present disclosure, the present disclosure also provides a computer program product, including a computer program, wherein, when the computer program is executed by a processor, it implements the steps of the speech recognition error correction method described in the above embodiment of the present disclosure, or implements the steps of the model training method described in the above embodiment of the present disclosure.

[0159] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0160] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.< / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / tt> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / ne> < / tt> < / ne>

Claims

1. A speech recognition error correction method, comprising: Obtaining the original speech recognition results of the speech information; Identifying that the target text exists in the original speech recognition result; Replacing the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result; Obtaining the text position of the target text in the original speech recognition result; Inputting the original speech recognition result and the candidate speech recognition result into a target masked language model MLM, the target masked language model outputting a first score of the target text at the text position and a second score of the preset label, wherein the first score is positively correlated with the predicted probability that the text position is the target text, and the second score is positively correlated with the predicted probability that the text position is the preset label; Correct the target text based on the first score and the second score.

2. The method according to claim 1, wherein Correcting the target text based on the first score and the second score includes: Identifying that the first score is less than the second score, and acquiring a target reference text corresponding to the target text; The target text in the original speech recognition result is replaced with the target reference text to generate a target speech recognition result.

3. The method according to claim 2, wherein: The obtaining of a target reference text corresponding to the target text includes: Access to a library of reference texts; A first similarity between the target text and a reference text is obtained, and the reference text corresponding to the maximum first similarity is determined as the target reference text.

4. The method according to claim 1, wherein The method further comprises: It is recognized that the first score is greater than or equal to the second score, and the original speech recognition result is used as a target speech recognition result.

5. The method according to any one of claims 1 to 4, wherein: The method further comprises: Access to a library of reference texts; Dividing the original speech recognition result into at least one word segmentation; Comparing the segmented words with reference texts to identify candidate segmented words that are inconsistent with any of the reference texts; A second similarity between the candidate segmented word and the reference text is obtained, and the candidate segmented word whose second similarity is greater than a first preset threshold is determined as the target text.

6. The method according to claim 5, wherein: The obtaining of the second similarity between the candidate word segmentation and the reference text includes: Acquire a first phoneme feature of the candidate word segmentation and a second phoneme feature of the reference text; A third similarity between the first phoneme feature and the second phoneme feature is obtained, and the third similarity is used as the second similarity.

7. A model training method comprising: Obtaining a training sample, wherein the training sample includes a sample original speech recognition result and a sample candidate speech recognition result, the sample original speech recognition result includes a sample target text, the training sample also includes a text label of the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label; The masked language model is trained based on the training sample, wherein the masked language model is used to obtain a first training score of the sample target text at a sample text position and a second training score of the preset label, the sample text position being the text position of the sample target text in the sample original speech recognition result, the first training score is positively correlated with the predicted probability that the sample text position is the sample target text, and the second training score is positively correlated with the predicted probability that the sample text position is the preset label; In response to the model training end condition not being met, returning to use the next training sample to continue training the masked language model until the model training end condition is met, thereby generating a target masked language model.

8. The method according to claim 7, wherein the training of the masked language model based on the training samples comprises: Inputting the sample original speech recognition result and the sample candidate speech recognition result into the masked language model, and having the masked language model output the first training score and the second training score; In response to the text label being a target text label, adjusting the model parameters of the masked language model with the second training score being greater than the first training score as a training target; or In response to the text label being a non-target text label, the model parameters of the masked language model are adjusted with the second training score being smaller than the first training score as a training target.

9. A speech recognition error correction device comprising: An acquisition module is used to obtain the original speech recognition results of the speech information; A recognition module, configured to recognize that a target text exists in the original speech recognition result; A replacement module, configured to replace the target text in the original speech recognition result with a preset label to generate a candidate speech recognition result; An error correction module, configured to correct errors in the target text based on the original speech recognition result and the candidate speech recognition results; Wherein, the error correction module includes: An acquiring unit, configured to acquire a text position of the target text in the original speech recognition result; an acquisition unit, configured to input the original speech recognition result and the candidate speech recognition result into a target masked language model MLM, and output, by the target masked language model, a first score of the target text at the text position and a second score of the preset label, wherein the first score is positively correlated with the predicted probability that the text position is the target text, and the second score is positively correlated with the predicted probability that the text position is the preset label; An error correction unit is configured to correct errors in the target text based on the first score and the second score.

10. The device according to claim 9, wherein The error correction unit is further configured to: Identifying that the first score is less than the second score, and acquiring a target reference text corresponding to the target text; The target text in the original speech recognition result is replaced with the target reference text to generate a target speech recognition result.

11. The device according to claim 10, wherein The error correction unit is further configured to: Access to a library of reference texts; A first similarity between the target text and a reference text is obtained, and the reference text corresponding to the maximum first similarity is determined as the target reference text.

12. The device according to claim 9, wherein The error correction module is further used to: It is recognized that the first score is greater than or equal to the second score, and the original speech recognition result is used as a target speech recognition result.

13. The device according to any one of claims 9 to 12, wherein: The identification module is further used to: Access to a library of reference texts; Dividing the original speech recognition result into at least one word segmentation; Comparing the segmented words with reference texts to identify candidate segmented words that are inconsistent with any of the reference texts; A second similarity between the candidate segmented word and the reference text is obtained, and the candidate segmented word whose second similarity is greater than a first preset threshold is determined as the target text.

14. The device according to claim 13, wherein The identification module is further used to: Acquire a first phoneme feature of the candidate word segmentation and a second phoneme feature of the reference text; A third similarity between the first phoneme feature and the second phoneme feature is obtained, and the third similarity is used as the second similarity.

15. A model training device comprising: an acquisition module, configured to acquire a training sample, wherein the training sample includes a sample original speech recognition result and a sample candidate speech recognition result, the sample original speech recognition result includes a sample target text, the training sample also includes a text label of the sample target text, and the sample candidate speech recognition result is generated by replacing the sample target text in the sample original speech recognition result with a preset label; a training module, configured to train a masked language model based on the training sample, wherein the masked language model is configured to obtain a first training score of the sample target text at a sample text position and a second training score of the preset label, the sample text position being the text position of the sample target text in the sample original speech recognition result, the first training score being positively correlated with a predicted probability that the sample text position is the sample target text, and the second training score being positively correlated with a predicted probability that the sample text position is the preset label; The training module is further configured to, in response to a failure to meet a model training end condition, return to continue training the masked language model using the next training sample until the model training end condition is met, thereby generating a target masked language model.

16. The apparatus according to claim 15, wherein the training module is further configured to: Inputting the sample original speech recognition result and the sample candidate speech recognition result into the masked language model, and having the masked language model output the first training score and the second training score; In response to the text label being a target text label, adjusting the model parameters of the masked language model with the second training score being greater than the first training score as a training target; or In response to the text label being a non-target text label, the model parameters of the masked language model are adjusted with the second training score being smaller than the first training score as a training target.

17. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the speech recognition error correction method described in any one of claims 1-6, or execute the model training method described in any one of claims 7-8.

18. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the speech recognition error correction method as described in any one of claims 1-6, or to execute the model training method as described in any one of claims 7-8.

19. A computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the speech recognition error correction method as described in any one of claims 1 to 6, or implements the steps of the model training method as described in any one of claims 7 to 8.

Images