Method, device, and system providing artificial intelligence solution for detecting deep voice generated by generative artificial intelligence and preventing related incidents

Abstract

According to one embodiment of the present invention, provided is a system providing an artificial intelligence solution for detecting a deep voice generated by generative artificial intelligence and preventing incidents, the system comprising: a memory for storing instructions; and one or more processors for executing the instructions, wherein the processors are configured to: extract a feature vector including frequency characteristics and temporal change patterns from inputted voice data; input the extracted feature vector into an ensemble artificial intelligence model including a first AI model which is an acoustic model, a second AI model which is a language model, and a third AI model which is a pattern recognition model, so as to calculate respective individual deep voice probabilities; measure signal quality of the inputted voice data to calculate a signal-to-noise ratio (SNR); calculate a final deep voice probability through an adaptive weight allocation algorithm that varies, in real time, weights to be applied to each AI model on the basis of the calculated SNR value; and generate a deep voice suspicion alert if the final deep voice probability exceeds a threshold value.

Description

Method, device, and system for providing an AI solution that detects deep voices generated by generative AI and prevents related accidents

[0001] The present invention relates to a method, apparatus, and system for providing an artificial intelligence solution that detects a Deep Voice generated using Generative AI technology and prevents related accidents.

[0002] With the recent advancement of artificial intelligence technology, deep voice technology has emerged that mimics the human voice with great sophistication. While this deep voice technology can be utilized in various fields such as speech synthesis and entertainment, if used for malicious purposes, it can cause serious social problems such as voice phishing, impersonation, and defamation.

[0003] Existing deep voice detection technologies primarily use methods that analyze the characteristics of speech signals. For example, they extract features found in deep voice through Mel frequency cepstrum (MFCC) analysis and linear predictive coding (LPC) analysis, and determine whether a voice is deep voice based on these findings. However, these methods are showing limitations in effectively detecting deep voices that are becoming increasingly sophisticated due to the advancement of generative AI technology.

[0004] The present invention aims to provide a solution that effectively detects deep voices, which have become more sophisticated due to the advancement of generative AI technology, and prevents related accidents.

[0005] According to one embodiment of the present invention, a system for detecting deep voice generated by generative artificial intelligence is provided, comprising: a memory for storing instructions; and one or more processors for executing said instructions, wherein the processor extracts a feature vector including frequency characteristics and temporal change patterns from input voice data, inputs said extracted feature vectors to an ensemble artificial intelligence model including a first AI model which is an acoustic model, a second AI model which is a language model, and a third AI model which is a pattern recognition model to calculate each individual deep voice probability, measures the signal quality of said input voice data to calculate a signal-to-noise ratio (SNR), calculates a final deep voice probability through an adaptive weight allocation algorithm that varies the weights to be applied to said AI model in real time based on said calculated SNR value, and generates a deep voice suspicion alarm when said final deep voice probability exceeds a threshold value.

[0006] According to one embodiment of the present invention, the processor may be configured to perform dynamic resource switching to control the load of computing resources, and to perform a first analysis of the input voice data using the third AI model, wherein the amount of computation is less than a preset standard, and, only when a section in which the deep voice probability exceeds a preliminary threshold is detected as a result of the first analysis, to load the deep learning-based first AI model and the second AI model into the active area of ​​the memory to perform a precise analysis.

[0007] According to one embodiment of the present invention, the processor may be configured to perform liveness detection feedback by adding a penalty weight to the final deep voice probability and updating it regardless of the content of the response when the final deep voice probability exceeds the threshold, and when the final deep voice probability exceeds the threshold, the processor presents an additional question to verify the identity of the caller through a user terminal, receives the caller's response voice signal to the additional question, measures the latency from the time of the question to the time of the response and the jitter and shimmer fluctuation rates of the response voice waveform, and when the latency exceeds a preset human response range or the jitter fluctuation rate falls within a preset artificial synthesized voice pattern range.

[0008] According to one embodiment of the present invention, the adaptive weight assignment algorithm, when the SNR value is a low-quality signal below a threshold, reduces the weight of the noise-sensitive first AI model (acoustics) and increases the weight of the context-dependent second AI model (language), and the final deep voice probability (P final ) is calculated based on the following formula, where w i is the weight of each model that varies according to SNR, p i can be the output probability of each model.

[0009]

[0010] According to one embodiment of the present invention, the first AI model is a Convolutional Neural Network (CNN)-based model that receives a Mel-Spectrogram as input, the second AI model is a Transformer-based language model that receives utterance content converted into text as input, and the third AI model may be a pattern recognition model that receives statistical values ​​of pitch and energy changes of speech as input.

[0011] According to another embodiment of the present invention, a method for deep voice detection performed by a computing device comprises: a step of extracting a feature vector including frequency characteristics and temporal change patterns from input voice data; a step of inputting the extracted feature vector into an ensemble model including a first AI model which is an acoustic model, a second AI model which is a language model, and a third AI model which is a pattern recognition model to calculate individual probabilities; a step of analyzing the signal quality of the input voice data to measure the signal-to-noise ratio (SNR); a step of performing adaptive weight assignment to dynamically determine weights to be applied to each of the plurality of AI models according to the measured SNR value to calculate a final deep voice probability; and a step of displaying a warning message to a user terminal or transmitting a report signal to an investigative agency server when the final deep voice probability exceeds a threshold value.

[0012] According to another embodiment of the present invention, a computer-readable recording medium is provided that records a program for executing the method on a computer.

[0013] Deep voices can be detected more accurately than existing methods by using generative AI models. Through training on various generative AI models, various types of deep voices can be effectively detected.

[0014] Damage can be prevented by immediately detecting deep voices occurring in real-time, such as during voice calls and online meetings. Additionally, crimes such as voice phishing and impersonation using deep voices can be prevented.

[0015] FIG. 1 is a diagram illustrating an artificial intelligence solution providing system that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0016] FIG. 2 is a diagram illustrating the learning of a neural network according to one embodiment.

[0017] FIG. 3 is a diagram illustrating the configuration of an artificial intelligence model according to one embodiment.

[0018] FIG. 4 is a block diagram showing the configuration of an artificial intelligence solution providing system that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0019] FIG. 5 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0020] FIG. 6 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0021] FIG. 7 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0022] FIG. 8 illustrates an adaptive weight allocation process according to one embodiment of the present invention.

[0023] Hereinafter, embodiments will be described in detail with reference to the attached drawings.

[0024] However, various modifications may be made to the embodiments, so the scope of the patent application is not limited or restricted by these embodiments. It should be understood that all modifications, equivalents, and substitutions to the embodiments are included within the scope of the rights.

[0025] Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, the embodiments are not limited to the specific disclosed forms, and the scope of this specification includes modifications, equivalents, or substitutions that fall within the technical concept.

[0026] Terms such as "first" or "second" may be used to describe various components, but these terms should be interpreted solely for the purpose of distinguishing one component from another. For example, the first component may be named the second component, and similarly, the second component may be named the first component.

[0027] When it is stated that a component is "connected" to another component, it should be understood that it may be directly connected to or joined to that other component, or that there may be other components in between.

[0028] The terms used in the embodiments are for illustrative purposes only and should not be interpreted as intended to be limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0029] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the embodiments pertain. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0030] In addition, when describing with reference to the attached drawings, identical components are assigned the same reference numeral regardless of drawing symbols, and redundant descriptions thereof are omitted. In describing the embodiments, if it is determined that a detailed description of related prior art could unnecessarily obscure the essence of the embodiments, such detailed description is omitted.

[0031] The embodiments can be implemented in various forms of products such as personal computers, laptop computers, tablet computers, smartphones, televisions, smart home appliances, intelligent vehicles, kiosks, and wearable devices.

[0032] Artificial Intelligence (AI) systems are computer systems that implement human-level intelligence. Unlike existing rule-based smart systems, they are systems in which machines learn and make decisions autonomously. As AI systems improve in recognition accuracy and gain a more accurate understanding of user preferences with continued use, existing rule-based smart systems are gradually being replaced by deep learning-based AI systems.

[0033] Artificial intelligence technology consists of machine learning and component technologies utilizing machine learning. Machine learning is an algorithmic technology that autonomously classifies and learns the characteristics of input data, while component technologies are technologies that mimic the cognitive and judgmental functions of the human brain by utilizing machine learning algorithms such as deep learning, and are comprised of technological fields such as linguistic understanding, visual understanding, reasoning / prediction, knowledge representation, and motion control.

[0034] The various fields where artificial intelligence technology is applied are as follows. Linguistic understanding refers to technologies that recognize, apply, and process human language and text, including natural language processing, machine translation, dialogue systems, question answering, and speech recognition / synthesis. Visual understanding refers to technologies that perceive and process objects like human vision, including object recognition, object tracking, image search, person recognition, scene understanding, spatial understanding, and image enhancement. Inference and prediction refers to technologies that logically reason and predict by judging information, including knowledge / probability-based inference, optimization prediction, preference-based planning, and recommendation. Knowledge representation refers to technologies that automatically process human experiential information into knowledge data, including knowledge construction (data generation / classification) and knowledge management (data utilization). Motion control refers to technologies that control the autonomous driving of vehicles and the movement of robots, including motion control (navigation, collision, driving) and manipulation control (behavior control).

[0035] Generally, to apply machine learning algorithms to real-world situations, training is performed using a trial-and-error method due to the inherent characteristics of machine learning methodologies. In particular, deep learning requires hundreds of thousands of iterations. Since it is impossible to execute this in a real physical external environment, training is instead performed through simulations that virtually recreate the actual physical environment on a computer.

[0036] In the present invention, Artificial Intelligence (AI) refers to a technology that imitates human learning ability, reasoning ability, perceptual ability, etc., and implements them on a computer, and may include concepts such as machine learning and symbolic logic. Machine Learning (ML) is an algorithmic technology that autonomously classifies or learns the characteristics of input data. The technology of artificial intelligence, as a machine learning algorithm, [uses] input data

[0037] It can analyze, learn from the results of that analysis, and make judgments or predictions based on the results of that learning. Furthermore, technologies that utilize machine learning algorithms to mimic functions of the human brain, such as cognition and judgment, can also be understood within the category of artificial intelligence. For example, this may include technological fields such as linguistic understanding, visual understanding, reasoning / prediction, knowledge representation, and motion control.

[0038] An artificial intelligence learning model or neural network model can be designed to implement the structure of the human brain on a computer and may include multiple network nodes that have weights and simulate neurons of a human neural network. The multiple network nodes may have interconnected relationships by simulating the synaptic activity of neurons, where neurons exchange signals through synapses. In an artificial intelligence learning model, multiple network nodes may be located in layers of different depths and exchange data according to convolutional connections. The artificial intelligence learning model may be, for example, an Artificial Neural Network (ANN) or a Convolutional Neural Network (CNN). As an embodiment, the artificial intelligence learning model may be machine learned according to methods such as supervised learning, unsupervised learning, and reinforcement learning. Machine learning algorithms for performing machine learning may include Decision Tree, Bayesian Network, Support Vector Machine, Artificial Neural Network, Ada-boost, Perceptron, Genetic Programming, and Clustering.

[0039] Among these, CNNs are a type of multilayer perceptron designed to use minimal preprocessing. CNNs consist of one or more convolutional layers and standard artificial neural network layers stacked on top, additionally utilizing weights and pooling layers. Thanks to this structure, CNNs can fully utilize two-dimensional input data. Compared to other deep learning architectures, CNNs demonstrate good performance in both image and audio fields. CNNs can also be trained using standard backpropagation. CNNs have the advantage of being easier to train than other feedforward artificial neural network techniques and using a small number of parameters.

[0040] Convolutional networks are neural networks comprising sets of nodes with bounded parameters. Many computer vision tasks have been significantly improved, driven by the increased size of available training data and the availability of computational power, combined with algorithmic advancements such as discriminative linear units and dropout training. In the case of massive datasets, such as those available for many tasks today, outfitting is not critical, and increasing the network size improves test accuracy. Optimal utilization of computing resources becomes a limiting factor. To address this, distributed, scalable implementations of deep neural networks can be employed.

[0041] FIG. 1 is a diagram illustrating an artificial intelligence solution providing system that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0042] As illustrated in FIG. 1, an artificial intelligence solution providing system (100) for detecting deep voices generated by generative artificial intelligence and preventing related accidents may include a plurality of user terminals (110-1, ... 110-n), a server (120), and a database (130). According to one embodiment, the database (130) is shown as being configured separately from the server (120), but is not limited thereto, and the database (130) may be provided within the server (120). For example, the server (120) may include a plurality of artificial intelligences for performing machine learning algorithms. According to one embodiment, the plurality of user terminals (110-1, ... 110-n), the server (120), and the database (130) may be connected to communicate with each other through a network (N).

[0043] A network (N) can perform wireless or wired communication between multiple user terminals (110-1, ... 110-n), a server (120), a database (130), etc. For example, the network can perform wireless communication according to methods such as LTE (long-term evolution), LTE-A (LTE Advanced), CDMA (code division multiple access), WCDMA (wideband CDMA), WiBro (Wireless BroadBand), WiFi (wireless fidelity), Bluetooth, NFC (near field communication), GPS (Global Positioning System), or GNSS (global navigation satellite system). For example, the network (N) can perform wired communication according to methods such as USB (universal serial bus), HDMI (high definition multimedia interface), RS 232 (recommended standard 232), or POTS (plain old telephone service).

[0044] The database (130) can store various data. Data stored in the database (130) may include software (e.g., programs) as data acquired, processed, or used by at least one component of a plurality of user terminals (110-1, ... 110-n) and a server (120). The database (130) may include volatile and / or non-volatile memory.

[0045] FIG. 2 is a diagram illustrating the learning of a neural network according to one embodiment.

[0046] As illustrated in FIG. 2, the learning device can train a neural network (123) to process review responses received from a plurality of user terminals (110-1, ... 110-n) by item. Additionally, the learning device can train a neural network (123) to extract user stay history from user movement path information. According to one embodiment, the learning device may be a separate entity from the server (120), but is not limited thereto.

[0047] The neural network (123) includes an input layer (121) into which training samples are input and an output layer (125) that outputs training outputs, and can be trained based on the difference between the training outputs and the labels. Here, the labels are defined based on items corresponding to review responses and can be defined based on user dwell history corresponding to movement path information. The neural network (123) is connected as a group of multiple nodes and is defined by weights between the connected nodes and an activation function that activates the nodes.

[0048] The learning device can train the neural network (123) using the GD (Gradient Descent) technique or the SGD (Stochastic Gradient Descent) technique. The learning device can use a loss function designed by the outputs and labels of the neural network.

[0049] The learning device can calculate the training error using a predefined loss function. The loss function can be predefined with labels, outputs, and parameters as input variables, where the parameters can be set by weights within the neural network (123). For example, the loss function can be designed in the form of Mean Square Error (MSE), entropy, etc., and various techniques or methods may be employed in the embodiments in which the loss function is designed.

[0050] The learning device can find weights that affect the training error using the backpropagation technique. Here, the weights are relationships between nodes within the neural network (123). The learning device can use the SGD technique with labels and outputs to optimize the weights found through the backpropagation technique. For example, the learning device can update the weights of the loss function defined based on the labels, outputs, and weights using the SGD technique.

[0051] According to one embodiment, the learning device extracts first objects from review responses, obtains first labels which are items corresponding to the first objects, applies the first objects to a first neural network to generate first training outputs corresponding to the first objects, and can train the first neural network based on the first training outputs and the first labels.

[0052] The learning device extracts second objects from movement path information, obtains second labels which are user dwell history corresponding to the second objects, applies the second objects to a second neural network to generate second training outputs corresponding to the second objects, and can train the second neural network based on the second training outputs and second labels.

[0053] According to one embodiment, the learning device can generate first training feature vectors based on the constituent features, location features, and pattern features of the review response.

[0054] Various methods can be employed to extract features.

[0055] According to one embodiment, the learning device can generate second training feature vectors based on the constituent features, length features, and pattern features of the movement path information. Various methods may be employed for extracting features.

[0056] According to one embodiment, the learning device may obtain training outputs by applying first training feature vectors to the neural network (123). The learning device may train a review item extraction algorithm of the neural network (123) based on the training outputs and first labels. The learning device may train a review item extraction algorithm of the neural network (123) by calculating training errors corresponding to the training outputs and optimizing the connection relationships of nodes within the neural network (123) to minimize the training errors. The server (120) may extract items from review responses using the first neural network that has been trained.

[0057] According to one embodiment, the learning device can obtain training outputs by applying second training feature vectors to the neural network (123). The learning device can train the user stay history acquisition algorithm of the neural network (123) based on the training outputs and second labels. The learning device can train the user stay history acquisition algorithm of the neural network (123) by calculating training errors corresponding to the training outputs and optimizing the connection relationships of nodes within the neural network (123) to minimize the training errors. The server (120) can obtain user stay history from movement path information using the second neural network that has been trained.

[0058] FIG. 3 is a diagram illustrating the configuration of an artificial intelligence model according to one embodiment.

[0059] An artificial intelligence model according to one embodiment may include an input layer, a hidden layer, and an output layer.

[0060] The input layer is the layer associated with the input values ​​fed into the artificial intelligence model.

[0061] In the hidden layer, a feature map can be output by performing MAC (multiply-accumulate) and activation operations on the input values.

[0062] A MAC operation can be an operation that multiplies each input value by its corresponding weight and sums the multiplied values.

[0063] An activation operation may be an operation that inputs the result of a MAC operation into an activation function and outputs a result value. The activation function may be of various types. For example, the activation function may include a sigmoid function, a tangent function, a ReLU function, a Leaky ReLU function, a Max Out function, and / or an ELU function, but there are no restrictions on the types.

[0064] A hidden layer may consist of at least one layer. For example, if the hidden layer consists of a first hidden layer and a second hidden layer, the first hidden layer performs MAC operations and activation operations based on input values ​​of an input system to output a feature map, and the feature map, which is the result value of the first hidden layer, may become the input value of the second hidden layer. The second hidden layer may perform MAC operations and activation operations based on the feature map, which is the result value of the first hidden layer.

[0065] The output layer may be a layer associated with the result of an operation performed in the hidden layer.

[0066] In one embodiment, the learning model learns syllable (character) patterns that are frequently combined and used in a given corpus to automatically learn the boundaries of compound words and named entities, integrates object information from a first UI source and object information rendered in a browser to create a learning object information file, uses the learning object information file to create learning data for the learning of a deep learning network, receives data from various domains of a support system, standardizes the data from the various domains into an integrated format based on at least one standardization method corresponding to each of the various domains, learns and infers data from a specific domain, determines information to be transmitted for standardization from the data of the specific domain, and can perform post-processing on the data from the various domains. The first UI source includes an XML file, and the learning object information file includes an input JSON file for feature learning and an output JSON file that is the correct answer (label) data during learning, and the output JSON file includes a file containing HTML DOM Tree information implemented in compliance with web standards, and the various domains include at least one of a RAN (radio access network), a transport, or a core, and post-processing may include a correlation function.

[0067] FIG. 4 is a block diagram showing the configuration of an artificial intelligence solution providing system that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0068] A system (400) according to one embodiment may include a processor (420) and a memory (430), and some of the illustrated components may be omitted or substituted. A system (400) according to one embodiment may be a server or a terminal. According to one embodiment, the processor (420) is a component capable of performing operations or data processing regarding the control and / or communication of each component of the system (400), and may be composed of one or more processors. The memory (430) may store information related to the method described above or store a program in which the method described above is implemented. The memory (430) may be volatile memory or non-volatile memory. The memory (430) may store various file data, and the stored file data may be updated according to the operation of the processor (120).

[0069] According to one embodiment, the processor (420) can execute a program and control the device (400). The code of the program executed by the processor (120) can be stored in memory (430). Operations of the processor (420) can be performed by loading instructions stored in memory (430). The system (400) can be connected to an external device (e.g., a personal computer or a network) through an input / output device (not shown in the drawing) and exchange data.

[0070] A system (400) according to one embodiment includes a processor (420) and a memory (430). A system (400) according to one embodiment may be the server or terminal described above. The processor (420) may include at least one device described above through FIGS. 1 to 3 or may perform at least one method described above through FIGS. 1 to 3. The memory (430) may store information related to the method described above or store a program in which the method described above is implemented. The memory (430) may be volatile memory or non-volatile memory.

[0071] A system (400) according to one embodiment includes a processor (420) and a memory (430). A system (400) according to one embodiment may be the server or terminal described above. The processor (420) may include at least one device described above through FIGS. 1 to 3 or may perform at least one method described above through FIGS. 1 to 3. The memory (430) may store information related to the method described above or store a program in which the method described above is implemented. The memory (430) may be volatile memory or non-volatile memory.

[0072]

[0073] Final Deep Voice Probability (P final ) is calculated by applying weights (w1, w2, w3) to each output probability (p1, p2, p3) of the first AI model (acoustics), the second AI model (language), and the third AI model (pattern).

[0074]

[0075] At this time, each weight w i It is not a fixed value, but changes in real time based on the signal-to-noise ratio (SNR) of the input speech signal and the confidence score of each model. For example, if the background noise of the input speech is severe and the SNR is below a threshold (e.g., 10 dB), an adaptive weight allocation algorithm is applied in which the weight w1 of the first AI model (acoustics), which is vulnerable to noise, is reduced by 0.5 times, and the weight w2 of the second AI model (language), which is context-dependent, is increased by 1.5 times.

[0076]

[0077] FIG. 5 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0078] Although process steps, method steps, algorithms, etc. are described in a sequential order in the flowchart of FIG. 5, such processes, methods, and algorithms may be configured to operate in any suitable order. In other words, the steps of the processes, methods, and algorithms described in various embodiments of the present invention do not need to be performed in the order described in the present invention. Furthermore, even if some steps are described as being performed asynchronously, in other embodiments, such steps may be performed simultaneously. Also, the example of a process by the depiction in the drawings does not mean that the exampled process excludes other variations and modifications therefrom, does not mean that any of the exampled process or any of its steps is essential to one or more of the various embodiments of the present invention, and does not mean that the exampled process is desirable.

[0079] In operation S510, a system for detecting deep voices generated through generative artificial intelligence (AI) and preventing related crimes (e.g., system (400) of FIG. 4) can extract deep voice suspected features under the control of a processor (e.g., processor (420) of FIG. 4). The system (400) receives voice data (e.g., phone call) in real time and can extract deep voice suspected features by analyzing frequency characteristics, temporal change patterns, and fine noise from the input voice data.

[0080] For example, when phone call content is input into the system, the system collects this voice data in real time. By analyzing the frequency characteristics of the input voice data, if abnormal patterns are detected in specific frequency bands, this can be identified as a suspected deep voice feature. For instance, when a user says, "Hello, I am Kim Cheol-su," the system analyzes the frequency pattern of this voice to identify differences from a normal speaker's voice. Additionally, it analyzes temporal variation patterns to detect changes in pronunciation speed or intonation. If the user speaks faster than usual or uses an abnormal intonation, this is considered a suspected deep voice feature. Finally, it analyzes micro-noise to evaluate background noise or the degree of voice distortion. For instance, if mechanical noise or distorted voice is detected during the call, this serves as an additional suspicious factor.

[0081] It can be assumed that user A is on a call with voice phishing criminal B. In this case, criminal B may use a voice modulated to sound like a woman's voice using deep voice technology. The system (400) receives the content of the call between A and B in real time. The system (400) can detect that energy in the lower frequency band appears higher in B's voice than in a typical female voice. The system (400) detects that the spacing between words or changes in intonation in B's voice are unnatural or show a consistent pattern. The system (400) can detect specific noise patterns present in the deep learning model training data or subtle digital noise that does not generally appear in a human voice in B's voice.

[0082] In operation S520, the system (400) can determine whether it is a deep voice. The system (400) can determine whether it is a deep voice by analyzing the extracted features using a generative AI model.

[0083] The AI ​​model is specialized for deep voice detection and has high accuracy in determining whether something is deep voice by learning a large amount of deep voice voice data and actual human voice data. As a result of analyzing the input features, the AI ​​model discovers characteristic patterns that occur during the deep voice generation process in B's voice and identifies it as deep voice.

[0084] According to one embodiment, the system (400) determines whether it is a deep voice. In this step, a generative AI model is used to analyze the features extracted in the previous step. For example, the system uses a deep learning-based acoustic model to compare and analyze the frequency patterns and temporal change patterns of the voice. If the system reaches the conclusion that "this voice is likely generated by a generative AI," it is determined to be a deep voice. For example, if abnormal patterns appear repeatedly in a specific frequency band or if there is a lack of consistency in pronunciation, the system may recognize it as a deep voice. This determination process is performed in real time, and the system makes a final decision by comprehensively analyzing the output results of each model.

[0085] In operation S530, the system (400) can display a warning message on the user terminal and automatically report it. If the system (400) is identified as deep voice, it can display a warning message on the user terminal saying "Caution! A voice suspected of being deep voice has been detected." and control the system to automatically report information related to the voice suspected of being deep voice data to the Cyber ​​Safety Bureau of the National Police Agency.

[0086] According to one embodiment, the system (400) analyzes extracted features using a generative artificial intelligence model (AI), calculates a deep voice probability based on the analysis results, and may display a warning message if the deep voice probability exceeds a specified level and is determined to be a deep voice.

[0087] The system can prevent crime by displaying a warning message on the user's terminal and automatically reporting it if it is identified as Deep Voice. For example, the system displays a warning message on User A's smartphone screen stating, "Caution! Suspicious Deep Voice detected." This message informs the user that the voice has been identified as Deep Voice and helps the user immediately recognize the situation. Additionally, the system controls the automatic reporting of information related to the suspected Deep Voice data to the Cyber ​​Safety Bureau of the National Police Agency. The information reported includes the caller's phone number, call duration, probability of Deep Voice, and a portion of the call content. For example, the system proceeds with the report in the form of, "Suspicious Deep Voice occurring at 15:30 on October 17, 2024, from User A's phone number 010-1234-5678."

[0088] When the AI ​​model identifies B's voice as deep voice, the system (400) immediately displays a warning message on A's smartphone screen saying, "Caution! A voice suspected of being deep voice has been detected." At the same time, it automatically reports relevant information, such as B's deep voice voice data, the time the call started, and the content of the call, to the Cyber ​​Safety Bureau of the National Police Agency. Through this, A can be vigilant about voice phishing crimes and prevent damage, and the police can track down criminals based on the reported information and use it for investigation.

[0089] FIG. 6 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0090] Although process steps, method steps, algorithms, etc. are described in a sequential order in the flowchart of FIG. 6, such processes, methods, and algorithms may be configured to operate in any suitable order. In other words, the steps of the processes, methods, and algorithms described in various embodiments of the present invention do not need to be performed in the order described in the present invention. Furthermore, even if some steps are described as being performed asynchronously, in other embodiments, such steps may be performed simultaneously. Also, the example of a process by the depiction in the drawings does not mean that the exampled process excludes other variations and modifications thereof, nor does it mean that any of the exampled process or any of its steps is essential to one or more of the various embodiments of the present invention, nor does it mean that the exampled process is desirable.

[0091] In operation S610, the system (e.g., the system (400) of FIG. 4) can extract frequency characteristics from input voice data under the control of a processor (e.g., the processor (420) of FIG. 4). The system (400) can extract frequency characteristics including Mel frequency cepstrum coefficients (MFCC), linear predictive coding (LPC) coefficients, and spectrograms from the input voice data, and can extract temporal change patterns including pitch change, formant change, and energy change.

[0092] First, the system (400) divides the voice data into small time units (e.g., 20ms) and performs a Fourier transform on each unit to obtain a frequency spectrum. From this spectrum, Mel frequency cepstrum coefficients (MFCC) are extracted, which compress frequency information by mimicking the way the human auditory system perceives frequencies. MFCC is a widely used feature in speech recognition and has the advantage of being particularly robust against noise.

[0093] The system (400) then extracts linear prediction coding (LPC) coefficients. LPC is a technique that efficiently represents the spectral envelope of a speech signal, which predicts the current speech sample as a linear combination of past samples and analyzes the difference. Through this, the resonance characteristics of the vocal tract can be identified, and the characteristics of the speech can be effectively represented.

[0094] A spectrogram is a visual representation of frequency changes over time and is useful for analyzing the energy distribution of a speech signal. The processor (420) analyzes the spectrogram to track pitch changes, that is, changes in the height of the speech. It also analyzes formant changes, which are the resonance frequencies of the vocal tract and are important factors in determining the characteristics of vowels. Finally, it analyzes energy changes over time to identify changes in speech intensity.

[0095] In operation S620, the system (400) can calculate the final deep voice probability by weighted averaging the output results of each model. The system (400) can extract fine noise including background noise, breathing sounds and lip sounds, extract pronunciation patterns including phoneme duration, resonance features and articulation modes, extract speech pause times including intervals between words and intervals between sentences, extract contextual information including speech content, conversation topics and situational information, input the extracted features into an ensemble AI model including a deep learning-based acoustic model, a natural language processing-based language model, and a statistical-based pattern recognition model to determine whether it is deep voice, and calculate the final deep voice probability by weighted averaging the output results of each model.

[0096] First, the system (400) separates fine noises such as background noise, breathing sounds, and lip sounds from voice data. Since these noises can interfere with deep voice identification, they are removed or attenuated for accurate analysis.

[0097] The system (400) analyzes pronunciation patterns such as phoneme duration, resonance characteristics, and manner of articulation. Since deep voice may differ slightly from human pronunciation, such pronunciation

[0098] It detects differences in patterns and uses them to determine whether it is a deep voice. The system (400) analyzes voice pause times, that is, the intervals between words and the intervals between sentences. Since deep voice does not require breathing unlike humans, the pause time patterns may appear differently from humans.

[0099] The system (400) analyzes contextual information such as speech content, conversation topic, and situational information. Deep Voice may make speech that is out of context or may not properly understand the conversation topic. This contextual information is used to determine whether it is Deep Voice. The extracted features are input into an ensemble AI model to determine whether it is Deep Voice.

[0100] The ensemble AI model combines a deep learning-based acoustic model, a natural language processing-based language model, and a statistics-based pattern recognition model.

[0101] The acoustic model determines whether it is a deep voice by analyzing the acoustic characteristics of the speech data. The language model determines whether it is a deep voice by analyzing grammatical and semantic errors in the speech content. The pattern recognition model determines whether it is a deep voice by analyzing the statistical patterns of the speech data. Each model independently determines whether it is a deep voice, and the system (400) calculates the final deep voice probability by weighting the output results of each model. The weights are determined based on the performance of each model, and higher weights are assigned to models with better performance. Through this, the strengths of each model are combined and the weaknesses are compensated for, thereby increasing the accuracy of deep voice determination.

[0102] In operation S630, the system (400) generates a warning when the voice is identified as deep voice and presents additional questions to verify the identity of the caller. Specifically, the system (400) identifies the voice as deep voice if the calculated deep voice probability is greater than or equal to a preset threshold (e.g., user setting value or system default value). When the voice is identified as deep voice, the system immediately displays a visual warning message on the user terminal stating, "Caution! A voice suspected of being deep voice has been detected," and simultaneously generates a vibration or warning sound so that the user can immediately recognize the danger. Subsequently, the system (400) presents additional questions, such as "Does the caller know their date of birth?" or "What was the topic of conversation you recently had with the caller?" to verify if the caller is an actual acquaintance. If the caller is unable to provide an accurate answer (e.g., "It is January 1, 1990") or hesitates, the system may use this as additional grounds for identifying the voice as deep voice.

[0103] In operation S640, the system (400) can re-evaluate the deep voice probability by analyzing the collected user response. The system (400) collects the user's response in voice or text format, analyzes the collected user response to re-evaluate the deep voice probability, and can provide the re-evaluated deep voice probability to the user in the form of a numerical value or graph, such as "Deep voice probability: XX%".

[0104] The system analyzes the collected user responses to re-evaluate the Deep Voice probability. When a user responds in voice or text format, the system collects and analyzes this response. For example, if a user answers "I know my date of birth" to a follow-up question, the system processes this response to obtain the information necessary to verify the caller's identity.

[0105] During this process, the system evaluates the consistency of the user's response. If a user answers, "My recent conversation topic is travel," the system determines whether this response is reliable by comparing it to previous conversation records. Subsequently, the system re-evaluates the Deep Voice probability based on the collected user responses. For example, if the initial Deep Voice probability was 85%, it may drop to 50% due to the accurate information provided by the user. The re-evaluated Deep Voice probability is presented to the user in the form of a numerical value or a graph, such as "Deep Voice Probability: XX%." This allows the user to clearly understand the current situation.

[0106] The user may respond to additional questions presented by the system (400) in voice or text format. The system (400) collects the user's response and analyzes the content of the response using natural language processing technology. Based on the analysis results, the deep voice probability is re-evaluated. For example, if the user fails to answer the question accurately, or if the content of the answer is ambiguous or inconsistent, the deep voice probability is increased. Conversely, if the user answers the question accurately and clearly, the deep voice probability is decreased. The re-evaluated deep voice probability is provided to the user by displaying it numerically, such as "Deep voice probability: XX%", or by visually representing it in the form of a graph, such as a bar graph or a pie chart. Through this, the user can more intuitively determine whether it is a deep voice.

[0107]

[0108] In addition to simply checking whether the correct answer matches, the system analyzes the 'Liveness Response' from the received user's response voice signal and reflects it in the probability re-evaluation. Specifically, it measures the latency from the time the question is presented to the time the response begins in milliseconds (ms) and calculates the jitter and shimmer fluctuation rates of the response voice waveform. If the response latency exceeds the typical human response speed range (e.g., 200–500ms) or if the waveform exhibits a flat pattern characteristic of artificial synthesized voice (e.g., less than 0.5%), a feedback loop is performed to update the final probability upward by adding a penalty weight (e.g., +0.2) to the Deep Voice probability, even if the answer is correct.

[0109]

[0110] In operation S650, the system (400) may periodically update the deep voice detection model. The system (400) may, depending on the user's choice, immediately block a call suspected of being deep voice, record the call content and store it in the user's cloud storage, analyze the recording file to determine whether it is deep voice, and periodically update the deep voice detection model. The update may include retraining the model using new data that reflects the latest deep voice generation technology. The system (400) may automatically distribute the updated model to the user terminal and control the encryption of information related to the suspected deep voice voice data, including at least one of the caller number, call duration, deep voice probability, user response, and recording file, to automatically report to the Cyber ​​Safety Bureau of the National Police Agency.

[0111] The system periodically updates the deep voice detection model. If a user chooses to immediately block a call suspected of being deep voice in a specific situation, the system blocks the call. For example, when a user decides to end a call because they suspect the caller's voice is suspicious, the system executes this immediately.

[0112] Additionally, if the user configures the system to record and save call content to cloud storage, the system records the call content and saves it to the user's cloud. Subsequently, the saved recording file is used to determine whether it is a deep voice. For example, the system can analyze this recording file to recalculate the probability of a deep voice.

[0113] The process of periodically updating the deep voice detection model involves retraining the model using new data that reflects the latest deep voice generation technology. For example, the system collects data on the latest speech synthesis technology and updates the model based on this to improve performance. Once this update process is complete, the system automatically deploys the updated model to user terminals.

[0114] Finally, the system encrypts information related to suspected deep voice voice data, including at least one of the caller number, call duration, deep voice probability, user response, and recording file, and automatically reports it to the Cyber ​​Safety Bureau of the National Police Agency. For example, the system can make a report including information such as "Caller number: 010-1234-5678, Call duration: 15:30 on October 17, 2024, Deep voice probability: 85%".

[0115] Users can immediately block calls suspected of being deep voice or record the call content to check later. The recorded call content is securely stored in the user's cloud storage. The recording file can be used to determine whether it is deep voice and can also be used to improve the deep voice detection model in the future. The system (400) periodically updates the deep voice detection model. The update includes retraining the model using new data that reflects the latest deep voice generation technology. For example, when new deep voice generation technology emerges, deep voice voice data generated by that technology is collected to retrain the model. Additionally, the model can be improved by analyzing deep voice suspected voice data recorded by users. The updated model is automatically distributed to the user terminal. The system (400) collects information related to deep voice suspected voice data, including at least one of the caller number, call duration, deep voice probability, user response, and recording file. The collected information is encrypted to protect personal information and is automatically reported to the Cyber ​​Safety Bureau of the National Police Agency. This can contribute to the prevention and investigation of deep voice crimes.

[0116] FIG. 7 is a flowchart of a method for providing an artificial intelligence solution that detects deep voices generated by generative artificial intelligence according to one embodiment and prevents related accidents.

[0117] Although process steps, method steps, algorithms, etc. are described in a sequential order in the flowchart of FIG. 7, such processes, methods, and algorithms may be configured to operate in any suitable order. In other words, the steps of the processes, methods, and algorithms described in various embodiments of the present invention do not need to be performed in the order described in the present invention. Furthermore, even if some steps are described as being performed asynchronously, in other embodiments, such steps may be performed simultaneously. Also, the example of a process by the depiction in the drawings does not mean that the exampled process excludes other variations and modifications therefrom, does not mean that any of the exampled process or its steps is essential to one or more of the various embodiments of the present invention, and does not mean that the exampled process is desirable.

[0118] In operation S710, the system (e.g., the system (400) of FIG. 4) can collect an analog voice signal under the control of a processor (e.g., the processor (420) of FIG. 4). The system (400) can collect the analog voice signal by acquiring a sound source through a plurality of array microphones spatially separated and arranged in a distributed control area. Each array microphone may include a plurality of microphone units arranged to detect the direction of the sound.

[0119] The processor (420) collects analog voice signals by controlling multiple array microphones spatially separated and arranged in a distributed control area. That is, by recording sound using multiple microphones, richer and more accurate voice information is obtained than using a single microphone.

[0120] Each array microphone includes multiple microphone units arranged to detect the direction of sound. Through this microphone array, the direction from which the sound is coming can be identified, and the sound from that direction can be collected more clearly. For example, beamforming technology can be used to determine the location of the sound source by analyzing the time difference of signals collected from multiple microphone units. By utilizing this spatial information, the system (400) can effectively remove ambient noise and selectively collect only the desired voice signal.

[0121] In operation S720, the system (400) can divide the voice segment using a voice activity detection algorithm. The system (400) converts the collected analog voice signal into a digital voice signal in real time using an analog-to-digital converter, adjusts the sampling rate and bit depth, removes noise from the digital voice signal using at least one of adaptive filtering, spectral subtraction, and beamforming, adjusts the volume using automatic gain control, and divides the voice segment using a voice activity detection algorithm.

[0122] The system segments voice segments using a voice activity detection algorithm. At this stage, the system converts collected analog voice signals into digital voice signals in real time using an analog-to-digital converter. For example, when a user says "Hello," the system optimizes the sound quality by adjusting the sampling rate and bit depth while converting the voice into a digital signal. A higher sampling rate allows for more detailed reproduction of the voice, while a deeper bit depth allows for greater timbre diversity.

[0123] Subsequently, the system removes noise from the digital voice signal using at least one of adaptive filtering, spectral subtraction, and beamforming. For example, in an environment with significant background noise, adaptive filtering can be applied to clearly extract only the user's voice.

[0124] In addition, the system adjusts the volume using automatic gain control. This feature maintains a consistent volume by automatically adjusting if the user's voice is too quiet or too loud. Finally, it segments voice segments using a voice activity detection algorithm. This algorithm detects the beginning and end of speech to clearly distinguish between periods with and without voice. For example, if a user says "Hello" and pauses briefly, the system recognizes this pause as a voice segment and identifies the subsequent period as a voice-free segment.

[0125] The system (400) converts collected analog voice signals into digital voice signals in real time using an analog-to-digital converter (ADC). At this time, the quality of the digital voice signal is determined by adjusting the sampling rate and bit depth. The sampling rate indicates how many times the analog signal is measured per second, and the bit depth indicates the number of bits used to express the amplitude of each sample. Various techniques are applied to remove noise from the digital voice signal. Adaptive filtering is a technique that removes noise by automatically adjusting the coefficients of a filter according to the noise characteristics of the surrounding environment. Spectral subtraction is a technique that removes noise by estimating the noise spectrum and subtracting it from the spectrum of the original signal. Beamforming is a technique that uses an array microphone to strengthen sound in a specific direction and attenuate sound in other directions.

[0126] The system (400) controls the volume using automatic gain control (AGC).

[0127] AGC is a technology that automatically amplifies or attenuates the input signal based on its magnitude to maintain a constant output signal magnitude.

[0128] The system (400) divides speech segments using a speech activity detection (VAD) algorithm. The VAD algorithm analyzes the energy, frequency characteristics, and temporal changes of the speech signal to distinguish between segments where speech exists and segments where speech is silent. This can improve the efficiency of speech recognition and analysis.

[0129] In operation S730, the system (400) can calculate a first deep voice probability by weighted averaging the output results of each model. The system (400) can calculate a first deep voice probability by extracting acoustic features including Mel frequency cepstrum coefficients (MFCC), linear predictive coding (LPC) coefficients, pitch, formants, and spectrograms on a frame-by-frame basis from a preprocessed digital voice signal, adjusting the frame size and frame shift interval, and inputting the extracted acoustic features into a first generative AI model, which is an ensemble AI model including a deep learning-based acoustic model, a natural language processing-based language model, and a statistical-based pattern recognition model, and weighted averaging the output results of each model.

[0130] The system (400) calculates a first deep voice probability by weighting the output results of each model. In this process, the system extracts various acoustic features from the preprocessed digital voice signal. For example, the system extracts acoustic features including Mel frequency cepstrum coefficients (MFCC), linear predictive coding (LPC) coefficients, pitch, formants, and spectrograms on a frame-by-frame basis. Each feature is used to analyze the voice in various ways. For example, MFCC represents the frequency components of the voice, and LPC models the characteristics of the timbre.

[0131] Pitch represents the height of a voice, and formants represent the frequency band of a specific vowel.

[0132] Spectrograms visually represent changes in frequency over time, making them useful for analyzing pronunciation patterns.

[0133] The system (400) optimizes each acoustic feature by adjusting the frame size and frame movement interval. For example, by setting the frame size to 20ms and the movement interval to 10ms, detailed analysis can be performed at short time intervals.

[0134] Subsequently, the system (400) inputs the extracted acoustic features into a first generative AI model, which is an ensemble AI model including a deep learning-based acoustic model, a natural language processing-based language model, and a statistical-based pattern recognition model. Each model determines whether it is a deep voice based on the input acoustic features, and calculates the first deep voice probability by weighting the output results of each model. For example, if the acoustic model determines it to be a deep voice with a probability of 75%, the language model with 80%, and the pattern recognition model with 70%, the system combines these to calculate the final probability.

[0135] The system (400) divides the preprocessed digital voice signal into frames and extracts acoustic features from each frame, including Mel frequency cepstrum coefficients (MFCC), linear predictive coding (LPC) coefficients, pitch, formants, and spectrograms. The precision of the analysis can be adjusted by controlling the frame size and frame shift interval. The extracted acoustic features are input into a first generative AI model, which is an ensemble AI model. The first generative AI model includes a deep learning-based acoustic model, a natural language processing-based language model, and a statistical-based pattern recognition model. Each model analyzes the input acoustic features to determine whether it is a deep voice.

[0136] The system (400) calculates a first deep voice probability by weighting the output results of each model. The weights are determined based on the performance of each model, and higher weights are assigned to models with better performance. This combines the strengths of each model and compensates for its weaknesses to increase the accuracy of deep voice detection.

[0137] In operation S740, the system (400) can calculate the second deep voice probability by inputting it into a second generative AI model, which is a time series analysis model. The system (400) can calculate the second deep voice probability by using extracted acoustic features to analyze temporal changes including at least one of pitch change, energy change, pronunciation speed change, formant change, and spectrum change of a speech signal, and inputting it into a second generative AI model, which is a time series analysis model including at least one of a recurrent neural network (RNN) based model, a long short-term memory (LSTM) based model, and a gated recurrent unit (GRU) based model.

[0138] The system calculates the final Deep Voice probability. In this process, the system fuses the first Deep Voice probability and the second Deep Voice probability using at least one of a weighted average, rule-based combination, or machine learning-based combination. For example, if the first Deep Voice probability is 70% and the second Deep Voice probability is 80%, the system can calculate the final probability as 75% by weighting these two probabilities. At this time, if the final Deep Voice probability is greater than a pre-set threshold that is user-configurable and can be dynamically adjusted depending on the situation, the system identifies the voice as Deep Voice and starts an alert mode. For example, if the threshold set by the user is 75%, the system will immediately issue a warning if the final probability exceeds this criterion.

[0139] The system (400) analyzes the pattern of change over time, such as pitch change, energy change, pronunciation speed change, formant change, and spectrum change of the voice signal, using the extracted acoustic features. Unlike human voice, deep voice may show subtle differences in the pattern of change over time. For example, deep voice may show a tendency for pitch change to be monotonous or pronunciation speed to remain constant compared to human voice.

[0140] The analyzed temporal change information is input into a second generative AI model, which is a time series analysis model. The second generative AI model includes at least one of a recurrent neural network (RNN)-based model, a long short-term memory (LSTM)-based model, and a gated recurrent unit (GRU)-based model.

[0141] RNN, LSTM, and GRU are all deep learning models specialized for time-series data analysis, which remember information from previous times and use it to analyze data at the current time.

[0142] The second generative AI model analyzes input temporal change information to calculate the second Deep Voice probability. That is, it determines whether it is Deep Voice by analyzing how similar the temporal change pattern of the voice signal is to the pattern appearing in Deep Voice.

[0143] In operation S750, the system (400) can calculate a final deep voice probability. The system (400) calculates the final deep voice probability by fusing the first deep voice probability and the second deep voice probability using at least one of a weighted average, a rule-based combination, and a machine learning-based combination, and if the final deep voice probability is greater than a preset threshold that is user-configurable and can be dynamically adjusted according to the situation, it can determine it as a deep voice and start an alarm mode.

[0144]

[0145] FIG. 8 is a flowchart illustrating a deep voice detection method using signal quality (SNR)-based adaptive weight allocation and resource optimization according to an embodiment of the present invention.

[0146] In step S810, the system (400) extracts a composite feature vector necessary for deep voice discrimination from input voice data. The extracted features include a Mel-Spectrogram representing frequency domain characteristics, MFCC and LPC coefficients representing temporal changes and vocal tract characteristics, and text data including the semantic context of the utterance.

[0147] In step S820, the system (400) inputs the extracted feature vector into an ensemble model comprising a plurality of AI models to calculate individual probabilities. According to one embodiment of the present invention, the ensemble model may be composed of a first AI model based on a Convolutional Neural Network (CNN) that analyzes Mel spectrograms (acoustics), a second AI model based on a Transformer that analyzes the context of text (language), and a third AI model that analyzes pitch / energy statistics (patterns). In particular, during this process, the system (400) optimizes hardware resources by performing 'Dynamic Resource Switching'. Specifically, the processor performs a first screening by activating only the third AI model (pattern recognition), which has a low computational load, during normal operation. If no suspicious section where the deep voice probability exceeds a preliminary threshold is detected as a result of the first analysis, the first AI model (CNN) and the second AI model (Transformer), which occupy large amounts of memory, remain in an unloaded state to minimize power consumption and memory usage. If a suspicious section is detected, the processor immediately loads the first and second AI models into active memory to perform a detailed analysis.

[0148] In step S830, the system (400) analyzes the signal quality of the input voice data and measures the signal-to-noise ratio (SNR). This is to prevent malfunction of a specific model in environments with high ambient noise or poor communication conditions.

[0149] In step S840, the system (400) performs 'Adaptive Weight Allocation', which changes the weights to be applied to each AI model in real time based on the measured SNR value. For example, if the measured SNR value is a low-quality signal (noisy) below a preset threshold (e.g., 10 dB), the reliability of the first AI model (acoustics) that analyzes fine frequency patterns may decrease. Accordingly, the system performs a correction by decreasing the weight of the first AI model (e.g., w1 = 0.2) and increasing the weight of the second AI model (language / context) that is relatively less affected by noise (e.g., w2 = 0.5).

[0150] In step S850, the system (400) calculates the determined dynamic weights and the individual probabilities of each model to obtain the final Deep Voice probability (P final It calculates the probability of each model (p i Weights (w) that vary depending on ) and SNR i Weighted average operation of )

[0151]

[0152] It follows. This enables precise detection that is robust to environmental variables.

[0153] Finally, if the final deep voice probability exceeds a threshold (e.g., S860), the system immediately triggers an alarm or performs a reporting procedure.

[0154] Furthermore, as another embodiment of the present invention, a 'liveness detection feedback' process may be performed after or in parallel with step S860. When deep voice is suspected, the system (400) presents an additional challenge to the caller through a user terminal and analyzes the caller's response to it. At this time, the system measures not only whether the answer is correct, but also the latency from the time of the question to the time of the response, and the jitter and shimmer fluctuation rates of the response voice waveform in milliseconds (ms). If the measured latency deviates from the average human reaction speed (e.g., 200–500ms) (AI computational delay), or if the jitter fluctuation rate shows an excessively flat pattern of less than 0.5% (characteristic of artificial synthesized voice), the system considers this as 'mechanical speech'. Accordingly, the system effectively blocks cleverly learned deep voice attacks by adding a penalty weight (e.g., +0.2) to the final deep voice probability and updating the probability upward, even if the answer content is accurate.

[0155]

[0156] The first Deep Voice probability and the second Deep Voice probability are fused using at least one of a weighted average, rule-based combination, and machine learning-based combination to calculate the final Deep Voice probability. A weighted average is a method of calculating the average by assigning weights to each probability.

[0157] Rule-based combining is a method of combining two probabilities according to predefined rules. For example, a rule such as "If either of the two probabilities is 0.8 or higher, set the final probability to 0.9" can be used. Machine learning-based combining is a method of combining two probabilities using a machine learning model. For example, a classification model can be trained that takes two probabilities as input and outputs a final probability. If the final Deep Voice probability exceeds a pre-set threshold that is user-configurable and can be dynamically adjusted depending on the situation, it is identified as Deep Voice. The threshold can be set directly by the user or automatically adjusted by the system based on the situation. For example, if the user sets a high security level, the threshold is lowered, increasing the probability of being identified as Deep Voice. Conversely, if the user prioritizes convenience, the threshold is raised, decreasing the probability of being identified as Deep Voice.

[0158] According to one embodiment, when the alarm mode starts, the system (400) may play a user-configurable preset warning voice that may include messages such as "Deep Voice detected," "Caution," and "Warning! Stop Deep Voice use," and output it to a loudspeaker through digital audio power amplification, and simultaneously with the warning voice output, display a visual warning message on a display that may be displayed in various forms such as text, images, and animations, generate vibrations or warning sounds in various patterns that may attract the user's attention, and transmit an alarm system including suspected Deep Voice voice data, caller information, Deep Voice probability, current time, location information, microphone ID, and occurrence time information to a management center host via a network.

[0159] When the alarm mode starts, a warning voice is played containing messages such as "Deep Voice detected," "Caution," and "Warning! Stop Deep Voice use." The warning voice may be a user-configurable preset voice. The warning voice is output to a loudspeaker via digital audio power amplification to convey the warning message to people around the user. Simultaneously with the output of the warning voice, a visual warning message is displayed on the display, which can be displayed in various forms such as text, images, and animations. For example, a text message saying "Deep Voice detected!" can be displayed on the screen along with a Deep Voice warning image. The system (400) generates vibrations or warning sounds in various patterns to attract the user's attention. For example, it can generate a warning sound that vibrates briefly multiple times or changes in pitch.

[0160] The system (400) transmits an alarm signal (or alarm message) containing voice data suspected of being deep voice, caller information, deep voice probability, current time, location information, microphone ID, and occurrence time information to a management center host via a network. Through this, the administrator can identify the occurrence of deep voice in real time and take necessary measures.

[0161] According to one embodiment, the system (400) controls the reception of alarm information from the management center host and displays the caller's voice profile information, which may include the caller's gender, age, region of origin, language used, voice characteristics, etc., along with the alarm occurrence location, on an alarm display module. After receiving the alarm information, the system starts on-site monitoring and comprehensively analyzes sensor data, which may include information collected from CCTV video, temperature, humidity, illuminance, motion detection sensors, etc., and user information, which may include the user's location, authority, access records, etc., to verify on-site information. After verifying the on-site information, the system can perform voice intercom reconfirmation with noise removal and voice enhancement technology applied, which supports bidirectional communication through a microphone and a distributed control area.

[0162] The system (400) receives alarm information transmitted from the management center host.

[0163] Alert information includes the location of the deep voice occurrence and the caller's voice profile information. The caller's voice profile information may include the caller's gender, age, region of origin, language used, voice characteristics, etc. This information can help identify and track deep voice users.

[0164] The system (400) displays the received alarm information on the alarm display module. The alarm display module visually displays the location where the deep voice occurred and provides the caller's voice profile information to the user. After receiving the alarm information, the system (400) starts on-site monitoring. On-site monitoring is a process of verifying on-site information by comprehensively analyzing sensor data, including information collected from CCTV video, temperature, humidity, illuminance, motion detection sensors, etc., and user information, including the user's location, authority, access records, etc.

[0165] The system (400) analyzes collected field information to more accurately identify the situation in which Deep Voice occurred. For example, it can identify the location and time where Deep Voice was used through CCTV video analysis and track the movement path of the Deep Voice user through motion detection sensor data. After verifying the field information, the system (400) supports bidirectional communication through a microphone and a distributed control area and performs reconfirmation through a voice intercom with noise removal and voice enhancement technology applied. That is, it directly converses with a person at the scene where Deep Voice occurred to understand the situation and finally confirm whether Deep Voice was used.

[0166] According to one embodiment, the system controls the reception of alert information from a management center host. In this process, the system displays caller voice profile information, which may include the caller's gender, age, region of origin, language used, voice characteristics, etc., on an alert display module. For example, when a user detects a suspicious voice during a call, the system displays information on the screen such as "The caller is presumed to be male, appears to be 30 years old, is from Seoul, and speaks Korean" to aid in a quick judgment.

[0167] Upon receiving the aforementioned alarm information, the system initiates on-site monitoring. At this time, the system comprehensively analyzes the situation using sensor data, which may include information collected from CCTV footage, temperature, humidity, illuminance, and motion detection sensors. For example, it detects movement in a specific space via CCTV and identifies abnormal signs by analyzing environmental conditions through temperature and humidity data. Additionally, the system verifies on-site information by comprehensively analyzing user data, which may include user location, authority, and access records. For instance, if a specific user has a history of unauthorized access, the system may issue a warning based on this information.

[0168] After verifying on-site information, the system supports two-way communication through microphones and distributed control areas, and performs voice intercom reconfirmation with noise removal and voice enhancement technologies applied. For example, it eliminates background noise that may occur during a call, allowing the other party's voice to be transmitted more clearly. This ensures smooth communication with the user.

[0169] According to one embodiment, the system (400) pauses the real-time voice intercom during the voice intercom reconfirmation process and delivers a warning message to the deepfake voice user, which can be generated using voice synthesis technology and provided in various languages ​​and voices, stating, "Deep voice usage has been detected. Legal liability may apply." The system receives information about the deepfake voice user from the management center and issues an alert including the deepfake voice user's location information, risk level, response manual, access path, surrounding situation information, and relevant legal information to notify security personnel. Based on the deep voice detection results, the system analyzes user behavior patterns including voice usage frequency, time, place, conversation partner, words used, and emotional changes to predict potential dangerous situations in advance and controls the system to suggest warning and preventive measures to the user terminal, such as deep voice prevention education, security system reinforcement, and user authentication procedure reinforcement.

[0170] During the voice intercom reconfirmation process, the system (400) pauses the real-time voice intercom and delivers a warning message generated using voice synthesis technology to the deepfake voice user. The warning message includes content such as "Deep voice usage detected. Legal liability may apply," and can be provided in various languages ​​and voices.

[0171] The system (400) receives information about deepfake voice users from the management center. The information about deepfake voice users includes location information, risk level, response manual, access route, surrounding situation information, and relevant legal information. Based on the received information, the system (400) issues an alert to notify security personnel. Based on the deep voice detection results, the system (400) analyzes user behavior patterns including voice usage frequency, time, place, conversation partner, words used, and emotional changes. Through the analyzed behavior patterns, potential dangerous situations are predicted in advance. For example, if the frequency of using a specific word is high at a specific place during a specific time period, this can be determined as a potential dangerous situation.

[0172] The system (400) suggests warning and preventive measures to the user terminal, such as deep voice prevention education, security system reinforcement, and user authentication procedure reinforcement. For example, it may provide deep voice prevention education materials or recommend security reinforcement measures such as changing passwords or setting up two-factor authentication.

[0173] According to one embodiment, the system (400) collects voice data in real time through a sensor, performs a preprocessing process including noise removal, volume control, and voice segment segmentation on the collected voice data, determines voice segments and noise segments in the preprocessed voice data through endpoint detection, removes noise segments, and performs framing using a Hamming window on the voice segments.

[0174] According to one embodiment, the system (400) performs high-frequency pre-emphasis processing using a first-order high-pass filter for each frame, and for the voice data of the voice segment, removes the influence of vocal cord excitation from the fundamental voice signal frequency spectrum and removes the influence of formants through a lifting method to obtain a glottal waveform signal, constructs an acoustic model using linear predictive coding and a discrete complete or collected work model, and finally obtains a glottal waveform signal using lifting.

[0175] According to one embodiment, the system (400) extracts feature parameters including normalized amplitude, glottal closure time, pitch change, formant change, spectrogram, and Mel frequency cepstrum coefficient (MFCC) from the glottal waveform signal, performs classification by providing the extracted feature parameters as input to an SVM model trained using training data including deep voice data generated by a generative AI model, determines the state of a speaker including one of normal voice, suspected deep voice, or confirmed deep voice using the classification result, outputs a speaker state label, and performs at least one operation among displaying a warning message, blocking a call, recording, or reporting to the Cyber ​​Safety Bureau of the National Police Agency when a suspected deep voice or confirmed deep voice label is received based on the speaker state label.

[0176] According to one embodiment, the system pauses the real-time voice intercom during the voice intercom reconfirmation process. At this time, the system delivers a warning message to the deepfake voice user that can be generated using voice synthesis technology. For example, the system provides a warning message such as "Deep voice usage detected. Legal liability may apply" in various languages ​​and voices. This can be usefully applied in multinational environments so that users can understand it. After the warning message is delivered, the system receives information about the deepfake voice user from the management center. This information includes the deepfake voice user's location information, risk level, response manual, access route, surrounding situation information, and relevant legal information. For example, it enables the user to respond quickly to security personnel through information such as "Risk Level: High, Access Route: Entrance 3."

[0177] Subsequently, the system issues a warning to security personnel to enable a rapid response to the situation. Based on the results of deep voice detection, the system analyzes user behavior patterns—including voice usage frequency, time, location, conversation partners, words used, and emotional changes—to predict potential risk situations in advance. For example, if a pattern is detected where a user frequently makes suspicious remarks at a specific time, the system records this and considers it a risk factor. Finally, the system controls the user terminal to suggest warnings and preventive measures, such as deep voice prevention training, security system reinforcement, and enhanced user authentication procedures. For instance, the system sends a notification to the user stating, "Please participate in the training program to prevent deep voice usage," thereby encouraging the user to take preventive measures. This contributes to raising user awareness and reducing potential risks.

[0178] According to one embodiment, the system (400) emits an acoustic signal in response to a profile processing request associated with a device, receives a series of customized echo acoustic signals based on acoustic signals reflected from unique contours of one or more depth parts associated with a user, extracts one or more region segments associated with the echo acoustic signals to train a classification model, and can generate a classification model based on one or more extracted region segments.

[0179] In response to a profile processing request associated with a device, the system (400) emits an acoustic signal with a specific frequency through a speaker. This acoustic signal is reflected back from the unique contours of one or more depth parts of the user's body, particularly the face. The system (400) receives the reflected acoustic signal through a microphone and converts it into a customized series of echo acoustic signals. A classification model is used to authenticate the user. To train the classification model, the system (400) extracts one or more region segments from the echo acoustic signals. These region segments contain information representing the user's unique physical characteristics. The classification model is trained using the extracted region segments.

[0180] For example, characteristic patterns can be extracted from echo acoustic signals containing depth information of the user's nose, mouth, ears, etc., and a classification model can be created to distinguish "User A", "User B", etc. based on this.

[0181] According to one embodiment, the system (400) extracts a combined feature representation based on a classification model, wherein the combined feature representation includes acoustic features and visual landmark features of a user obtained simultaneously during a discrete epoch, generates a vector-based classification model used for predicting the combined feature representation, determines whether the combined feature representation is associated with an echo signature based on the prediction of the combined feature representation, wherein the combined feature representation associated with the user profile includes extracted landmark coordinates and acoustic features associated with unique contours of one or more depth portions for a discrete epoch, and when generating the combined feature representation associated with the user profile, the combined feature representation may be augmented with synthesized acoustic features and augmented landmark coordinates associated with unique contours of one or more depth portions associated with the user for a changed discrete epoch.

[0182] The system (400) extracts a combined feature representation based on a trained classification model. The combined feature representation includes the user's acoustic features and visual landmark features obtained simultaneously over a discrete epoch, that is, over a specific period of time. The acoustic features are features extracted from the echo acoustic signal described earlier. The visual landmark features are location information of the eyes, nose, mouth, etc., extracted from the user's face image obtained through a camera.

[0183] The system (400) generates a vector-based classification model used to predict a combined feature representation. This classification model determines whether the combined feature representation matches an echo signature, that is, the user's unique acoustic and visual features. The user profile includes extracted landmark coordinates and acoustic features associated with unique contours of one or more depth portions for discrete epochs. When generating the user profile, the system (400) may augment the combined feature representation using the user's synthesized acoustic features and augmented landmark coordinates for a changed discrete epoch. This is to maintain authentication accuracy by taking into account that the user's appearance or acoustic features may change over time.

[0184] According to one embodiment, the system (400) may extract one or more altered vector features associated with synthesized acoustic features and / or augmented landmark coordinates to generate a combined feature representation for an altered discrete epoch. A profile processing request associated with a computing device may include initial registration of an original user profile or authentication of a user profile for a current epoch for access to the computing device. A profile processing request associated with a computing device may also include authenticating a current user profile for access to the computing device by comparing a vector feature associated with a combined feature representation of the current user profile for a current epoch with a vector feature associated with a combined feature representation of the original user profile.

[0185] The system (400) extracts one or more altered vector features associated with synthesized acoustic features and / or augmented landmark coordinates to generate a combined feature representation for the altered discrete epoch. A profile processing request may include the initial registration of the original user profile or the authentication of the user profile for the current epoch for access to the computing device. That is, the system (400) creates a profile when the user is first registered and subsequently authenticates the user profile for the current epoch whenever the user attempts to access the device.

[0186] User profile authentication is performed by comparing the vector features associated with the combined feature representation of the current user profile for the current epoch with the vector features associated with the combined feature representation of the original user profile. That is, the system (400) compares the features of the current user with the features of the registered original user to check if they match. If the similarity between the two feature vectors is sufficiently high, user authentication is successful and device access is allowed.

[0187] According to one embodiment, the system emits an acoustic signal in response to a profile processing request associated with a device. In this process, the system generates an acoustic signal of a specific frequency tailored to the user's environment. For example, if a user requests an acoustic signal to listen to music, the system emits a high-quality audio signal to enhance the clarity of the music.

[0188] It receives a customized series of echo acoustic signals based on acoustic signals reflected from the unique contours of one or more depth parts associated with the user. For example, when a user emits an acoustic signal in a specific space, the echo acoustic signals returning after reflecting off walls or furniture are collected. These echo signals contain information about the user's location or the structure of the space, which the system analyzes to enhance its understanding of the environment.

[0189] The system extracts one or more region segments associated with the echo acoustic signal to train a classification model. For example, the system identifies key segments of the signal by analyzing changes in specific frequency bands within the received echo signal. Since each segment contains important information about the user's environment or location, a classification model can be generated based on these segments.

[0190] According to one embodiment, the system extracts a combined feature representation based on a classification model. In this process, the system generates a combined feature representation by simultaneously analyzing acoustic features and visual landmark features collected from a user. For example, when a user speaks in a specific space, visual information of that space (e.g., an image captured by a camera) is collected along with the voice.

[0191] A vector-based classification model is created for the prediction of combined feature representations. This model is designed to analyze a combination of acoustic and visual data. For example, the system trains the model by combining frequency components of speech with landmark information that visually indicates the user's location.

[0192] Based on the prediction of the combined feature representation, it determines whether the combined feature representation is associated with an echo signature. For example, the system evaluates the likelihood of a specific acoustic pattern occurring in a specific environment to determine whether this pattern is similar to previously collected echo signatures.

[0193] The combined feature representation associated with the user profile includes extracted landmark coordinates and acoustic features associated with unique contours of one or more depth parts for discrete epochs. For example, the system analyzes contours based on the user's physical shape or features to generate landmark coordinates for said contours.

[0194] When generating a combined feature representation associated with a user profile, the combined feature representation can be augmented with synthesized acoustic features and augmented landmark coordinates associated with unique contours of one or more depth parts associated with the user for a changed discrete epoch. For example, the system can learn acoustic features generated by the user in a new environment and combine them with existing profile data to create a more sophisticated user model.

[0195] According to one embodiment, the system extracts one or more modified vector features associated with synthesized acoustic features and / or augmented landmark coordinates to generate a combined feature representation for a modified discrete epoch. For example, when a user generates speech in a new environment, the system collects acoustic features specific to that environment and updates vector features based on them.

[0196] Profile processing requests associated with a computing device may include the initial registration of the original user profile or user profile authentication for the current epoch for accessing the computing device. For example, when a user uses a new device, the system verifies the original profile information and proceeds with the process of granting access rights.

[0197] A profile processing request associated with a computing device may also involve authenticating the current user profile for accessing the computing device by comparing vector features associated with the combined feature representation of the current user profile for the current epoch with vector features associated with the combined feature representation of the original user profile. For example, based on an acoustic pattern generated by a user at a specific time and place, the system authenticates the user's identity by verifying whether this pattern matches the data of the original profile. Through this process, the system can perform more secure and reliable user authentication.

[0198] The embodiments described above may be implemented as hardware components, software components, and / or combinations of hardware components and software components. For example, the devices, methods, and components described in the embodiments may be implemented using one or more general-purpose computers or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions.

[0199]

[0200] A processor according to one embodiment of the present invention performs 'Dynamic Resource Switching' for the efficient use of computing resources. Normally, it performs primary screening by activating only a statistical-based model (third AI model) with low computational load, and performs precise analysis only when a suspicious section where the primary deep voice probability exceeds a threshold is detected, by loading a deep learning model (first and second AI models) that occupies a large amount of memory into memory. Through this, hardware optimization is applied to minimize CPU occupancy and power consumption in an always-on monitoring environment and to guarantee real-time detection speed (latency).

[0201]

[0202] The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0203] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0204] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0205] [Available for Industrial Use]

[0206] The present invention is applicable in the information and communications security industry, voice phishing prevention systems, and financial security service fields that require artificial intelligence-based voice analysis technology.

[0207] Alternatively, the present invention has industrial applicability in various digital content security industries and communication service fields where deep voice detection and user identity verification are required.

Claims

1. In a system for detecting deep voices generated by generative artificial intelligence, A memory for storing instructions; and one or more processors for executing said instructions, A deep voice detection and accident prevention AI solution providing system configured such that the processor extracts a feature vector including frequency characteristics and temporal change patterns from input voice data, inputs the extracted feature vector into an ensemble AI model including a first AI model which is an acoustic model, a second AI model which is a language model, and a third AI model which is a pattern recognition model to calculate each individual deep voice probability, measures the signal quality of the input voice data to calculate the signal-to-noise ratio (SNR), calculates the final deep voice probability through an adaptive weight allocation algorithm that varies the weights to be applied to each AI model in real time based on the calculated SNR value, and generates a deep voice suspicion alert when the final deep voice probability exceeds a threshold value.

2. In Paragraph 1, The processor is configured to perform dynamic resource switching to control the load of computing resources, perform a first analysis of the input voice data using the third AI model, wherein the amount of computation is less than a preset standard, and, only when a section in which the deep voice probability exceeds a preliminary threshold is detected as a result of the first analysis, load the deep learning-based first AI model and the second AI model into the active area of ​​the memory to perform a precise analysis. Deep voice detection and accident prevention AI solution provision system.

3. In Paragraph 1, The processor is configured to perform liveness detection feedback by adding a penalty weight to the final deep voice probability and updating it regardless of the content of the response when the final deep voice probability exceeds the threshold, if the final deep voice probability exceeds the threshold, if the processor presents an additional question to verify the identity of the caller through the user terminal, receives the caller's response voice signal to the additional question, measures the latency from the time of the question to the time of the response and the jitter and shimmer fluctuation rates of the response voice waveform, and if the latency exceeds a preset human response range or the jitter fluctuation rate falls within a preset artificial synthesized voice pattern range. Deep voice detection and accident prevention AI solution provision system.

4. In Paragraph 1, The above adaptive weight assignment algorithm, when the SNR value is a low-quality signal below a threshold, reduces the weight of the noise-sensitive first AI model (acoustics) and increases the weight of the context-dependent second AI model (language), and the final deep voice probability (P final ) is calculated based on the following formula, where w i is the weight of each model that varies according to SNR, p i is the output probability of each model, Deep voice detection and accident prevention AI solution provision system.

5. In Paragraph 1, The first AI model is a Convolutional Neural Network (CNN)-based model that receives a Mel-Spectrogram as input, the second AI model is a Transformer-based language model that receives utterance content converted into text as input, and the third AI model is a pattern recognition model that receives statistical values ​​of pitch and energy changes of speech as input. Deep voice detection and accident prevention AI solution provision system.

6. A deep voice detection method performed by a computing device, A step of extracting a feature vector including frequency characteristics and temporal change patterns from input voice data; A step of calculating individual probabilities by inputting the extracted feature vectors into an ensemble model comprising a first AI model which is an acoustic model, a second AI model which is a language model, and a third AI model which is a pattern recognition model; A step of analyzing the signal quality of the input voice data and measuring the signal-to-noise ratio (SNR); A step of calculating the final deep voice probability by performing adaptive weight assignment to dynamically determine the weights to be applied to each of the plurality of AI models according to the measured SNR value; and The step of displaying a warning message on the user terminal or transmitting a report signal to an investigative agency server when the above final deep voice probability exceeds a threshold, Method for providing an AI solution for deep voice detection and accident prevention.

7. A computer-readable recording medium storing a program for executing the method of paragraph 6 on a computer.

Images