A method, device, computer equipment and storage medium for optimizing speech recognition

Abstract

The present application discloses a method, device, equipment and storage medium for optimizing speech recognition, which relates to the field of speech recognition technology. The method includes: obtaining audio to be recognized in a target scenario that includes at least one human language; inputting the audio to be recognized into a speech-to-text STT model to obtain all candidate texts converted from the same speech to be converted and the corresponding confidence information; based on the confidence information of the candidate texts and the target high-frequency word library in the target scenario, correcting the converted candidate texts, and outputting the language text converted from the audio to be recognized. The present application loads a high-frequency word library and a dynamic parameter library based on personal language and language ratio in the STT model speech recognition process, which can realize the correction and optimization of speech recognition results according to personal language expression habits in the corresponding scenario, and ensure the accuracy of long-term speech and multi-language recognition.

Description

Technical Field

[0001] The embodiments of the present application relate to speech recognition technology, and more particularly to a speech recognition optimization method, apparatus, computer equipment and storage medium. Background Art

[0002] Speech recognition technology collects external audio and converts human speech into text for output or display. It is widely used in scenarios such as online translation, recording recognition, and artificial intelligence.

[0003] In the relevant technology, in the scenario of long-term speech recognition, the speech recognition technology used mostly adopts the Speech-to-text speech recognition STT model. However, in some long meetings, especially international meetings and multi-person meetings, multilingual expressions and professional vocabulary are involved, or multilingual mixed expressions are used due to personal habits, such as "I don't care about this matter" and "the activity of catalase", which are recognized as mixed outputs of Chinese and English and professional phrases. However, the recognition results of the STT model for this type of speech will have different degrees of deviation, and the recognition accuracy is not high. Summary of the invention

[0004] The present application provides a method, device, computer equipment and storage medium for optimizing speech recognition. The invention solves the problem in the related art that speech recognition accuracy is not high in long-term speech recognition scenarios. The technical solution is as follows:

[0005] In one aspect, a method for optimizing speech recognition is provided, the method comprising:

[0006] Acquire audio to be recognized in a target scene that is composed of at least one human language; the audio to be recognized is obtained by on-site collection or audio import;

[0007] Input the audio to be recognized into a speech-to-text STT model to obtain all candidate texts converted from the same speech to be converted and the corresponding confidence information; wherein the STT model contains all language recognition models in the audio to be recognized;

[0008] Based on the confidence information of the candidate text and the target high-frequency vocabulary in the target scenario, the converted candidate text is corrected, and the language text converted from the audio to be recognized is output; wherein the target high-frequency vocabulary is obtained by converting and extracting historical speech in the same target scenario, and the output language text includes at least one language.

[0009] On the other hand, a speech recognition optimization device is provided, the device comprising:

[0010] A speech acquisition module, used to acquire audio to be recognized in a target scene, which is composed of at least one human language; the audio to be recognized is obtained by on-site collection or audio import;

[0011] A speech recognition module, used to input the audio to be recognized into a speech-to-text STT model, and obtain all candidate texts converted from the same speech to be converted and corresponding confidence information; wherein the STT model contains all language recognition models in the audio to be recognized;

[0012] A text output module is used to correct the converted candidate text based on the confidence information of the candidate text and the target high-frequency vocabulary in the target scenario, and output the language text converted from the audio to be recognized; wherein the target high-frequency vocabulary is obtained by converting and extracting historical speech in the same target scenario, and the output language text includes at least one language.

[0013] On the other hand, a computer device is provided, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the above-mentioned optimization method for speech recognition.

[0014] On the other hand, a computer-readable storage medium is provided, in which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the above-mentioned optimization method for speech recognition.

[0015] On the other hand, a computer program product or a computer program is provided, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the method for optimizing speech recognition described in any one of the above aspects.

[0016] The beneficial effects brought by the above technical solution include at least: the audio to be recognized in the target scene can be selected in a targeted manner, and the STT model of speech recognition integrates all the language recognition models in the audio to be recognized, so that candidate texts in multiple languages ​​can be generated during speech recognition, and the candidate texts are matched with confidence. In the correction stage of the output results, the target high-frequency word library corresponding to the target scene is enabled to correct the recognized candidate texts, because the target high-frequency word library is obtained by converting and extracting historical speech in the same target scene, and multi-language output can be achieved in the correction process, and special words in special scenes can be output, which has higher recognition accuracy than the existing speech recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of a method for optimizing speech recognition provided by an embodiment of the present application;

[0018] Figure 2 It is a schematic diagram of loading a high-frequency word library for speech recognition and optimization provided by an embodiment of the present application;

[0019] Figure 3 is a flow chart of a method for optimizing speech recognition provided by another embodiment of the present application;

[0020] Figure 4 is a schematic diagram of loading a dynamic parameter library and a high-frequency word library provided in an embodiment of the present application;

[0021] Figure 5 It is a schematic diagram of loading and updating personalized parameters based on voiceprint ID provided in an embodiment of the present application;

[0022] Figure 6 is a flow chart of a method for optimizing speech recognition provided by another embodiment of the present application;

[0023] Figure 7 is a flow chart of updating a high-frequency word library provided in an embodiment of the present application;

[0024] Figure 8 It is a structural block diagram of the speech recognition optimization device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0026] The term "multiple" as used herein refers to two or more than two. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.

[0027] In the related technology, speech recognition technology adopts the Speech-to-text method, which mainly trains different types of language models, and the application is mainly monolingual STT. However, in actual application scenarios, many people are accustomed to multilingual mixed expressions. For example, some people prefer to express in Chinese alone, while some people prefer a mixture of Chinese and English. The speech recognition equipment cannot be specially optimized and recognized for individuals, so the effect of speech recognition is not ideal. In addition, for some long-term meeting scenarios, such as multi-person meetings, the conversion method of STT is audio sampling and sentence segmentation. Each paragraph or phrase conversion is independent, and there is no correlation between the previous and the next, and the output text will also be isolated. At present, there is technology that optimizes specific scenarios, but the limitation is that this specific scenario is fixed. For example, in order to optimize the STT method for medical and health-related topics, it is necessary to deliberately collect and organize the keywords of this topic, and compare the text library obtained by special training with word frequency algorithms such as tfidf in the application to improve accuracy. Therefore, this type of technology needs to be trained separately for different scenarios, lacking some versatility.

[0028] Figure 1 This is a flow chart of the method for optimizing speech recognition provided by the embodiment of the present application. Specifically, it includes:

[0029] Step 101: Acquire audio to be recognized in a target scene that includes at least one human language.

[0030] The computer device in this application can be a personal terminal, a PC, a recording translation pen or other devices. The target scenario can be specifically used for personal voice scenarios or multi-person meeting scenarios. The on-site audio is collected as the audio to be recognized, and then it is recognized online on the spot, or the audio is recorded in the target scenario and recognized by importing the audio after the meeting.

[0031] Step 102, input the audio to be recognized into the speech-to-text STT model to obtain all candidate texts for the same speech to be converted and the corresponding confidence information.

[0032] The STT model is a basic model for speech recognition, which includes all language recognition models in the audio to be recognized, such as Chinese, English, French, and Japanese. It is particularly suitable for long-term multi-person speeches at international conferences. By using various language recognition models to perform speech recognition and conversion on the audio to be recognized, candidate texts corresponding to each audio segment are obtained. Different candidate texts correspond to different languages ​​or texts. For example, the candidate texts recognized by the speech recognition device for a certain audio segment are "incredible", "Bruce Lee", and "Bruce Lee". Usually, the speech recognition device will recognize it according to a specific language model, and obtain the confidence information of each candidate text while generating multiple candidate texts. The size of the confidence is used to judge the accuracy of speech recognition. For example, the accuracy of matching text will be higher if the pronunciation is standard.

[0033] Step 103, based on the confidence information of the candidate text and the target high-frequency vocabulary in the target scenario, the converted selected text is corrected, and the language text converted from the audio to be recognized is output.

[0034] For habitual multilingual expressions and professional vocabulary, the speech recognition model has recognition bias, and the results of single sentence recognition cannot be associated with the previous and next sentences, and the recognition results are isolated. Simply relying on the confidence value will lead to language recognition bias and recognition errors. The present invention establishes a high-frequency word library in the speech recognition process, and uses the high-frequency word library and confidence information for dual identification. Figure 2 As shown, the high-frequency word library is the high-frequency words recognized and collected in the historical audio through multiple language models. It selectively collects keywords according to personal language habits and conference topics. For example, specific medical vocabulary is collected during medical conferences, and recognition and correction can be performed according to the collected word library content during subsequent speeches. And the high-frequency word library can contain multiple types, such as medical word library, biological word library and other specific word libraries. The target high-frequency word library is selected according to the determined target scene, and the corrected language and text are finally output. The language and text can contain one or more language words. In addition, the word library can also judge the recognized language and text in real time, update the word library database, etc., for subsequent recognition and correction.

[0035] In summary, in the embodiments of the present application, the audio to be recognized in the target scene can be selected in a targeted manner, and the STT model of speech recognition integrates all the language recognition models in the audio to be recognized, so that candidate texts in multiple languages ​​can be generated during speech recognition, and the candidate texts are matched accordingly with confidence. In the correction stage of the output results, the target high-frequency vocabulary corresponding to the target scene is enabled to correct the recognized candidate texts, because the target high-frequency vocabulary is obtained by converting and extracting historical speech in the same target scene, and multi-language output can be achieved during the correction process, and special vocabulary in special scenes can be output, which has higher recognition accuracy than the existing speech recognition results.

[0036] Figure 3 : is a flow chart of a method for optimizing speech recognition provided by another embodiment of the present application, which specifically includes the following steps:

[0037] Step 301: Acquire audio to be recognized in a target scene that includes at least one human language.

[0038] Step 302: When the target scene is a single-person scene, a dynamic parameter library under the target scene and personalized parameter information of speech recognition contained in the dynamic parameter library are obtained.

[0039] Taking the voice recorder used for speech recognition in this method as an example, the voice recorder can collect external audio and display the speaker's language and text through the display screen. When the voice recorder is used for personal online speech recognition. For example, when a game anchor records a game tutorial, the target scene of the voice recorder can be set to a single-person scene in advance, that is, for personal recognition, and the voice recorder will obtain the dynamic parameter library under the scene and the personalized parameter information of speech recognition contained in the dynamic parameter library. The dynamic parameter library is generated when voice conversion is performed based on the audio to be converted obtained in the historical target scene, that is, the personalized parameter information generated by the anchor in the previous game recording process. The personalized parameter information includes at least one of language preference, language ratio and vocabulary type. That is, according to the language habits of the anchor and the ratio of each language in the sentence, a personal data model is established, and the model and parameters are directly called for speech recognition in the subsequent process. For example, through the historical speech recognition process, it is detected that the anchor's language is Chinese, but there is a mixed habit of Chinese and English, with the proportion of Chinese at 80% and the proportion of English at 20%. The personalized parameter information is used for subsequent screening and correction of candidate texts.

[0040] Step 303: When the target scene is a multi-person scene, the acquired audio to be converted is input into a voiceprint recognition model for voiceprint recognition, and the voiceprint ID of the speech is determined according to the audio frequency.

[0041] In addition to the above-mentioned use for personal identification, when used in international conferences and multi-person conferences, in order to improve the accuracy of multi-person identification, a temporary parameter library can be established. Figure 4 and Figure 5 As shown, in a multi-person conference scenario, unique personalized parameters are established based on the audio frequency characteristics of each person. Assuming a multi-person conference consisting of n people participating in the speech, n groups of personalized parameters for n people are established in the temporary parameter library. During speech recognition, speech recognition is performed based on the personalized parameters of each person. The temporary data is directly released from the memory after the meeting to avoid occupying space. "Audio recognition of people" can be determined by voiceprint recognition. The voiceprint recognition model first identifies the sound frequency of the input audio to be converted, and determines each person's voiceprint ID based on the audio frequency, and then establishes their respective personalized parameter information based on the voiceprint ID. During the recognition process, the personalized parameters corresponding to the voiceprint ID are extracted from the dynamic parameter library, and then the output is corrected.

[0042] Step 304: Obtain a dynamic parameter library corresponding to the voiceprint ID and personalized parameter information of speech recognition contained in the dynamic parameter library. Different voiceprint IDs correspond to their own personalized parameter information.

[0043] The personalized parameter information based on the voiceprint ID includes each person's language expression habits, including languages ​​and language ratios, etc. It should be noted that in a multi-person conference scenario, the same conference topic is usually expressed, so the same set of temporary vocabulary is used, which can be released at the end of the meeting and does not occupy memory space.

[0044] In some embodiments, a voiceprint ID can also be established for the holder of the voice recorder. After the voiceprint ID is confirmed to be the holder in a single-person scenario, the personalized parameter library and voice recognition can be loaded to improve data security and privacy. The personal parameter library and personal vocabulary library are different from the temporary parameter library and temporary vocabulary library. They can be permanently saved and loaded multiple times for subsequent voice recognition, and the recognition accuracy is higher.

[0045] Step 305, arrange the confidences of the converted candidate characters in descending order to obtain the first confidence, the second confidence to the nth confidence, and obtain the candidate characters ranked in the top m in confidence.

[0046] The STT model contains multiple language recognition models. For any segment of the audio to be converted, the recognition result will output multiple candidate texts, covering multiple languages ​​and texts. Before outputting the target text, it is necessary to sort them according to the confidence value to obtain the first confidence, the second confidence to the nth confidence. In order to reduce the amount of calculation, you can set a threshold or the number of candidate texts. For example, select the top m candidate texts with the highest confidence from n. Subsequent operations are to screen the top m candidate texts with the highest confidence.

[0047] In some embodiments, the confidence levels may also be normalized, that is, the sum of the first m confidence levels is 1, to facilitate subsequent comparison processing.

[0048] Step 306: In response to the first confidence being less than the confidence threshold, a target high-frequency word library corresponding to the target scene is acquired.

[0049] The confidence level is used to indicate the accuracy of speech conversion. When the confidence level is higher than a certain ratio, it indicates that the recognition accuracy meets the requirements. For example, if the confidence threshold is set to 0.5, when the first confidence level exceeds 0.5, the sum of the second confidence level to the mth confidence level is only 0.5, that is, the accuracy of the candidate text conversion corresponding to the first confidence level far exceeds the accuracy of other candidate texts, and should be directly used as the output. However, when the maximum first confidence level is 0.4, the second confidence level is 0.35, indicating that the speech pronunciation is inaccurate and the conversion accuracy is not high. Since the first two confidence levels are similar, no matter the first two candidate texts are different texts of the same language, or different texts of different languages, the accuracy of the conversion cannot be guaranteed. At this time, the target high-frequency vocabulary corresponding to the target scene should be obtained, and further judgment should be made through the target high-frequency vocabulary.

[0050] Step 307, traverse the high-frequency words in the target high-frequency word library, match them one by one with the candidate words ranked in the top m in confidence, and screen them according to the personalized parameter information to obtain the target conversion text.

[0051] The high-frequency word library is screened and optimized by weighted calculation, that is, matching high-frequency words, and then making a weighted calculation on the confidence of the candidate words that are successfully matched. However, in order to improve the accuracy of speech recognition and the accuracy of multi-language mixed recognition, the confidence can also be weighted according to personalized parameters. For example, the confidence of English in a language with mixed Chinese and English expressions can be weighted. Figure 6 As shown, step 307 also includes the following steps:

[0052] Step 307A, obtaining the language preference and language ratio in the target scenario, and performing a first weighted calculation on the confidence according to a first weighted coefficient of the confidence in the target language.

[0053] Indicatively, for the matched voiceprint ID and personalized parameters, determine the target language type of the speaker, that is, the language type with the highest proportion in the historical speech. For example, when the language read (in a multi-person meeting, a temporary parameter library is established for the first speech) is Chinese, a mixed Chinese and English expression is used, with a Chinese ratio of 70% and an English ratio of 30%, that is, the frequency of the speaker's expression in Chinese mixed with English. It is necessary to perform a first weighted calculation on the confidence that the candidate text is Chinese (target language).

[0054] In a possible implementation manner, the first weighting coefficient may be obtained by nonlinear calculation, for example, by logarithmic operation, and the calculation formula of the first weighting coefficient is:

[0055] M=log B A

[0056] Wherein, M represents the first weighting coefficient, A represents the percentage of the target language, and B represents the threshold value for the first weighted calculation. Because in the actual speech recognition scenario, when multiple languages ​​are involved and the proportions are similar (for example, 30%, 30% and 40%), in order to ensure the rationality of the personalized parameter optimization and avoid the phenomenon of overfitting prediction, a threshold value needs to be set, and the optimization mechanism is triggered only when the proportion of the target language exceeds the threshold value.

[0057] In this solution, the threshold value is set to 60, which means that when the target language needs to exceed 60%, the confidence of the target language is first weighted. B A=log 60 74=105.12%, which means the confidence of the candidate characters of Chinese is multiplied by 105.12%.

[0058] When the second candidate text is Chinese, the first weighting coefficient is set to 105.12%, and the corresponding second confidence is 0.4, then the confidence after the first weighted calculation = 0.4*105.12% = 0.4208. When the first confidence of the first candidate text is 0.4 and it is English, the second candidate text should be selected. When all the candidate texts are in the same language, there is no need to perform the first weighting operation. Continue to step 307B.

[0059] Step 307B, respectively match the first m candidate characters with the high-frequency words in the target high-frequency word library, and perform a second weighted calculation on the confidence of the successfully matched candidate characters.

[0060] The personalized parameters are to optimize the output multilingual coherence according to personal habits, and the purpose of the high-frequency word library is to improve the recognition accuracy of fuzzy words and professional words, avoid the isolation of output text, and also establish and improve professional word libraries. For the temporary word library of a multi-person meeting, it is established based on the speech topics of multiple former speakers of the meeting, including high-frequency words that appear during speeches. This process matches the first m and candidate texts with the high-frequency words of the target high-frequency word library one by one. For the candidate texts that are successfully matched, the confidence is calculated by the second weighted calculation (including the confidence after the first weighted calculation). For example, in a single input, the highest confidence ranking of a word is A (database), with a confidence of 0.4, and the second is B (dimethylbenzene ring), with a confidence of 0.38. However, because B is in the target high-frequency word library, it is weighted by 10%, so 0.38*110%>0.4, which will cause the output to be corrected to B (dimethylbenzene ring) output. For the candidate texts that are not successfully matched, the second weighted calculation is not performed, and the output is directly output according to the result of step 307A.

[0061] Step 307C, comparing the confidences of the candidate characters that have been weighted and those that have not been weighted, and determining the candidate character with the highest confidence as the target conversion character.

[0062] After the first weighted calculation and the second weighted calculation, the calculated confidences are compared again, and the candidate text with the highest execution degree is used as the target conversion text.

[0063] Step 308, correcting the converted candidate text based on the target conversion text, and outputting the language text of the audio conversion to be recognized, the language text including at least one language.

[0064] The above embodiment is a process of performing correction and optimization according to personalized parameters and high-frequency vocabulary during the speech recognition conversion process, wherein the personalized parameters and high-frequency vocabulary also need to be updated in real time during the speech recognition process. Figure 7 This is a flowchart of updating a high-frequency word library provided by an embodiment of the present application, which specifically includes the following steps:

[0065] Step 701, count the number of times the same voice text is converted after voice conversion to obtain the corresponding conversion confidence.

[0066] like Figure 5 As shown, for a single-person scenario, since it is the operation behavior of the holder himself, the high-frequency word library and personalized parameters can be permanently saved, corresponding to the personal parameter library and personal word library. For a multi-person conference scenario, due to the existence of contingency, a temporary storage method is adopted. During a long speech process, the temporary word library is gradually improved and established based on multiple speakers, and the same speaker will also continuously adjust the corresponding personalized parameters during the speech process. Regardless of the target scenario, after obtaining the target conversion text, it is necessary to count the language, language ratio and converted voice and text of the sentence. The purpose of counting the same voice and text is to determine the high-frequency words, and in order to ensure the accuracy of the statistics, it is also necessary to record the corresponding conversion confidence (not the confidence calculated by weighted calculation). The purpose is to only include words with accurate and clear pronunciation to improve the accuracy of subsequent corrections.

[0067] Step 702, in response to the number of spoken text reaching a statistical threshold, calculate the cumulative confidence of the spoken text.

[0068] In a possible implementation, the number of times the voice text reaches a statistical threshold of 3 in a multi-person or single-person scenario can be set, which means that the same word is mentioned 3 times in total during this recognition process, and the conversion confidence of the word is summed up to obtain the cumulative confidence, and whether to include the word is determined based on the cumulative confidence. However, it should be noted that the included voice text must be nouns and / or entity words that meet the statistical conditions to avoid word library redundancy.

[0069] Step 703, in response to the cumulative confidence exceeding the cumulative threshold, the spoken text is determined as a high-frequency word and entered into a target high-frequency word library.

[0070] Indicatively, the cumulative threshold is set to 2.5, and word A appears 3 times in total during recognition, with confidence levels of 0.85, 0.95, and 0.82, respectively. Then 0.85+0.95+0.82>2.5, and word A can be saved in the target high-frequency word library.

[0071] In other embodiments, a threshold may be set for the conversion confidence, that is, the conversion confidence needs to exceed a certain threshold (such as the threshold is set to 0.5 to ensure the accuracy of recognition) before statistics can be performed.

[0072] Step 704, determine the number and proportion of each language in the language text after the current audio conversion, and update the language preference, language proportion and vocabulary type in the personalized parameter information according to the number and proportion of the language words in the historical voice conversion content.

[0073] For the converted language and text, the language, word count and proportion of each person's speech are counted, and the language preference and language proportion in the personalized parameter information are updated according to the language, word count and proportion of the historical voice conversion content. For example, in special scenarios, if the proportion of English in the voice content of an individual or speaker increases, the first weighting coefficient is increased accordingly to optimize the accuracy of mixed Chinese and English speeches.

[0074] This application introduces a high-frequency word library and a personalized parameter library based on the use of the STT model, and uses voiceprint recognition technology to identify the voiceprint ID in multi-person scenarios, and uses the voiceprint ID to establish a temporary parameter library to optimize the voice recognition of personal behavior; for single-person scenarios, a personal parameter library is directly established for the optimization of personal voice recognition;

[0075] During speech recognition, the confidence level of candidate words is judged, and the first weighted calculation is performed to optimize the output in combination with the language and proportion in the personalized parameters; for high-frequency words that are successfully matched, the second weighted calculation is performed on the confidence level to improve the recognition accuracy of ambiguous words and professional words;

[0076] At the same time, this solution also updates the high-frequency word library and personalized parameters synchronously through the recognition results of historical speech, so as to achieve rapid recognition when switching personal language expressions in different scenarios. Compared with the method of directly outputting according to the confidence level of STT model conversion in related technologies, the optimized adaptation scenario is more extensive, and it can support mixed output of multiple languages, and the accuracy of the corrected output is higher.

[0077] Figure 8 : is a structural block diagram of a speech recognition optimization device provided in an embodiment of the present application, the device comprising:

[0078] The speech acquisition module 801 is used to acquire the audio to be recognized in the target scene, which is composed of at least one human language; the audio to be recognized is obtained by on-site collection or audio import;

[0079] The speech recognition module 802 is used to input the audio to be recognized into a speech-to-text STT model to obtain all candidate texts converted from the same speech to be converted and the corresponding confidence information; wherein the STT model contains all language recognition models in the audio to be recognized;

[0080] The text output module 803 is used to correct the converted candidate text based on the confidence information of the candidate text and the target high-frequency vocabulary in the target scenario, and output the language text converted from the audio to be recognized; wherein the target high-frequency vocabulary is obtained by converting and extracting historical speech in the same target scenario, and the output language text includes at least one language.

[0081] In an embodiment of the present application, a computer device is also provided, which includes a processor and a memory; the memory stores at least one instruction, and the at least one instruction is used to be executed by the processor to implement the optimization method of speech recognition provided by the above-mentioned various method embodiments.

[0082] In an embodiment of the present application, a computer-readable storage medium is also provided, in which at least one instruction, at least one program, a code set or an instruction set is stored. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the optimization method of speech recognition provided by the above-mentioned method embodiments.

[0083] The preferred embodiments of the present invention are described above; it should be understood that the present invention is not limited to the above-mentioned specific embodiments, and the devices and structures not described in detail should be understood to be implemented in a common manner in the art; any technician familiar with the art can make many possible changes and modifications without departing from the technical solution of the present invention, or modify them into equivalent embodiments with equivalent changes, which does not affect the essential content of the present invention; therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of protection of the technical solution of the present invention.

Claims

1. A method for optimizing speech recognition, characterized in that: The method comprises: Acquire audio to be recognized in a target scene that is composed of at least one human language; the audio to be recognized is obtained by on-site collection or audio import; Input the audio to be recognized into a speech-to-text STT model to obtain all candidate texts converted from the audio to be converted and the corresponding confidence information; wherein the STT model contains all language recognition models in the audio to be recognized; Based on the confidence information of the candidate text and the target high-frequency word library in the target scenario, the converted candidate text is corrected, and the language text converted from the audio to be recognized is output; the target high-frequency word library is obtained by historical speech conversion and extraction in the same target scenario, and the output language text includes at least one language; specifically, the confidences of the converted candidate texts are arranged in descending order to obtain a first confidence, a second confidence to an nth confidence, and the candidate texts ranked in the top m in confidence are obtained; wherein different confidences correspond to candidate texts of the same or different languages; In response to the first confidence being less than a confidence threshold, the target high-frequency word library corresponding to the target scenario is obtained; the high-frequency words in the target high-frequency word library are traversed, and the candidate words ranked in the top m in confidence are matched one by one, and the target conversion words are obtained by screening according to the personalized parameter information; the personalized parameter information includes at least one of language preference, language proportion and word library type; the language preference and language proportion in the target scenario are specifically obtained, and the confidence is first weighted calculated according to the first weighting coefficient of the confidence in the target language; wherein different language proportions correspond to different first weighting coefficients; After performing the first weighted calculation, the first m candidate characters are matched with the high-frequency words in the target high-frequency word library respectively, and the confidence of the successfully matched candidate characters is further subjected to the second weighted calculation; the confidence of the successfully matched candidate characters corresponds to the second weighted coefficient; Comparing the confidences of the candidate characters that have been weighted and those that have not been weighted, and determining the candidate character with the highest confidence as the target conversion character; The candidate characters after conversion are corrected based on the target conversion characters, and the language characters of the audio conversion to be recognized are output.

2. The method according to claim 1, characterized in that When the target scene is a single-person scene, the input audio to be recognized is the voice information of the same person; Before inputting the audio to be recognized into the speech-to-text STT model, the method further includes: A dynamic parameter library under the target scenario and personalized parameter information of speech recognition contained in the dynamic parameter library are obtained; wherein the dynamic parameter library is generated when speech conversion is performed on the audio to be converted obtained under the historical target scenario.

3. The method according to claim 2, characterized in that When the target scene is a multi-person scene, the input audio to be recognized is voice information of multiple people: Before inputting the audio to be recognized into the speech-to-text STT model, the method further includes: Inputting the acquired audio to be converted into a voiceprint recognition model for voiceprint recognition, and determining the voiceprint ID of the speech according to the audio frequency; The dynamic parameter library corresponding to the voiceprint ID and the personalized parameter information of speech recognition contained in the dynamic parameter library are obtained; wherein different voiceprint IDs correspond to respective personalized parameter information.

4. The method according to claim 2 or 3, characterized in that: After obtaining the target conversion text, the method further includes: Counting the number of times the same voice text is used after voice conversion, and obtaining the corresponding conversion confidence; the voice text is a noun and / or entity word that meets the statistical conditions; In response to the number of times the voiced text reaches a statistical threshold, calculating the cumulative confidence of the voiced text; In response to the cumulative confidence exceeding a cumulative threshold, the spoken text is determined as a high-frequency word and entered into the target high-frequency word library.

5. The method according to claim 4, characterized in that After outputting the language and text converted from the audio to be recognized, the method further includes: The number and proportion of words in each language in the language text after the current audio conversion are determined, and the language preference, language proportion and vocabulary type in the personalized parameter information are updated according to the number and proportion of words in the language of the historical voice conversion content.

6. A speech recognition optimization device, characterized in that: The device comprises: A speech acquisition module, used to acquire audio to be recognized in a target scene, which is composed of at least one human language; the audio to be recognized is obtained by on-site collection or audio import; A speech recognition module, used to input the audio to be recognized into a speech-to-text STT model, and obtain all candidate texts converted from the same audio to be converted and corresponding confidence information; wherein the STT model contains all language recognition models in the audio to be recognized; A text output module, used to correct the candidate text after conversion based on the confidence information of the candidate text and the target high-frequency word library in the target scenario, and output the language text converted from the audio to be recognized; the target high-frequency word library is obtained by converting and extracting the historical speech in the same target scenario, and the output language text includes at least one language; specifically, the confidences of the candidate text after conversion are arranged in descending order to obtain a first confidence, a second confidence to an nth confidence, and the candidate text ranked in the top m in confidence; wherein different confidences correspond to candidate texts of the same or different languages; In response to the first confidence being less than a confidence threshold, acquiring the target high-frequency vocabulary corresponding to the target scene; Traversing the high-frequency words in the target high-frequency word library, matching them one by one with the candidate words ranked in the top m in confidence, and screening them according to personalized parameter information to obtain target conversion words; the personalized parameter information includes at least one of language preference, language ratio and word library type; specifically, obtaining the language preference and language ratio in the target scenario, and performing a first weighted calculation on the confidence according to a first weighted coefficient of the confidence in the target language; wherein different language ratios correspond to different first weighted coefficients; After performing the first weighted calculation, the first m candidate characters are matched with the high-frequency words in the target high-frequency word library respectively, and the confidence of the successfully matched candidate characters is further subjected to the second weighted calculation; the confidence of the successfully matched candidate characters corresponds to the second weighted coefficient; Comparing the confidences of the candidate characters that have been weighted and those that have not been weighted, and determining the candidate character with the highest confidence as the target conversion character; The candidate characters after conversion are corrected based on the target conversion characters, and the language characters of the audio conversion to be recognized are output.

7. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the optimization method for speech recognition as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the optimization method for speech recognition as described in any one of claims 1 to 5.

Images