Method, apparatus, and computer program for providing acoustic recognition results with improved reliability

Abstract

To provide a method, device and computer program for providing acoustic recognition results with improved accuracy according to various surrounding conditions, by dividing acoustic data into time-based segments and reducing the amount of calculation while improving a recognition rate in process of performing recognition for each segment.SOLUTION: In a system in which a user terminal, which is a wireless communication device that ensures portability and mobility, and an external server are connected to a server through a network to improve the recognition accuracy of acoustic data, a method implemented by the server includes the steps of: acquiring acoustic data using an artificial intelligence model; dividing the acoustic data to generate multiple pieces of acoustic sub-data; and processing the multiple pieces of acoustic sub-data as input to an acoustic recognition model to generate acoustic recognition result information corresponding to each acoustic sub-data.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to a method for improving the recognition rate of acoustic data. More specifically, the method divides acoustic data into intervals in time units, and while reducing the amount of calculation in the process of performing recognition for each interval, it aims to improve the recognition rate and provide an acoustic recognition result with improved accuracy corresponding to various surrounding situations.

Background Art

[0002] Hearing-impaired people who cannot completely hear sounds or cannot distinguish sounds well have many difficulties not only in daily life because it is difficult to judge the situation by listening to sounds, but also cannot recognize urgent situations in indoor and outdoor environments using sound information, so they cannot respond immediately. Not only hearing-impaired people, but also in situations where hearing is absent or restricted, such as pedestrians wearing earphones and the elderly, the sound generated around the user can be blocked. Additionally, in situations where it is difficult to perceive sounds, such as when the user is sleeping, the surrounding situation cannot be recognized, and the user may be placed in an urgent situation or at risk of an accident. On the other hand, the need for technology development to detect and recognize acoustic events in such environments is emerging. The technology for detecting and recognizing acoustic events is a technology applicable to various fields such as context recognition in real-life environments, urgent situation recognition, media content recognition, and situation analysis on wired communications, and is continuously being studied. Regarding acoustic event recognition technology, research mainly focuses on extracting various feature values such as MFCC, energy, spectral flux, and zero crossing rate from audio signals to verify excellent features, and research on classification methods such as Gaussian mixture model or rule-based methods. Recently, in order to improve the above methods, machine learning methods based on deep learning have been studied. However, such methods have a limitation in that the accuracy of acoustic detection is guaranteed at a low signal-to-noise ratio, and it is difficult to distinguish between surrounding noise and the sound of an event.

[0003] In other words, in real-world environments containing diverse ambient noise, reliable detection of acoustic events can be difficult. Specifically, in order to detect valid acoustic events, it is necessary to determine whether or not an acoustic event occurred in acoustic data acquired chronologically (i.e., continuously), and at the same time, the type of event that occurred must also be recognized, which can make it difficult to ensure high reliability. Furthermore, when two or more events occur simultaneously, the problem of recognizing multiple events (polyphonic) must be solved in addition to the problem of recognizing single events (monophonic), which can further reduce the recognition rate of acoustic events. Therefore, there may be a demand for acoustic recognition that improves accuracy and reliability by responding to acoustic data acquired over time in real-world environments. [Patent Document 1] Registered Patent No. 10-2014-0143069, Republic of Korea [Overview of the project] [Problems that the invention aims to solve]

[0004] The problem that this invention aims to solve is the aforementioned problem, and aims to improve the recognition rate while reducing the computational load in the process of dividing acoustic data into time-unit intervals and performing recognition for each interval, thereby providing acoustic recognition results with improved accuracy that can respond to various surrounding conditions. The problems that this invention aims to solve are not limited to those mentioned above, and any other problems not mentioned can be clearly understood by an ordinary person from the following description. [Means for solving the problem]

[0005] A method for providing acoustic recognition results with improved reliability, relating to various embodiments of the present invention for solving the aforementioned problems, is disclosed. The method may include the steps of acquiring acoustic data, dividing the acoustic data to generate a plurality of acoustic subdata, and processing the plurality of acoustic subdata as input to an acoustic recognition model to generate acoustic recognition result information corresponding to each acoustic subdata. In an alternative embodiment, the step of generating the plurality of acoustic subdata includes a step of dividing the acoustic data into predetermined size units to generate the plurality of acoustic subdata, and the acoustic recognition model may include a first recognition model that takes each of the plurality of acoustic subdata as input and provides a plurality of outputs corresponding to each acoustic subdata, and a second recognition model that takes combination re-verification acoustic subdata generated through combinations of the plurality of acoustic subdata as input and provides an output corresponding to the combination re-verification acoustic subdata. In an alternative embodiment, the step of generating the acoustic recognition result information includes a step of selecting re-verification acoustic subdata that require re-verification based on the acoustic recognition result information output by the first recognition model corresponding to each of the acoustic subdata, a step of generating combined re-verification acoustic subdata based on the selected re-verification acoustic subdata, and a step of processing the combined re-verification acoustic subdata as input to the second recognition model to generate acoustic recognition result information, wherein the second recognition model is implemented through a cloud API. In an alternative embodiment, the step of selecting the re-verification acoustic subdata may include the step of deriving similarity scores between recognition items related to the output of the first recognition model and the step of selecting the re-verification acoustic subdata based on the calculated similarity scores. In an alternative embodiment, the step of selecting the re-verification acoustic subdata may include the step of identifying whether the acoustic recognition result information related to the output of the first recognition model is included in a pre-set re-verification item, and, if the acoustic recognition result information is included in the re-verification item, selecting the acoustic subdata that forms the basis for calculating the acoustic recognition result information as re-verification acoustic subdata that requires re-verification. In an alternative embodiment, the first recognition model is characterized by outputting probability values ​​for each recognition item and generating acoustic recognition result information based on the recognition item corresponding to the probability value with the maximum value among the calculated probability values, and the step of selecting the re-verification acoustic subdata may include a step of selecting the acoustic subdata that forms the basis for calculating the acoustic recognition result information as re-verification acoustic subdata that requires re-verification if there are multiple recognition items among the probability values ​​for each recognition item calculated through the first recognition model that exceed a preset critical threshold value. In an alternative embodiment, the step of generating the acoustic recognition result information may include the step of generating relationship information between the acoustic recognition result information and the step of performing a correction on at least one of the acoustic recognition result information corresponding to each acoustic subdata based on the relationship information. In an alternative embodiment, the step of performing a correction on the at least one acoustic recognition result information may be characterized by performing a correction on at least one of the first acoustic recognition result and the second acoustic recognition result based on relationship information corresponding to the first acoustic recognition result and the second acoustic recognition result information, if the first acoustic recognition result information and the second acoustic recognition result information are generated within a predetermined time. In an alternative embodiment, the method further includes the step of generating mood information corresponding to the acoustic recognition result information by utilizing a mood sensing model, wherein the mood information is predictive information for the atmosphere related to the space in which the acoustic data is acquired, and includes location prediction information and emotion prediction information, and the step of generating the acoustic recognition result information includes the step of generating relationship information between the first acoustic recognition result information corresponding to the first acoustic subdata and the mood information corresponding to the first acoustic subdata, the step of not performing a correction to the first acoustic recognition result information if the relationship information is greater than or equal to a preset reference value, and the step of performing a correction to the first acoustic recognition result information if the relationship information is less than a preset reference value, and the mood sensing model is a neural network model that has been trained to recognize the acoustic recognition result information and output the mood information corresponding to the surrounding situation at each point in time.

[0006] According to another embodiment of the present invention, an acoustic recognition result providing device having improved reliability is disclosed. The device includes a memory for storing one or more instructions and a processor for executing the one or more instructions stored in the memory, the processor being able to perform the acoustic recognition result providing method having improved reliability described above by executing the one or more instructions. According to yet another embodiment of the present invention, a computer program stored on a computer-readable recording medium is disclosed. The computer program is coupled with a computer, which is hardware, to perform the aforementioned method for providing acoustic recognition results with improved reliability. Further specific details of the present invention are included in the detailed description and drawings. [Effects of the Invention]

[0007] Through various embodiments of the present invention, it is possible to divide acoustic data into time-based intervals, reduce the computational load while improving the recognition rate during the recognition process for each interval, and provide acoustic recognition results with improved accuracy that can handle diverse surrounding conditions. The effects of the present invention are not limited to those mentioned above, and any further effects not mentioned can be clearly understood by an ordinary person from the following description. [Brief explanation of the drawing]

[0008] [Figure 1] This diagram schematically shows a system for carrying out an improved reliability acoustic recognition result provision method according to one embodiment of the present invention. [Figure 2] This is a hardware configuration diagram of a server that performs an improved reliability acoustic recognition result provision method related to one embodiment of the present invention. [Figure 3] A flowchart illustrating an exemplary method for providing acoustic recognition results with improved reliability, related to one embodiment of the present invention, is shown. [Figure 4A] This is an illustrative diagram illustrating the process of dividing acoustic data related to one embodiment of the present invention. [Figure 4B] This is an illustrative diagram illustrating the process of dividing acoustic data related to one embodiment of the present invention. [Figure 5] This flowchart exemplifies the process of generating acoustic recognition result information through re-verification using multiple artificial intelligence models related to one embodiment of the present invention. [Figure 6] This is an illustrative diagram illustrating the process by which an acoustic recognition model related to one embodiment of the present invention derives an acoustic recognition result. [Figure 7] This is an illustrative diagram illustrating the process of recognizing acoustic data using a first recognition model and a second recognition model related to one embodiment of the present invention. [Figure 8] This flowchart exemplifies the process of performing corrections to acoustic recognition result information based on the relationships between acoustic subdata related to one embodiment of the present invention. [Figure 9A] This is an illustrative diagram illustrating the process of correcting acoustic recognition result information related to one embodiment of the present invention. [Figure 9B]This is an illustrative diagram illustrating the process of correcting acoustic recognition result information related to one embodiment of the present invention. [Figure 10] This flowchart exemplifies the process of generating acoustic recognition result information while considering the surrounding environment and atmosphere, related to one embodiment of the present invention. [Figure 11] This is an illustrative diagram illustrating the limitations of acoustic recognition result information derived through a general recognition process related to one embodiment of the present invention. [Figure 12] This is an illustrative diagram illustrating a process in which the overall atmosphere is grasped through corresponding acoustic subdata at points before and after specific acoustic subdata related to one embodiment of the present invention, and acoustic recognition result information with improved accuracy is output considering the grasped atmosphere. [Modes for carrying out the invention]

[0009] Various embodiments are described with reference to the drawings. Various descriptions are presented herein to provide an understanding of the invention. However, it is evident that such embodiments may be carried out without such specific descriptions. As used herein, terms such as "component", "module", "system", etc. refer to computer-related entities, hardware, firmware, software, combinations of software and hardware, or execution of software. For example, a component can be a procedure executed on a processor, a processor, an object, an execution thread, a program, and / or a computer, but is not limited thereto. For example, an application executed on a computing device and the computing device can both be components. One or more components can reside within a processor and / or execution thread. A component can be localized within one computer. A component can be distributed between two or more computers. Also, such components can execute from various computer-readable media having various data structures stored therein. A component can communicate through local and / or remote processing via, for example, a signal having one or more data packets (e.g., data and / or signals transmitted through a network such as the Internet from one component interacting with other components in a local system or a distributed system). Also, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, it is intended to mean one of the natural inclusive substitutions of "X uses A or B" when not otherwise specified or clear from the context. That is, "X uses A or B" can apply to any of these cases when X uses A; X uses B; or X uses both A and B. Also, the term "and / or" used herein should be understood to refer to and include any and all possible combinations of one or more of the listed related items. Furthermore, the terms “contains” and / or “contains” should be understood to mean that the features and / or components in question are present. However, the terms “contains” and / or “contains” should be understood not to exclude the presence or addition of one or more other features, components and / or groups thereof. Also, where not otherwise specified or clearly indicated in context, the singular should generally be interpreted as “one or more” in this specification and claims.

[0010] Those skilled in the art should recognize that the various exemplary logical blocks, configurations, modules, circuits, means, logic, and algorithmic stages described in relation to the embodiments disclosed herein can be embodied in electronic hardware, computer software, or a combination of both. To clearly illustrate the interoperability of hardware and software, the various exemplary components, blocks, configurations, means, logic, modules, circuits, and stages have been generally described in terms of their functional aspects. Whether such functionality is embodied in hardware or software depends on the specific application and design constraints imposed on the overall system. A skilled technician can embodied the described functionality in various ways for each specific application; however, the decision to embodied such functionality should not be construed as departing from the scope of the invention. The descriptions of the presented embodiments are provided so that a person with ordinary skill in the art of the invention may utilize or practice the invention. Various modifications of such embodiments will be obvious to a person with ordinary skill in the art of the invention. The general principles defined herein may be applied to other embodiments without departing from the scope of the invention. The invention is not limited to the embodiments presented herein. The invention should be interpreted in the broadest sense, consistent with the principles and novel features presented herein. In this specification, "computer" means all types of hardware devices including at least one processor, and can be understood to also include software configurations that operate on the corresponding hardware devices according to embodiments. For example, "computer" can be understood to include all smartphones, tablet PCs, desktops, laptop computers, user clients and applications driven by each device, and is not limited thereto. Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Although each step described in this specification is described as being performed by a computer, the subject of each step is not limited thereto, and at least a part of each step may be performed by different devices according to embodiments. The method for providing an acoustic recognition result with improved reliability according to various embodiments of the present invention can include a method of performing additional verification on acoustic data recognized so that the recognition rate of the acoustic data is improved, a method of correcting at least a part of the acoustic recognition results based on the acoustic recognition results output within a certain time, or a method of grasping the surrounding environment and atmosphere in real time and deriving the acoustic recognition result in consideration of this together with the acoustic data. In an embodiment, the improvement in the recognition accuracy of the acoustic data may mean that the recognition accuracy of sensing a specific event with the acoustic data is improved.

[0011] A method for performing additional verification on acoustic data may, for example, involve using multiple neural network models (e.g., acoustic recognition models) to derive and verify additional acoustic recognition results multiple times, thereby improving the reliability of the final acoustic recognition result. In this case, each of the multiple neural network models may have a different output process. For example, each neural network model may use acoustic data of different lengths as input; one model may derive an acoustic recognition result based on 1-second intervals of acoustic data, while another model may derive an acoustic recognition result based on 5-second intervals of acoustic data. However, this is not limited to this, and the multiple neural network models may have different performance characteristics. As a concrete example, a relatively low-performance, lightweight first recognition model can be used to perform acoustic recognition initially and output an acoustic recognition result. If the reliability of the initially output acoustic recognition result is judged to be somewhat inferior, acoustic recognition will be performed again through a second recognition model with higher performance than the first recognition model, and the first acoustic recognition result will be re-verified. This configuration utilizes the first recognition model under normal circumstances and only uses the second recognition model for re-verification when necessary. This has the advantage of effectively reducing computational complexity while improving the reliability of the recognition results. Furthermore, a method of correcting at least a portion of the acoustic recognition results based on the acoustic recognition results output within a certain time period may mean, for example, post-processing correction of the acoustic recognition results corresponding to the acoustic data. Acoustic data may be input in segments at regular time intervals, and the present invention outputs an acoustic recognition result corresponding to each segment of data. In this case, it is possible to understand the relationships between acoustic recognition results output within a certain time period and correct a specific acoustic recognition result to another acoustic recognition result. As a concrete example, if an acoustic recognition result of sound A (e.g., gunshot) is output, and then an acoustic recognition result of sound B (e.g., applause) is output within n seconds, the recognition result of sound A can be corrected to sound C (e.g., firecracker sound). In other words, by considering the relationships between acoustic recognition results output (or recognized) within a certain time period and performing a correction on at least one of the acoustic recognition results, the recognition accuracy can be improved. Furthermore, the present invention can grasp the surrounding environment and atmosphere in real time and consider this together with the acquired acoustic data to derive acoustic recognition results, thereby improving the accuracy of acoustic data recognition. By utilizing a separate neural network model (e.g., a mood sensing model) that senses the surrounding environment and atmosphere at the time the acoustic data is acquired, it is possible to recognize the situation before and after the input of the acoustic data, and by considering this together with the acoustic data during the acoustic analysis process, the reliability of the output acoustic recognition results can be improved. As a specific example, in the case of sound A (e.g., a firecracker sound), a general acoustic recognition model may recognize it as a gunshot. In the case of a short explosion sound, it is difficult to output acoustic recognition results with high accuracy. The present invention can grasp the atmosphere corresponding to the situation in which the acoustic data is acquired based on the acoustic recognition results sensed before and after the sound is detected, and derive acoustic recognition results by considering this together with the acoustic data. For example, by identifying that singing, applause, and cheering were detected before and after the recognition of sound A, and understanding that the atmosphere associated with sound A is celebratory (or party, cheerful, or happy), the system will recognize the explosion sound as a "firecracker" rather than a "gunshot." In other words, by grasping the overall atmosphere based on the sound recognition results output before and after the sound data to be recognized, and by performing recognition of the sound data while considering the grasped atmosphere, the accuracy of the output of the sound recognition results can be improved. A more specific explanation related to methods for improving the accuracy of sound data recognition will be given in detail below.

[0012] Figure 1 is a schematic diagram illustrating a system for carrying out an improved reliability acoustic recognition result provision method according to one embodiment of the present invention. As shown in Figure 1, the system for carrying out an improved reliability acoustic recognition result provision method according to one embodiment of the present invention may include a server 100 for improving the recognition accuracy of acoustic data, a user terminal 200, and an external server 300. Here, the system for carrying out an improved reliability acoustic recognition result provision method illustrated in Figure 1 is according to one embodiment, and its components are not limited to the embodiment shown in Figure 1, and may be added, modified, or deleted as necessary. In one embodiment, a server 100 for improving the accuracy of acoustic data recognition can determine whether a specific event has occurred based on the acoustic data. Specifically, the server 100 for improving the accuracy of acoustic data recognition can acquire acoustic data related to real life and determine whether a specific event has occurred through analysis of the acquired acoustic data. In one embodiment, the specific event is related to security, safety, or the occurrence of a hazard, and may be related to the occurrence of, for example, an alarm sound, a child crying, the sound of glass breaking, or the sound of a tire puncturing. The specific description of the acoustics related to the aforementioned specific event is merely an example, and the present invention is not limited thereto. According to the embodiment, the server 100 for improving the accuracy of acoustic data recognition can include any server embodied by an API (Application Programming Interface). For example, a user terminal 200 can acquire acoustic data and transmit it to the server 100 via a Cloud API. For example, the server 100 can acquire acoustic data from the user terminal 200 and determine, through analysis of the acoustic data, that an emergency alarm sound (e.g., a siren) has occurred. In the embodiment, the server 100 for improving the accuracy of acoustic data recognition can perform analysis of the acoustic data through an acoustic recognition model (e.g., an artificial intelligence model). In one embodiment, an acoustic recognition model (e.g., an artificial intelligence model) consists of one or more network functions, which may consist of a set of interconnected computational units that can generally be referred to as “nodes.” Such “nodes” may also be referred to as “neurons.” One or more network functions consist of at least one or more nodes. The nodes (or neurons) that make up one or more network functions may be connected to one or more “links.”

[0013] Within an artificial intelligence model, one or more nodes connected through links can form a relative input-output node relationship. The concepts of input and output nodes are relative; any node that is an output node to another node can be an input node to another node, and vice versa. As mentioned above, the input-output node relationship can be generated primarily through links. One input node can be connected to one or more output nodes via links, and vice versa. In a relationship between input and output nodes connected through a single link, the value of the output node can be determined based on the data input to the input node. Here, the nodes connecting the input and output nodes to each other may have weights. These weights can be variable and can be changed by the user or algorithm to enable the artificial intelligence model to perform the desired function. For example, if one or more input nodes are connected to a single output node by their respective links, the output node's value can be determined based on the values ​​input to the input nodes connected to the output node and the weights set for the links corresponding to each input node. As mentioned above, an artificial intelligence model consists of one or more nodes connected to each other through one or more links, forming the relationship between input and output nodes within the AI ​​model. The characteristics of an AI model can be determined by the number of nodes and links within the AI ​​model, the relationships between nodes and links, and the weight values ​​assigned to each link. For example, if there are two AI models with the same number of nodes and links but different weight values ​​between links, the two AI models may be perceived as different from each other. Some of the nodes that make up an artificial intelligence model can form a layer based on their distance from the initial input node. For example, a set of nodes whose distance from the initial input node is n can form an n-layer. The distance from the initial input node can be defined by the minimum number of links that must be traversed to reach that node. However, such a definition of a layer is arbitrary for illustrative purposes, and the order of layers within an artificial intelligence model can be defined in ways other than those described above. For example, the layer of nodes may be defined by their distance from the final output node. The initial input node may mean one or more nodes in the artificial intelligence model to which data is directly input without going through links in relation to other nodes. Alternatively, it may mean a node in the artificial intelligence model network that does not have other input nodes connected to links in relation to nodes based on links. Similarly, the final output node may mean one or more nodes in the artificial intelligence model that do not have output nodes in relation to other nodes. Furthermore, a hidden node may mean a node that constitutes the artificial intelligence model that is neither the initial input node nor the final output node. An artificial intelligence model according to one embodiment of the present invention may have more nodes in the input layer than in the hidden layer which is closer to the output layer, and the number of nodes may decrease as one progresses from the input layer to the hidden layer.

[0014] An artificial intelligence model may contain one or more hidden layers. The hidden nodes of a hidden layer can take the output of the previous layer and the output of surrounding hidden nodes as input. The number of hidden nodes in each hidden layer may be the same or different. The number of nodes in an input layer may be determined based on the number of data fields in the input data, and may be the same or different as the number of hidden nodes. The input data input to an input layer may be processed by the hidden nodes of the hidden layers and output by a fully connected layer (FCL), which is the output layer. In various implementations, artificial intelligence models can be trained using supervised learning, with multiple acoustic data sets and the corresponding feature information for each set of acoustic data as training data. However, this is not the only possible approach, and a variety of learning methods can be applied. Here, supervised learning typically involves labeling specific data and related information to generate training data, and then using this data for learning. In other words, it refers to a method where two causally related data are labeled to generate training data, and learning is performed through the generated training data. In one embodiment, a server 100 for improving the accuracy of acoustic data recognition can use validation data to determine whether to interrupt training if the training of one or more network functions has been completed for a predetermined number of epochs or more. The predetermined number of epochs may be part of the overall training target epoch. Validation data can consist of at least a portion of the labeled training data. That is, the server 100 for improving the recognition accuracy of acoustic data can perform training of an artificial intelligence model through the training data, and after the training of the artificial intelligence model has been repeated for a predetermined number of epochs or more, it can use the validation data to determine whether the learning effect of the artificial intelligence model is above a predetermined level. For example, if the server 100 for improving the recognition accuracy of acoustic data performs training with a target number of iterations of 10 using 100 training data, after performing 10 iterations of training, which is a predetermined number of epochs, it can perform 3 iterations of training using 10 validation data. If the change in the output of the artificial intelligence model during the 3 iterations of training is below a predetermined level, it can determine that further training is meaningless and terminate the training. In other words, validation data can be used to determine the completion of learning in the iterative learning of an artificial intelligence model based on whether the effect of epoch-by-epoch learning is above or below a certain level. The number of training data and validation data and the number of iterations mentioned above are merely examples, and the present invention is not limited thereto.

[0015] Server 100 for improving the accuracy of acoustic data recognition can generate an artificial intelligence model by using test data to test the performance of one or more network functions and determine whether or not to activate one or more network functions. Test data can be used to verify the performance of the artificial intelligence model and may consist of at least a portion of the training data. For example, 70% of the training data may be used for training the artificial intelligence model (i.e., training to adjust weights to output result values ​​similar to labels), and 30% may be used as test data to verify the performance of the artificial intelligence model. Server 100 for improving the accuracy of acoustic data recognition can input the test data into the trained artificial intelligence model, measure the error, and determine whether or not to activate the artificial intelligence model based on whether the performance is above a predetermined level. Server 100, which improves the accuracy of acoustic data recognition, can verify the performance of a trained artificial intelligence model using test data. If the performance of the trained artificial intelligence model exceeds a predetermined standard, the server can activate the artificial intelligence model for use in other applications. Furthermore, the server 100 for improving the recognition accuracy of acoustic data can deactivate and discard an artificial intelligence model if its performance falls below a predetermined standard after training. For example, the server 100 for improving the recognition accuracy of acoustic data can determine the performance of a generated artificial intelligence model based on factors such as accuracy, precision, and recall. The performance evaluation criteria mentioned above are merely examples, and the present invention is not limited thereto. The server 100 for improving the recognition accuracy of acoustic data can generate multiple artificial intelligence models by independently training each artificial intelligence model, evaluate their performance, and use only those artificial intelligence models that meet or exceed a certain performance standard. However, it is not limited thereto. Throughout this specification, the terms computational model, neural network, network function, and neural network may be used interchangeably (hereinafter, "neural network" will be used consistently). A data structure may include a neural network, and a data structure including a neural network may be stored on a computer-readable medium. A data structure including a neural network may also include data input to the neural network, weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, activation functions associated with each node or layer of the neural network, and a loss function for training the neural network. A data structure including a neural network may include any of the components of the configurations disclosed above. That is, a data structure including a neural network may consist of all or any combination thereof of data input to the neural network, weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, activation functions associated with each node or layer of the neural network, and a loss function for training the neural network. In addition to the configurations described above, a data structure including a neural network may include any other information that determines the properties of the neural network. Furthermore, the data structure may include, and is not limited to, all forms of data used or generated in the computational processes of the neural network. Computer-readable media may include computer-readable recording media and / or computer-readable transmission media. A neural network may consist of a set of interconnected computational units that may generally be referred to as nodes. Such nodes may also be referred to as neurons. A neural network consists of at least one node.

[0016] According to one embodiment of the present invention, the server 100 for improving the accuracy of acoustic data recognition may be a server that provides cloud computing services. More specifically, the server 100 for improving the accuracy of acoustic data recognition may be a type of internet-based computing service that provides cloud computing services where information is processed on other computers connected to the internet rather than on the user's computer. The cloud computing service may be a service that stores materials on the internet and allows users to access necessary materials and programs anytime, anywhere via an internet connection without having to install them on their own computers, and allows for easy sharing and transmission of materials stored on the internet with simple operations and clicks. Furthermore, the cloud computing service may not only simply store materials on a server on the internet, but also allow users to perform desired tasks using the functions of applications provided on the web without having to install separate programs, and may be a service that allows many people to share documents and work on them simultaneously. In addition, the cloud computing service may be embodied in at least one form from among IaaS (Infrastructure as a Service), PaaS (Platform as a Service), SaaS (Software as a Service), virtual machine-based cloud servers, and container-based cloud servers. In other words, the server 100 for improving the recognition accuracy of acoustic data according to the present invention can be embodied in at least one form of the cloud computing service described above. The specific description of the cloud computing service described above is merely illustrative, and the present invention may include any platform for building a cloud computing environment. In various embodiments, the server 100 for improving the accuracy of acoustic data recognition can be connected to a user terminal 200 via a network, generate and provide an acoustic recognition model for analyzing acoustic data, and not only that, can provide the user terminal with information (e.g., acoustic recognition result information) obtained by analyzing the acoustic data through the acoustic recognition model. Here, a network can refer to a connected structure in which information can be exchanged between nodes such as multiple terminals and servers. For example, networks include local area networks (LANs), wide area networks (WANs), the internet (WWW), wired / wireless data networks, telephone networks, and wired / wireless television networks. Furthermore, wireless data communication networks here include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), LTE (Long Term Evolution), WiMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth networks, NFC (Near-Field Communication) networks, satellite broadcasting networks, analog broadcasting networks, and DMB (Digital Multimedia Broadcasting) networks.

[0017] In one embodiment, the user terminal 200 can be connected via a network to a server 100 for improving the accuracy of acoustic data recognition, and can provide acoustic data to the server 100 for improving the accuracy of acoustic data recognition. In response to the provided acoustic data, the user can receive acoustic recognition result information regarding the occurrence of various events (for example, the occurrence of gunshots, alarms, children crying, the sound of glass breaking, the sound of a tire puncturing, etc.). Here, the user terminal 200 is a wireless communication device that ensures portability and mobility, and can include, but is not limited to, all types of handheld-based wireless communication devices such as navigation systems, PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminals, smartphones, smartpads, and tablet PCs. For example, the user terminal 200 may be installed in a specific area to perform sensing related to that area. For example, the user terminal 200 may be installed in a vehicle to acquire acoustic data generated while the vehicle is parked or in motion. As another example, the user terminal 200 may be provided for sensing an urgent situation in the area where the children are located (e.g., a playground or facility space). The above description of the specific location or place where the user terminal may be provided is illustrative and the present invention is not limited thereto. In one embodiment, the external server 300 may be connected to the server 100 for improving the recognition accuracy of acoustic data via a network. The external server 300 can provide various information / data necessary for the server 100 to analyze acoustic data using an artificial intelligence model, and can receive, store, and manage the result data derived as the acoustic data analysis using the artificial intelligence model is carried out. For example, the external server 300 may be, but is not limited to, a storage server provided separately outside the server 100 for improving the recognition accuracy of acoustic data. The hardware configuration of the server 100 for improving the recognition accuracy of acoustic data will be described below with reference to Figure 2. Figure 2 is a hardware configuration diagram of a server for improving the recognition accuracy of acoustic data related to one embodiment of the present invention. Referring to Figure 2, a server 100 for improving the recognition accuracy of acoustic data according to one embodiment of the present invention (hereinafter, "server 100") may include one or more processors 110, memory 120 for loading computer programs 151 executed by the processors 110, a bus 130, a communication interface 140, and storage 150 for storing the computer programs 151. Here, only components related to embodiments of the present invention are illustrated in Figure 2. Therefore, a person of the art who is an ordinary person to the art to which the present invention belongs will see that other general-purpose components may be included in addition to the components illustrated in Figure 2.

[0018] The processor 110 controls the overall operation of each configuration of the server 100. The processor 110 may be composed of a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphics Processing Unit), or any form of processor widely known in the art of the present invention. The processor 110 can read a computer program stored in memory 120 and perform data processing for an artificial intelligence model according to one embodiment of the present invention. According to one embodiment of the present invention, the processor 110 can perform calculations for training a neural network. The processor 110 can perform calculations for training a neural network, such as processing input data for training using deep learning (DL), extracting features from the input data, calculating errors, and updating the weight values ​​of the neural network using backpropagation. Furthermore, the processor 110 can process network function learning using at least one of the CPU, GPGPU, and TPU. For example, the CPU and GPGPU can both process network function learning and data classification using network functions. Also, in one embodiment of the present invention, the processors of multiple computing devices can be used together to process network function learning and data classification using network functions. Moreover, a computer program executed by a computing device according to one embodiment of the present invention may be a CPU, GPGPU, or TPU executable program. In this specification, network functions may be used interchangeably with artificial neural networks and neural networks. In this specification, a network function may include one or more neural networks, in which case the output of the network function may be an ensemble of the outputs of one or more neural networks. The processor 110 can read a computer program stored in memory 120 and provide an acoustic recognition model according to one embodiment of the present invention. According to one embodiment of the present invention, the processor 110 can perform calculations to train the acoustic recognition model. According to one embodiment of the present invention, the processor 110 can typically handle the overall operation of the server 100. The processor 110 can provide or process appropriate information or functions to the user or user terminal by processing signals, data, information, etc. that are input or output through the components described in detail above, or by driving applications stored in the memory 120. Furthermore, the processor 110 can perform calculations for at least one application or program to carry out the method according to the embodiment of the present invention, and the server 100 may be equipped with one or more processors. In various embodiments, the processor 110 may further include RAM (Random Access Memory, not shown) and ROM (Read-Only Memory, not shown) for temporarily and / or permanently storing signals (or data) processed within the processor 110. Furthermore, the processor 110 may be implemented as a system-on-chip (SoC) including at least one of the graphics processing unit, RAM, and ROM. Memory 120 stores various data, instructions, and / or information. Memory 120 can load a computer program 151 from storage 150 to perform a method / operation according to various embodiments of the present invention. Once the computer program 151 is loaded into memory 120, the processor 110 can perform the method / operation by executing one or more instructions that constitute the computer program 151. Memory 120 may be embodied in volatile memory such as RAM, but the technical scope of the present invention is not limited thereto.

[0019] Bus 130 provides communication functionality between the components of server 100. Bus 130 can be implemented as various types of buses, such as an address bus, a data bus, and a control bus. The communication interface 140 supports wired / wireless internet communication for the server 100. The communication interface 140 may also support a variety of communication methods other than internet communication. For this purpose, the communication interface 140 may be configured to include communication modules widely known in the art of the present invention. In some embodiments, the communication interface 140 may be omitted. The storage 150 can temporarily store the computer program 151. When a process to improve the recognition accuracy of acoustic data is performed through the server 100, the storage 150 can store various information necessary to provide the process to improve the recognition accuracy of acoustic data. The storage 150 may consist of non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium widely known in the art to which the present invention belongs. When the computer program 151 is loaded into memory 120, it may include one or more instructions that cause the processor 110 to perform methods / operations according to various embodiments of the present invention. That is, the processor 110 can perform the methods / operations according to various embodiments of the present invention by executing the one or more instructions. In one embodiment, the computer program 151 may include one or more instructions to perform a method for providing acoustic recognition results with improved reliability, which includes the steps of acquiring acoustic data, dividing the acoustic data to generate multiple acoustic subdata, and processing the multiple acoustic subdata as input to an acoustic recognition model to generate acoustic recognition result information corresponding to each acoustic subdata. Steps of the methods or algorithms described in relation to embodiments of the present invention may be embodied directly in hardware, in software modules executed by hardware, or in combination thereof. The software modules may reside on RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, hard disk, removable disk, CD-ROM, or any form of computer-readable recording medium widely known in the art to which the present invention belongs.

[0020] The components of the present invention can be embodied in a program (or application) and stored on a medium for execution in conjunction with a computer, which is hardware. The components of the present invention can be executed by software programming or software elements, and similarly, embodiments include a variety of algorithms embodied in combinations of data structures, processes, routines or other programming configurations, which can be embodied in programming or scripting languages ​​such as C, C++, Java, and assembler. Functional aspects can be embodied in algorithms executed on one or more processors. The method for providing acoustic recognition results with improved reliability, performed by server 100, will now be described with reference to Figures 3 to 12. Figure 3 is an illustrative flowchart illustrating a method for providing acoustic recognition results with improved reliability related to one embodiment of the present invention. The steps shown in Figure 3 may be changed in order as necessary, and at least one or more steps may be omitted or added. In other words, the following steps represent only one embodiment of the present invention, and the scope of the present invention is not limited thereto. According to one embodiment of the present invention, a method for providing acoustic recognition results with improved reliability may include a step of acquiring acoustic data (S100). The server 100 can acquire acoustic data 10. The acoustic data 10 may include information related to sounds acquired in real life. Acquisition of acoustic data according to one embodiment of the present invention may involve receiving or loading acoustic data 10 stored in memory 120. Alternatively, acquisition of acoustic data may involve receiving or loading data from other storage media, other computing devices, or separate processing modules within the same computing device based on wired / wireless means. According to one embodiment, the acoustic data 10 can be acquired through a user terminal 200 associated with the user. For example, the user terminal 200 associated with the user may include all types of handheld wireless communication devices such as smartphones, smartpads, and tablet PCs, or electronic devices installed in a specific space (e.g., the user's living space) (e.g., devices capable of receiving acoustic data through a microphone). According to one embodiment of the present invention, a method for providing acoustic recognition results with improved reliability may include a step (S200) of dividing acoustic data to generate a plurality of acoustic sub-data. In the embodiment, the step of generating a plurality of acoustic sub-data may include a step of dividing acoustic data into a predetermined size unit to generate a plurality of acoustic sub-data. Server 100 can divide the acoustic data 10 into fixed time units to generate multiple acoustic subdata. Multiple acoustic subdata may be obtained by dividing the time-series acoustic data into multiple frames based on specific time units. Specifically, Server 100 can construct multiple acoustic subdata by dividing the acoustic data into units of a predetermined size for a first time unit. In the case of acoustics (or sound), since it is time-series data that occurs continuously, it must be divided into fixed time intervals and recognition must be performed for each interval. For example, if the first acoustic data is acquired corresponding to a time period of 1 minute, the server 100 can set the first time unit to 1 second and divide the first acoustic data to 60 acoustic sub-data (or acoustic subframes). The specific numerical descriptions related to the first time unit and one or more acoustic sub-data mentioned above are merely illustrative, and the present invention is not limited thereto. In the embodiment, a benefit of dividing acoustic data into multiple acoustic sub-data is that the smaller the time unit in which it is divided, the less computational load the neural network (i.e., the acoustic recognition model) has, and the better the performance. If the computational load is significantly reduced by shortening the length of the acoustic sub-data corresponding to the input, the acoustic recognition model can be used as a short-interval recognition model and can therefore be used as an Edge SDK. That is, by dividing acoustic data into multiple short intervals, the acoustic recognition model can be driven through a single application that can be executed on the server itself. In the examples, for certain sounds, it may be difficult to derive acoustic recognition result information with improved reliability through an acoustic recognition model that performs short-interval recognition. For example, in the case of sounds with short duration, there is a risk that the sound in question may not be accurately recognized by referring only to a specific interval. As a more concrete example, referring to Figure 4A, in the case of the first acoustic data 10a associated with a short-duration gunshot, it can be divided into 1-second intervals and separated into the first acoustic sub-data 11a to the eighth acoustic sub-data 18a. In this case, the gunshot occurring in the third acoustic sub-data 13a is entirely contained within the interval corresponding to the third acoustic sub-data 13a, so accurate recognition may be possible even when considering only one interval (i.e., the third acoustic sub-data 13a).

[0021] On the other hand, referring to Figure 4B, the second acoustic data 10b associated with a short-duration gunshot can also be similarly divided into 1-second intervals and separated into the first acoustic sub-data 11b to the eighth acoustic sub-data 18b. In this case, as illustrated in Figure 4B, the gunshot may span the second acoustic sub-data 12b and the third acoustic sub-data 13b. Since acoustic data is acquired chronologically and divided through a certain magnitude, as mentioned above, the explosion sound and echo of the gunshot may exist in different intervals. In this case, when recognizing the second acoustic sub-data 12b and the third acoustic sub-data 13b respectively through an acoustic recognition model, accurate recognition may be impossible because they do not contain the entire gunshot, and consequently, the accuracy of the output (i.e., recognition accuracy) may decrease. In other words, the accuracy of sound recognition can decrease when short-duration sounds are divided into different intervals. To overcome this, the time unit for dividing the sound can be increased (for example, 2 seconds), but in such cases, the computational complexity of the sound recognition model increases, making it impossible to use in inexpensive devices and thus inefficient. Accordingly, the server 100 of the present invention can perform acoustic recognition again in situations where recognition accuracy is lacking by additionally configuring a separate neural network model. That is, by utilizing multiple neural network models to derive additional acoustic recognition results multiple times and performing re-verification, the reliability of the final acoustic recognition result can be improved. In particular, rather than simply utilizing multiple neural network models multiple times, by performing additional verification in response to specific situations where some degree of recognition accuracy is predicted to be lacking, it is possible to improve the reliability of the recognition result while effectively reducing the computational load. An explanation of the configuration for improving recognition accuracy by performing re-verification using multiple neural network models will be described in detail below with reference to Figure 5. Figure 5 is a flowchart illustrating the process of generating acoustic recognition result information through re-verification using multiple artificial intelligence models related to one embodiment of the present invention. The steps shown in Figure 5 may be changed in order as necessary, and at least one or more steps may be omitted or added. In other words, the following steps represent only one embodiment of the present invention, and the scope of the present invention is not limited thereto. According to one embodiment, the step of generating acoustic recognition result information may include a step (S311) of selecting re-verification acoustic subdata that require re-verification of acoustic subdata based on the acoustic recognition result information output by the first recognition model 210 corresponding to each acoustic subdata. According to an embodiment, the acoustic recognition model of the present invention may include a first recognition model 210 that takes a plurality of acoustic subdata as input and provides a plurality of outputs corresponding to each acoustic subdata. In one embodiment, the server 100 can train an autoencoder through unsupervised learning. Specifically, the server 100 can train the dimensionality reduction network function (e.g., encoder) and dimensionality restoration network function (e.g., decoder) that constitute the autoencoder to output output data similar to the input data. More specifically, through the dimensionality reduction network function, only the core feature data (or features) of the input acoustic data can be learned through the hidden layer during the encoding process, while the remaining information is lost. In this case, the output data of the hidden layer during the decoding process through the dimensionality restoration network function may not be a perfect copy but an approximation of the input data (i.e., acoustic data). That is, the server 100 can train the autoencoder by adjusting the weights so that the output data and the input data are as similar as possible.

[0022] An autoencoder can be a type of neural network that outputs output data similar to input data. An autoencoder may include at least one hidden layer, and an odd number of hidden layers may be placed between the input and output layers. The number of nodes in each layer may be reduced from the number of nodes in the input layer to an intermediate layer called a bottleneck layer (encoding), and then expanded symmetrically from the bottleneck layer to the output layer (symmetric to the input layer). The number of input and output layers may correspond to the number of input data items remaining after preprocessing of the input data. An autoencoder structure may have a structure in which the number of nodes in the hidden layers included in the encoder decreases as it moves away from the input layer. The number of nodes in the bottleneck layer (the layer with the fewest nodes located between the encoder and decoder) may be maintained above a certain number (e.g., more than half of the input layer) because if it is too small, a sufficient amount of information may not be transmitted. Server 100 can match and store object-specific feature data output from a training dataset containing multiple training data each tagged with object information, using the training dataset as input to a reduced-dimensionality network. Specifically, Server 100 can use a reduced-dimensionality network function to obtain feature data of the first object for the training data included in the first training data subset, using a first training data subset tagged with first acoustic discrimination information (e.g., the sound of glass breaking) as input to the reduced-dimensionality network function. The obtained feature data can be represented as vectors. In this case, the feature data output corresponding to each of the multiple training data included in the first training data subset can be located relatively close to each other in the vector space because they are outputs through the training data related to the first acoustic. Server 100 can match and store the first acoustic discrimination information (i.e., the sound of glass breaking) with the feature data related to the first acoustic represented as vectors. For a trained autoencoder's dimensionality reduction network function, it can be trained to extract features that allow the dimensionality restoration network function to well reconstruct the input data. Furthermore, for example, multiple training data points included in each of the first training data subsets tagged with a second acoustic identification information (e.g., the sound of a siren) can be transformed into feature data (i.e., features) through a dimensionality reduction network function and displayed in a vector space. In this case, since the relevant feature data is the output through the training data associated with the second acoustic identification information (i.e., the sound of a siren), it can be located relatively close to each other in the vector space. In this case, the feature data corresponding to the second acoustic identification information may be displayed in a different vector space than the feature data corresponding to the first acoustic identification information (e.g., the sound of glass breaking). In this embodiment, the server 100 can construct an acoustic recognition model that includes a dimensionality reduction network function using a trained autoencoder. That is, when an acoustic recognition model constructed that includes a dimensionality reduction network function generated through the training process described above takes acoustic subdata as input, it can extract feature information (i.e., features) corresponding to the acoustic subdata through calculations utilizing the dimensionality reduction network function. In this case, the acoustic recognition model can evaluate the similarity of acoustic styles by comparing the distance between the area where the features corresponding to the acoustic subdata are displayed and the vector space of the object-specific feature data, and can output acoustic recognition result information corresponding to the acoustic subdata based on the similarity evaluation. In one embodiment, the acoustic recognition result information may include multiple recognition items and probability values ​​corresponding to each of the multiple recognition items.

[0023] Specifically, an acoustic recognition model (or first recognition model) can output feature information (i.e., features) by processing acoustic subdata using a dimensionality reduction network function. In this case, the acoustic recognition model can output multiple recognition items corresponding to the acoustic subdata and probability values ​​corresponding to each recognition item, based on the position between the output feature information corresponding to the acoustic subdata and the acoustic discrimination information feature data pre-recorded in a vector space through learning. Multiple recognition items are used to identify what items an acoustic sound is associated with, and may include, but are not limited to, gunshots, firecrackers, screams, flat tires, sirens, puppy barks, and rain. Such multiple recognition items can be generated based on acoustic identification information that is close in position in the vector space to the feature information output in correspondence with the acoustic subdata. For example, an acoustic recognition model can construct multiple recognition items through acoustic identification information matched to feature information that is close in position to the first feature information output in correspondence with the first acoustic subdata. The specific descriptions of multiple recognition items mentioned above are merely illustrative, and the present invention is not limited thereto. The probability values ​​corresponding to each recognition item may represent information about the accuracy predicted for each recognition item. For example, an acoustic recognition model can construct multiple recognition item information through acoustic identification information matched to feature information located close to the first feature information output in relation to the first acoustic subdata. In this case, a higher probability value may be calculated if the first feature information is close to the feature information corresponding to each acoustic identification information, and a lower probability value may be calculated if the first feature information is farther from the feature information corresponding to each acoustic identification information. As a concrete example, as illustrated in Figure 6, the acoustic recognition model (or first recognition model) can output multiple recognition items 310 corresponding to the first acoustic subdata, such as "siren sound," "scream," "glass breaking sound," and "other sound." The acoustic recognition model can also output probability values ​​320 corresponding to each recognition item 310, such as "1," "95," "3," and "2." That is, the acoustic recognition model can output probability values ​​320 corresponding to the first acoustic subdata, with a probability of "1" for the siren sound, a probability of "95" for the scream, a probability of "3" for the glass breaking sound, and a probability of "2" for the other sound. The specific numerical values ​​for each of the recognition items and probability values ​​described above are merely examples, and the present invention is not limited thereto. In other words, the server 100 can output recognition items and probability values ​​corresponding to each of the multiple acoustic subdata generated based on acoustic data through the acoustic recognition model. For example, the acoustic recognition model can output a first recognition item and the probability value corresponding to the first recognition item for the first acoustic subdata, and can output a second recognition item and the probability value corresponding to the second recognition item for the second acoustic subdata. According to one embodiment, the first recognition model 210 calculates probability values ​​for each recognition item and generates acoustic recognition result information based on the recognition item corresponding to the maximum probability value among the calculated probability values. For example, referring to Figure 6, acoustic recognition result information indicating that the recognition result of the acoustic subdata is "scream" is generated through the recognition item corresponding to the maximum probability value of 95%. Furthermore, in the embodiment, the first recognition model 210 may be characterized by calculating probability values ​​for each recognition item and generating acoustic recognition result information based on whether the calculated probability values ​​exceed critical probability values ​​set in advance for each recognition item. According to various embodiments, the acoustic recognition model of the present invention may comprise multiple models. The acoustic recognition model may include a first recognition model 210 and a second recognition model 220. In this case, the second recognition model 220 may be a model in which the length of the interval of the acoustic subdata used as input differs from that of the first recognition model 210. The second recognition model 220 may be a neural network model having higher performance than the first recognition model 210. For example, the first recognition model 210 may be a neural network model trained to derive an acoustic recognition result based on acoustic subdata in a 1-second interval, and the second recognition model 220 may be a neural network model that derives an acoustic recognition result based on acoustic subdata in a 2-second interval. That is, the first recognition model 210 and the second recognition model 220 are neural network models with different performance levels, and the second recognition model 220, which has better performance, will be used for re-verification. In the embodiment, the first recognition model 210 may be a neural network model that recognizes acoustic subdata over a relatively short interval and outputs the result, and the second recognition model 220 may be a neural network model that recognizes acoustic subdata over a relatively long interval and outputs the result.

[0024] The first recognition model 210 has the advantage of requiring less computation and thus less computing power consumption because it analyzes acoustics over a short period. On the other hand, the second recognition model 220 has the disadvantage of requiring more computation and being less likely to be installed and used as an application program on inexpensive devices because it analyzes acoustics over a relatively long period. However, because it recognizes an extended period, it has the advantage of higher output accuracy. In various embodiments, the server 100 can decide whether to perform re-verification using other neural network models (e.g., a second recognition model) based on the output results corresponding to the first recognition model 210, i.e., acoustic recognition result information. This is to reduce computational complexity and computing power used by utilizing the lightweight first recognition model 210 under normal circumstances, and only using additional neural network models when output accuracy is insufficient. Server 100 can select acoustic subdata that require re-verification based on the acoustic recognition result information output by the first recognition model 210 corresponding to each acoustic subdata. Server 100 can perform re-verification when it determines that the acoustic recognition result information is related to a pre-classified item (for example, an item requiring re-verification) or when it determines that the accuracy of the acoustic recognition result information is ambiguous, and can select acoustic subdata to be re-verified. In a specific embodiment, the server 100 can calculate similarity scores between each recognition item calculated through the first recognition model 210, and select re-verification acoustic subdata based on the calculated similarity scores. For example, the first recognition model can output recognition items such as "gunshot," "balloon bursting," and "fireworks display" in response to specific acoustic subdata, and can calculate probability values ​​of 85%, 90%, and 94% for each recognition item. In this case, acoustic recognition result information can be generated based on the fireworks display, which has the highest probability value. Meanwhile, the server 100 can calculate the similarity score between the recognition items output by the first recognition model 210. In this embodiment, the similarity score between recognition items can be calculated based on the similarity evaluation between the sounds corresponding to each item. As an example, by inputting representative acoustic subdata corresponding to each recognition item into a dimensionality reduction network function, an embedding corresponding to each recognition item can be obtained, and the similarity between recognition items can be evaluated through the embedding similarity. Alternatively, for example, the similarity between each recognition item may be evaluated using cosine similarity, which measures similarity by calculating the cosine angle between embedded word vectors. As another example, a Siamese network and transformer-based model can be used to determine semantic text similarity. The specific descriptions of keyword similarity evaluation methods mentioned above are merely illustrative, and the present invention is not limited thereto. In additional embodiments, Jacquard similarity, Euclidean distance, and WordNet may be used to measure similarity between keywords or words. The similarity score is derived based on the similarity evaluation between the recognition items output by the first recognition model 210. A higher similarity score is calculated when the similarity between the recognition items is high, and a lower similarity score is calculated when the similarity between the recognition items is low. According to the embodiment, the higher the similarity between the recognition items, the lower the output accuracy of the first recognition model 210 may be. For example, if the recognition items output by the first recognition model 210 corresponding to the A acoustic subdata are "scream," "collision sound," and "sound of rain," the acoustic similarity between these items is low, so the similarity score will be calculated to be low. In other words, the more different the candidate recognition items for predicting the result are, the more clearly different the features of each item are, thus ensuring the reliability of the output of the first recognition model. On the other hand, if the recognition items output by the first recognition model 210 corresponding to the B acoustic subdata are "gunshot," "balloon bursting sound," and "fireworks," the acoustic similarity between these items is high, so the similarity score may be calculated to be high. In other words, the candidate recognition items for predicting the result are very similar, so the difficulty in distinguishing the features of each item is high, which may slightly lower the reliability of the output of the first recognition model. In other words, when recognition items with completely dissimilar acoustic features (e.g., screams, cracking sounds, and rain sounds) are output, the final recognition result is more likely to be accurate (because the acoustic features of each recognition item are completely different, so the neural network has a high probability of providing the correct output), but when recognition items with a high degree of similarity (e.g., gunshots, firecrackers, and fireworks sounds) are output, the final recognition result is more likely to be somewhat inaccurate.

[0025] In other words, the server 100 can determine whether the acoustic recognition result information is reliable by calculating the similarity score of the recognition items output by the first recognition model 210. If the similarity score between recognition items is low (i.e., below a certain threshold), the server 100 determines that the acoustic recognition result information is reliable and does not perform separate re-verification of the acoustic sub-data. If the similarity score between recognition items is high (i.e., exceeds a certain threshold), the server 100 determines that the reliability of the acoustic recognition result information is low and selects the acoustic sub-data as re-verification acoustic sub-data that requires re-verification. In the example, re-verification of the acoustic subdata may mean using the second recognition model 220 to derive the acoustic recognition result information again. Furthermore, if multiple recognition items exceed a predetermined critical threshold value among the probability values ​​for each recognition item calculated through the first recognition model 210, the server 100 can select the acoustic subdata that forms the basis for calculating acoustic recognition result information as re-verified acoustic subdata that requires re-verification. For example, the first recognition model can output recognition items such as "gunshot," "balloon bursting," and "fireworks display" in response to specific acoustic subdata, and can calculate probability values ​​of 85%, 90%, and 94% for each recognition item. In this case, the acoustic recognition result information will be generated based on the fireworks display, which has the highest probability value. However, all other recognition items can exceed the pre-set critical threshold values ​​(e.g., 82%, 85%, and 88%) corresponding to each recognition item. This means that the first recognition model 210 has determined that the other recognition items also have high probabilities, which may mean that the reliability of the acoustic recognition result information is somewhat reduced. For example, if the probability values ​​of recognition items a, b, and c are 97%, 6%, and 3%, only recognition item a's probability value exceeds the pre-set critical threshold value (e.g., 80%), so the reliability of the acoustic recognition result information may be high. In other words, the fewer the number of recognition items whose probability values ​​exceed the critical threshold corresponding to each recognition item, the more likely the final recognition result to be accurate (for example, if there is only one recognition item whose probability value exceeds the critical threshold, the more likely the reliability of the recognition result is to be the highest). However, the more recognition items whose probability values ​​exceed the critical threshold corresponding to each recognition item, the more likely the final recognition result to be somewhat inaccurate. In other words, the server 100 can determine whether the acoustic recognition result information is reliable by identifying whether multiple recognition items exceed a predetermined critical threshold among the probability values ​​for each recognition item calculated through the first recognition model 210. If only one recognition item exceeds the critical threshold in probability value, the server 100 determines that the acoustic recognition result information is reliable and does not perform a separate re-verification of the acoustic sub-data. If multiple recognition items (i.e., two or more) exceed the critical threshold in probability value, the server 100 determines that the acoustic recognition result information is poor and selects the acoustic sub-data as re-verification acoustic sub-data that requires re-verification. Furthermore, in the embodiment, the step of selecting re-verification acoustic subdata may include the step of identifying whether the acoustic recognition result information related to the output of the first recognition model 210 is included in the pre-set re-verification items, and, if the acoustic recognition result information is included in the re-verification items, the step of selecting the acoustic subdata that forms the basis for calculating the acoustic recognition result information as re-verification acoustic subdata that requires re-verification. The pre-configured re-verification items may be a combination of predefined items that are expected to slightly reduce the accuracy of the output. For example, in the case of an explosion, it is difficult to distinguish between a gunshot, a balloon bursting, a firecracker exploding, and a fireworks display because their acoustic characteristics are similar. If the acoustic recognition result information is generated based on only one of these sounds, it is difficult to guarantee sufficient reliability. Therefore, the present invention makes it possible to construct pre-configured re-verification items by pre-saving items that are difficult to classify acoustically, and when the first recognition model 210 outputs acoustic recognition result information included in the pre-configured re-verification items, the server 100 can select the acoustic subdata that formed the basis for generating the corresponding acoustic recognition result information as re-verification acoustic subdata that requires re-verification. In other words, pre-configured re-verification items can be set based on similar sounds that are difficult to distinguish, and if the output of the first recognition model (i.e., sound recognition result information) corresponds to a pre-configured re-verification item, the corresponding sound sub-data can be determined as the re-verification sound sub-data.

[0026] According to the embodiment, the step of generating acoustic recognition result information may include the step of generating combined re-verification acoustic sub-data based on selected re-verification acoustic sub-data (S312). Specifically, when re-verified acoustic subdata is selected, server 100 can identify additional acoustic subdata related to the preceding and succeeding data based on the re-verified acoustic subdata. Furthermore, server 100 can combine the re-verified acoustic subdata and the additional acoustic subdata to generate combined re-verified acoustic subdata. Furthermore, in the embodiment, the step of generating acoustic recognition result information may include the step of processing the combined re-verification acoustic subdata as input to the second recognition model 220 to generate acoustic recognition result information (S313). According to an embodiment, the acoustic recognition model of the present invention may include a second recognition model 220 that takes combined re-validated acoustic subdata generated through combinations of multiple acoustic subdata as input and provides an output corresponding to the combined re-validated acoustic subdata. In this case, the second recognition model 220 may be a model in which the length of the interval of the acoustic subdata used as input differs from that of the first recognition model 210. The second recognition model 220 may be a neural network model with higher performance than the first recognition model 210. For example, the first recognition model 210 may be a neural network model trained to derive acoustic recognition results based on acoustic subdata in a 1-second interval, while the second recognition model 220 may be a neural network model that derives acoustic recognition results based on acoustic subdata in a 2-second interval. That is, the first recognition model 210 and the second recognition model 220 are neural network models with different performance levels, and the second recognition model 220, which has better performance, will be used for re-verification. In the embodiment, the first recognition model 210 may be a neural network model that recognizes acoustic subdata in a relatively short interval and outputs the result, while the second recognition model 220 may be a neural network model that recognizes acoustic subdata in a relatively long interval and outputs the result. In this embodiment, the second recognition model 220 is characterized by being implemented through a cloud API. Because the second recognition model 220 is implemented through a cloud API, it may be possible to utilize a relatively high-performance second recognition model 220 even with low-performance equipment. As a concrete example, referring to Figure 7, it is possible to output acoustic recognition result information by first utilizing a relatively low-performance and lightweight first recognition model 210 to take as input each of the multiple acoustic sub-data (first acoustic sub-data 11, second acoustic sub-data 12, and third acoustic sub-data 13) that were initially generated corresponding to the acoustic data 10. If the reliability of the acoustic recognition result information output temporarily in response to the second acoustic subdata 12 is deemed somewhat low, the acoustic subdata corresponding to the acoustic recognition result information (i.e., the second acoustic subdata) can be selected as re-verified acoustic subdata that requires re-examination. In this case, the server 100 can identify additional acoustic subdata generated before and after the second acoustic subdata 12. For example, it can identify the third acoustic subdata 13, which is acoustic subdata generated at a point adjacent to the second acoustic subdata 12, as additional acoustic subdata.

[0027] Server 100 will generate combined re-verified acoustic subdata by combining the re-verified acoustic subdata and the additional acoustic subdata. That is, combined re-verified acoustic subdata 20 can be generated through the combination of the second acoustic subdata 12 and the third acoustic subdata 13. In this case, the combined re-verification acoustic subdata 20 can have a longer length than the acoustic subdata input to the first recognition model 210 because it is generated through a combination of acoustic subdata. That is, if it is determined that the recognition result for the acoustic subdata is likely to be inaccurate, the combined re-verification acoustic subdata can be generated to perform a re-verification of the acoustic subdata that leads to the inaccurate output. Server 100 can process the combinatorial re-verification acoustic subdata 20 as input to the second recognition model 220 and output acoustic recognition result information. In this case, since the second recognition model is a neural network model that performs recognition on acoustic subdata over a longer interval than the first recognition model, it derives acoustic recognition results through analysis of acoustic subdata over a larger range than the first recognition model (i.e., combinatorial re-verification acoustic subdata), thus generating acoustic recognition result information with higher accuracy. In other words, server 100 utilizes a relatively lightweight first recognition model and, when necessary, uses the second model only to perform re-verification, thereby effectively reducing computational complexity while improving the reliability of the recognition results. Figure 8 is a flowchart illustrating the process of performing corrections to acoustic recognition result information based on the relationships between acoustic subdata related to one embodiment of the present invention. According to the embodiment, the step of generating acoustic recognition result information may include a step of generating correlation information between multiple pieces of acoustic recognition result information (S321). Correlation information may be information indicating the relationship between each piece of acoustic recognition result information. For example, if the first acoustic recognition result information generated as a recognition result for the first acoustic subdata corresponding to the first section is "gunshot," and the second acoustic recognition result information generated as a recognition result for the second acoustic subdata corresponding to the second section is "applause," then the server 100 can generate correlation information indicating that the first acoustic recognition result information and the second acoustic recognition result information are unrelated. The specific descriptions of each piece of acoustic recognition result information and the specific descriptions of the correlation information described above are merely examples, and the present invention is not limited thereto. Furthermore, the step of generating acoustic recognition result information may include a step (S322) of performing a correction on at least one acoustic recognition result information corresponding to each acoustic subdata based on the relationship information. The server 100 can perform a correction on at least one acoustic recognition result information based on the relationship information between the acoustic recognition result information. The server 100 can decide whether or not to correct the acoustic recognition result information based on the relationship information between each acoustic recognition information. According to a specific embodiment, the step of performing a correction on at least one acoustic recognition result information can be characterized by performing a correction on at least one of the first and second acoustic recognition results based on relationship information corresponding to the first and second acoustic recognition results information, if the first and second acoustic recognition result information are generated within a predetermined time. In other words, the server 100 can perform post-processing corrections on at least a portion of the acoustic recognition result information based on the acoustic recognition result information output within a certain time period. The acoustic data is divided into frames at regular time intervals and input into the neural network, and acoustic recognition result information is derived corresponding to each divided data. The server 100 generates correlation information between acoustic recognition results output within a certain time period, and through this, can correct specific acoustic recognition result information to other acoustic recognition results based on the correlation relationships between generated acoustic recognition results corresponding to adjacent time points. This configuration is intended to improve accuracy when generating acoustic recognition result information based solely on acoustic subdata. For example, if an acoustic recognition result of sound A (e.g., gunshot) is output, and then an acoustic recognition result of sound B (e.g., applause) is output within n seconds, the recognition result of sound A can be corrected to sound C (e.g., firecracker sound). In other words, by considering the relationship between acoustic recognition results output (or recognized) within a certain time period and performing a correction on at least one of the acoustic recognition results, recognition accuracy can be improved.

[0028] As a more concrete example, if a single frame (i.e., one acoustic subdata) like that shown in Figure 9A is input to an acoustic recognition model, the acoustic recognition result information "gunshot" may be generated. This result is derived simply by analyzing one acoustic subdata related to an explosion sound, so high accuracy cannot be guaranteed. For example, in the case of "gunshot," there are many sounds with similar acoustic characteristics, such as the sound of firecrackers, fireworks, or balloons bursting, so there is a risk of some lack of accuracy. As shown in Figure 9B, based on a specific acoustic recognition result (i.e., gunshot), the server 100 can identify that the acoustic recognition result information corresponding to frames existing within a certain time period is the sound of singing a birthday song, applause, and sounds of joy. The server 100 can generate association information indicating that the specific acoustic recognition result (i.e., gunshot) is unrelated to other acoustic recognition result information through other acoustic recognition results output within a certain time period. Accordingly, the specific acoustic recognition result information is post-processed and corrected to match the other acoustic recognition results. That is, the acoustic recognition result information "gunshot" that was initially output can be corrected to "the sound of a firecracker exploding," which is another acoustic recognition result information, based on the association information generated based on its relationship with other acoustic recognition result information (e.g., the sound of singing a birthday song, applause, and sounds of joy) that were output within a certain time period. In this embodiment, the server 100 can identify that the initially output gunshot is included in a specific item (e.g., an item corresponding to an explosion sound), and can select other types of keywords included in the specific item to perform corrections to the acoustic recognition result information. Server 100 can select keywords from other types of sounds included in a specific item that are related to other sound recognition results (for example, the sound of firecrackers exploding) and perform corrections to the initially output sound recognition result information. Accordingly, the initially output sound recognition result information, "gunshot," will be corrected to "the sound of firecrackers exploding." In other words, the server 100 can improve the accuracy of recognition by performing post-processing correction on at least one of the acoustic recognition result information, taking into account the relationships between the acoustic recognition result information output (or recognized) within a certain period of time. To put it another way, the accuracy of the acoustic recognition result can be improved by identifying specific acoustic recognition result information that does not have similarity to the acoustic recognition result information output at a nearby time, and correcting it to other acoustic recognition result information that is related to the other acoustic recognition result information. Figure 10 is a flowchart illustrating the process of generating acoustic recognition result information considering the surrounding environment and atmosphere related to one embodiment of the present invention. According to the embodiment, the server 100 can generate mood information corresponding to the acoustic recognition result information through a mood sensing model. The mood sensing model may be a neural network model that has been trained to recognize the acoustic recognition result information and output mood information corresponding to the surrounding situation at each point in time. Mood information is predictive information about the atmosphere associated with the space in which acoustic data is acquired, and may include location prediction information and emotion prediction information. A mood sensing model can be a neural network model that grasps the real-time atmosphere based on acoustic recognition result information derived from each acoustic subdata. The mood sensing model can generate mood information, including predictive information about where the current acoustic data is acquired and predictive information about the atmosphere in which the acoustic data is acquired, through the acoustic recognition result information. The mood sensing model can analyze the user's emotions through keyword analysis and pitch analysis of speech, and can recognize the location where the dialogue is taking place. In some embodiments, the mood sensing model can include a natural language processing model and an emotion analysis model. The natural language processing model can be used to understand and process text data (e.g., acoustic recognition result information), and the emotion analysis model can be used to sense the emotions contained in the text.

[0029] For example, by sensing the continuous acquisition of acoustic recognition results related to vehicle noise, it can be determined that the location where the current acoustic data is being acquired is a "vehicle in motion." As another example, by sensing the acquisition of acoustic recognition results with progressively higher tones, it can be determined that the real-time atmosphere is related to the emotion of "anger." The specific descriptions of location and emotion mentioned above are merely illustrative, and the present invention is not limited thereto. Server 100 can generate acoustic recognition result information based on mood information and acoustic subdata generated through the mood sensing model. This is because it recognizes the situation or atmosphere before and after inputting the acoustic subdata and takes this into account to generate more accurate acoustic recognition result information. Referring to Figure 11, specific acoustic subdata can be processed as input to an acoustic recognition model to derive acoustic recognition result information such as "child scream." However, it can be difficult to distinguish whether the child's scream is related to an urgent situation or whether the child is screaming because they are having fun while playing. In other words, not all screams indicate an urgent situation, so it is necessary to distinguish whether a scream is related to an urgent situation or occurred in a fun situation. The present invention can generate mood information corresponding to the atmosphere at the time surrounding specific acoustic subdata through a mood model, thus enabling the deriving of acoustic recognition result information that is more appropriate to the situation. Server 100 can utilize a mood sensing model to grasp the atmosphere before and after, i.e., mood information, and use this to generate more appropriate acoustic recognition result information. For example, as shown in Figure 12, Server 100 can sense sounds such as "the sound of splashing water," "children's laughter," and "a mother's voice calling her child" at adjacent moments, and generate acoustic recognition result information such as "a scream a child makes in a fun situation" corresponding to specific acoustic subdata. In a specific embodiment, the step of generating acoustic recognition result information may include a step (S331) of generating relationship information between the first acoustic recognition result information corresponding to the first acoustic subdata and the mood information corresponding to the first acoustic subdata. In the embodiment, the relevance information may be information that quantifies the degree of relevance between mood information and acoustic recognition result information, which are generated in real time. For example, if the mood information is related to "happy feelings" and the acoustic recognition result information is related to "gunshots," the server 100 can determine that the degree of relevance between the mood information and the acoustic recognition result information is low and generate relevance information indicating a relevance of 20%. As another example, if the mood information is related to "anger and fear" and the acoustic recognition result information is related to "gunshots," the server 100 can determine that the degree of relevance between the mood information and the acoustic recognition result information is high and generate relevance information indicating a relevance of 93%. The specific description of the generation of relevance information described above is merely illustrative, and the present invention is not limited thereto. The stage of generating acoustic recognition result information may include a stage (S332) in which no correction is performed on the first acoustic recognition result information if the relevance information is equal to or greater than a preset threshold value. The preset threshold value is determined in advance by the user and can serve as a criterion for determining whether the acoustic recognition result information output in real time is appropriate. If the relevance information between the acoustic recognition result information and the mood information is equal to or greater than a preset threshold value, it may be determined that the currently output acoustic recognition result information is appropriate and no further correction may be performed. The step of generating acoustic recognition result information may include a step (S333) of performing a correction on the first acoustic recognition result information if the relevance information is below a preset threshold value. If the relevance information between the acoustic recognition result information and the mood information is below a preset threshold value, the accuracy of the currently output acoustic recognition result information has decreased, and therefore a separate correction must be performed. More specifically, the server 100 can identify multiple keywords that have similarity to the first acoustic recognition result information. For example, if the first acoustic recognition result information is "gunshot," the server 100 can identify multiple keywords such as "firecracker sound," "fireworks sound," and "balloon bursting sound." The specific descriptions of the first acoustic recognition result information and multiple keywords mentioned above are merely illustrative, and the present invention is not limited thereto. Furthermore, server 100 can generate multiple relevance sub-informations between each of the multiple keywords and the mood information. That is, it can calculate the degree of relevance between each keyword and the mood information and generate multiple relevance sub-informations. For example, if the mood information is "happy atmosphere," a first relevance sub-information of 93% may be generated corresponding to the sound of firecrackers, a second relevance sub-information of 82% corresponding to the sound of fireworks being launched, and a third relevance sub-information of 76% corresponding to the sound of balloons bursting. The specific numerical descriptions for each relevance sub-information are merely examples, and the present invention is not limited thereto.

[0030] Server 100 can identify the maximum relevance sub-information among multiple relevance sub-informations and perform corrections to the first acoustic recognition result information based on the keyword corresponding to the maximum relevance sub-information. For example, Server 100 can perform corrections to the first acoustic recognition result information based on the keyword corresponding to the maximum relevance sub-information (93%) (i.e., firecracker sound). In other words, if the server 100 determines that the specific acoustic recognition result information does not match the mood information (i.e., if the correlation information between the acoustic recognition result information and the mood information is below a predetermined threshold value), it can perform a correction to the specific acoustic recognition result information through the keyword that is most highly related to the mood information among keywords similar to the acoustic recognition result. Steps of the methods or algorithms described in relation to embodiments of the present invention may be embodied directly in hardware, in software modules executed by hardware, or in combination thereof. The software modules may reside on RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, hard disk, removable disk, CD-ROM, or any form of computer-readable recording medium widely known in the art to which the present invention belongs. The components of the present invention may be embodied in a program (or application) and stored on a medium for execution in conjunction with a computer, which is hardware. The components of the present invention may be executed by software programming or software elements; similarly, embodiments include a variety of algorithms embodied in combinations of data structures, processes, routines, or other programming configurations, and may be embodied in programming or scripting languages ​​such as C, C++, Java, and assembler. Functional aspects may be embodied in algorithms executed on one or more processors. Although embodiments of the present invention have been described above with reference to the attached drawings, a person of ordinary skill in the art to which the present invention pertains will understand that the present invention can be carried out in other specific forms without changing its technical idea or essential features. Therefore, the embodiments described above should be understood to be illustrative and not restrictive in all respects.

Claims

1. In a method performed by one or more processors of a computing device, The stage of acquiring acoustic data; The stage of dividing acoustic data to generate multiple acoustic sub-data; and The step includes processing the plurality of acoustic subdata as input to an acoustic recognition model to generate acoustic recognition result information corresponding to each acoustic subdata; The step of generating the aforementioned multiple acoustic subdata is: The step includes dividing the aforementioned acoustic data into predetermined size units to generate multiple acoustic sub-data; The aforementioned acoustic recognition model, The system includes a first recognition model that takes each of the aforementioned plurality of acoustic subdata as input and provides a plurality of outputs corresponding to each acoustic subdata, and a second recognition model that takes combination re-verification acoustic subdata generated through combinations of the plurality of acoustic subdata as input and provides outputs corresponding to the combination re-verification acoustic subdata. The step of generating the aforementioned acoustic recognition result information is: The process includes the steps of: selecting re-verification acoustic subdata based on acoustic recognition result information output by the first recognition model corresponding to each of the acoustic subdata; generating combined re-verification acoustic subdata based on the selected re-verification acoustic subdata; and inputting the combined re-verification acoustic subdata into the second recognition model for processing and generating acoustic recognition result information. The step of generating combined re-verification acoustic subdata based on the selected re-verification acoustic subdata includes identifying the selected re-verification acoustic subdata and acoustic subdata generated at adjacent time points as additional acoustic subdata, and generating combined re-verification acoustic subdata by combining the selected re-verification acoustic subdata and the additional acoustic subdata. The second recognition model is a neural network model that performs recognition of acoustic subdata over a longer interval than the first recognition model, and, A method for providing acoustic recognition results with improved reliability, comprising the step of selecting the re-verification acoustic subdata based on acoustic recognition result information output by the first recognition model in correspondence to each of the acoustic subdata, the step of identifying whether the acoustic recognition result information output by the first recognition model is included in a pre-set re-verification item, and, if the acoustic recognition result information is included in the re-verification item, selecting the acoustic subdata that formed the basis for calculating the acoustic recognition result information as re-verification acoustic subdata that requires re-verification.

2. The second recognition model is A method for providing acoustic recognition results with improved reliability according to claim 1, characterized in that it is implemented via a cloud API.

3. The step of selecting the aforementioned re-verified acoustic subdata is, A step of deriving similarity scores between recognition items related to the output of the first recognition model; and A method for providing an acoustic recognition result with improved reliability according to claim 1, comprising the step of selecting re-verification acoustic subdata based on the calculated similarity score.

4. The first recognition model is It is characterized by outputting probability values ​​for each recognition item and generating acoustic recognition result information based on the recognition item corresponding to the probability value that corresponds to the maximum value among the calculated probability values. The step of selecting the aforementioned re-verified acoustic subdata is, A method for providing acoustic recognition results with improved reliability according to claim 1, further comprising the step of selecting acoustic subdata that form the basis for calculating the acoustic recognition result information as re-verified acoustic subdata that requires re-verification if there are multiple recognition items among the probability values ​​for each recognition item calculated through the first recognition model that exceed a predetermined critical threshold;

5. Memory to store one or more instructions; and Includes a processor that executes one or more instructions stored in the memory, An apparatus comprising a processor that performs the method described in claim 1 by executing one or more instructions.

6. A computer program stored on a computer-readable recording medium so that it can be coupled with a computer, which is hardware, to perform the method described in claim 1.

Images