Speech-to-text acquisition method and device
The cross-validation method among multiple ASR systems addresses inaccuracies and slow speeds in speech recognition by calculating edit distances and weights, enhancing accuracy and speed in speech-to-text processing.
Patent Information
- Application Number
- CN202210842959.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-18
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-07-18
AI Technical Summary
Existing speech recognition systems face inaccuracies and slow execution speeds due to the need for extensive post-processing of multiple ASR system outputs, limiting their applicability in various scenarios.
A method and device that utilize a cross-validation approach among multiple ASR systems by calculating edit distances, weights, and loss functions to select the most accurate speech recognition text, incorporating scoring and iterative weight updates to enhance accuracy and speed.
Improves the accuracy and speed of speech-to-text processing by leveraging multiple ASR systems, ensuring high applicability and efficient text output.
Smart Images

Figure CN115223564B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of speech and text recognition, and in particular to a speech and text acquisition method and device. Background Art
[0002] Speech recognition technology is also known as Automatic Speech Recognition (ASR), which is a technology that converts the vocabulary content in human speech into computer-readable input, such as keystrokes, binary codes or character sequences (such as text). To achieve speech recognition, a common approach is to use an ASR system to convert audio into text. Due to factors such as audio quality, language habits and the specific implementation method of the ASR system, the ASR system cannot guarantee that the speech recognition text is completely correct. Based on this, there is currently a use of multiple ASR systems to retain the same parts of the multiple speech recognition texts converted by multiple ASR systems for the speech recognition texts of different ASR systems, while the different parts of the multiple speech recognition texts need to be further processed according to some custom rules.
[0003] The above-mentioned customized rules are usually considered to be set and include a large number of judgment operations, and the application of such solutions still requires further modification of the text after the ASR system recognizes the speech recognition text, resulting in slow execution speed of such solutions and few applicable scenarios. Summary of the invention
[0004] The purpose of the present invention is to provide a method and device for acquiring speech text, so as to solve the above-mentioned problems existing in the prior art.
[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] In a first aspect, at least one embodiment of the present application provides a method for acquiring speech text, comprising: for a first audio, acquiring the edit distance between any two speech recognition texts in a plurality of speech recognition texts, the first audio being any one of the plurality of audios, and the plurality of speech recognition texts being speech recognition texts recognized by the first audio via a plurality of speech recognition systems; for the first audio, calculating the sum of the edit distances of each speech recognition system in the plurality of speech recognition systems; for the first audio, calculating the value of a first weight of each speech recognition system based on the sum of the edit distances, the first weight being an initial weight; calculating the loss function value of each speech recognition system based on the value of a second weight and a plurality of edit distances, and selecting the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, the second weight being the first weight or an updated weight, and the updated weight being determined based on the number of output texts corresponding to each speech recognition system.
[0007] For example, in a method provided according to at least one embodiment of the first aspect of the present invention, the value of the loss function of each speech recognition system is calculated based on the value of the second weight and multiple edit distances, and the speech recognition text corresponding to the speech recognition system with the smallest loss function value is selected as the output text of the first audio, including: counting the number of output texts corresponding to each speech recognition system based on the output texts corresponding to multiple audios, and the output texts corresponding to each speech recognition system in the multiple speech recognition systems constitute the output texts corresponding to multiple audios; using the number of output texts corresponding to each speech recognition system, updating the second weight to obtain an updated weight; judging whether the difference between the value of the updated weight and the value of the second weight is less than or equal to a preset threshold; when the judgment result is yes, stopping updating the second weight, and selecting the output texts corresponding to the current multiple audios as the final output text; and when the judgment result is no, repeatedly calculating the loss function value of each speech recognition system and updating the second weight.
[0008] For example, in the method provided according to at least one embodiment of the first aspect of the present invention, before calculating the loss function value of each speech recognition system based on the value of the second weight and multiple edit distances, and selecting the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, the method also includes: scoring the multiple speech recognition texts based on the length and edit distance of the speech recognition texts to obtain scoring scores; and discarding one or more audios with scoring scores less than a preset score value based on the scoring score of each speech recognition text in the multiple speech recognition texts.
[0009] For example, in the method provided according to at least one embodiment of the first aspect of the present invention, scoring multiple speech recognition texts includes: scoring the multiple speech recognition texts using a first scoring function or a second scoring function, wherein the first scoring function is used to reflect the maximum error between the multiple speech recognition texts, and the second scoring function is used to reflect the median value of the errors between the multiple speech recognition texts.
[0010] For example, in the method provided according to at least one embodiment of the first aspect of the present invention, the first weight of each speech recognition system represents the proportion of the sum of the edit distances of other speech recognition systems other than the sum of the edit distances of the corresponding speech recognition system in the sum of the edit distances of all speech recognition systems, wherein, the higher the proportion of the sum of the edit distances of a speech recognition system in the sum of the edit distances of all speech recognition systems, the smaller the value of the first weight of the speech recognition system.
[0011] In a second aspect, at least one embodiment of the present invention provides a speech text acquisition device, comprising: an acquisition unit, for acquiring, for a first audio, an edit distance between any two speech recognition texts in a plurality of speech recognition texts, wherein the first audio is any one of the plurality of audios, and the plurality of speech recognition texts are speech recognition texts recognized by the first audio via a plurality of speech recognition systems; a first calculation unit, for calculating, for the first audio, the sum of the edit distances of each speech recognition system in the plurality of speech recognition systems; a second calculation unit, for calculating, for the first audio, a first weight value of each speech recognition system according to the sum of the edit distances, wherein the first weight is an initial weight; and an output unit, for calculating, for the first audio, a loss function value of each speech recognition system according to the value of the second weight and a plurality of edit distances, and selecting the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, wherein the second weight is the first weight or an updated weight, and the updated weight is determined according to the number of output texts corresponding to each speech recognition system.
[0012] For example, in a speech-text acquisition device provided according to at least one embodiment of the second aspect of the present invention, the text output unit is used to: count the number of output texts corresponding to each speech recognition system according to the output texts corresponding to the multiple audios, the output texts corresponding to each speech recognition system in the multiple speech recognition systems constituting the output texts corresponding to the multiple audios; update the second weight using the number of output texts corresponding to each speech recognition system to obtain the updated weight; determine whether the difference between the value of the updated weight and the value of the second weight is less than or equal to a preset threshold; when the determination result is yes, stop updating the second weight, and select the output text corresponding to the current multiple audios as the final output text; and when the determination result is no, repeatedly calculate the loss function value of each speech recognition system and update the second weight.
[0013] For example, in a speech text acquisition device provided according to at least one embodiment of the second aspect of the present invention, the device also includes: a scoring unit, used to score the multiple speech recognition texts according to the length and edit distance of the speech recognition texts to obtain scoring scores; and according to the scoring score of each speech recognition text in the multiple speech recognition texts, discarding one or more audios whose scoring scores are less than a preset score value.
[0014] In a third aspect, at least one embodiment of the present invention provides a speech text acquisition device, including a processor and a memory, wherein the memory stores instructions executable by the processor, and when the instructions are executed by the processor, the processor executes any speech text acquisition method as described in the first aspect above.
[0015] In a fourth aspect, at least one embodiment of the present application provides a computer-readable storage medium, which stores a program code for execution by a device, wherein the program code includes a method for executing any one of the methods in the first aspect described above.
[0016] The beneficial effects of the present invention are:
[0017] The present invention discloses a method and device for acquiring speech text. The method adopts a cross-validation method of multiple ASR systems for the same audio. Compared with the traditional speech text acquisition method, the method can not only improve the accuracy of the output text, but also has higher applicability and can make the speech text processing speed faster. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a flow chart of a method for acquiring speech text provided by the present invention;
[0019] Figure 2 is a schematic diagram of a speech text acquisition device provided by the present invention;
[0020] Figure 3 It is a schematic diagram of another voice text acquisition device provided by the present invention. DETAILED DESCRIPTION
[0021] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the described embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without creative work all belong to the protection scope of the present invention.
[0022] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. The words "A system", "B system", "C system" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different identification systems. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.
[0023] This embodiment provides a method for acquiring speech-to-text, which acquires speech-to-text by cross-validating a plurality of ASR systems, thereby ensuring the accuracy of the speech-to-text, having high applicability and fast processing speed.
[0024] Figure 1 The figure is a flow chart of a method for acquiring speech text according to an embodiment of the present invention.
[0025] exist Figure 1 The voice-based acquisition method includes the following steps:
[0026] Step S110, for the first audio, obtaining the edit distance between any two speech recognition texts in multiple speech recognition texts, the first audio is any one of the multiple audios, and the multiple speech recognition texts are speech recognition texts recognized by the first audio via multiple speech recognition systems.
[0027] Step S120: For the first audio, calculate the sum of the edit distances of each speech recognition system in the plurality of speech recognition systems.
[0028] Step S130: for the first audio, calculate the value of the first weight of each speech recognition system according to the sum of the edit distances, where the first weight is the initial weight.
[0029] Step S140, calculate the loss function value of each speech recognition system according to the value of the second weight and multiple edit distances, and select the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, the second weight is the first weight or the updated weight, and the updated weight is determined according to the number of output texts corresponding to each speech recognition system.
[0030] The following describes the implementation of the speech text acquisition method in the present invention by taking the multiple speech recognition systems as three speech recognition systems as an example.
[0031] Multiple speech recognition systems are 3 ASR systems: ASR A 、ASR B 、ASR C (hereinafter also referred to as A, B and C systems respectively). It can be understood that the multiple speech recognition systems can include up to 3 ASR systems, for example 5, 6, 10, etc.
[0032] Assuming there are 1,000 audio data, these three ASR systems recognize the 1,000 audio data and obtain a total of 3,000 speech recognition texts.
[0033] For each audio i, the edit distances between the three speech recognition texts recognized by the ASR system are calculated and the text length of each speech recognition text is obtained. i is a positive integer. There are three edit distances in total, namely, i_AB 、dis i_BC 、dis i_AC .dis i_AB Indicates the ASR of the i-th audio A and ASR B The edit distance between two speech recognition texts recognized by the system. The larger the edit distance, the greater the difference between the two speech recognition texts. The method for obtaining or calculating the edit distance in the present invention can refer to the existing methods and is not limited here. The three text lengths are denoted as len(A), len(B) and len(C) respectively.
[0034] Optionally, all speech recognition texts obtained from each audio recognition may be scored based on the text length of the obtained speech recognition text and the edit distance of the speech recognition text. Each speech recognition text has 2 edit distances.
[0035] The score indicates the degree of certainty of the ASR system's recognition of the audio. The higher the score, the more similar the recognized texts obtained by the three ASR systems, which also shows that the result of the audio being recognized as text is relatively certain. The high degree of certainty of the ASR system's recognition of the audio also indicates that the quality of the audio is relatively high (that is, the audio is easy to be correctly recognized by the ASR system), thus ensuring that the results obtained by using any ASR system for recognition are almost the same.
[0036] Optionally, for the scoring process, the available scoring functions are as follows:
[0037]
[0038] In the above formula The meaning of is: the case where the edit distance per unit length is the largest. This formula can simply and directly reflect the worst error of the three recognized texts of audio i, so the higher the score, the higher the quality of the audio.
[0039] For example, for audio i:
[0040] The recognition result of system A: We will be neighbors from now on.
[0041] The recognition result of system B: from now on it will be 09.
[0042] The recognition result of C system: we will be neighbors from now on.
[0043] In the above formula So the score is 4 / 7 = 0.57. However, as can be seen from the above text, the actual recognition text is very certain and should not be scored like this.
[0044] Therefore, in a more preferred embodiment, the score can be calculated using the following formula:
[0045]
[0046]
[0047] Using the above formula to calculate the recognition result of audio i in the previous example, we can get:
[0048] dis i_AB =3,dis i_AC =0,dis i_BC =3. len(A)=8, len(B)=7, len(C)=8.
[0049] Substituting into the formula we get:
[0050]
[0051] The final score of this audio is scorei=1-median(0, 3 / 7, 0)=1. Obviously, the second scoring formula can get better results.
[0052] After all 1000 audios have been scored, they can be screened. Since a higher score indicates a higher audio quality, and vice versa, audios with scores lower than a preset score can be discarded or manually annotated. For example, the preset score is 0.6, and among the 1000 audios, the number of audios with scores less than or equal to the score is 974, and 26 audios are discarded.
[0053] For all n audios, for example, 1000 audios or 974 audios, the edit distance of each speech recognition text has been obtained. Where n is a positive integer. For these n audios, the edit distance of each ASR system and all other ASR systems is accumulated to obtain the sum of the edit distance of each ASR system. The calculation formula is as follows:
[0054]
[0055]
[0056]
[0057] The initial weight of each system can be calculated based on the sum of the edit distances. The calculation formula is as follows:
[0058]
[0059]
[0060]
[0061] ∝ A , ∝ B , ∝ C are the initial weights of system A, system B, and system C. Since the magnitude of all weights needs to be unified, they are normalized, and bias is a bias that can keep the weights consistent in magnitude during subsequent weight iterations. For example, bias is equal to 0, 0.05, etc.
[0062] For each audio i, the loss function of each system is calculated using the following formula:
[0063]
[0064]
[0065]
[0066] loss iA , loss iB , loss iC These are the loss functions of system A, system B, and system C. The scoring in the previous article is to evaluate which recognized text of the same audio is the most reliable among the recognized texts of different ASR systems, so the higher the score, the better. The loss function represents the recognition effect of different systems on the same audio, and the smaller the value, the better.
[0067] You can select according to the formula i =min(lossiA , loss iB , loss iC ), select the speech recognition text of the ASR system with the smallest loss function value for the i-th audio as the speech recognition text of the audio.
[0068] The following process can be iterated:
[0069] For the recognized text after n audios are recognized, the number of audios of the recognized texts of each ASR system is counted. Based on the statistical results, the updated weights ∝′ of each ASR system can be calculated using the following formula: A , ∝′ B ,∝′ C :
[0070]
[0071]
[0072]
[0073] n = select_num A +select_num B +select_num C , select_num A select_num is the number of audio files for text recognition of system A. B and select_num C are the number of audios for the recognized texts of the selection B and C systems, respectively.
[0074] The bias in this formula has the same value as the bias in the initial weight calculation formula. Its function is to prevent the denominator from being 0 when calculating the loss function.
[0075] For example, n = 900, select_num A =180,select_num B =270,select_num C =450, bias = 0.01. So the updated weight ∝′ A =0.05, ∝′ B =0.10, ∝′ C =0.26.
[0076] Compare the initial weights and updated weights of each ASR system. If the difference is not big, stop the iteration and take the current result directly as the final speech recognition result. If the difference is too big, exceeding a certain threshold, such as 5% or 0.01, repeat the above iterative process until the initial weights and updated weights converge, stop and output the output text selected for all audios when the weights converge as the final output text.
[0077] Optionally, the initial weight values of each ASR system may be set to 1. In this case, the weights will eventually converge, but at a slower speed.
[0078] This embodiment provides a device for acquiring speech text. Figure 2 A schematic diagram of a speech text acquisition device provided according to an embodiment of the present invention is shown.
[0079] exist Figure 2 In the invention, the speech text acquisition device 200 includes: an acquisition unit 210, which is used to acquire the edit distance between any two speech recognition texts in a plurality of speech recognition texts for a first audio, wherein the first audio is any one of the plurality of audios, and the plurality of speech recognition texts are speech recognition texts recognized by the first audio via a plurality of speech recognition systems; a first calculation unit 220, which is used to calculate the sum of the edit distances of each speech recognition system in the plurality of speech recognition systems for the first audio; a second calculation unit 230, which is used to calculate the value of a first weight of each speech recognition system according to the sum of the edit distances for the first audio, wherein the first weight is an initial weight; and an output unit 240, which is used to calculate the loss function value of each speech recognition system according to the value of the second weight and the plurality of edit distances, and select the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, wherein the second weight is the first weight or an updated weight, and the updated weight is determined according to the number of output texts corresponding to each speech recognition system.
[0080] Optionally, the text output unit 240 is used to count the number of output texts corresponding to each speech recognition system based on the output texts corresponding to the multiple audios, and the output texts corresponding to each speech recognition system in the multiple speech recognition systems constitute the output texts corresponding to the multiple audios; using the number of output texts corresponding to each speech recognition system, update the second weight to obtain the updated weight; determine whether the difference between the value of the updated weight and the value of the second weight is less than or equal to a preset threshold; when the judgment result is yes, stop updating the second weight and select the current output text corresponding to the multiple audios as the final output text; and when the judgment result is no, repeat the calculation of the loss function value of each speech recognition system and update the second weight.
[0081] Optionally, the speech text acquisition device also includes: a scoring unit, used to score the multiple speech recognition texts according to the length and edit distance of the speech recognition texts to obtain scoring scores; and discard one or more audios whose scoring scores are less than a preset score value according to the scoring score of each speech recognition text in the multiple speech recognition texts.
[0082] It can be understood that the speech text acquisition device 200 can implement the speech text acquisition method described in the above method embodiment. For details, please refer to the above description, which will not be repeated here.
[0083] This embodiment provides another device for acquiring speech text. Figure 3 A schematic diagram of a speech text acquisition device provided according to an embodiment of the present invention is shown.
[0084] like Figure 3 As shown, the speech text acquisition device 300 includes a processor 301 and a memory 302. The memory 302 stores instructions executable by the processor 301. When the processor 301 executes the instructions, the processor 301 executes the speech text acquisition method described in the above method embodiment.
[0085] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:
[0086] The present invention discloses a method and device for acquiring speech text. The device adopts a cross-validation method of multiple ASR systems for the same audio. Compared with the traditional speech text acquisition method, the device can not only improve the accuracy of the output text, but also has higher applicability and can make the speech text processing speed faster.
[0087] The various illustrative logic blocks, modules, circuits, and algorithmic operations described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and operations have been described above generally in terms of their functions. Whether such functions are implemented as hardware or software depends on the specific application and the design constraints imposed on the entire system. A skilled person may implement the described functions in different ways for each specific application, but such implementation decisions should not be interpreted as causing a deviation from the scope of the claims.
[0088] The hardware for implementing the various illustrative logics, logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general-purpose processor, a digital signal processor (DSP), a dedicated application integrated circuit (TCUASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in an alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in combination with a DSP core, or any other such configuration. Alternatively, some operations or methods may be performed by circuits specific to a given function.
[0089] In one or more embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored as one or more instructions or codes on a non-transitory computer-readable medium or a non-transitory processor-readable medium. The operation of the method or algorithm disclosed herein may be embodied in a processor-executable software module, which may reside on a non-transitory computer-readable or processor-readable storage medium. A non-transitory computer-readable or processor-readable storage medium may be any storage medium accessible by a computer or processor. As an example and not limitation, such non-transitory computer-readable or processor-readable media may include RAM, ROM, EEPROM, flash memory, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the required program code in the form of instructions or data structures and can be accessed by a computer. As used herein, disks and discs include compact discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks, and Blu-ray discs, where the discs typically reproduce data magnetically, and the discs reproduce data optically by lasers. The above combinations are also included within the scope of non-transitory computer-readable and processor-readable media. Furthermore, the operations of a method or algorithm may reside as one or any combination or set of codes and / or instructions on a non-transitory processor-readable medium and / or computer-readable medium, which may be incorporated into a computer program product.
[0090] The foregoing description of the disclosed embodiments is provided to enable any person skilled in the art to formulate or use the claims. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments without departing from the scope of the claims. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be consistent with the widest scope consistent with the claims and the principles and novel features disclosed herein.
Claims
1. A method for acquiring speech text, characterized in that: include: For a first audio, obtaining an edit distance between any two speech recognition texts in a plurality of speech recognition texts, wherein the first audio is any one of the plurality of audios, and the plurality of speech recognition texts are speech recognition texts recognized by the first audio via a plurality of speech recognition systems; For the first audio, calculating a sum of the edit distances of each speech recognition system in the plurality of speech recognition systems; For the first audio, calculate a first weight value of each speech recognition system according to the sum of the edit distances, where the first weight is an initial weight; Calculate the loss function value of each speech recognition system according to the value of the second weight and multiple edit distances, and select the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, wherein the second weight is the first weight or an updated weight, and the updated weight is determined according to the number of output texts corresponding to each speech recognition system; The step of calculating the loss function value of each speech recognition system according to the value of the second weight and the plurality of edit distances, and selecting the speech recognition text corresponding to the speech recognition system having the smallest loss function value as the output text of the first audio includes: According to the output texts corresponding to the multiple audios, counting the number of output texts corresponding to each speech recognition system, the output texts corresponding to each speech recognition system in the multiple speech recognition systems constituting the output texts corresponding to the multiple audios; Using the number of output texts corresponding to each speech recognition system, updating the second weight to obtain the updated weight; Determine whether a difference between the value of the updated weight and the value of the second weight is less than or equal to a preset threshold; When the judgment result is yes, stop updating the second weight, and select the output text corresponding to the current multiple audios as the final output text; and When the judgment result is no, the loss function value of each speech recognition system is repeatedly calculated and the second weight is updated.
2. The method for acquiring speech text according to claim 1, characterized in that: Before calculating the loss function value of each speech recognition system according to the value of the second weight and the plurality of edit distances, and selecting the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, the method further includes: Scoring the plurality of speech recognition texts according to the length and the edit distance of the speech recognition texts to obtain scoring scores; According to the scoring score of each of the multiple speech recognition texts, one or more audios whose scoring score is less than a preset score value are discarded.
3. The method for acquiring speech text according to claim 2, characterized in that: The step of scoring the plurality of speech recognition texts comprises: The plurality of speech recognition texts are scored using a first scoring function or a second scoring function, wherein the first scoring function is used to reflect the maximum error between the plurality of speech recognition texts, and the second scoring function is used to reflect the median value of the error between the plurality of speech recognition texts.
4. The method for acquiring speech text according to claim 1, characterized in that: The first weight of each speech recognition system represents the proportion of the sum of edit distances of other speech recognition systems except the sum of edit distances of the corresponding speech recognition system in the sum of edit distances of all speech recognition systems, wherein the higher the proportion of the sum of edit distances of a speech recognition system in the sum of edit distances of all speech recognition systems, the smaller the value of the first weight of the speech recognition system.
5. A speech text acquisition device, characterized in that: include: an acquiring unit, configured to acquire, for a first audio, an edit distance between any two speech recognition texts in a plurality of speech recognition texts, wherein the first audio is any one of the plurality of audios, and the plurality of speech recognition texts are speech recognition texts recognized by the first audio via a plurality of speech recognition systems; A first calculation unit, configured to calculate, for the first audio, a sum of edit distances of each speech recognition system in the plurality of speech recognition systems; A second calculation unit is used to calculate a first weight value of each speech recognition system according to the sum of the edit distances for the first audio, where the first weight is an initial weight; and The output unit is used to calculate the loss function value of each speech recognition system according to the value of the second weight and the multiple edit distances, and select the speech recognition text corresponding to the speech recognition system with the smallest loss function value as the output text of the first audio, wherein the second weight is the first weight or the updated weight, and the updated weight is determined according to the number of output texts corresponding to each speech recognition system; The text output unit is used for: According to the output texts corresponding to the multiple audios, counting the number of output texts corresponding to each speech recognition system, the output texts corresponding to each speech recognition system in the multiple speech recognition systems constituting the output texts corresponding to the multiple audios; Using the number of output texts corresponding to each speech recognition system, updating the second weight to obtain the updated weight; Determine whether a difference between the value of the updated weight and the value of the second weight is less than or equal to a preset threshold; When the judgment result is yes, stop updating the second weight, and select the output text corresponding to the current multiple audios as the final output text; as well as When the judgment result is no, the loss function value of each speech recognition system is repeatedly calculated and the second weight is updated.
6. The speech text acquisition device according to claim 5, characterized in that: The device also includes: A scoring unit is used to score the multiple speech recognition texts according to the length and edit distance of the speech recognition texts to obtain scoring scores; and according to the scoring score of each speech recognition text in the multiple speech recognition texts, discard one or more audios whose scoring scores are less than a preset score value.
7. A speech text acquisition device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions to execute the method according to any one of claims 1 to 4.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes for execution by a device, the program codes including instructions for executing the method according to any one of claims 1 to 4.
