KR102509007B1 - Method for training speech recognition model based on the importance of tokens in sentences - Google Patents

Abstract

Disclosed is a method for enabling learning of a speech recognition model, performed by a computing device, according to one embodiment of the present disclosure. The method may comprise: a step of obtaining ground truth text information; a step of determining an importance for each of a plurality of tokens included in the ground truth text information; and a step of enabling learning of the speech recognition model in consideration of the importance for each of the plurality of tokens included in the ground truth text information.

Description

A method for training a speech recognition model based on the importance of tokens in a sentence {METHOD FOR TRAINING SPEECH RECOGNITION MODEL BASED ON THE IMPORTANCE OF TOKENS IN SENTENCES}

The present invention relates to a method for training a speech recognition model, and more particularly, to a technique for training a speech recognition model based on analyzing importance in text information of a ground truth.

In the field of Speech-to-Text (STT) technology, it is important that semantically important speech parts during speech are accurately converted into text, while less important speech parts such as slang are omitted or dictated incorrectly. No big deal. For reference, gantoor here refers to exclamations that are less important in meaning among expressions included in dialogue.

However, when dictation is performed in units of tokens (eg, mainly syllables or character tokens), existing general speech recognition models treat all tokens (eg, syllables or character tokens) equally in importance. Because of this, writing down frequently appearing syllables or letters has a great impact on accuracy and loss functions, and existing general speech recognition models are semantically insignificant, such as Gantour, but very much included in actual voice recordings. There is a problem that learning to write down the parts well occurs.

Korean Patent Publication No. 10-2003-0060588 (July 16, 2003) discloses a method for selecting a recorded sentence for voice synthesis based on a corpus.

An object of the present disclosure is to provide a method for learning a speech recognition model based on correct answer text information whose importance is determined in units of tokens. For example, the present disclosure determines the importance of each token in the correct answer text information and trains the speech recognition model by differentiating according to the importance, so that a speech recognition model that can better dictate an important speech part rather than an unimportant speech part It aims to implement

On the other hand, the technical problem to be achieved by the present disclosure is not limited to the above-mentioned technical problem, and may include various technical problems within a range apparent to those skilled in the art from the contents to be described below.

A method performed by a computing device according to an embodiment of the present disclosure for realizing the above object is disclosed. The method may include obtaining ground truth text information; determining an importance for each of a plurality of tokens included in the correct answer text information; and learning the voice recognition model in consideration of importance of each of the plurality of tokens included in the correct answer text information.

Alternatively, determining the importance of each of the plurality of tokens may include deriving a reference feature vector for the correct answer text information; deriving each comparison feature vector for each of the plurality of tokens; and determining an importance of each of the plurality of tokens based on the reference feature vector and each comparison feature vector.

Alternatively, deriving the reference feature vector for the correct answer text information may include extracting the reference feature vector by performing sentence embedding on the correct answer text information.

Alternatively, extracting each comparison feature vector for each of the plurality of tokens may include performing masking on one token of the plurality of tokens; extracting a comparison feature vector corresponding to the masked token by performing sentence embedding on the correct answer text information to which the masking is reflected; and performing the masking operation and the extraction of a comparison feature vector corresponding to the masked token for each of the plurality of tokens.

Alternatively, extracting each comparison feature vector for each of the plurality of tokens may include performing masking on a first token of the plurality of tokens; extracting a first comparison feature vector corresponding to the first token by performing sentence embedding on the correct answer text information in which the masking of the first token is reflected; performing masking on a second token among the plurality of tokens; and extracting a second comparison feature vector corresponding to the second token by performing sentence embedding on the correct answer text information in which the masking of the second token is reflected.

Alternatively, determining an importance for each of the plurality of tokens based on the reference feature vector and each comparison feature vector compares the reference feature vector and each comparison feature vector, calculating each degree of similarity for each comparison feature vector; and determining an importance of each token associated with each comparison feature vector based on each similarity.

Alternatively, determining the importance of each token associated with each comparison feature vector based on the respective degree of similarity may include determining the importance of each token as the degree of similarity increases as the degree of importance of each token decreases. can include

Alternatively, the step of determining the importance of each of a plurality of tokens included in the correct answer text information may include calculating a frequency of each of the plurality of tokens in learning data to which the correct answer text information belongs. step; and determining an importance of each of the plurality of tokens based on the frequency of each.

Alternatively, the step of learning the speech recognition model in consideration of the importance of each of the plurality of tokens included in the correct answer text information, based on the importance of each of the plurality of tokens, the plurality of determining a respective weight for each of the tokens; and constructing a loss function in consideration of each of the weights.

Alternatively, constructing a loss function in consideration of each weight may include applying each weight to a cross entropy (CE) loss value for each of the plurality of tokens; and performing a weighted average operation based on each of the cross entropy loss values to which the respective weights are applied.

A computer program stored in a computer readable storage medium is disclosed according to an embodiment of the present disclosure for realizing the above object. When the computer program is executed on one or more processors, the following operations for training a speech recognition model are performed, and the operations include: obtaining ground truth text information; determining an importance for each of a plurality of tokens included in the correct answer text information; and learning the voice recognition model by considering the importance of each of the plurality of tokens included in the correct answer text information.

A computing device according to an embodiment of the present disclosure for realizing the above object is disclosed. The device may include at least one processor; and a memory, wherein the at least one processor is configured to obtain ground truth text information; determining importance for each of a plurality of tokens included in the correct answer text information; And it may be configured to train the voice recognition model in consideration of the importance of each of the plurality of tokens included in the correct answer text information.

A method for performing voice recognition performed by a computing device according to an embodiment of the present disclosure for realizing the above object, the method comprising: inputting voice data to a voice recognition model; and generating text information corresponding to the input voice data, wherein the voice recognition model may correspond to a model learned based on correct answer text information for which importance information is determined in units of tokens.

According to the present disclosure, a voice recognition model with improved performance may be implemented by learning a voice recognition model based on correct text information whose importance is determined in units of tokens. For example, the present disclosure determines the importance of each token in the correct answer text information and differentiates it according to the importance and trains a speech recognition model, thereby improving performance of dictating important speech parts better than unimportant speech parts. A speech recognition model of can be implemented. In addition, there is an effect of improving the accuracy of the user's experience by utilizing the voice recognition model implemented in this way.

On the other hand, the effects of the present disclosure are not limited to the effects mentioned above, and various effects may be included within a range apparent to those skilled in the art from the contents to be described below.

1 is a block diagram of a computing device for training a speech recognition model according to an embodiment of the present disclosure.
2 is a schematic diagram illustrating a network function according to an embodiment of the present disclosure.
3 is a block diagram schematically showing a plurality of modules for learning a voice recognition model according to an embodiment of the present disclosure.
4 is a flowchart illustrating a method for training a voice recognition model according to an embodiment of the present disclosure.
5 is a simplified and general schematic diagram of an exemplary computing environment in which embodiments of the present disclosure may be implemented.

Various embodiments are now described with reference to the drawings. In this specification, various descriptions are presented to provide an understanding of the present disclosure. However, it is apparent that these embodiments may be practiced without these specific details.

The terms “component,” “module,” “system,” and the like, as used herein, refer to a computer-related entity, hardware, firmware, software, a combination of software and hardware, or an execution of software. For example, a component may be, but is not limited to, a procedure, processor, object, thread of execution, program, and/or computer running on a processor. For example, both an application running on a computing device and a computing device may be components. One or more components may reside within a processor and/or thread of execution. A component can be localized within a single computer. A component may be distributed between two or more computers. Also, these components can execute from various computer readable media having various data structures stored thereon. Components may be connected, for example, via signals with one or more packets of data (e.g., data and/or signals from one component interacting with another component in a local system, distributed system) to other systems and over a network such as the Internet. data being transmitted) may communicate via local and/or remote processes.

In addition, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or”. That is, unless otherwise specified or clear from the context, “X employs A or B” is intended to mean one of the natural inclusive substitutions. That is, X uses A; X uses B; Or, if X uses both A and B, "X uses either A or B" may apply to either of these cases. Also, the term "and/or" as used herein should be understood to refer to and include all possible combinations of one or more of the listed related items.

Also, the terms "comprises" and/or "comprising" should be understood to mean that the features and/or components are present. However, it should be understood that the terms "comprises" and/or "comprising" do not exclude the presence or addition of one or more other features, elements, and/or groups thereof. Also, unless otherwise specified or where the context clearly indicates that a singular form is indicated, the singular in this specification and claims should generally be construed to mean "one or more".

In addition, the term “at least one of A or B” should be interpreted as meaning “when only A is included”, “when only B is included”, and “when A and B are combined”.

Those skilled in the art will further understand that the various illustrative logical blocks, components, modules, circuits, means, logics, and algorithm steps described in connection with the embodiments disclosed herein may be implemented using electronic hardware, computer software, or combinations of both. It should be recognized that it can be implemented as To clearly illustrate the interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented in hardware or as software depends on the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application. However, such implementation decisions should not be interpreted as causing a departure from the scope of this disclosure.

The description of the presented embodiments is provided to enable any person skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art of this disclosure. The general principles defined herein may be applied to other embodiments without departing from the scope of this disclosure. Thus, the present invention is not limited to the embodiments presented herein. The present invention is to be accorded the widest scope consistent with the principles and novel features set forth herein.

In the present disclosure, network functions, artificial neural networks, and neural networks may be used interchangeably.

1 is a block diagram of a computing device for training a speech recognition model according to an embodiment of the present disclosure.

The configuration of the computing device 100 shown in FIG. 1 is only a simplified example. In one embodiment of the present disclosure, the computing device 100 may include other components for performing a computing environment of the computing device 100, and only some of the disclosed components may constitute the computing device 100.

The computing device 100 may include a processor 110 , a memory 130 , and a network unit 150 .

The processor 110 may include one or more cores, and includes a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU) of a computing device. unit), data analysis, and processors for deep learning. The processor 110 may read a computer program stored in the memory 130 and process data for machine learning according to an embodiment of the present disclosure. According to an embodiment of the present disclosure, the processor 110 may perform an operation for learning a neural network. The processor 110 is used for neural network learning, such as processing input data for learning in deep learning (DL), extracting features from input data, calculating errors, and updating neural network weights using backpropagation. calculations can be performed. At least one of the CPU, GPGPU, and TPU of the processor 110 may process learning of the network function. For example, the CPU and GPGPU can process learning of network functions and data classification using network functions. In addition, in an embodiment of the present disclosure, the learning of a network function and data classification using a network function may be processed by using processors of a plurality of computing devices together. In addition, a computer program executed in a computing device according to an embodiment of the present disclosure may be a CPU, GPGPU or TPU executable program.

According to an embodiment of the present disclosure, the processor 110 may be configured to train a voice recognition model by considering the importance of each of a plurality of tokens included in the correct answer text information. Specifically, the processor 110 may evaluate the importance of each token included in correct answer text information in a basic spoken sentence-correct answer text information pair, which is training data, and train a speech recognition model based on the evaluation. . For example, the processor 110 may train a speech recognition model by utilizing a loss function that implements a difference between significant tokens and unimportant tokens within correct text information. In addition, the processor 110 determines the importance by using a sentence embedding and masking method to determine the importance of each token, or judges the importance of a token that does not appear frequently (eg, a syllable) to be high importance can be judged using the That is, the processor 110 may learn a voice recognition model by differentiating a plurality of tokens included in the correct text information (sentence) according to semantic importance.

According to an embodiment of the present disclosure, the memory 130 may store any type of information generated or determined by the processor 110 and any type of information received by the network unit 150 .

According to an embodiment of the present disclosure, the memory 130 is a flash memory type, a hard disk type, a multimedia card micro type, or a card type memory (eg, SD or XD memory, etc.), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory) -Only Memory), a magnetic memory, a magnetic disk, and an optical disk may include at least one type of storage medium. The computing device 100 may operate in relation to a web storage that performs a storage function of the memory 130 on the Internet. The above description of the memory is only an example, and the present disclosure is not limited thereto.

The network unit 150 according to an embodiment of the present disclosure includes a Public Switched Telephone Network (PSTN), x Digital Subscriber Line (xDSL), Rate Adaptive DSL (RADSL), Multi Rate DSL (MDSL), and VDSL ( Various wired communication systems such as Very High Speed DSL), Universal Asymmetric DSL (UADSL), High Bit Rate DSL (HDSL), and Local Area Network (LAN) may be used.

In addition, the network unit 150 presented in this specification includes Code Division Multi Access (CDMA), Time Division Multi Access (TDMA), Frequency Division Multi Access (FDMA), Orthogonal Frequency Division Multi Access (OFDMA), SC-FDMA ( Single Carrier-FDMA) and other systems.

In the present disclosure, the network unit 150 may be configured regardless of its communication mode, such as wired and wireless, and may include a local area network (LAN), a personal area network (PAN), and a wide area network (WAN: It can be composed of various communication networks such as Wide Area Network). In addition, the network may be the known World Wide Web (WWW), or may use a wireless transmission technology used for short-range communication, such as Infrared Data Association (IrDA) or Bluetooth.

The techniques described herein may be used in the networks mentioned above as well as other networks.

2 is a schematic diagram illustrating a network function according to an embodiment of the present disclosure.

Throughout this specification, computational model, neural network, network function, and neural network may be used interchangeably. A neural network may consist of a set of interconnected computational units, which may generally be referred to as nodes. These nodes may also be referred to as neurons. A neural network includes one or more nodes. Nodes (or neurons) constituting neural networks may be interconnected by one or more links.

In a neural network, one or more nodes connected through a link may form a relative relationship of an input node and an output node. The concept of an input node and an output node is relative, and any node in an output node relationship with one node may have an input node relationship with another node, and vice versa. As described above, an input node to output node relationship may be created around a link. More than one output node can be connected to one input node through a link, and vice versa.

In a relationship between an input node and an output node connected through one link, the value of data of the output node may be determined based on data input to the input node. Here, a link interconnecting an input node and an output node may have a weight. The weight may be variable, and may be changed by a user or an algorithm in order to perform a function desired by the neural network. For example, when one or more input nodes are interconnected by respective links to one output node, the output node is set to a link corresponding to values input to input nodes connected to the output node and respective input nodes. An output node value may be determined based on the weight.

As described above, in the neural network, one or more nodes are interconnected through one or more links to form an input node and output node relationship in the neural network. Characteristics of the neural network may be determined according to the number of nodes and links in the neural network, an association between the nodes and links, and a weight value assigned to each link. For example, when there are two neural networks having the same number of nodes and links and different weight values of the links, the two neural networks may be recognized as different from each other.

A neural network may be composed of a set of one or more nodes. A subset of nodes constituting a neural network may constitute a layer. Some of the nodes constituting the neural network may form one layer based on distances from the first input node. For example, a set of nodes having a distance of n from the first input node may constitute n layers. The distance from the first input node may be defined by the minimum number of links that must be passed through to reach the corresponding node from the first input node. However, the definition of such a layer is arbitrary for explanation, and the order of a layer in a neural network may be defined in a method different from the above. For example, a layer of nodes may be defined by a distance from a final output node.

An initial input node may refer to one or more nodes to which data is directly input without going through a link in relation to other nodes among nodes in the neural network. Alternatively, in a relationship between nodes based on a link in a neural network, it may mean nodes that do not have other input nodes connected by a link. Similarly, the final output node may refer to one or more nodes that do not have an output node in relation to other nodes among nodes in the neural network. Also, the hidden node may refer to nodes constituting the neural network other than the first input node and the last output node.

In the neural network according to an embodiment of the present disclosure, the number of nodes in the input layer may be the same as the number of nodes in the output layer, and the number of nodes decreases and then increases again as the number of nodes progresses from the input layer to the hidden layer. can In addition, the neural network according to another embodiment of the present disclosure may be a neural network in which the number of nodes of the input layer may be less than the number of nodes of the output layer and the number of nodes decreases as the number of nodes increases from the input layer to the hidden layer. there is. In addition, the neural network according to another embodiment of the present disclosure is a neural network in which the number of nodes in the input layer may be greater than the number of nodes in the output layer, and the number of nodes increases as the number of nodes increases from the input layer to the hidden layer. can A neural network according to another embodiment of the present disclosure may be a neural network in the form of a combination of the aforementioned neural networks.

A deep neural network (DNN) may refer to a neural network including a plurality of hidden layers in addition to an input layer and an output layer. Deep neural networks can reveal latent structures in data. In other words, it can identify the latent structure of a photo, text, video, sound, or music (e.g., what objects are in the photo, what the content and emotion of the text are, what the content and emotion of the audio are, etc.). . Deep neural networks include convolutional neural networks (CNNs), recurrent neural networks (RNNs), auto encoders, generative adversarial networks (GANs), and restricted boltzmann machines (RBMs). machine), a deep belief network (DBN), a Q network, a U network, a Siamese network, a Generative Adversarial Network (GAN), and the like. The description of the deep neural network described above is only an example, and the present disclosure is not limited thereto.

In one embodiment of the present disclosure, the network function may include an autoencoder. An autoencoder may be a type of artificial neural network for outputting output data similar to input data. An auto-encoder may include at least one hidden layer, and an odd number of hidden layers may be disposed between input and output layers. The number of nodes of each layer may be reduced from the number of nodes of the input layer to an intermediate layer called the bottleneck layer (encoding), and then expanded symmetrically with the reduction from the bottleneck layer to the output layer (symmetrical to the input layer). Autoencoders can perform non-linear dimensionality reduction. The number of input layers and output layers may correspond to dimensions after preprocessing of input data. In the auto-encoder structure, the number of hidden layer nodes included in the encoder may decrease as the distance from the input layer increases. If the number of nodes in the bottleneck layer (the layer with the fewest nodes located between the encoder and decoder) is too small, a sufficient amount of information may not be conveyed, so more than a certain number (e.g., more than half of the input layer, etc.) ) may be maintained.

The neural network may be trained using at least one of supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Learning of the neural network may be a process of applying knowledge for the neural network to perform a specific operation to the neural network.

A neural network can be trained in a way that minimizes output errors. In the learning of the neural network, the learning data is repeatedly input into the neural network, the output of the neural network for the training data and the error of the target are calculated, and the error of the neural network is transferred from the output layer of the neural network to the input layer in the direction of reducing the error. It is a process of updating the weight of each node of the neural network by backpropagating in the same direction. In the case of teacher learning, the learning data in which the correct answer is labeled is used for each learning data (ie, the labeled learning data), and in the case of comparative teacher learning, the correct answer may not be labeled in each learning data. That is, for example, learning data in the case of teacher learning regarding data classification may be data in which each learning data is labeled with a category. Labeled training data is input to a neural network, and an error may be calculated by comparing an output (category) of the neural network and a label of the training data. As another example, in the case of comparative history learning for data classification, an error may be calculated by comparing input learning data with a neural network output. The calculated error is back-propagated in a reverse direction (ie, from the output layer to the input layer) in the neural network, and the connection weight of each node of each layer of the neural network may be updated according to the back-propagation. The amount of change in the connection weight of each updated node may be determined according to a learning rate. The neural network's computation of input data and backpropagation of errors can constitute a learning cycle (epoch). The learning rate may be applied differently according to the number of iterations of the learning cycle of the neural network. For example, a high learning rate may be used in the early stage of neural network training to increase efficiency by allowing the neural network to quickly obtain a certain level of performance, and a low learning rate may be used in the late stage to increase accuracy.

In neural network learning, generally, training data can be a subset of real data (ie, data to be processed using the trained neural network), and therefore, errors for training data are reduced, but errors for real data are reduced. There may be incremental learning cycles. Overfitting is a phenomenon in which errors for actual data increase due to excessive learning on training data. For example, a phenomenon in which a neural network that has learned a cat by showing a yellow cat does not recognize that it is a cat when it sees a cat other than yellow may be a type of overfitting. Overfitting can act as a cause of increasing the error of machine learning algorithms. Various optimization methods can be used to prevent such overfitting. To prevent overfitting, methods such as increasing the training data, regularization, inactivating some nodes in the network during learning, and using a batch normalization layer should be applied. can

In the present disclosure, the "reinforcement learning control model" may be an artificial neural network model having the configuration of the neural network described above with reference to FIG. 2, but may be an artificial neural network model learned by a reinforcement learning method.

Reinforcement learning is an artificial neural network based on the reward that an artificial neural network model calculates for an action selected so that the artificial neural network model can determine a better action based on the input state. It is a kind of learning method to train a model. Reinforcement learning methods can be understood as "learning through trial and error" in that decisions (i.e., actions) are rewarded. A reward given to an artificial neural network model in a reinforcement learning process may be a reward obtained by accumulating results of various actions. Reinforcement learning creates an artificial neural network model that maximizes the reward itself or the return (sum of rewards) by considering rewards according to various states and actions through learning. In the present disclosure, a "reinforcement learning control model" is a subject that determines behavior and may be used interchangeably with the term "agent". In the technical field related to reinforcement learning, the term "environment (Env)" or "environment model" is used as a term to refer to "a model that returns a result considering the behavior of an agent." The environment model may be a model that returns output data (e.g. state information) for given input data (e.g. control information). The environment model may be a model structure for reaching from input to output or a model in which the causal relationship between input and output data is unknown. Agents and environments can work by exchanging data.

Speech To Text, or Automatic Speech Recognition (STT or ASR), which will be described below, is a dictation technology that converts voice into text. The input of the speech recognition (STT) may include at least one of a speech signal, a spectrogram obtained by converting the speech signal, or a speech feature. In addition, the output of speech recognition (STT) is text in the form of a character string. Meanwhile, the speech recognition (STT) model may be implemented in various types of models including the neural network model described above. In addition, the speech recognition (STT) model may be divided into a modular method and a non-modular end-to-end (e2e) method according to an implemented method. Here, the modularized method is divided into an acoustic model (a model that indicates how a voice signal can be expressed), a language model (a model that assigns a probability of occurrence to a word based on a given sentence and word), and a pronunciation dictionary, etc. It may include, but is not limited to, traditional models for performing recognition (eg, some models of Kaldi toolkit-based ASR models, Hybrid-ASR models, etc.). On the other hand, the non-modularized method mainly means an e2e model (eg, a transformer-based encoder decoder model, etc.), and the model can be created by learning a lot of data without having sub-modules. On the other hand, the beam search technique is representative of the decoding technique, and the beam search technique does not predict just one word that is closest to the correct answer according to the situation, but opens up various possibilities and considers the entire sentence to find the most optimal way to find the correct answer.

According to an embodiment of the present disclosure, an unimportant token in the correct answer text information may include a token corresponding to interjection. For reference, interjection may refer to an interjection or the like of which the importance of a meaning included in a dialogue in the field of speech recognition is low. In Korean, ‘um, uh, me, ah, eh, ??’ In English, uh, huh, mhm, mm, ??. etc. belong to this. In addition, although interjection is not semantically important, it is very often included in actual voice recordings, and as a result, appears in many texts resulting from speech recognition (STT). On the other hand, an example of an unimportant token is not limited to a gantoor, and other words, sub-words, syllables, and the like may also correspond to unimportant tokens.

According to an embodiment of the present disclosure, the computing device 100 may input voice data to a voice recognition model. Also, the computing device 100 may generate text information corresponding to the input voice data. In this case, the voice recognition model may correspond to a model learned based on correct answer text information for which importance information is determined in units of tokens. More specifically, the computing device 100 obtains correct answer text information, determines the importance of each of a plurality of tokens included in the correct answer text information, and determines the importance of each of the plurality of tokens included in the correct answer text information. A voice recognition model can be trained by taking this into account. That is, the computing device 100 determines the importance of each token in the correct answer text information, and differentiates and learns according to the importance, so that the speech recognition model can dictate important parts better than unimportant parts. .

At this time, the computing device 100 extracts a “reference feature vector” for the correct answer text information, extracts “each comparison feature vector” for each of a plurality of tokens included in the correct answer text information, and extracts the extracted “reference feature vector”. The importance of each token can be determined based on the "feature vector" and "each comparison feature vector". For example, the computing device 100 may calculate a similarity between the “reference feature vector” and “each comparison feature vector” and determine the importance of each token based on the calculated similarity information. In addition, the computing device 100 may determine the importance of each token as the degree of similarity increases. Meanwhile, in relation to extraction of the reference feature vector, the computing device 100 may extract the reference feature vector by performing sentence embedding on the correct answer text information. In addition, in relation to the extraction of each comparison feature vector, the computing device 100 may perform "masking on one token among a plurality of tokens in the correct answer text information", "the masking reflected Operation of extracting a comparison feature vector corresponding to the masked token by performing sentence embedding on the correct answer text information", "operation of performing the masking and deriving a comparison feature vector corresponding to the masked token , "Operation performed on each of the plurality of tokens" may be performed. In other words, in relation to the extraction of each comparison feature vector, the computing device 100 performs "masking on a first token among a plurality of tokens included in the correct answer text information", "the first token Operation of deriving a first comparison feature vector corresponding to the first token by performing sentence embedding on the correct answer text information in which the masking for is reflected, "Performing masking on a second token among the plurality of tokens operation", "deriving a second comparison feature vector corresponding to the second token by performing sentence embedding on the correct answer text information in which the masking of the second token is reflected", etc. can be performed.

Also, the computing device 100 may determine the importance of each of the plurality of tokens based on frequency information of each of the plurality of tokens included in the correct answer text information. For example, the computing device 100 may perform "calculating a frequency of each of the plurality of tokens in the learning data to which the correct answer text information belongs", "based on the respective frequency, An operation of determining importance for each of a plurality of tokens" may also be performed.

In addition, the computing device 100 may perform an operation of constructing a loss function in consideration of the importance of each token in relation to learning a voice recognition model in consideration of the importance of each token. For example, the computing device 100 may include "determining a weight of each of the plurality of tokens based on the importance of each of the plurality of tokens", "taking into account the weight of each of the plurality of tokens" An operation of constructing a loss function" and the like may be performed. In addition, the computing device 100, in relation to the configuration of the loss function, "applies the respective weight to each cross entropy (CE) loss value for each of the plurality of tokens", " An operation of performing a weighted average operation based on each of the cross entropy loss values to which each of the weights are applied" may be performed.

3 is a block diagram schematically illustrating a plurality of modules for learning a voice recognition model according to an embodiment of the present disclosure.

According to an embodiment of the present disclosure, the computing device 100 may include an input module 111 , an importance determination module 112 and a loss function construction module 113 . Meanwhile, a plurality of modules that may be included in the computing device 100 may be controlled by the processor 110 or implemented by an operation of the processor 110 . In addition, modules that may be included in the computing device 100 in relation to a technique for learning a voice recognition model are not limited to the plurality of modules as described above, and additional modules may be included. A plurality of exemplary modules for training a speech recognition model will be described in more detail below.

According to an embodiment of the present disclosure, the input module 111 may obtain ground truth text information. In addition, the input module 111 may obtain a pair of "basic spoken sentence (voice) - correct answer (Ground truth) text information", which is data for learning. For example, the input module 111 may obtain a basic spoken sentence (voice) of “Hello, I am Dongchan Shin.” Also, the input module 111 says "Hello, uh, I'm Shin Dong-chan." It is possible to obtain correct answer text information.

According to an embodiment of the present disclosure, the importance determination module 112 may determine the importance of each of the plurality of tokens included in the correct answer text information. More specifically, the importance determination module 112 may determine the importance of each of the plurality of tokens based on the reference feature vector and each comparison feature vector. For example, a 'token' may include a word, a sub-word, a syllable, a grapheme, a consonant, and a vowel unit. . A sub-word means a semantic unit that can be subdivided into semantic units within a word. ”, “rice”, “eul”, “eat”, and “had” sub-words.

Illustratively, the importance determination module 112 may extract a reference feature vector for correct text information. In this case, the importance determination module 112 may extract the reference feature vector by performing sentence embedding on the correct text information. For reference, sentence embedding is a task of vectorizing sentence by sentence so that the meaning of the sentence can be expressed well. For example, sentence embedding may be performed based on BERT (Bidirectional Encoder Representations from Transformers), and Google's LaBSE model (Language-agnostic BERT sentence embedding model) may be used, but is not limited thereto.

Also, the importance determination module 112 may extract each comparison feature vector for each of the plurality of tokens. For example, if the correct answer text information is "Hello, I am Shin Dong-chan", the importance determination module 112 divides the correct answer text information into token units (eg, sub-words) and selects [Hello , Do it, _Uh, _I, _Shin Dongchan, I am] can be divided into a total of six tokens. For example, the plurality of tokens may include a first token (“Hello”), a second token (“Do”), a third token (“uh”), a fourth token (“Me”), and a fifth token (“Shin Dong-chan”). ), and a sixth token ("is"). However, this is just an example. Even if tokens are similarly classified in sub-word units for the same correct answer text information above, the number of tokens may vary depending on a tokenizer that distinguishes tokens. For example, if a tokenizer learned to distinguish more frequently is used, even if the correct text information of the same "Hello, uh, I'm Shin Dong-chan" is divided into sub-word units as well, [Hello, do it, _uh, _ It can be divided into 7 tokens such as _Shin Dong-chan, Ip, Nida].

The importance determination module 112 may perform masking on one token among the plurality of tokens in order to extract a respective comparison feature vector for each of the plurality of tokens. For example, the importance determination module 112 may perform masking on a first token among the plurality of tokens. For example, the importance determination module 112 may perform masking on a first token among a plurality of tokens, such as [MASK, do it, _uh, _me, _Shin Dongchan, is]. Also, the importance determination module 112 may perform masking on a second token among the plurality of tokens. For example, the importance determination module 112 may perform masking on a second token among a plurality of tokens, such as [Hello, MASK, _uh, _I, _Shin Dongchan, is]. In addition, the importance determination module 112 performs masking on the third token, such as [Hello, do it, MASK, _I am, _Shin Dong-chan, is], [Hello, do it, _uh, MASK, _Shin Dong-chan, Masking is performed on the 4th token as in [Hello, do, _uh, _I, MASK, is], and masking is performed on the 5th token as in [Hello, do, _uh, _I, _Shin Dong-chan, MASK], masking can be performed on the 6th token.

In addition, the importance determination module 112 may extract a comparison feature vector corresponding to the masked token by performing sentence embedding on the correct answer text information to which the masking is reflected. For example, the importance determination module 112 may extract a first comparison feature vector corresponding to the first token by performing sentence embedding on the correct answer text information in which the masking of the first token is reflected. In addition, the importance determination module 112 may extract a second comparison feature vector corresponding to the second token by performing sentence embedding on the correct answer text information in which the masking of the second token is reflected. The importance determination module 112 performs the masking operation and the extraction of a comparison feature vector corresponding to the masked token for each of the plurality of tokens (eg, first to sixth tokens). can be done At this time, a plurality of comparison feature vectors are extracted. Even for the same correct answer text information, the number of tokens may vary according to the token unit, and thus the number of comparison feature vectors may also vary.

According to an embodiment of the present disclosure, the importance determination module 112 may compare the reference feature vector and each comparison feature vector to calculate a similarity for each comparison feature vector. Illustratively, the importance determination module 112 compares the "reference feature vector" and "the first comparison feature vector to the sixth comparison feature vector" to determine the similarity (eg, the first comparison feature vector) for each comparison feature vector. A first similarity corresponding to a vector, ..., a sixth similarity corresponding to a sixth comparison feature vector) may be calculated. Meanwhile, as the similarity between vectors, cosine similarity may be used, but is not limited thereto.

In addition, the importance determination module 112 may determine the importance of each token associated with each comparison feature vector based on each degree of similarity. At this time, the importance determination module 112 may determine the importance of each token as the importance of each token increases as the degree of similarity increases. In addition, the importance determination module 112 performs sentence embedding on the correct answer text information to which masking is reflected, and extracts the "comparison feature vector corresponding to the masked token" and the "standard" extracted by sentence embedding on the correct answer text information The similarity can be calculated by comparing the feature vectors. For example, if the similarity between the first comparison feature vector and the reference feature vector is high, since it means that the "masked first token" corresponding to the first comparison feature vector is not important, the importance of the first token may be determined to be low. . That is, the similarity of the comparison feature vector and the importance of the masked token in the comparison feature vector may be inversely proportional.

For example, if the basic utterance sentence and correct answer text is "Hello, uh, I'm Shin Dong-chan", and the tokens are separated by sub-word units, the correct answer text is [Hello, do it, _uh, _ I am, _Dongchan Shin, I am]. Since the meaning of the third token "_word" is maintained even if it is masked, the similarity between the sentence masked with the third token and the correct answer text may be determined to be high, and the importance of the third token may be determined to be low. For example, the importance determining module 112 may determine the importance of each token as [0.9, 0.9, 0.1, 0.9, 0.9, 0.9] after calculating the similarity of the sentence masked with each token.

According to an embodiment of the present disclosure, the importance determination module 112 may calculate each frequency of each of the plurality of tokens in learning data to which the correct answer text information belongs. In addition, the importance determination module 112 may determine the importance of each of the plurality of tokens based on each frequency. For example, the importance determination module 112 may determine that a token that does not frequently appear on answer text information, which is learning data, is more important.

According to an embodiment of the present disclosure, the loss function configuration module 113 may determine a weight for each of the plurality of tokens based on the importance of each of the plurality of tokens. In addition, the loss function construction module 113 may configure a loss function in consideration of each weight. For example, the loss function construction module 113 may construct a loss function from a weighted average operation based on the importance of each token of the correct answer text information determined by the importance determination module 112 described above. More specifically, the loss function configuration module 113 may apply the respective weight to each cross entropy (CE) loss value for each of the plurality of tokens. Also, the loss function constructing module 113 may perform a weighted average operation based on each cross entropy (CE) loss value to which each weight is applied.

Illustratively, the loss function construction module 113 may set the "token cross-entropy loss value between the prediction predicted by the speech recognition model for each of a plurality of tokens and the cross-entropy loss value between the tokens of the correct answer text information matching the prediction" A loss function may be configured by reflecting the "weight determined based on the importance of". As an example, if the prediction predicted by the speech recognition model for each token was [p1, p2, p3, ..., p6], "loss function = 0.9 * <probability value set for the correct answer text 'Hello' ( CE loss between normal one-hot embedding, i.e. vector such as [1 0 0 0 0 0] and p1> + 0.9 * <Probability value set for the correct answer text 'Do it' (usually one-hot embedding, i.e. [0 1 0 0 0 0] and the CE loss between p2> + 0.1 * <the probability value set for the correct answer text 'uh' (usually a one-hot embedding, that is, a vector such as [0 0 1 0 0 0]) and p3 CE loss> + 0.9 * <CE loss between the probability value set for the correct answer text 'Me' (usually a one-hot embedding, that is, a vector such as [0 0 0 1 0 0]) and p4> + 0.9 * <Correct answer text 'Dongchan Shin' CE loss between p5 and the probability value set for ' (usually one-hot embedding, that is, a vector such as [0 0 0 0 1 0]) > + 0.9 * <probability value set for the correct answer text 'is' (usually a one-hot embedding , that is, a vector such as [0 0 0 0 0 1]) and the CE loss between p6>". Here, p3 corresponding to the token with low importance can be learned relatively less “to be close to the probability value set for the correct answer text 'uh'”. On the other hand, p1, p2, p4, p5 corresponding to the remaining tokens with high importance , p6 can be learned relatively better to "close to the set probability value for the correct text".

According to an embodiment of the present disclosure, since the existing speech recognition model is trained to minimize the loss function, the predicted probability (eg, p1 to p6) for each token is learned to approach the probability value set for each correct text information. can However, the present disclosure weights the tokens according to their importance so that the speech recognition model can be trained to dictate better for important tokens.

According to an embodiment of the present disclosure, the processor 110 may train a voice recognition model by considering the importance of each of a plurality of tokens included in the correct answer text information. In other words, the processor 110 considers the importance of each of the plurality of tokens determined through the input module 111, the importance determination module 112, and the loss function construction module 113, and trains a voice recognition model. can That is, the processor 110 may determine the importance of each token in the correct answer text information, and differentiate and learn according to the importance, so that the voice recognition model dictates the important part better than the unimportant part.

4 is a flowchart illustrating a method for training a voice recognition model according to an embodiment of the present disclosure.

The method for training the voice recognition model shown in FIG. 4 may be performed by the computing device 100 described above. Even if not mentioned in detail below, the above description of the computing device 100 may be equally applied to a description of a method for learning a voice recognition model.

Referring to FIG. 4 , a method for learning a speech recognition model according to an embodiment of the present disclosure includes obtaining ground truth text information (S110), a plurality of tokens included in the ground truth text information It may include determining the importance of each (S120) and learning the voice recognition model by considering the importance of each of the plurality of tokens included in the correct answer text information (S130), In addition to these steps, various steps may be included. Also, a method for learning a voice recognition model according to an embodiment of the present disclosure may be performed by the computing device 100 .

Step S110 is a step of obtaining correct answer text information.

The step S120 is a step of determining the importance of each of a plurality of tokens included in the correct answer text information. The step S120 includes extracting a reference feature vector for the correct answer text information, extracting each comparison feature vector corresponding to each of the plurality of tokens, and the reference feature vector and each comparison feature vector. Based on , determining the importance of each of the plurality of tokens may be included. Here, the extracting of the reference feature vector for the correct answer text information may include extracting the reference feature vector by performing sentence embedding on the correct answer text information.

Meanwhile, the step of extracting each comparison feature vector for each of the plurality of tokens may include performing masking on one token among the plurality of tokens; extracting a comparison feature vector corresponding to the masked token by performing sentence embedding on the correct answer text information to which the masking is reflected; and performing the masking operation and the extraction of a comparison feature vector corresponding to the masked token for each of the plurality of tokens. Specifically, the extracting of each comparison feature vector for each of the plurality of tokens may include performing masking on a first token among the plurality of tokens; extracting a first comparison feature vector corresponding to the first token by performing sentence embedding on the correct answer text information in which the masking of the first token is reflected; performing masking on a second token among the plurality of tokens; and extracting a second comparison feature vector corresponding to the second token by performing sentence embedding on the correct answer text information in which the masking of the second token is reflected.

In addition, the step of determining the importance of each of the plurality of tokens based on the reference feature vector and each comparison feature vector compares the reference feature vector and each comparison feature vector, calculating a degree of similarity for each comparison feature vector; and determining an importance of each token associated with each comparison feature vector based on each similarity. Here, the step of determining the importance of each token associated with each comparison feature vector based on the degree of similarity may include determining the importance of each token as the degree of similarity increases. can

In addition, step S120 may include calculating each frequency of each of the plurality of tokens in the learning data to which the correct answer text information belongs; and determining an importance of each of the plurality of tokens based on the frequency of each.

The step S130 is a step of learning the voice recognition model by considering the importance of each of the plurality of tokens included in the correct answer text information. The step S130 may include determining a weight for each of the plurality of tokens based on the importance of each of the plurality of tokens; and constructing a loss function in consideration of each of the weights.

The steps mentioned in the foregoing description may be further divided into additional steps or combined into fewer steps, depending on the implementation of the present disclosure. Also, some steps may be omitted if necessary, and the order of steps may be changed.

According to an embodiment of the present disclosure, a computer readable medium storing a data structure is disclosed.

Data structure can refer to the organization, management, and storage of data that enables efficient access and modification of data. Data structure may refer to the organization of data to solve a specific problem (eg, data retrieval, data storage, data modification in the shortest time). A data structure may be defined as a physical or logical relationship between data elements designed to support a specific data processing function. A logical relationship between data elements may include a connection relationship between user-defined data elements. A physical relationship between data elements may include an actual relationship between data elements physically stored in a computer-readable storage medium (eg, a persistent storage device). The data structure may specifically include a set of data, a relationship between data, and a function or command applicable to the data. Through an effectively designed data structure, a computing device can perform calculations while using minimal resources of the computing device. Specifically, the computing device can increase the efficiency of operation, reading, insertion, deletion, comparison, exchange, and search through an effectively designed data structure.

The data structure can be divided into a linear data structure and a non-linear data structure according to the shape of the data structure. A linear data structure may be a structure in which only one data is connected after one data. Linear data structures may include lists, stacks, queues, and decks. A list may refer to a series of data sets in which order exists internally. The list may include a linked list. A linked list may be a data structure in which data are connected in such a way that each data is connected in a single line with a pointer. In a linked list, a pointer can contain information about connection to the next or previous data. A linked list can be expressed as a singly linked list, a doubly linked list, or a circular linked list depending on the form. A stack can be a data enumeration structure that allows limited access to data. A stack can be a linear data structure in which data can be processed (eg, inserted or deleted) at only one end of the data structure. The data stored in the stack may be a LIFO-Last in First Out (Last in First Out) data structure. A queue is a data listing structure that allows limited access to data, and unlike a stack, it can be a data structure (FIFO-First in First Out) in which data stored later comes out later. A deck can be a data structure that can handle data from either end of the data structure.

The nonlinear data structure may be a structure in which a plurality of data are connected after one data. The non-linear data structure may include a graph data structure. A graph data structure can be defined as a vertex and an edge, and an edge can include a line connecting two different vertices. A graph data structure may include a tree data structure. The tree data structure may be a data structure in which one path connects two different vertices among a plurality of vertices included in the tree. That is, it may be a data structure that does not form a loop in a graph data structure.

Throughout this specification, computational model, neural network, network function, and neural network may be used interchangeably. Hereinafter, a neural network is unified and described. The data structure may include a neural network. And the data structure including the neural network may be stored in a computer readable medium. The data structure including the neural network may also include preprocessed data for processing by the neural network, data input to the neural network, weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, activation function associated with each node or layer of the neural network, and neural network It may include a loss function for learning of . A data structure including a neural network may include any of the components described above. That is, the data structure including the neural network includes preprocessed data for processing by the neural network, data input to the neural network, weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, activation function associated with each node or layer of the neural network, and neural network. It may be configured to include all or any combination thereof, such as a loss function for learning of . In addition to the foregoing configurations, the data structure comprising the neural network may include any other information that determines the characteristics of the neural network. In addition, the data structure may include all types of data used or generated in the computational process of the neural network, but is not limited to the above. A computer readable medium may include a computer readable recording medium and/or a computer readable transmission medium. A neural network may consist of a set of interconnected computational units, which may generally be referred to as nodes. These nodes may also be referred to as neurons. A neural network includes one or more nodes.

The data structure may include data input to the neural network. A data structure including data input to the neural network may be stored in a computer readable medium. Data input to the neural network may include training data input in a neural network learning process and/or input data input to the neural network after learning has been completed. Data input to the neural network may include data that has undergone pre-processing and/or data to be pre-processed. Pre-processing may include a data processing process for inputting data to a neural network. Accordingly, the data structure may include data subject to preprocessing and data generated by preprocessing. The foregoing data structure is only an example, and the present disclosure is not limited thereto.

The data structure may include the weights of the neural network. (In this specification, weights and parameters may be used in the same meaning.) Also, a data structure including weights of a neural network may be stored in a computer readable medium. A neural network may include a plurality of weights. The weight may be variable, and may be changed by a user or an algorithm in order to perform a function desired by the neural network. For example, when one or more input nodes are interconnected by respective links to one output node, the output node is set to a link corresponding to values input to input nodes connected to the output node and respective input nodes. A data value output from an output node may be determined based on the weight. The foregoing data structure is only an example, and the present disclosure is not limited thereto.

As a non-limiting example, the weights may include weights that are varied during neural network training and/or weights for which neural network training has been completed. The variable weight in the neural network learning process may include a weight at the time the learning cycle starts and/or a variable weight during the learning cycle. The weights for which neural network learning has been completed may include weights for which learning cycles have been completed. Accordingly, the data structure including the weights of the neural network may include a data structure including weights that are variable during the neural network learning process and/or weights for which neural network learning is completed. Therefore, it is assumed that the above-described weights and/or combinations of weights are included in the data structure including the weights of the neural network. The foregoing data structure is only an example, and the present disclosure is not limited thereto.

The data structure including the weights of the neural network may be stored in a computer readable storage medium (eg, a memory or a hard disk) after going through a serialization process. Serialization can be the process of converting a data structure into a form that can be stored on the same or another computing device and later reconstructed and used. A computing device may serialize data structures to transmit and receive data over a network. The data structure including the weights of the serialized neural network may be reconstructed on the same computing device or another computing device through deserialization. The data structure including the weights of the neural network is not limited to serialization. Furthermore, the data structure including the weights of the neural network is a data structure for increasing the efficiency of operation while minimizing the resource of the computing device (for example, B-Tree, Trie, m-way search tree, AVL tree, Red-Black Tree). The foregoing is only an example, and the present disclosure is not limited thereto.

The data structure may include hyper-parameters of the neural network. Also, the data structure including the hyperparameters of the neural network may be stored in a computer readable medium. A hyperparameter may be a variable variable by a user. Hyperparameters include, for example, learning rate, cost function, number of learning cycle iterations, weight initialization (eg, setting the range of weight values to be targeted for weight initialization), hidden unit number (eg, the number of hidden layers and the number of nodes in the hidden layer). The foregoing data structure is only an example, and the present disclosure is not limited thereto.

5 is a simplified and general schematic diagram of an exemplary computing environment in which embodiments of the present disclosure may be implemented.

Although the present disclosure has been described above as being generally embodied by a computing device, those skilled in the art will understand that the present disclosure may be combined with computer-executable instructions and/or other program modules that may be executed on one or more computers and/or may be implemented in hardware and software. It will be appreciated that it can be implemented as a combination.

Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In addition, those skilled in the art will understand that the methods of the present disclosure can be applied to single-processor or multiprocessor computer systems, minicomputers, mainframe computers as well as personal computers, handheld computing devices, microprocessor-based or programmable consumer electronics, and the like ( It will be appreciated that each of these may be implemented with other computer system configurations, including those that may be operative in connection with one or more associated devices.

The described embodiments of the present disclosure may also be practiced in distributed computing environments where certain tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

Computers typically include a variety of computer readable media. Computer readable media can be any medium that can be accessed by a computer, including volatile and nonvolatile media, transitory and non-transitory media, removable and non-transitory media. Includes removable media. By way of example, and not limitation, computer readable media may include computer readable storage media and computer readable transmission media. Computer readable storage media are volatile and nonvolatile media, transitory and non-transitory, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. includes media Computer readable storage media may include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital video disk (DVD) or other optical disk storage device, magnetic cassette, magnetic tape, magnetic disk storage device or other magnetic storage device. device, or any other medium that can be accessed by a computer and used to store desired information.

A computer readable transmission medium typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism. Including all information delivery media. The term modulated data signal means a signal that has one or more of its characteristics set or changed so as to encode information within the signal. By way of example, and not limitation, computer readable transmission media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of any of the above are also intended to be included within the scope of computer readable transmission media.

An exemplary environment 1100 implementing various aspects of the present disclosure is shown including a computer 1102, which includes a processing unit 1104, a system memory 1106, and a system bus 1108. do. System bus 1108 couples system components, including but not limited to system memory 1106 , to processing unit 1104 . Processing unit 1104 may be any of a variety of commercially available processors. Dual processor and other multiprocessor architectures may also be used as the processing unit 1104.

System bus 1108 may be any of several types of bus structures that may additionally be interconnected to a memory bus, a peripheral bus, and a local bus using any of a variety of commercial bus architectures. System memory 1106 includes read only memory (ROM) 1110 and random access memory (RAM) 1112 . A basic input/output system (BIOS) is stored in non-volatile memory 1110, such as ROM, EPROM, or EEPROM, and is a basic set of information that helps transfer information between components within computer 1102, such as during startup. contains routines. RAM 1112 may also include high-speed RAM, such as static RAM, for caching data.

The computer 1102 may also include an internal hard disk drive (HDD) 1114 (eg, EIDE, SATA) - the internal hard disk drive 1114 may also be configured for external use within a suitable chassis (not shown). Yes—a magnetic floppy disk drive (FDD) 1116 (e.g., for reading from or writing to a removable diskette 1118), and an optical disk drive 1120 (e.g., a CD-ROM) for reading disc 1122 or reading from or writing to other high capacity optical media such as DVDs). The hard disk drive 1114, magnetic disk drive 1116, and optical disk drive 1120 are connected to the system bus 1108 by a hard disk drive interface 1124, magnetic disk drive interface 1126, and optical drive interface 1128, respectively. ) can be connected to The interface 1124 for external drive implementation includes at least one or both of Universal Serial Bus (USB) and IEEE 1394 interface technologies.

These drives and their associated computer readable media provide non-volatile storage of data, data structures, computer executable instructions, and the like. In the case of computer 1102, drives and media correspond to storing any data in a suitable digital format. Although the description of computer readable media above refers to HDDs, removable magnetic disks, and removable optical media such as CDs or DVDs, those skilled in the art can use zip drives, magnetic cassettes, flash memory cards, cartridges, etc. It will be appreciated that other tangible media readable by the computer, such as the like, may also be used in the exemplary operating environment and any such media may include computer executable instructions for performing the methods of the present disclosure.

A number of program modules may be stored on the drive and RAM 1112, including an operating system 1130, one or more application programs 1132, other program modules 1134, and program data 1136. All or portions of the operating system, applications, modules and/or data may also be cached in RAM 1112. It will be appreciated that the present disclosure may be implemented in a variety of commercially available operating systems or combinations of operating systems.

A user may enter commands and information into the computer 1102 through one or more wired/wireless input devices, such as a keyboard 1138 and a pointing device such as a mouse 1140. Other input devices (not shown) may include a microphone, IR remote control, joystick, game pad, stylus pen, touch screen, and the like. Although these and other input devices are often connected to the processing unit 1104 through an input device interface 1142 that is connected to the system bus 1108, a parallel port, IEEE 1394 serial port, game port, USB port, IR interface, may be connected by other interfaces such as the like.

A monitor 1144 or other type of display device is also connected to the system bus 1108 through an interface such as a video adapter 1146. In addition to the monitor 1144, computers typically include other peripheral output devices (not shown) such as speakers, printers, and the like.

Computer 1102 may operate in a networked environment using logical connections to one or more remote computers, such as remote computer(s) 1148 via wired and/or wireless communications. Remote computer(s) 1148 may be a workstation, computing device computer, router, personal computer, handheld computer, microprocessor-based entertainment device, peer device, or other common network node, and generally includes It includes many or all of the components described for, but for simplicity, only memory storage device 1150 is shown. The logical connections shown include wired/wireless connections to a local area network (LAN) 1152 and/or a larger network, such as a wide area network (WAN) 1154 . Such LAN and WAN networking environments are common in offices and corporations and facilitate enterprise-wide computer networks, such as intranets, all of which can be connected to worldwide computer networks, such as the Internet.

When used in a LAN networking environment, computer 1102 connects to local network 1152 through wired and/or wireless communication network interfaces or adapters 1156. Adapter 1156 may facilitate wired or wireless communications to LAN 1152, which also includes a wireless access point installed therein to communicate with wireless adapter 1156. When used in a WAN networking environment, computer 1102 may include a modem 1158, be connected to a communicating computing device on WAN 1154, or establish communications over WAN 1154, such as over the Internet. have other means. A modem 1158, which may be internal or external and a wired or wireless device, is connected to the system bus 1108 through a serial port interface 1142. In a networked environment, program modules described for computer 1102, or portions thereof, may be stored on remote memory/storage device 1150. It will be appreciated that the network connections shown are exemplary and other means of establishing a communication link between computers may be used.

Computer 1102 is any wireless device or entity that is deployed and operating in wireless communication, eg, printers, scanners, desktop and/or portable computers, portable data assistants (PDAs), communication satellites, wireless detectable tags associated with It operates to communicate with arbitrary equipment or places and telephones. This includes at least Wi-Fi and Bluetooth wireless technologies. Thus, the communication may be a predefined structure as in conventional networks or simply an ad hoc communication between at least two devices.

Wi-Fi (Wireless Fidelity) makes it possible to connect to the Internet without wires. Wi-Fi is a wireless technology, such as a cell phone, that allows such devices, eg, computers, to transmit and receive data both indoors and outdoors, i.e. anywhere within coverage of a base station. Wi-Fi networks use a radio technology called IEEE 802.11 (a, b, g, etc.) to provide secure, reliable, and high-speed wireless connections. Wi-Fi can be used to connect computers to each other, to the Internet, and to wired networks (using IEEE 802.3 or Ethernet). Wi-Fi networks can operate in the unlicensed 2.4 and 5 GHz radio bands, for example, at 11 Mbps (802.11a) or 54 Mbps (802.11b) data rates, or in products that include both bands (dual band) .

Those skilled in the art will understand that information and signals may be represented using any of a variety of different technologies and techniques. For example, data, instructions, instructions, information, signals, bits, symbols and chips that may be referenced in the above description are voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields s or particles, or any combination thereof.

Those skilled in the art will understand that the various illustrative logical blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein are electronic hardware, (for convenience) , may be implemented by various forms of program or design code (referred to herein as software) or a combination of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the particular application and the design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

Various embodiments presented herein may be implemented as a method, apparatus, or article of manufacture using standard programming and/or engineering techniques. The term article of manufacture includes a computer program, carrier, or media accessible from any computer-readable storage device. For example, computer-readable storage media include magnetic storage devices (eg, hard disks, floppy disks, magnetic strips, etc.), optical disks (eg, CDs, DVDs, etc.), smart cards, and flash memory devices (eg, EEPROM, cards, sticks, key drives, etc.), but are not limited thereto. Additionally, various storage media presented herein include one or more devices and/or other machine-readable media for storing information.

It is to be understood that the specific order or hierarchy of steps in the processes presented is an example of exemplary approaches. Based upon design priorities, it is to be understood that the specific order or hierarchy of steps in the processes may be rearranged within the scope of this disclosure. The accompanying method claims present elements of the various steps in a sample order, but are not meant to be limited to the specific order or hierarchy presented.

The description of the presented embodiments is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art of this disclosure, and the general principles defined herein may be applied to other embodiments without departing from the scope of this disclosure. Thus, the present disclosure is not to be limited to the embodiments presented herein, but is to be interpreted in the widest scope consistent with the principles and novel features presented herein.

Claims (17)

A method for training a speech recognition model, performed by a computing device, comprising:
obtaining ground truth text information;
extracting a reference feature vector for the correct text information;
extracting a comparison feature vector for each of a plurality of tokens included in the correct answer text information;
determining an importance for each of the plurality of tokens based on the reference feature vector and each comparison feature vector; and
Learning the speech recognition model in consideration of the importance of each of the plurality of tokens included in the correct answer text information
including,
method.
delete According to claim 1,
The step of extracting a reference feature vector for the correct answer text information,
Extracting the reference feature vector by performing sentence embedding on the correct answer text information,
method.
According to claim 1,
Extracting each comparison feature vector for each of the plurality of tokens,
performing masking on one token among the plurality of tokens;
extracting a comparison feature vector corresponding to the masked token by performing sentence embedding on the correct answer text information to which the masking is reflected; and
Performing the masking operation and extracting a comparison feature vector corresponding to the masked token for each of the plurality of tokens.
including,
method.
According to claim 1,
Extracting each comparison feature vector for each of the plurality of tokens,
performing masking on a first token among the plurality of tokens;
extracting a first comparison feature vector corresponding to the first token by performing sentence embedding on the correct answer text information in which the masking of the first token is reflected;
performing masking on a second token among the plurality of tokens; and
extracting a second comparison feature vector corresponding to the second token by performing sentence embedding on the correct answer text information in which the masking of the second token is reflected;
including,
method.
According to claim 1,
Determining the importance of each of the plurality of tokens based on the reference feature vector and each comparison feature vector,
comparing the reference feature vector and each comparison feature vector, and calculating a similarity for each comparison feature vector; and
Based on each similarity, determining the importance of each token associated with each comparison feature vector.
including,
method.
According to claim 6,
Based on each degree of similarity, determining the importance of each token associated with each comparison feature vector,
Determining the importance of each token as the degree of similarity increases as the degree of similarity increases.
method.
A method for training a speech recognition model, performed by a computing device, comprising:
obtaining ground truth text information;
calculating a frequency of each of a plurality of tokens included in the correct answer text information in learning data to which the correct answer text information belongs;
determining an importance for each of the plurality of tokens based on the respective frequency; and
learning the speech recognition model in consideration of importance of each of the plurality of tokens included in the correct answer text information;
including,
method.
A method for training a speech recognition model, performed by a computing device, comprising:
obtaining ground truth text information;
determining an importance for each of a plurality of tokens included in the correct answer text information;
determining a weight for each of the plurality of tokens based on an importance of each of the plurality of tokens; and
Constructing a loss function in consideration of each weight
including,
method.
According to claim 9,
Constructing a loss function in consideration of each weight,
applying the weight to each cross entropy (CE) loss value for each of the plurality of tokens; and
Performing a weighted average operation based on each cross entropy loss value to which each weight is applied.
including,
method.
A computer program stored on a computer-readable storage medium, which, when executed on one or more processors, causes the computer program to perform the following operations for training a speech recognition model, the operations comprising:
obtaining ground truth text information;
extracting a reference feature vector for the correct answer text information;
extracting comparison feature vectors for each of a plurality of tokens included in the correct answer text information;
determining an importance of each of the plurality of tokens based on the reference feature vector and each comparison feature vector; and
Learning the voice recognition model by considering the importance of each of the plurality of tokens included in the correct answer text information
including,
A computer program stored on a computer readable storage medium.
A computing device for training a speech recognition model,
at least one processor; and
Memory
including,
The at least one processor,
obtaining ground truth text information;
extracting a reference feature vector for the correct answer text information;
extracting each comparison feature vector for each of a plurality of tokens included in the correct answer text information;
determine an importance for each of the plurality of tokens based on the reference feature vector and each comparison feature vector; and
In consideration of the importance of each of the plurality of tokens included in the correct answer text information, configured to learn the speech recognition model,
Device.
A method of performing speech recognition, performed by a computing device, comprising:
inputting voice data into a voice recognition model; and
generating text information corresponding to the input voice data;
including,
The voice recognition model,
a reference feature vector extracted for correct answer text information;
respective comparison feature vectors extracted for each of a plurality of tokens included in the correct answer text information;
Corresponding to a model learned based on correct answer text information in which importance information is determined in units of tokens based on the reference feature vector and each comparison feature vector,
method.
A computer program stored on a computer-readable storage medium, which, when executed on one or more processors, causes the computer program to perform the following operations for training a speech recognition model, the operations comprising:
obtaining ground truth text information;
calculating a frequency of each of a plurality of tokens included in the correct answer text information in learning data to which the correct answer text information belongs;
determining a degree of importance for each of the plurality of tokens based on the respective frequency; and
Learning the voice recognition model by considering the importance of each of the plurality of tokens included in the correct answer text information
including,
A computer program stored on a computer readable storage medium.

A computing device for training a speech recognition model,
at least one processor; and
Memory
including,
The at least one processor,
obtaining ground truth text information;
Calculate each frequency of each of a plurality of tokens included in the correct answer text information in the training data to which the correct answer text information belongs;
determine an importance for each of the plurality of tokens based on the respective frequency; and
In consideration of the importance of each of the plurality of tokens included in the correct answer text information, configured to learn the speech recognition model,
computing device.
A computer program stored on a computer-readable storage medium, which, when executed on one or more processors, causes the computer program to perform the following operations for training a speech recognition model, the operations comprising:
obtaining ground truth text information;
determining an importance for each of a plurality of tokens included in the correct answer text information;
determining a weight for each of the plurality of tokens based on the importance of each of the plurality of tokens; and
constructing a loss function in consideration of each of the weights;
including,
A computer program stored on a computer readable storage medium.
A computing device for training a speech recognition model,
at least one processor; and
Memory
including,
The at least one processor,
obtaining ground truth text information;
determining importance for each of a plurality of tokens included in the correct answer text information;
determine a respective weight for each of the plurality of tokens based on the importance of each of the plurality of tokens; and
Computing device configured to construct a loss function in consideration of each of the weights.

Images