A method and apparatus for direction estimamtion using artificial intelligence

The combination of curriculum learning and deep supervision methods addresses the limitations of fixed parameters in AI-based speech direction estimation, enhancing learning speed and accuracy in estimating voice direction.

KR103045292B1Active Publication Date: 2026-11-06INDUSTRY UNIVERSITY COOPERATION FOUNDATION HANYANG UNIVERSITY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
KR1020230054346
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2023-04-25
Publication Date
2026-11-06
Estimated Expiration
2043-04-25

AI Technical Summary

Technical Problem

Existing AI-based speech direction estimation systems are limited by fixed heuristic parameters during training, restricting their effectiveness, necessitating more systematic and efficient training methods.

Method used

A method combining curriculum learning and deep supervision to improve speech direction estimation, involving initial and termination parameters, and multiple stages of training with varying parameter values to enhance learning accuracy.

Benefits of technology

Enhances speech direction estimation capabilities in complex environments by improving learning speed and accuracy through a systematic training approach.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112023046644306-PAT00023_ABST
    Figure 112023046644306-PAT00023_ABST
Patent Text Reader

Abstract

A method performed by an electronic device using artificial intelligence according to an embodiment of the present invention may include: a step of determining a first initial parameter and a first termination parameter; a step of outputting a first data from a first artificial intelligence module that has been previously trained based on the first initial parameter and the first termination parameter, using voice data as input information; a step of determining a second initial parameter and a second termination parameter; a step of outputting a second data from a second artificial intelligence module that has been previously trained based on the second initial parameter and the second termination parameter, using the first data as input information; and a step of estimating the voice direction in the voice data based on the second data.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present disclosure describes a method and apparatus for estimating direction using artificial intelligence. Specifically, it describes a method for estimating the direction of a speaker by training artificial intelligence using speech direction estimation technology. Background Technology

[0003] The present disclosure relates to an apparatus and method for estimating the direction in which a voice is received for a voice signal. Recently, speech recognition technology, which collects voice signals and processes them into data, has emerged. Speech recognition technology refers to a technology that enables human speech to be processed into text format through a program. A speech recognition program can understand and process the speaker's grammar, syntax, structure, etc. To improve the functionality of speech recognition programs, an efficient method for receiving voice signals has been required, and methods such as specifying the direction of the voice and concentrating the reception of signals from that direction have been proposed.

[0004] While there are various types of speech direction estimation technologies, AI-based speech direction estimation systems are receiving particular attention. However, regarding AI-based speech direction estimation systems, there was a limitation in that the effectiveness was restricted because, although AI was previously trained using hard labeling and soft labeling methods, the parameters were fixed heuristically during training. Therefore, there is a need for methods that enable AI to be trained more systematically and efficiently through various approaches to deep learning models. The problem to be solved

[0006] The present disclosure aims to provide a method and apparatus for improving speech direction estimation technology using a curriculum learning method.

[0007] The present disclosure aims to provide a method for improving speech direction estimation capabilities by combining a curriculum learning method with a deep supervision method. means of solving the problem

[0009] A method performed by an electronic device using artificial intelligence according to an embodiment of the present invention may include: a step of determining a first initial parameter and a first termination parameter; a step of outputting a first data from a first artificial intelligence module that has been previously trained based on the first initial parameter and the first termination parameter, using voice data as input information; a step of determining a second initial parameter and a second termination parameter; a step of outputting a second data from a second artificial intelligence module that has been previously trained based on the second initial parameter and the second termination parameter, using the first data as input information; and a step of estimating the voice direction in the voice data based on the second data.

[0010] In one embodiment, the parameter may be a parameter representing a predetermined range centered on the angle of the voice direction.

[0011] In one embodiment, the first termination parameter and the second termination parameter may have the same value.

[0012] In one embodiment, the second initial parameter may have a value smaller than the first initial parameter.

[0013] In one embodiment, the first artificial intelligence module may be trained based on: a step of determining a parameter value as the first initial parameter value; a step of performing a calculation on the voice data based on the first initial parameter; a step of determining a first reduction value; and a step of repeatedly performing a calculation on the voice data while reducing the parameter value by the first reduction value until the parameter value becomes the first termination parameter value.

[0014] In one embodiment, the previously learned first artificial intelligence module may further include a step of calculating a first loss by comparing the first data with a specific data value corresponding to the first initial parameter value.

[0015] In one embodiment, the previously trained second artificial intelligence module may further include a step of calculating a second loss by comparing a specific data value corresponding to the second initial parameter value with the third data output using the first loss and the first data as input information.

[0016] An electronic device according to an embodiment of the present invention comprises: a memory; a modem; and a processor connected to the modem and the memory, wherein the processor is configured to: determine a first initial parameter and a first termination parameter, output a first data from a first artificial intelligence module that has been trained based on the first initial parameter and the first termination parameter using voice data as input information, determine a second initial parameter and a second termination parameter, output a second data from a second artificial intelligence module that has been trained based on the second initial parameter and the second termination parameter using the first data as input information, and estimate the voice direction from the voice data based on the second data.

[0017] A program stored in a medium for performing a direction estimation method through an artificial intelligence algorithm executable by a processor according to an embodiment of the present invention may perform the steps of: determining a first initial parameter and a first termination parameter; outputting first data from a first artificial intelligence module that has been previously trained based on the first initial parameter and the first termination parameter, using voice data as input information; determining a second initial parameter and a second termination parameter; outputting second data from a second artificial intelligence module that has been previously trained based on the second initial parameter and the second termination parameter, using the first data as input information; and estimating a voice direction from the voice data based on the second data. Effects of the invention

[0019] According to one embodiment of the present disclosure, a learning method combining a curriculum learning method and a deep supervision method can improve speech direction estimation capabilities in complex environments. Brief explanation of the drawing

[0021] A brief description of each drawing is provided to help to more fully understand the drawings cited in the detailed description of the present invention. FIG. 1 is a conceptual diagram illustrating the basic principles of an artificial intelligence structure according to one embodiment of the present disclosure. FIG. 2 is a diagram showing result data according to artificial intelligence learned by applying one of the artificial intelligence learning methods for direction estimation according to one embodiment of the present disclosure. FIG. 3 is a diagram showing various result data according to artificial intelligence learned by applying a curriculum learning method according to an embodiment of the present disclosure. FIG. 4 is a diagram showing an artificial intelligence model learned through an artificial intelligence learning method according to one embodiment of the present disclosure. FIG. 5 shows the voice direction estimation result values ​​for an artificial intelligence model and a conventional model according to one embodiment of the present disclosure. FIG. 6 is a block diagram of an electronic device for a voice enhancement system according to one embodiment of the present disclosure. FIG. 7 is a flowchart for explaining a voice direction estimation method according to one embodiment of the present disclosure. Specific details for implementing the invention

[0022] The technical concept of the present invention is subject to various modifications and may have various embodiments. Specific embodiments are illustrated in the drawings and described in detail through the detailed description. However, this is not intended to limit the technical concept of the present invention to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the scope of the technical concept of the present invention.

[0023] In describing the technical concept of the present invention, detailed descriptions of related prior art are omitted if it is determined that such descriptions may unnecessarily obscure the essence of the invention. Furthermore, numbers used in the description of this specification (e.g., First, Second, etc.) are merely identification symbols to distinguish one component from another.

[0024] In addition, when a component is described in this specification as being "connected" or "connected" to another component, it should be understood that the component may be directly connected to or directly connected to the other component, but unless otherwise specifically stated, it may also be connected or connected through another component in between.

[0025] In addition, terms such as “~part,” “~device,” “~device,” and “~module” described in this specification refer to a unit that processes at least one function or operation, and may be implemented as hardware or software or a combination of hardware and software such as a processor, microprocessor, microcontroller, CPU (Central Processing Unit), GPU (Graphics Processing Unit), APU (Accelerate Processor Unit), DSP (Drive Signal Processor), ASIC (Application Specific Integrated Circuit), and FPGA (Field Programmable Gate Array), and may also be implemented in a form combined with memory that stores data necessary for processing at least one function or operation.

[0026] Furthermore, it is intended to clarify that the classification of components in this specification is merely based on the primary function each component is responsible for. That is, two or more components described below may be combined into a single component, or a single component may be divided into two or more components based on more subdivided functions. Additionally, each component described below may additionally perform some or all of the functions of other components in addition to the primary function it is responsible for, and it is obvious that some of the primary functions of each component may be exclusively performed by other components.

[0027] In describing the embodiments of the present disclosure, specific descriptions of related functions or configurations are omitted if it is determined that such detailed descriptions would unnecessarily obscure the essence of the present disclosure. Furthermore, terms used below are defined in consideration of their functions within the present disclosure, and these definitions may vary depending on the intent or practices of the user or operator. Therefore, such definitions should be based on the content throughout this specification.

[0028] For the same reason, some components in the attached drawings may be exaggerated, omitted, or schematically depicted. Additionally, the size of each component does not entirely reflect its actual size. Identical or corresponding components in each drawing have been assigned the same reference number.

[0029] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. The embodiments are provided merely to make the description of the present disclosure complete and to fully inform those skilled in the art of the scope of the invention, and the scope of the claims of the present disclosure is defined only by the scope of the claims.

[0030] At this point, it will be understood that each block of the drawings showing the process flow diagram and combinations of the process flow diagrams can be executed by computer program instructions. Since these computer program instructions can be loaded into the processor of a general-purpose computer, a specialized computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create means to perform the functions described in the flow diagram block(s). Since these computer program instructions can also be stored in computer-available or computer-readable memory that can be directed toward the computer or other programmable data processing equipment to implement the function in a specific way, the instructions stored in computer-available or computer-readable memory can also produce a manufactured item containing the means of instruction to perform the function described in the flow diagram block(s). Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0031] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). It should also be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order. For instance, two blocks described in succession may actually be executed substantially simultaneously, or the blocks may be executed in reverse order according to their corresponding functions.

[0032] As used in this disclosure, the term “unit or part” refers to a software or hardware component, such as a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC), and the “part” may be configured to perform specific roles. However, the “part” is not limited to software or hardware. The “part” may be configured to reside in an addressable storage medium or to execute one or more processors. Thus, by example, the “part” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and “parts” may be combined into a smaller number of components and “parts” or further separated into additional components and “parts.” In addition, the components and 'parts' may be implemented to utilize one or more CPUs within the device or secure multimedia card. Also, in the embodiments, 'parts' may include one or more processors and / or devices.

[0034] Hereinafter, embodiments according to the technical concept of the present invention will be described in detail in turn.

[0036] FIG. 1 is a conceptual diagram illustrating the basic principles of an artificial intelligence structure according to one embodiment of the present disclosure.

[0037] Referring to Figure 1, the basic principle of how learning is performed in an artificial intelligence structure is illustrated.

[0038] Artificial intelligence (AI) technology refers to techniques designed to solve cognitive problems primarily associated with human intelligence, such as learning, problem-solving, and perception. AI can be trained through machine learning (ML) and deep learning (DL). Machine learning is primarily used in techniques for pattern recognition and learning, representing algorithms that learn from recorded data to predict future data. It represents a technology that learns autonomously from data rather than relying on predefined rules or patterns. On the other hand, deep learning is a subfield of machine learning that differs in that it processes data based on Artificial Neural Networks (ANN). Because deep learning utilizes artificial neural networks, it can handle more complex and sophisticated computations than machine learning. Types of algorithms for deep learning include Convolutional Neural Networks (CNN), Deep Neural Networks (DNN), Artificial Neural Networks (ANN), and Recurrent Neural Networks (RNN).

[0039] Referring to FIG. 1, the artificial intelligence structure can be represented as an artificial intelligence module (110). The artificial intelligence module (110) receives predetermined input data (105), performs learning through a method predetermined in the module, and outputs output data (115) for the learning result. According to one embodiment, the input data (105) may include predetermined data, a voice signal, and an input sequence. The output data (115) may include an output sequence, an enhanced voice signal, etc.

[0041] FIG. 2 is a diagram showing result data according to artificial intelligence learned by applying one of the artificial intelligence learning methods for direction estimation according to one embodiment of the present disclosure.

[0042] The artificial intelligence algorithm used in Fig. 2 may be one of the types of artificial intelligence modules (110) of Fig. 1.

[0043] In the present disclosure, the term "speech direction estimation technology" may refer to a technology that estimates the direction in which a voice is received when voice is present in each frame. Speech direction estimation technology is generally used in conjunction with beamforming technology. Beamforming technology separates sound signals coming from a target (or a voice with a defined direction) voice direction, and speech direction estimation technology may be essential to perform this.

[0044] One method of utilizing deep learning in speech direction estimation technology is the hard label method, which trains an artificial intelligence algorithm by labeling a fixed target direction with an output value of 1 and other directions with an output value of 0. However, recently, it has been shown that training an artificial intelligence algorithm using the soft label method, which assigns an output value of 1 to a fixed target direction and assigns a non-zero output value to directions near the target direction, has a significant learning effect.

[0045] Referring to FIG. 2, the x-axis represents the azimuth direction, and the y-axis represents the output value. When using the hard label method, the output value is assigned only to the target directions, Source 1 (205) and Source 2 (210), and training proceeds. On the other hand, in the soft label method, in addition to the target directions, Source 1 (205) and Source 2 (210), surrounding value 1 (215) can be additionally assigned as a surrounding value of Source 1 (205) and surrounding value 2 (220) as a surrounding value of Source 2 (210). In this case, an example is shown in which one intermediate value is assigned per target, and as shown in FIG. 2, multiple surrounding values ​​can be assigned.

[0046] The following mathematical formulas can be used to assign soft labels.

[0047] [Mathematical Formula 1]

[0048]

[0049] [Mathematical Formula 2]

[0050]

[0051] [Mathematical Formula 3]

[0052]

[0053] Here, mathematical formula 1 represents a formula for creating targets by assigning 1 to the correct target direction and positive numbers smaller than 1 to the surrounding areas. As the distance from the correct answer increases, smaller values ​​closer to 0 are assigned, allowing recognition that the direction is moving away from the correct answer. Here, l represents the frame index, indicating which frame it is from the start. It represents a set formed by the direction in which the speaker is located.

[0055] FIG. 3 is a diagram showing various result data according to artificial intelligence learned by applying a curriculum learning method according to an embodiment of the present disclosure.

[0056] Figure 3 may show the result of an artificial intelligence algorithm learned using the soft label method among the curriculum learning methods described in Figures 1 and 2.

[0057] Curriculum learning is a method of training artificial intelligence through deep learning that improves the model's learning speed and performance by providing training data in a progressively increasing order of difficulty. In other words, it refers to a learning method in which training is conducted by setting relatively easy-to-understand datasets and targets in the early stages, and then gradually increasing the difficulty level to set relatively difficult datasets and targets as training progresses, thereby enabling the algorithm to be applied to complex tasks.

[0058] Referring to Fig. 3, conventionally, in the Soft label method, the attenuated level (in relation to Equations 1 to 3) Learning was conducted with ) and width (σ) selected and fixed based on experience, but in the following method Learning is performed by changing σ while keeping it fixed. Here, the reduction level is a parameter representing the degree of reduction of power, and the width may be a parameter representing a certain width or distance from the target direction. In order to learn to estimate the target direction (305), the value of σ can be set large in the early stages of learning. For example, when σ is set to 30 and learning is performed, a first result value (310) appears. In the first result value (310), a value close to 1 appears even in the periphery of the target direction (305), making learning easier. Next, when σ is set to 15 and learning is performed, a second result value (320) appears. In the second result value (320), a smaller value is assigned to the periphery of the target direction (305) than when σ is 30, making learning more difficult but allowing the target direction (305) to be estimated more clearly. Next, when training is performed by setting σ to 5, a third result value (330) appears. In the third result value (330), a smaller value is assigned to the periphery of the target direction (305) than when σ is 15, making training more difficult, but training is performed so that the target direction (305) can be estimated more clearly. As the σ value gradually decreases, the range of values ​​in the periphery of the target direction decreases, and there is an advantage that it becomes easier to recognize the target direction.

[0059] However, while assigning large values ​​to the periphery of the target direction in the curriculum learning method has the advantage of making deep learning easier, it has the disadvantage that the difference from the correct answer value of 1 becomes small, making it unclear regarding the accurate estimation of the target direction. On the other hand, assigning small values ​​to the periphery of the target direction makes learning more difficult, but it can be trained to estimate the accurate target direction. To utilize the advantages of both methods, we propose a method that combines the curriculum learning method and the deep supervision method, which are artificial intelligence algorithm learning methods.

[0060] FIG. 4 is a diagram showing an artificial intelligence model learned through an artificial intelligence learning method according to one embodiment of the present disclosure.

[0061] Figure 4 can be configured in a form that combines the curriculum learning method and the deep supervision method. Deep supervision refers to a method of providing additional supervised information (e.g., intermediate output) to intermediate layers, rather than a method of training that calculates the target and loss using only the output of the end layer, as is done with deep learning models. In a model using the deep supervision method, intermediate layers may have an auxiliary loss function. The auxiliary loss function can predict the error at each layer based on the output of the intermediate layers. The predicted error can be combined with the final loss function to facilitate learning.

[0062] Referring to FIG. 4, a short time Fourier transform (STFT) is performed on multi-channel audio (405), and then instantaneous relative transfer function (iRTF) features are extracted. The extracted iRTF features are passed through a shared module (410) composed of a convolutional neural network (CNN), batch normalization (BN), and an exponential linear unit (ELU). The features passed through the shared module (410) are transmitted to n skip-connection modules (SCM) (415a, 415b, 415c), undergo training, and then are transmitted to output layers (420a, 420b, 420c), respectively. Here, SCM(415a, 415b, 415c) may have a structure of CNN+BN+ELU, and output layer(420a, 420b, 420c) may consist of CNN+Sigmoid operation steps.

[0063] Each Cross-entropy loss (CE Loss) (425a, 425b, 425c) can be measured using labeling information (430a, 430b, 430c) (or map information, data values ​​corresponding to specific σ values) with different σ values ​​and output values ​​from each output layer (420a, 420b, 420c). CE Loss can represent the difference between the model's predicted probability distribution and the actual label probability distribution.

[0064] For example, in the first SCM (415a), the output result can be given as map information labeling information (430a) having the largest σ value of 16 to measure the loss (425a). In the next step, the second SCM (415b), the labeling information (430b) having a σ value of 6 can be given as map information to measure the loss (425b). In the next step, the third SCM (415c), the labeling information (430c) having a σ value of 2 can be given as map information to measure the loss (425c). This may be a method that induces a clearer output as the layer becomes deeper by assigning progressively smaller values ​​to the labeling information around the target.

[0065] In addition to assigning labeling information as intermediate information, curriculum learning can be performed by determining the initial σ value and the final σ value in one module so that the σ value gradually decreases. For example, a curriculum learning method can be performed together in which 16 is initially assigned as the labeling information (430a) in the first SCM, but 2 is finally assigned to learn to decrease at a constant rate.

[0066] In addition, the algorithm module of the first layer is trained through curriculum learning via scheduling by assigning a σ value such that the initial value is maximized until a predetermined final σ value is reached. In the module of the next layer, an initial σ value smaller than that of the first module is assigned, so that as the layers increase, the initial σ value gradually decreases, while the final σ value is set to be the same to train the artificial intelligence algorithm. Here, the scheduling may include a method of setting the σ value to decrease from the initial σ value to the final σ value by maintaining the rate of decrease of σ constant.

[0067] In relation to the deep supervision method, the result output from the first SCM (415a) is output to the output layer (420a), but is also output to the next layer, the second SCM (415b). In the second SCM, learning is performed by reflecting the result from the first SCM (415a). Therefore, the learning result of the second SCM (415b) derives a result that additionally reflects the learning result from the first SCM (415a). Even if the number of layers increases, the result can be derived by similarly reflecting the learning result up to the previous layer.

[0069] FIG. 5 shows the voice direction estimation result values ​​for an artificial intelligence model and a conventional model according to one embodiment of the present disclosure.

[0070] The dataset used in FIG. 5 is a public dataset. The table in FIG. 5 shows the results obtained by inputting the public dataset into the following algorithm modules: CRNNk (505), an existing algorithm module without any special methods added; CL (510), an algorithm module with a curriculum learning method added; DS (515, 520), an algorithm module with a deep supervision method added; and DSCL (Deeply supervised curriculum learning) module (525), an algorithm module that combines the curriculum learning method and the deep supervision method.

[0071] The measurement results presented were mean absolute error (MSE) and accuracy (ACC). Here, MSE represents the angular difference between the correct answer and the estimate, and a smaller value indicates better performance. ACC indicates the accuracy rate, treating a difference of 10 degrees or less between the correct answer and the estimate as the correct answer. A higher value indicates better results. Additionally, the Synthetic results are evaluated using synthesized data, while LOCATA results are evaluated using data recorded in a real-world environment. k represents the SCM index.

[0072] That is, referring to Fig. 5, CRNNk is the result when trained using only the output of the k-th SCM without using deep supervision and curriculum learning. CLk is the result when curriculum learning is applied using only the output of the k-th SCM. DSd is the result when trained with a target formed by the same σ for all SCMs. DSe is the result when trained with a target formed by a different σ for each SCM. DSCL is the result when deep supervision and curriculum learning are used simultaneously, and for each SCM, a different ' ', ' ' was used. Here, k can represent the index of a layer or module.

[0073] Looking at the results, the curriculum learning method shows high performance when using real-world data, while the deep supervision method shows high performance in environments with spatially coherent noise. For modules using both the curriculum learning and deep supervision methods, it can be confirmed that performance improved in both environments.

[0075] FIG. 6 is a block diagram of an electronic device for a voice enhancement system according to one embodiment of the present invention.

[0076] Referring to FIG. 6, the electronic device (610) may include a modem (MODEM, 620), a memory (MEMORY, 640), and a processor (PROCESSOR, 630).

[0077] The modem (620) may be a communication modem that is electrically connected to other electronic devices to enable mutual communication. In particular, the modem (620) may receive data input and transmit it to the processor (630), and the processor (630) may store the input data value in memory (640). Additionally, the data value output by the artificial intelligence algorithm learned in the system may be transmitted to another electronic device.

[0078] Memory (640) is a configuration in which various information and program instructions for the operation of the electronic device (610) are stored, and may be a storage device such as a hard disk or a solid state drive (SSD). In particular, memory (640) may store one or more data input values ​​input from the modem (620) under the control of the processor (630). Additionally, memory (640) may store program instructions such as an artificial intelligence algorithm for direction estimation that can be executed by the processor (630). Additionally, memory (640) may store programs (or program instructions) such as artificial intelligence algorithms learned through the curriculum learning method, deep supervision method, and deeply supervised curriculum learning described in this disclosure.

[0079] The processor (630) is composed of at least one processor and can learn a direction estimation artificial intelligence algorithm using data and program instructions stored in memory (640) and use it to calculate data. The processor (630) can control and calculate all artificial intelligence algorithms described in FIGS. 1 to 5. The processor (630) can then perform operations for the method described in FIG. 7.

[0081] FIG. 7 is a flowchart illustrating a voice direction estimation method according to an embodiment of the present invention.

[0082] With reference to FIG. 7, the learning operation of the artificial intelligence algorithm of the electronic device and the voice direction estimation method described with reference to FIGs. 1 to 6 will be summarized and explained below. Each operation is not an operation that must be necessarily included in the series of processes, and only some of them may be configured and operated depending on the situation.

[0083] In step S710, a first initial parameter and a first termination parameter can be determined to estimate the voice direction through artificial intelligence. Here, the parameter is the width described in FIGS. 2 to 4. It may represent ). Also, the initial parameter is of FIG. 5 , the termination parameter is of FIG. 5 It can represent.

[0084] In step S720, the first data can be output from the first artificial intelligence module, which has been trained based on the first initial parameter and the first termination parameter, using voice data as input information. The voice data may represent the voice data and multi-channel audio of FIGS. 1 to 4 of the present disclosure. Here, the first artificial intelligence module may be trained through a curriculum learning method based on the initial parameter and the termination parameter. Here, the first data may be the data output to the second SCM in FIG. 4.

[0085] In step S730, a second initial parameter and a second termination parameter can be determined. Here, the parameter is the width described in FIGS. 2 to 4. It may represent ). Also, the initial parameter is of FIG. 5 , the termination parameter is of FIG. 5 It can represent. Here, the second initial parameter and the second termination parameter are respectively those when k is 2 in FIG. 5. , and It may represent.

[0086] In step S740, the second data can be output from the second artificial intelligence module, which has been previously trained based on the second initial parameter and the second termination parameter, using the first data as input information. Here, the first data may be the data output from the first SCM to the second SCM in FIG. 4. Here, the second artificial intelligence module may be trained through a curriculum learning method based on the initial parameter and the termination parameter.

[0087] In step S750, the voice direction can be estimated from the voice data based on the second data. The voice data may include the speaker's voice, noise, echo signals, etc. The voice direction may be determined at a certain angle.

[0089] Although the technical concept of the present disclosure has been described in detail with reference to various embodiments, the technical concept of the present disclosure is not limited to the above embodiments, and various modifications and changes can be made by those skilled in the art within the scope of the technical concept of the present disclosure. Explanation of the symbols

[0091] 110: Artificial Intelligence Module 610: Electronic device 620 : Modem 630: Processor 640: Memory

Claims

Claim 1 A method performed by an electronic device using artificial intelligence, comprising: a step of determining a first initial parameter and a first termination parameter; a step of outputting a first data from a first artificial intelligence module that has been trained based on the first initial parameter and the first termination parameter, using voice data as input information; a step of determining a second initial parameter and a second termination parameter; a step of outputting a second data from a second artificial intelligence module that has been trained based on the second initial parameter and the second termination parameter, using the first data as input information; and a step of estimating a voice direction in the voice data based on the second data, wherein the second initial parameter has a value smaller than the first initial parameter. Claim 2 A method according to claim 1, wherein the parameter is a parameter representing a predetermined range centered on the angle of the voice direction. Claim 3 A method according to claim 1, wherein the first termination parameter and the second termination parameter have the same value. Claim 4 delete Claim 5 A method of learning based on the first artificial intelligence module, wherein the first artificial intelligence module comprises: a step of determining a parameter value as the value of the first initial parameter; a step of performing a calculation on the voice data based on the first initial parameter; a step of determining a first reduction value; and a step of repeatedly performing a calculation on the voice data while reducing the parameter value by the first reduction value until the parameter value becomes the value of the first termination parameter. Claim 6 A method according to claim 1, further comprising the step of calculating a first loss by comparing a specific data value corresponding to the value of the first data and the first initial parameter in the previously learned first artificial intelligence module. Claim 7 A method according to claim 6, further comprising the step of calculating a second loss by comparing a specific data value corresponding to the value of the second initial parameter with the third data output using the first loss and the first data as input information in the previously learned second artificial intelligence module. Claim 8 An electronic device comprising: a memory; a modem; and a processor connected to the modem and the memory, wherein the processor is configured to: determine a first initial parameter and a first termination parameter, output a first data from a first artificial intelligence module that has been trained based on the first initial parameter and the first termination parameter using voice data as input information, determine a second initial parameter and a second termination parameter, output a second data from a second artificial intelligence module that has been trained based on the second initial parameter and the second termination parameter using the first data as input information, and estimate a voice direction from the voice data based on the second data, wherein the second initial parameter has a value smaller than the first initial parameter. Claim 9 In paragraph 8, the processor is: an electronic device in which the parameter is a parameter representing a predetermined range centered on the angle of the voice direction. Claim 10 In claim 8, the electronic device having the same value for the first termination parameter and the second termination parameter. Claim 11 delete Claim 12 An electronic device that is learned based on the steps of: determining a parameter value as the first initial parameter value; performing calculations on voice data based on the first initial parameter; determining a first reduction value; and repeatedly performing calculations on the voice data while reducing the parameter value by the first reduction value until the parameter value becomes the first termination parameter value. Claim 13 In claim 8, the processor is further configured to calculate a first loss by comparing the first data with a specific data value corresponding to the value of the first initial parameter in the first artificial intelligence module that has been learned. Claim 14 In paragraph 13, the processor is further configured to calculate a second loss by comparing a specific data value corresponding to the value of the second initial parameter with the third data output using the first loss and the first data as input information in the previously learned second artificial intelligence module. Claim 15 A program stored in a medium for performing a direction estimation method through an artificial intelligence algorithm executable by a processor, comprising: a step of determining a first initial parameter and a first termination parameter; a step of outputting first data from a first artificial intelligence module that has been trained based on the first initial parameter and the first termination parameter, using voice data as input information; a step of determining a second initial parameter and a second termination parameter; a step of outputting second data from a second artificial intelligence module that has been trained based on the second initial parameter and the second termination parameter, using the first data as input information; and a step of estimating a voice direction from the voice data based on the second data, wherein the second initial parameter has a value smaller than the first initial parameter.

Citation Information

Patent Citations

  • A method for inferring of generating direction of sound using deep network and an apparatus for the same

    KR102329353B1