Speech recognition model optimization method and system

Abstract

The invention provides a voice recognition model optimization method and system, and belongs to the technical field of voice recognition, and the method specifically comprises the steps: employing the similar condition of the voiceprint characteristics of a voiceprint similar keyword, and determining that there is no deviation keyword whose recognition deviation probability does not meet the requirement when the use data of a matched service scene is determined to have no deviation keyword. Based on the use data of the deviation keywords in the business scene, determining the distribution data of the deviation keywords under different identification processing times, determining the identification accuracy under different identification processing times in combination with the identification deviation probabilities of the different deviation keywords, and based on the identification accuracy under different identification processing times, determining the identification accuracy of the deviation keywords under different identification processing times. And determining whether the business scene and the matched business scene can be subjected to unified voice recognition model establishment or not, thereby improving the accuracy of voice recognition processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech recognition, and in particular relates to a speech recognition model optimization method and system. Background Art

[0002] Speech recognition technology has been widely used in online conferences, online classes, voice customer service and other fields. Specifically, in the invention patent application CN202410595424.0 "An automatic speech recognition method and system for power grid communication dispatching", the first speech recognition result and the second speech recognition result are feature-fused to obtain the target speech recognition result of the speech data to be recognized. The present invention improves the recognition efficiency of power grid communication dispatching speech.

[0003] However, the above technical solutions ignore the following technical problems: In power companies, there are a large number of power-specific terms in voice recognition scenarios such as meeting recordings, dispatch instructions, and telephone recordings. The use of general voice recognition models not only has poor recognition accuracy, but also cannot distinguish voiceprints, making the information extraction results inaccurate.

[0004] In response to the above technical problems, the present invention provides a speech recognition model optimization method and system. Summary of the invention

[0005] To achieve the purpose of the present invention, the present invention adopts the following technical solutions: In order to solve the above technical problems, the present invention provides the following technical solutions to achieve the purpose of the present invention: According to one aspect of the present invention, a speech recognition model optimization method is provided.

[0006] A speech recognition model optimization method, specifically comprising: S1 determines deviation keywords based on the deviation of the keyword library between the business scenarios of speech recognition, and when it is determined that there are business scenarios with similar keywords through the usage data of the deviation keywords in the business scenarios, the business scenarios are used as matching business scenarios; S2: using the voiceprint features of the deviation keywords of the business scenario and the matching business scenario, determining the voiceprint similar keywords in the keywords of the matching business scenario, and using the similarity of the voiceprint features of the voiceprint similar keywords and the usage data of the matching business scenario to determine that there is no deviation keyword whose recognition deviation probability does not meet the requirements, proceed to the next step; S3 determines the distribution data of the deviation keywords under different recognition processing times based on the usage data of the deviation keywords in the business scenario, and determines the recognition accuracy under different recognition processing times in combination with the recognition deviation probabilities of different deviation keywords; S4 determines whether a unified speech recognition model can be constructed for the business scenario and the matching business scenario based on the recognition accuracy under different recognition processing times.

[0007] The beneficial effects of the present invention are: The similarity of the voiceprint features of voiceprint-similar keywords and the usage data in the matching business scenarios are used to determine whether there are deviation keywords whose recognition deviation probability does not meet the requirements, thereby achieving accurate assessment of the recognition deviation probability of different deviation keywords based on the similarity of the voiceprint features of voiceprint-similar keywords in the matching business scenarios and the frequency of use, avoiding the technical problem of low recognition accuracy caused by the use of a unified speech recognition model for business scenarios and matching business scenarios due to the existence of deviation keywords with a large recognition deviation probability, thereby ensuring the recognition accuracy and reliability of the speech recognition model.

[0008] Based on the recognition accuracy under different recognition processing times, it is determined whether a unified speech recognition model can be built for the business scenario and the matching business scenario. This avoids the technical problem of low recognition accuracy of a unified speech recognition model due to a large number of recognition processing times with low recognition accuracy in the business scenario. While reducing the difficulty of building a speech recognition model, it also ensures the recognition accuracy in different business scenarios.

[0009] A further technical solution is that the business scenarios include conference recording, scheduling instructions, and telephone recording.

[0010] A further technical solution is that the keyword library is constructed based on the analysis results of historical data in the business scenario.

[0011] A further technical solution is that the deviation keywords are keywords in the business scenario that have never been used in other business scenarios.

[0012] A further technical solution is that the usage data includes the number of times it is used in the business scenario.

[0013] A further technical solution is to determine whether a unified speech recognition model can be built, including: Determine the number of recognition processing times of the deviated keywords and use it as the number of deviated recognition processing times; Based on the recognition accuracy under different recognition deviation processing times, normalization processing is performed to obtain the accuracy weight coefficients of different recognition deviation processing times; The matching coefficient between the business scenario and the matching business scenario is determined by summing the accuracy weight coefficients of different recognition deviation processing times, and based on the matching coefficient, it is determined whether a unified speech recognition model can be built.

[0014] A further technical solution is that when the matching coefficient is less than a preset matching coefficient, it is determined that a unified speech recognition model cannot be constructed.

[0015] In a second aspect, the present invention provides a computer system comprising: a memory and a processor that are communicatively connected, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-mentioned speech recognition model optimization method when running the computer program.

[0016] Other features and advantages will be described in the following description, and partly become apparent from the description, or understood by practicing the invention. The purpose and other advantages of the invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings.

[0019] Figure 1 is a flow chart of a method for optimizing a speech recognition model; Figure 2 is a flow chart of a method for determining a probability of deviation in identifying a deviation keyword; Figure 3 is a flow chart of a method for determining a recognition accuracy rate under a recognition processing number; Figure 4 It is a framework diagram of a computer system. DETAILED DESCRIPTION

[0020] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that the present invention will be comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar structures, and thus their detailed description will be omitted.

[0021] The terms "a", "an", "the", and "said" are used to indicate the presence of one or more elements / components / etc.; the terms "comprising" and "having" are used to express an open-ended inclusive meaning and mean that additional elements / components / etc. may be present in addition to the listed elements / components / etc.

[0022] Example 1 To solve the above problems, according to one aspect of the present invention, Figure 1 According to one aspect of the present invention, a method for optimizing a speech recognition model is provided, which specifically includes: S1 determines deviation keywords based on the deviation of the keyword library between the business scenarios of speech recognition, and when it is determined that there are business scenarios with similar keywords through the usage data of the deviation keywords in the business scenarios, the business scenarios are used as matching business scenarios; S2: using the voiceprint features of the deviation keywords of the business scenario and the matching business scenario, determining the voiceprint similar keywords in the keywords of the matching business scenario, and using the similarity of the voiceprint features of the voiceprint similar keywords and the usage data of the matching business scenario to determine that there is no deviation keyword whose recognition deviation probability does not meet the requirements, proceed to the next step; S3 determines the distribution data of the deviation keywords under different recognition processing times based on the usage data of the deviation keywords in the business scenario, and determines the recognition accuracy under different recognition processing times in combination with the recognition deviation probabilities of different deviation keywords; S4 determines whether a unified speech recognition model can be constructed for the business scenario and the matching business scenario based on the recognition accuracy under different recognition processing times.

[0023] Furthermore, the business scenarios include conference recording, scheduling instructions, and telephone recording.

[0024] Optionally, the keyword library is built based on the analysis results of historical data in the business scenario.

[0025] It should be noted that the deviation keywords are keywords in the business scenario that have never been used in other business scenarios.

[0026] It is understandable that the usage data includes the number of times of usage in the business scenario.

[0027] It should be noted that the business scenarios that are determined to have similar keywords include: Determine the specific business scenario and the deviation keywords in the business scenario; Determine the usage times of different deviation keywords in the business scenarios based on usage data of different deviation keywords in the business scenarios, and build a preset mapping model based on the usage times to determine usage frequency coefficients of different deviation keywords; The frequently used keywords in the deviation keywords are determined based on the frequent use coefficient, and the number of the frequently used keywords is used to determine whether the specific business scenario is a business scenario with keywords similar to the business scenario.

[0028] Further, when the number of the frequently used keywords is greater than the preset number of keywords, it is determined that the specific business scenario does not belong to a business scenario with similar keywords to the business scenario.

[0029] Optionally, determining that there are business scenarios with similar keywords in the business scenarios includes steps S11-S13, which specifically include: S11 determines the proportion of the deviation keywords in the keyword library in the business scenario based on the specific business scenario and the deviation keywords in the business scenario, and determines the keyword deviation coefficient of the specific business scenario based on the proportion and the number of deviation keywords; S12: determining the usage times of different deviation keywords in the business scenarios based on usage data of different deviation keywords in the business scenarios, and constructing a preset mapping model based on the usage times to determine usage frequency coefficients of different deviation keywords; S13 determines the usage deviation coefficient of the specific business scenario through the usage frequency coefficients of different deviation keywords and the keyword deviation coefficient of the specific business scenario, and uses the usage deviation coefficient to determine whether the specific business scenario is a business scenario with similar keywords to the business scenario.

[0030] Further, the usage deviation coefficient of the specific business scenario is determined by using the usage frequency coefficients of different deviation keywords and the keyword deviation coefficient of the specific business scenario, specifically including: The usage deviation coefficient of the specific business scenario is determined based on the average value of the usage frequency coefficients of different deviation keywords and the product of the keyword deviation coefficient of the specific business scenario.

[0031] It should be noted that, when the usage deviation coefficient is not within the preset deviation interval, it is determined that the specific business scenario does not belong to a business scenario with similar keywords to the business scenario.

[0032] Optionally, the above step S11 includes steps S111 to S113, which are specifically: S111 determines the number of deviation keywords based on a specific business scenario and deviation keywords in the business scenario. When the number of deviation keywords is not within a preset number range, the process proceeds to step S113. When the number of deviation keywords is within the preset number range, the process proceeds to step S112. S112 determines the proportion of the deviation keyword in the keyword library in the business scenario. When the proportion of the deviation keyword in the keyword library in the business scenario meets the requirement, it is determined that the specific business scenario belongs to a business scenario with similar keywords to the business scenario. When the proportion of the deviation keyword in the keyword library in the business scenario meets the requirement or does not meet the requirement, it proceeds to step S113. S113 determines the keyword deviation coefficient of the specific business scenario based on the proportion and the number of deviation keywords. When the keyword deviation coefficient does not meet the requirements, it is determined that the specific business scenario does not belong to a business scenario with similar keywords to the business scenario. When the number of deviation keywords meets the requirements, proceed to step S12.

[0033] Optionally, the above step S12 includes steps S121 to S123, which are specifically: S121 determines the usage times of different deviation keywords in the business scenario based on the usage data of different deviation keywords in the business scenario, and determines the total usage times of the deviation keywords in the business scenario based on the usage times. When the total usage times do not meet the requirements, it is determined that the specific business scenario does not belong to a business scenario with similar keywords to the business scenario. When the total usage times meet the requirements, it proceeds to step S122. S122 determines the deviation keywords whose usage times are greater than the preset usage times based on the usage times of different deviation keywords in the business scenario. When there is no deviation keyword whose usage times are greater than the preset usage times, it is determined that the specific business scenario belongs to a business scenario similar to the keywords of the business scenario. When there is a deviation keyword whose usage times are greater than the preset usage times, the process proceeds to step S123. S123 constructs a preset mapping model based on the number of times used, and determines the frequency coefficients of use of different deviation keywords. When the number of deviation keywords whose frequency coefficients do not meet the requirements is greater than the preset number of keywords, it is determined that the specific business scenario belongs to a business scenario with similar keywords to the business scenario. When the number of deviation keywords whose frequency coefficients do not meet the requirements is not greater than the preset number of keywords, proceed to step S13.

[0034] Furthermore, the voiceprint feature of the deviation keyword is determined according to the analysis result of the voiceprint of the deviation keyword.

[0035] It can be understood that the voiceprint similar keyword is a keyword whose deviation amount from the voiceprint feature of the deviation keyword is within a preset range.

[0036] Specifically, the deviation amount of the voiceprint feature from the deviation keyword is determined according to a Euclidean distance function or a Mahalanobis distance function.

[0037] Optionally, the determination of the recognition deviation probability of the deviation keyword includes steps S21-S23, specifically: S21, determining a similarity coefficient of the voiceprint features of the voiceprint similar keywords based on the similarity of the voiceprint features of the voiceprint similar keywords; S22 determines the number of times the voiceprint similar keywords are used in the matching business scenario based on the usage data of different voiceprint similar keywords in the matching business scenario, and determines the frequency of use coefficients of different voiceprint similar keywords based on the usage number; S23 determines the recognition deviation interference coefficients of different voiceprint similar keywords based on the usage frequency coefficients of different voiceprint similar keywords and the similarity coefficients of voiceprint features, and determines the recognition deviation probability of the deviation keyword through the recognition deviation interference coefficients of different voiceprint similar keywords.

[0038] Optionally, before entering step S21, it is also necessary to determine whether the number of voiceprint similar keywords is greater than the preset number of keywords. When the number of voiceprint similar keywords is greater than the preset number of keywords, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirements. When the number of voiceprint similar keywords is not greater than the preset number of keywords, proceed to step S21.

[0039] Optionally, the above step S21 includes steps S211-S212, which are specifically: S211 determines the similarity coefficient of the voiceprint features of the voiceprint similar keywords based on the similarity of the voiceprint features of the voiceprint similar keywords. When the number of voiceprint similar keywords with similarity coefficients greater than the preset similarity coefficient does not meet the requirement, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirement. When the number of voiceprint similar keywords with similarity coefficients greater than the preset similarity coefficient meets the requirement, the process proceeds to step S212. S212 determines a comprehensive similarity coefficient based on the number of voiceprint similar keywords and the similarity coefficients of the voiceprint features of different voiceprint similar keywords. When the comprehensive similarity coefficient does not meet the requirements, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirements. When the comprehensive similarity coefficient meets the requirements, it proceeds to step S22.

[0040] Optionally, the above step S22 includes steps S221 and S222, which are specifically: S221 determines the number of times the voiceprint similar keywords are used in the matching business scenario through the usage data of different voiceprint similar keywords in the matching business scenario, and uses the usage number to determine the usage frequency coefficients of different voiceprint similar keywords. When the number of voiceprint similar keywords with usage frequency coefficients greater than a preset frequency coefficient threshold does not meet the requirement, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirement. When the number of voiceprint similar keywords with usage frequency coefficients greater than the preset frequency coefficient threshold meets the requirement, the process proceeds to step S222. S222 determines a comprehensive frequency coefficient based on the number of voiceprint similar keywords and the frequency coefficients of voiceprint features of different voiceprint similar keywords. When the comprehensive frequency coefficient does not meet the requirements, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirements. When the comprehensive frequency coefficient meets the requirements, it proceeds to step S23.

[0041] Optionally, in the above step S23, it is also necessary to determine whether there are voiceprint similar keywords whose recognition deviation interference coefficient does not meet the requirements. When there are voiceprint similar keywords whose recognition deviation interference coefficient does not meet the requirements, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirements. When there are no voiceprint similar keywords whose recognition deviation interference coefficient does not meet the requirements, the recognition deviation probability of the deviation keyword is determined by using the recognition deviation interference coefficients of different voiceprint similar keywords.

[0042] Furthermore, the recognition deviation interference coefficient of the voiceprint similar keyword is determined according to the product of the usage frequency coefficient of the voiceprint similar keyword and the similarity coefficient of the voiceprint feature.

[0043] Specifically, the value range of the recognition deviation probability of the deviation keyword is between 0 and 1. When the recognition deviation probability of the deviation keyword is greater than a preset probability threshold, it is determined that the recognition deviation probability of the deviation keyword does not meet the requirement.

[0044] It should be noted that when there are deviation keywords whose recognition deviation probability does not meet the requirements, it is determined that the business scenario and the matching business scenario cannot use the same speech recognition model.

[0045] Further, such as Figure 2 As shown, the method for determining the recognition deviation probability of the deviation keyword is: Determine the similarity coefficient of the voiceprint features of the voiceprint similar keywords based on the similarity of the voiceprint features of the voiceprint similar keywords, and determine the voiceprint similarity coefficient using the average value of the similarity coefficients; Determine the number of times the voiceprint similar keyword is used in the matching business scenario through the usage data of different voiceprint similar keywords in the matching business scenario, determine the total number of times the voiceprint similar keyword is used in the matching business scenario based on the number of times the voiceprint similar keyword is used in the matching business scenario, and determine the recognition processing frequency coefficient according to the total number of times; The recognition deviation probability of the deviation keyword is determined by multiplying the voiceprint similarity coefficient and the recognition processing frequency coefficient.

[0046] Specifically, Figure 3 As shown, the method for determining the recognition accuracy under the recognition processing times is: Based on the distribution data of the deviation keywords under the recognition processing times, determining the deviation keywords under the recognition processing times, and using them as matching deviation keywords; The recognition accuracy rate under the number of recognition processing times is determined based on the weights of the recognition deviation probabilities of different matching deviation keywords.

[0047] It should be noted that determining whether a unified speech recognition model can be built specifically includes: Determine the number of recognition processing times of the deviated keywords and use it as the number of deviated recognition processing times; Based on the recognition accuracy under different recognition deviation processing times, normalization processing is performed to obtain the accuracy weight coefficients of different recognition deviation processing times; The matching coefficient between the business scenario and the matching business scenario is determined by summing the accuracy weight coefficients of different recognition deviation processing times, and based on the matching coefficient, it is determined whether a unified speech recognition model can be built.

[0048] Furthermore, when the matching coefficient is less than a preset matching coefficient, it is determined that a unified speech recognition model cannot be constructed.

[0049] Optionally, determine whether a unified speech recognition model can be built, specifically: Determine the number of recognition processing times of the deviation keywords, and use it as the number of deviation recognition processing times. When the number of deviation recognition processing times does not meet the requirements, it is determined that the unified speech recognition model cannot be built; When the deviation identification processing times meet the requirements: Based on the recognition accuracy rates under different recognition deviation processing times, when it is determined that there is no deviation recognition processing times with a recognition accuracy rate less than a preset accuracy rate, it is determined that a unified speech recognition model can be constructed; When there is a deviation in the number of recognition processing times where the recognition accuracy is less than the preset accuracy: When the number of deviation recognition processing times with a recognition accuracy rate lower than a preset accuracy rate does not meet the requirement, it is determined that a unified speech recognition model cannot be constructed; When the recognition accuracy is less than the preset accuracy, and the number of deviation recognition processing times meets the requirements: Determine a basic recognition deviation amount according to the number of deviation recognition processing times and the proportion of the deviation recognition processing times in the recognition processing times in the business scenario; when the basic recognition deviation amount does not meet the requirement, determine that a unified speech recognition model cannot be constructed; When the basic identification deviation meets the requirements: Based on the recognition accuracy under different recognition deviation processing times, the recognition deviation processing times are divided into different recognition accuracy intervals, and the interval recognition deviation amounts of different recognition accuracy intervals are determined according to the recognition deviation processing times of different recognition accuracy intervals; The recognition deviation is determined by the interval recognition deviations of different recognition accuracy intervals, and based on the recognition deviation, it is determined whether a unified speech recognition model can be constructed.

[0050] Example 2 Second, as Figure 4 As shown, the present invention provides a computer system, comprising: a memory and a processor that are communicatively connected, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-mentioned speech recognition model optimization method when running the computer program.

[0051] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0052] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0053] The above description is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included in the scope of the claims of this specification.

Claims

1. A speech recognition model optimization method, characterized in that: Specifically include: Determine deviation keywords based on the deviation of the keyword library between the business scenarios of speech recognition, and when it is determined that there are business scenarios with similar keywords through the usage data of the deviation keywords in the business scenarios, use them as matching business scenarios; Determine the voiceprint similar keywords in the keywords of the matching business scenario by using the voiceprint features of the deviation keywords of the business scenario and the matching business scenario, and proceed to the next step when it is determined that there is no deviation keyword whose recognition deviation probability does not meet the requirements by using the similarity of the voiceprint features of the voiceprint similar keywords and the usage data of the matching business scenario; Based on the usage data of the deviation keyword in the business scenario, the distribution data of the deviation keyword under different recognition processing times is determined, and the recognition accuracy under different recognition processing times is determined in combination with the recognition deviation probabilities of different deviation keywords; Based on the recognition accuracy under different recognition processing times, it is determined whether a unified speech recognition model can be built for the business scenario and the matching business scenario.

2. The speech recognition model optimization method according to claim 1, characterized in that: The business scenarios include conference recording, scheduling instructions, and telephone recording.

3. The method for optimizing a speech recognition model according to claim 1, wherein: The keyword library is built based on the analysis results of historical data in the business scenario.

4. The method for optimizing a speech recognition model according to claim 1, wherein: The deviation keywords are keywords in the business scenario that have never been used in other business scenarios.

5. The method for optimizing a speech recognition model according to claim 1, wherein: Determine whether there are business scenarios with similar keywords, including: Determine the specific business scenario and the deviation keywords in the business scenario; Determine the usage times of different deviation keywords in the business scenarios based on usage data of different deviation keywords in the business scenarios, and build a preset mapping model based on the usage times to determine usage frequency coefficients of different deviation keywords; The frequently used keywords in the deviation keywords are determined based on the frequent use coefficient, and the number of the frequently used keywords is used to determine whether the specific business scenario is a business scenario with keywords similar to the business scenario.

6. The method for optimizing a speech recognition model according to claim 5, wherein: When the number of the frequently used keywords is greater than the preset number of keywords, it is determined that the specific business scenario does not belong to a business scenario with similar keywords to the business scenario.

7. The method for optimizing a speech recognition model according to claim 1, wherein: The voiceprint feature of the deviation keyword is determined according to the analysis result of the voiceprint of the deviation keyword.

8. The method for optimizing a speech recognition model according to claim 1, wherein: The voiceprint similar keyword is a keyword whose deviation amount from the voiceprint feature of the deviation keyword is within a preset range.

9. The method for optimizing a speech recognition model according to claim 1, wherein: Determine whether a unified speech recognition model can be built, including: Determine the number of recognition processing times of the deviated keywords and use it as the number of deviated recognition processing times; Based on the recognition accuracy under different recognition deviation processing times, normalization processing is performed to obtain the accuracy weight coefficients of different recognition deviation processing times; The matching coefficient between the business scenario and the matching business scenario is determined by summing the accuracy weight coefficients of different recognition deviation processing times, and based on the matching coefficient, it is determined whether a unified speech recognition model can be built.

10. A computer system comprising: A memory and a processor that are communicatively connected, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes a speech recognition model optimization method as described in any one of claims 1-9 when running the computer program.

Images