Method and apparatus for measuring uncertainty of speech recognition model

Abstract

One embodiment of the present disclosure provides a method for measuring uncertainty in a speech recognition model. The method comprises the steps of: generating a predicted text by converting a first output data, which is the result of recognizing input speech data using a first speech recognition model, into text; calculating a first uncertainty measurement value based on the predicted text and the first output data using a pre-set probability estimation algorithm; calculating a second uncertainty measurement value based on the predicted text and a second output data, which is the result of recognizing speech data using a second speech recognition model with dropout applied to the first speech recognition model, using the probability estimation algorithm; and calculating a final uncertainty estimate value based on the first uncertainty measurement value and the second uncertainty measurement value.

Description

Technology Field

[0001] The present invention relates to a method and apparatus for measuring uncertainty in a speech recognition model, and more specifically, to a technology for measuring uncertainty by converting speech data recognized through a speech recognition model into text. Background Technology

[0002] With the advent of smart devices, voice-based technologies are advancing. Technologies that recognize speech and convert it into text are being used in various ways, such as voice memos, voice commands, and conversing with AI via voice.

[0003] Measuring the uncertainty of speech recognition models has been studied from various perspectives. In the case of conventional encoder-decoder based speech recognition models, since the predicted value is output directly in the form of text, uncertainty can be estimated by utilizing the confidence of each text sequence. However, in encoder-based speech recognition models, the output sequence contains blank tokens—special tokens used in CTC algorithms—and the prediction sequence is not directly substituted for text, such as by compressing consecutive tokens. These characteristics act as factors that make it difficult to measure uncertainty; therefore, prior research has utilized the predicted text without using the confidence of the prediction sequence.

[0004] Accordingly, there is an increasing need for technology to accurately recognize speech and measure speech uncertainty. The problem to be solved

[0005] The present invention aims to solve the problems of the aforementioned prior art by providing a technology that measures uncertainty by converting voice data recognized through a voice recognition model into text.

[0006] However, the technical problem that this embodiment aims to solve is not limited to the technical problem described above, and other technical problems may exist. means of solving the problem

[0007] As a technical means for achieving the technical problem described above, an embodiment according to the first aspect of the present disclosure provides a method for measuring uncertainty in a speech recognition model. The method comprises the steps of: generating a predicted text by converting a first output data, which is the result of recognizing input speech data using a first speech recognition model, into text; calculating a first uncertainty measurement value based on the predicted text and the first output data using a pre-set probability estimation algorithm; calculating a second uncertainty measurement value based on the second output data, which is the result of recognizing speech data using the predicted text and a second speech recognition model with dropout applied to the first speech recognition model using the probability estimation algorithm; and calculating a final uncertainty estimate value based on the first uncertainty measurement value and the second uncertainty measurement value.

[0008] Additionally, an embodiment according to a second aspect of the present disclosure provides an uncertainty measurement device for a speech recognition model. The device comprises a communication module, at least one processor, and a memory electrically connected to the processor and storing at least one code executed by the processor. The memory stores a code that, when executed through the processor, causes the processor to convert a first output data, which is the result of recognizing input speech data using a first speech recognition model, into text to generate a predicted text, calculate a first uncertainty measurement value based on the predicted text and the first output data using a preset probability estimation algorithm, calculate a second uncertainty measurement value based on the predicted text and a second output data, which is the result of recognizing speech data using a second speech recognition model with dropout applied to the first speech recognition model using the probability estimation algorithm, and calculate a final uncertainty estimate value based on the first uncertainty measurement value and the second uncertainty measurement value. Effects of the invention

[0009] The present invention can estimate uncertainty more precisely by utilizing probability values ​​to estimate uncertainty.

[0010] The present invention can provide high-accuracy speech recognition by predicting final uncertainty through data uncertainty and model uncertainty. Brief explanation of the drawing

[0011] FIG. 1 is a drawing illustrating an uncertainty measurement device for a speech recognition model and a terminal connected to the device according to an embodiment of the present invention. Figure 2 is a diagram illustrating the detailed configuration of the server shown in Figure 1. FIG. 3 is a diagram illustrating the process of estimating an uncertainty prediction value according to one embodiment of the present invention. FIGS. 4 to 8 are drawings illustrating examples of CTC decoding according to an embodiment of the present invention. FIG. 9 is a flowchart illustrating the sequence of a method for measuring uncertainty in a speech recognition model according to another embodiment of the present invention. Specific details for implementing the invention

[0012] The present disclosure will be described in detail below with reference to the attached drawings. However, the present disclosure may be implemented in various different forms and is not limited to the embodiments described herein. Furthermore, the attached drawings are intended only to facilitate understanding of the embodiments disclosed herein, and the technical concept disclosed herein is not limited by the attached drawings. All terms used herein, including technical and scientific terms, should be interpreted in the sense generally understood by those skilled in the art to which the present disclosure pertains. Terms defined in advance should be interpreted as having additional meanings consistent with relevant technical literature and the present disclosure, and should not be interpreted in a highly ideal or restrictive sense unless otherwise defined.

[0013] In order to clearly explain the invention in the drawings, parts unrelated to the explanation have been omitted, and the size, form, and shape of each component shown in the drawings may be varied. Throughout the specification, identical or similar parts are denoted by identical or similar reference numerals.

[0014] Throughout the specification, when it is stated that a part is "connected (connected, contacted, or coupled)" to another part, this includes not only cases where they are "directly connected (connected, contacted, or coupled)," but also cases where they are "indirectly connected (connected, contacted, or coupled)" with other members interposed therebetween. Furthermore, when it is stated that a part "includes (provides, or provides)" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but rather allows for additional "included (provided, or provided)" of other components.

[0015] In this specification, the term "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Additionally, one unit may be realized using two or more pieces of hardware, and two or more units may be realized by one piece of hardware. Meanwhile, "part" is not limited to software or hardware, and "part" may be configured to reside in an addressable storage medium or configured to run on one or more processors. Accordingly, as an example, "part" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts." In addition, the components and '~parts' may be implemented to play one or more CPUs within the device or secure multimedia card.

[0016] In the following description, suffixes such as "module" and "part" for components are assigned or used interchangeably solely for the ease of drafting the specification, and do not inherently possess distinct meanings or roles. Furthermore, in describing the embodiments disclosed in this specification, detailed descriptions of related prior art have been omitted where it is determined that such detailed descriptions could obscure the essence of the embodiments disclosed in this specification.

[0017] Terms indicating ordinal numbers, such as first, second, etc., used herein are used solely for the purpose of distinguishing one component from another and do not limit the order or relationship of the components. For example, the first component of the present disclosure may be named the second component, and similarly, the second component may be named the first component. Singular forms used herein should be interpreted as including plural forms unless explicitly to the contrary.

[0018] The "user terminal" mentioned below may be implemented as a computer or portable terminal capable of connecting to a server or other terminal via a network. Here, the computer may include, for example, a laptop, desktop, or notebook equipped with a web browser, and VR HMDs (e.g., HTC VIVE, Oculus Rift, GearVR, DayDream, PSVR, etc.). Here, the VR HMD includes PC models (e.g., HTC VIVE, Oculus Rift, FOVE, Deepon, etc.), mobile models (e.g., GearVR, DayDream, Storm Mirror, Google Cardboard, etc.), console models (PSVR), and stand-alone models implemented independently (e.g., Deepon, PICO, etc.). Portable terminals are wireless communication devices that ensure portability and mobility, and may include, for example, smartphones, tablet PCs, and wearable devices, as well as various devices equipped with communication modules such as Bluetooth (BLE, Bluetooth Low Energy), NFC, RFID, Ultrasonic, Infrared, WiFi, and LiFi. Additionally, "network" refers to a connection structure capable of exchanging information between each node, such as terminals and servers, and includes Local Area Networks (LAN), Wide Area Networks (WAN), the World Wide Web (WWW), wired and wireless data communication networks, telephone networks, wired and wireless television communication networks, etc.Examples of wireless data communication networks include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Bluetooth communication, infrared communication, ultrasonic communication, Visible Light Communication (VLC), and LiFi.

[0019] FIG. 1 is a drawing illustrating an uncertainty measurement device for a speech recognition model and a terminal connected to the device according to an embodiment of the present invention.

[0020] Referring to FIG. 1, the uncertainty measurement device of the speech recognition model of the present invention can be connected to a terminal. At this time, the connection between the uncertainty measurement device of the speech recognition model and the terminal can be established via wired or wireless communication. The uncertainty measurement device of the speech recognition model may be referred to as a server. Hereinafter, the uncertainty measurement device of the speech recognition model will be referred to as a device.

[0021] The device can generate predicted text by converting first output data, which is the result of recognizing input voice data using a first voice recognition model, into text. Additionally, the device can calculate a first uncertainty measurement value based on the predicted text and the first output data using a preset probability estimation algorithm, and calculate a second uncertainty measurement value based on the second output data, which is the result of recognizing voice data using a second voice recognition model that applies dropout to the first voice recognition model and the predicted text using the probability estimation algorithm. Furthermore, the device can calculate a final uncertainty estimate value based on the first uncertainty measurement value and the second uncertainty measurement value.

[0022] The terminal can be referred to as a user terminal. The user terminal can transmit voice signals to the device.

[0023] Figure 2 is a drawing illustrating the detailed configuration of the device shown in Figure 1.

[0024] Referring to FIG. 2, the device of the present invention includes a communication module (110), a processor (120), and a memory (130).

[0025] The communication module (110) may include a device comprising hardware and software necessary to transmit and receive signals, such as control signals or data signals, through a wired or wireless connection with another network device.

[0026] The communication module (110) can receive a voice signal from a user terminal and transmit it to a device. However, the role of the communication module (110) is not limited to this.

[0027] The processor (120) may include various types of devices for controlling and processing data. The processor (120) may refer to a data processing device embedded in hardware having a physically structured circuit to perform a function expressed by code or instructions included in a program.

[0028] In one example, the processor (120) may be implemented in the form of a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc., but the scope of the invention is not limited thereto.

[0029] The processor (120) performs operations according to the code stored in memory (130).

[0030] The memory (130) can store at least one of the information and data input to the communication module (110), the information and data required for the function performed by the processor (120), and the data generated according to the execution of the processor (120).

[0031] The term "memory" (130) should be interpreted as a general term for a non-volatile storage device that retains stored information even when power is not supplied, and a volatile storage device that requires power to retain stored information. In addition to a volatile storage device that requires power to retain stored information, the memory (130) may include a magnetic storage medium or a flash storage medium, but the scope of the present invention is not limited thereto.

[0032] Memory (130) is electrically connected to the processor (120) and stores at least one code that is executed by the processor (120). Memory (130) stores code that causes the processor (120) to perform the following functions and procedures when executed through the processor (120).

[0033] The memory (130) stores code that causes the first output data, which is the result of recognizing input voice data using the first voice recognition model, to be converted into text and generate predicted text. For example, a voice signal may be received from a user terminal, voice data may be extracted by standardizing the voice signal through preprocessing, and the first output data, which is the result of recognizing voice data using the first voice recognition model, may be converted into text to generate predicted text. At this time, the first output data may be the probability value per unit text of the converted text after converting the voice data into text. And if the final uncertainty estimate is greater than or equal to a preset threshold, the user terminal may be requested to re-input voice data.

[0034] The memory (130) stores code that causes obtaining first output data based on voice data, selecting text per unit time based on the probability value per unit text of the first output data, and generating predicted text.

[0035] The memory (130) stores code that causes a first uncertainty measurement value to be calculated based on the predicted text and the first output data using a pre-set probability estimation algorithm, and a second uncertainty measurement value to be calculated based on the second output data, which is the result of recognizing voice data, using a second voice recognition model that applies dropout to the predicted text and the first voice recognition model using a probability estimation algorithm.

[0036] For example, the first uncertainty measure can be calculated by a probability estimation algorithm estimating the probability that the first output data produces a result identical to the predicted text, and calculating the uncertainty per unit text of the first output data based on this probability. That is, the uncertainty can be obtained by using the CTC probability estimation algorithm described later to estimate the probability value of the predicted text. Furthermore, if the text length of the first output data is greater than or equal to a preset length, the first uncertainty measure may be calculated by dividing the probability value by the text length.

[0037] And the second uncertainty measure may be the maximum value among the multiple second uncertainty candidate values, where multiple second output data generated by the probability estimation algorithm based on the second speech recognition model generates multiple second uncertainty candidate values ​​such that the same result as the predicted text is output.

[0038] Memory (130) stores code that causes the final uncertainty estimate to be calculated based on the first uncertainty measurement and the second uncertainty measurement. For example, a new training data set can be created that includes voice data satisfying the case where the final uncertainty estimate is below a preset second threshold and prediction texts corresponding to the voice data. The voice recognition model can be iteratively trained using the newly created training data set, and the voice recognition accuracy can be improved through iterative training.

[0039] FIG. 3 is a diagram illustrating the process of estimating an uncertainty prediction value according to one embodiment of the present invention.

[0040] Referring to FIG. 3, the uncertainty prediction involves the device receiving a voice signal from a user terminal (S101) and preprocessing it (S102). At this time, the voice signal may be voice data or voice information containing voice. The preprocessing may be a process of removing noise included in the voice signal or amplifying the voice if the volume is low. When the preprocessed voice signal (S102) passes through a first voice recognition model (S103), text for each time segment is produced as a probability value.

[0041] For example, the time of the voice signal can be divided from 1 to n, and the voice signal data can also exist in time units. Therefore, the output value of the voice signal can predict the probability value of each text for each time from 1 to n. This will be explained in more detail below through Fig. 4. When a voice signal is input as shown in Fig. 4, the time is divided into t1, t2, t3, and t4, and the text probability value can be predicted for each time. When the two largest predicted text probability values ​​are extracted, the probability of T at t1 is 0.9 and the probability of D is 0.1; the probability of E at t2 is 1; the probability of X at t3 is 0.8 and the probability of S is 0.2; and the probability of T at t4 is 0.9 and the probability of K is 0.1. Next, CTC decoding can be performed (S104) to extract the predicted text (S105). At this time, CTC decoding is a process of merging texts that are recognized as duplicates according to the length of the input signal.

[0042] For example, as shown in Fig. 5, when five input signals are recognized during the process of inputting a voice signal "TEXT," if t1 and t2 are recognized as T,T or D,D, they can be merged to represent T or D. Then, a predicted text, TEXT, can be obtained based on the probability value. Alternatively, as shown in Fig. 6, if a blank token exists between L and L, both Ls can be obtained. Therefore, when a voice signal is input as shown in Fig. 6, HELLO can be obtained as the predicted text. As another example, as shown in Fig. 7, if there is no blank token between E and E, HELLO can be obtained based on the probability value. Although not illustrated in the figure, even when there is a large number of texts, a predicted text can be obtained based on the probability value of each text. In the same process, for TELEPHONE in Fig. 8, only one L, E, and O are determined based on the probability values ​​at t3 and t4, t5 and t6, and t9 and t10, thereby obtaining the predicted text, TELEPHONE.

[0043] Furthermore, the first artificial intelligence model is trained to output probability values ​​for each unit of text in the speech data using speech data as a training dataset. This first speech recognition model can be implemented based on various neural network structures such as CNN, RNN, ANN, and LLM.

[0044] The CTC output probability is given by the formula below.

[0045]

[0046] Here, f(·) is a model that provides frame-by-frame output probabilities, α is frame-by-frame alignment, and represents the set of possible alignments for y. Based on this, the CTC loss function L(·) is given by the following formula.

[0047]

[0048] Next, data uncertainty can be estimated (S107) through a first probability estimation algorithm (S106) based on the predicted text and the probability value predicted through the first speech recognition model. At this time, the first probability estimation algorithm may be used to estimate the probability value of the predicted text to obtain the uncertainty. Alternatively, the data uncertainty can be obtained by dividing the predicted probability values ​​for each text by the text length according to the text length. Conventional CTC probability estimation algorithms are used only for model training, but the present invention presents a part that utilizes the CTC probability estimation algorithm for uncertainty estimation.

[0049] The first probability estimation algorithm is an algorithm programmed to output data uncertainty using predicted text and probability values ​​predicted through the first speech recognition model as input data. The input data may include predicted data and probability values ​​predicted through the first speech recognition model.

[0050] In a process that occurs simultaneously with processes S103 to S107, a voice signal received from a user terminal is preprocessed (S102), and then multiple probability values ​​can be predicted (S201) through a second voice recognition model to which dropout is applied. The second artificial intelligence model is trained to output multiple probability values ​​by performing dropout a predetermined number of times using voice data as a training data set. This second artificial intelligence model can be implemented based on various neural network structures such as CNN, RNN, ANN, and LLM. Furthermore, while Monte-Carlo Dropout (MCD) can be applied to the second voice recognition model, it is not limited thereto and can be implemented in a wider variety of forms.

[0051] Next, based on the predicted text extracted in S105 and the multiple probability values ​​predicted in S201, the minimum value with the smallest probability value (S203) can be estimated as the model uncertainty (S204) through a second probability estimation algorithm (S202). In this case, the minimum value may be the probability value closest to the incorrect answer among the multiple probability values ​​predicted by the second speech recognition model when compared with the predicted text. That is, it may be the maximum value among the uncertainty candidate values. The formula for calculating the probability value is as follows.

[0052]

[0053] Here, x is the speech signal, f is the speech recognition model, and is the decoded output signal with dropout not applied, and is a multiple-sampled speech recognition model for i=1,...,N, where N is the number of times dropout is applied. The formula for calculating uncertainty by dividing the predicted probability values ​​by the text length is as follows.

[0054]

[0055] The second probability estimation algorithm is an algorithm programmed to output data uncertainty using predicted text and probability values ​​predicted through the second speech recognition model as input data. The input data may include predicted data and probability values ​​predicted through the second speech recognition model.

[0056] Next, the final uncertainty prediction value can be estimated (S301) by adding the data uncertainty estimate obtained in S107 and the model uncertainty estimate obtained in S204.

[0057] As an embodiment of the present invention, during the service process of a speech recognition model, the user may be requested to re-enter voice input based on the estimated uncertainty, or the estimated uncertainty may be provided to the user. The figure below illustrates how, in a speech recognition service, the user may be requested to re-enter a voice signal when the prediction uncertainty is high.

[0058] As an embodiment of the present invention, unlabeled voice data can be used for semi-supervised learning by estimating text values ​​using a speech recognition model and then adding the voice data and estimated text to a training dataset to use as labels when the uncertainty of the model's prediction is small. Through this, the reliability of the model's prediction value is estimated as uncertainty, and only cases with high reliability can be utilized for future training.

[0059] More specifically, the prediction uncertainty estimation algorithm of the present invention can be utilized in an environment where text labeling of voice data has not been performed, that is, in cases where only voice data usable as training data exists. The prediction uncertainty estimation algorithm may use the corresponding prediction text as the result of text labeling so that voice data among the voice data, in which the final uncertainty estimate value is lower than the second threshold value, can be utilized for training the speech recognition model. This is because if the prediction uncertainty estimate value is high, incorrect results may be obtained; therefore, by utilizing voice data with low prediction uncertainty estimate values ​​for training, the accuracy of speech recognition performance can be improved.

[0060] FIG. 9 is a flowchart illustrating the sequence of a method for measuring uncertainty in a speech recognition model according to another embodiment of the present invention.

[0061] The method for measuring the uncertainty of a speech recognition model to be described below can be performed by the device for measuring the uncertainty of a speech recognition model described above with reference to FIGS. 1 to 8. Accordingly, the content of the embodiment of the present disclosure described above with reference to FIGS. 1 to 8 can be applied in the same way to the embodiment to be described below, and content that overlaps with the description above will be omitted. The steps described below do not necessarily have to be performed in order, the order of the steps can be set in various ways, and the steps may be performed almost simultaneously.

[0062] Referring to FIG. 9, the method for measuring uncertainty in a speech recognition model includes a prediction text generation step (S10), a first uncertainty measurement value and a second uncertainty measurement value calculation step (S20), and a final uncertainty estimate value calculation step (S30).

[0063] The predicted text generation step (S10) is a step of generating predicted text by converting the first output data, which is the result of recognizing input voice data using the first voice recognition model, into text. For example, a voice signal may be received from a user terminal, voice data may be extracted by standardizing the voice signal through preprocessing, and the first output data, which is the result of recognizing voice data using the first voice recognition model, may be converted into text to generate predicted text. At this time, the first output data may be the probability value per unit text of the converted text after converting the voice data into text. Furthermore, if the final uncertainty estimate is greater than or equal to a preset threshold, a request for re-input of voice data may be made to the user terminal.

[0064] The predicted text generation step (S10) is a step of obtaining first output data based on voice data, selecting text per unit time based on the probability value per unit text of the first output data, and generating predicted text.

[0065] The step of calculating the first uncertainty measurement and the second uncertainty measurement (S20) calculates the first uncertainty measurement based on the predicted text and the first output data using a pre-set probability estimation algorithm, and calculates the second uncertainty measurement based on the second output data, which is the result of recognizing voice data, using the predicted text and the second voice recognition model with dropout applied to the first voice recognition model using the probability estimation algorithm.

[0066] For example, the first uncertainty measure may be calculated by a probability estimation algorithm estimating the probability that the first output data outputs a result identical to the predicted text, calculating the uncertainty per unit text of the first output data based on the probability, and multiplying the uncertainty per unit text of the first output data. Additionally, if the text length of the first output data is greater than or equal to a preset length, the first uncertainty measure may be calculated by dividing the product of the uncertainty per unit text of the first output data by the text length.

[0067] And the second uncertainty measure may be the maximum value among the multiple second uncertainty candidate values, where multiple second output data generated by the probability estimation algorithm based on the second speech recognition model generates multiple second uncertainty candidate values ​​such that the same result as the predicted text is output.

[0068] The step of calculating the final uncertainty estimate (S30) is a step of calculating the final uncertainty estimate based on the first uncertainty measurement and the second uncertainty measurement. For example, a new training data set can be created that includes voice data satisfying the case where the final uncertainty estimate is less than or equal to a preset second threshold, and prediction texts corresponding to the voice data. The voice recognition model can be iteratively trained using the newly created training data set, and the voice recognition accuracy can be improved through iterative training.

[0069] One embodiment of the present invention may also be implemented in the form of a recording medium comprising computer-executable instructions, such as program modules executed by a computer. A computer-readable medium may be any available medium accessible by a computer and includes both volatile and non-volatile media, and both removable and non-removable media. Additionally, a computer-readable medium may include all computer storage media. A computer storage medium includes both volatile and non-volatile, removable and non-removable media implemented by any method or technique for storing information, such as computer-readable instructions, data structures, program modules, or other data.

[0070] Although the method and system of the present invention have been described in relation to specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

[0071] A person skilled in the art to which this disclosure pertains will understand that, based on the foregoing description, other specific forms can be easily modified without altering the technical spirit or essential features of this disclosure. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of this disclosure is defined by the claims set forth below, and all modifications or variations derived from the meaning and scope of the claims and equivalents thereof should be interpreted as being included within the scope of this disclosure. The scope of this application is defined by the claims set forth below rather than by the foregoing detailed description, and all modifications or variations derived from the meaning and scope of the claims and equivalents thereof should be interpreted as being included within the scope of this application. Explanation of the symbols

[0072] 100: Device 110: Communication module 120: Processor 130: Memory 200: Terminal

Claims

Claim 1 A method for measuring uncertainty in a speech recognition model performed by at least one processor, comprising: a) a step of generating a predicted text by converting a first output data, which is the result of recognizing input speech data using a first speech recognition model, into text; b) a step of calculating a first uncertainty measurement value based on the predicted text and the first output data using a pre-set probability estimation algorithm, and calculating a second uncertainty measurement value based on the second output data, which is the result of recognizing speech data using a second speech recognition model that applies dropout to the first speech recognition model using the probability estimation algorithm; and c) a step of calculating a final uncertainty estimate value based on the first uncertainty measurement value and the second uncertainty measurement value. Claim 2 A method for measuring uncertainty in a speech recognition model according to claim 1, wherein the first output data is a probability value per unit text of the converted voice data, which is converted into text. Claim 3 A method for measuring uncertainty in a speech recognition model according to claim 2, wherein step a) comprises the steps of obtaining the first output data based on the voice data, selecting text per unit time based on the probability value per unit text of the first output data, and generating the predicted text. Claim 4 A method for measuring the uncertainty of a speech recognition model according to claim 3, wherein step a) comprises: a-1) a step of receiving a voice signal from a user terminal and structuring the voice signal through preprocessing to extract voice data; and a-2) a step of converting a first output data, which is the recognition result of the voice data, into text using the first speech recognition model to generate predicted text. Claim 5 A method for measuring uncertainty in a speech recognition model according to claim 3, wherein in step b) above, the first uncertainty measurement value is calculated by the probability estimation algorithm estimating the probability that the first output data outputs a result identical to the predicted text, and calculating the uncertainty per unit text of the first output data based on the probability. Claim 6 A method for measuring uncertainty of a speech recognition model, wherein in step b) above, the second uncertainty measurement value is the maximum value among the plurality of second uncertainty candidate values, wherein the probability estimation algorithm generates a plurality of second output data based on the second speech recognition model and the second output data generates a result identical to the predicted text. Claim 7 A method for measuring uncertainty in a speech recognition model according to claim 4, wherein if the final uncertainty estimate is greater than or equal to a pre-set first threshold, the re-input of the voice data is requested to the user terminal. Claim 8 A method for measuring uncertainty in a speech recognition model according to claim 7, further comprising the step of generating a new training data set including speech data satisfying the case where the final uncertainty estimate is less than or equal to a preset second threshold value, and prediction texts corresponding to said speech data. Claim 9 A speech recognition model uncertainty measurement device comprising: a communication module; at least one processor; and a memory electrically connected to the processor and storing at least one code executed by the processor, wherein the memory stores a code that, when executed through the processor, causes the processor to convert first output data, which is the result of recognizing input speech data using a first speech recognition model, into text to generate a predicted text, calculate a first uncertainty measurement value based on the predicted text and the first output data using a preset probability estimation algorithm, calculate a second uncertainty measurement value based on the predicted text and the second output data, which is the result of recognizing speech data using a second speech recognition model with dropout applied to the first speech recognition model using the probability estimation algorithm, and calculate a final uncertainty measurement value based on the first uncertainty measurement value and the second uncertainty measurement value. Claim 10 In claim 9, the first output data is a probability value per unit text of the converted voice data, and is an uncertainty measuring device for a voice recognition model. Claim 11 In claim 10, the above memory stores code that causes the processor, when executed through the above processor, to obtain the first output data based on the voice data, select a text per unit time based on the probability value per unit text of the first output data, and generate the predicted text. Claim 12 A speech recognition model uncertainty measurement device according to claim 11, wherein the memory stores code that causes the processor, when executed through the processor, to receive a voice signal from a user terminal, standardize the voice signal through preprocessing to extract voice data, and convert a first output data, which is the recognition result of the voice data, into text using the first speech recognition model to generate predicted text. Claim 13 In claim 11, the first uncertainty measurement device of a speech recognition model is such that the probability estimation algorithm estimates the probability that the first output data outputs a result identical to the predicted text, and calculates the uncertainty per unit text of the first output data based on the probability. Claim 14 In claim 13, the second uncertainty measurement value is a speech recognition model uncertainty measurement device in which the second uncertainty measurement value is the maximum value among the second uncertainty candidate values, wherein the second output data generated by the probability estimation algorithm based on the second speech recognition model generates a plurality of second uncertainty candidate values ​​in which a result identical to the predicted text is output. Claim 15 A voice recognition model uncertainty measurement device according to claim 12, wherein if the final uncertainty estimate is greater than or equal to a preset threshold, the device requests the re-input of the voice data to the user terminal. Claim 16 An uncertainty measurement device for a speech recognition model, wherein, in claim 15, it further comprises the step of generating a new training data set including speech data satisfying the case where the final uncertainty estimate is less than or equal to a preset second threshold value and prediction texts corresponding to the speech data.

Images