Apparatus and method for contaminating voice data in real time

Abstract

An embodiment of the present invention may provide a method for contaminating voice data in real time, the method comprising the steps of: pre-processing voice data; adding contamination noise in a process of upsampling the pre-processed voice data in a frequency domain; generating contaminated voice data by inputting the voice data to which the contamination noise is added into a deep learning model; and generating a voice signal by converting the contaminated voice data into a time domain by using a regression neural network.

Description

Device and method for polluting voice data in real time

[0001] According to one embodiment of the present invention, a real-time voice data contamination device and method are provided, and more specifically, a device and method for contaminating voice data by intentionally adding errors and modifications to protect the voice data of a creator are provided.

[0002] With the rapid advancement of artificial intelligence (AI) technology, the value of voice data is increasing day by day, and the voice data trading market is steadily expanding. This voice data is playing a crucial role as a core resource in various AI applications, such as voice recognition systems, conversational AI, and natural language processing.

[0003] In the field of speech recognition, securing user voices containing human speech, as well as preprocessing them to protect personal information when using them as training data, are emerging as critical issues. While user voices contain a wealth of information that can be utilized in a variety of fields, their unprocessed use can significantly limit their utility, as it can potentially violate personal privacy.

[0004] Unauthorized use of a creator's voice data or the user's voice can lead to privacy violations. This can lead to the unauthorized collection of a user's personal information and its use for other purposes, or the exploitation of voice recognition technology to impersonate or deceive the user. Furthermore, sensitive information can be leaked, potentially violating an individual's privacy.

[0005] The technical task of the present invention is to provide a real-time voice data contamination device and method that prevents unauthorized learning by modifying a user's voice signal through a real-time high-speed computation algorithm and minimizes the impact on actual use through efficient storage and compression of voice.

[0006] According to one embodiment of the present invention, a method for polluting voice data in real time can be provided, including a step of preprocessing voice data, a step of adding contamination noise in a process of upsampling the preprocessed voice data in the frequency domain, a step of inputting the voice data with the contamination noise added into a deep learning model to generate polluted voice data, and a step of converting the polluted voice data into the time domain using a recurrent neural network to generate a voice signal.

[0007] According to one embodiment of the present invention, a step of generating noise data to be added to voice data may be further included, taking into account the frequency range and intensity of the voice data.

[0008] Here, the noise data is characterized in that it does not include the hearing range of 20 Hz to 20,000 Hz.

[0009] According to one embodiment of the present invention, the step of generating contaminated voice data may include the step of setting the contaminated voice data as a predictive variable and setting the original voice data as a target signal, and the step of generating the contaminated voice data using a deep learning model according to the set conditions.

[0010] According to one embodiment of the present invention, a device for polluting voice data in real time can be provided, including a memory for storing computer executable instructions, and a processor for preprocessing voice data by executing the computer executable instructions, adding pollution noise in the process of upsampling the preprocessed voice data in the frequency domain, inputting the voice data with the added pollution noise into a deep learning model to generate polluted voice data, and converting the polluted voice data into the time domain using a recurrent neural network to generate a voice signal.

[0011]

[0012] According to one embodiment of the present invention, a method may include a step of converting voice data into a frequency domain using a short-time Fourier transform (STFT), a step of calculating an output estimate using a plurality of STFT vectors including continuous noise generated in the frequency domain conversion process as input values, a step of modifying voice data by adding noise data to the frequency domain voice data, and a step of converting the modified voice data into voice data in a time domain using an inverse short-time Fourier transform (ISTFT).

[0013] According to one embodiment of the present invention, the step of converting into a frequency domain is characterized by including the steps of dividing time domain voice data into blocks with a window length of 256 samples, applying a Hamming window to each block, maintaining continuity of overlapping voice data between each block using an overlap of 75%, applying a short-time Fourier transform to convert each block into a frequency domain vector, and removing samples corresponding to negative frequencies from the frequency domain vector to reduce the size of the vector in the frequency domain to a window length of 129 samples.

[0014] According to one embodiment of the present invention, the plurality of STFT vectors are characterized by being eight consecutive STFT vectors including an STFT of the current time point and seven STFT vectors prior to the current time point.

[0015] According to one embodiment of the present invention, the method may include a step of forming a blockchain network based on a smart contract between a seller and a consumer, a step of creating a blockchain wallet in which a secret key between the seller and the consumer is set, a step of registering voice data including noise data in the blockchain network, and a step of transmitting the voice data including noise data to the consumer upon confirmation of payment from the consumer.

[0016] According to one embodiment of the present invention, when voice data is downloaded by a third party other than a consumer through a smart contract, a step of lowering the quality of the voice data by expanding the range of noise data included in the voice data may be included.

[0017] According to one embodiment of the present invention, the processor can obtain a speech recognition error rate of contaminated data using at least one of CER and WER.

[0018] CER measures the error rate at the character level, and is calculated by dividing the number of incorrectly recognized characters by the total number of characters. CER = (number of incorrectly recognized characters) / (total number of characters)

[0019] WER measures the error rate at the word level and is calculated by dividing the number of incorrectly recognized words by the total number of words. WER = (Number of incorrectly recognized words) / (Total number of words)

[0020] Our invention protects creators' voice data and prevents privacy violations. Furthermore, it prevents creators' voice data from being used in indiscriminate AI training data.

[0021] According to the present invention, a service based on artificial intelligence learning data in which individual privacy is guaranteed can be provided.

[0022] By inserting a specific pattern of noise into voice data according to the present invention, unauthorized use is prevented, thereby making unauthorized copying and use difficult without affecting the quality of the original data.

[0023] FIG. 1 is a flowchart illustrating a method for polluting voice data in real time according to one embodiment of the present invention.

[0024] FIG. 2 is a drawing specifically explaining a voice data contamination technology based on real-time frequency conversion and audio signal modification technology according to one embodiment of the present invention.

[0025] FIG. 3 is a diagram illustrating a voice signal upsampling algorithm for strengthening data security according to one embodiment of the present invention.

[0026] FIG. 4 is a diagram for explaining a neural network output precision maximization model that integrates audio signal transformation according to one embodiment of the present invention.

[0027] FIG. 5 is a diagram for explaining a regression neural network algorithm for precise conversion of audio data including noise data according to one embodiment of the present invention.

[0028] FIG. 6 is a drawing for explaining the configuration of a device for polluting voice data in real time according to one embodiment of the present invention.

[0029] FIG. 7 is a diagram for explaining technical features of a blockchain platform that distributes voice data according to one embodiment of the present invention.

[0030] Hereinafter, the present disclosure will be described in detail with reference to the attached drawings.

[0031] The terms used in this specification will be briefly explained, and the present disclosure will be described in detail.

[0032] The terms used in the embodiments of this disclosure have been selected from widely used, current terms, taking into account the functions of this disclosure. However, these terms may vary depending on the intentions of those skilled in the art, precedents, the emergence of new technologies, etc. Furthermore, in certain cases, terms may be arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the description of the relevant disclosure. Therefore, the terms used in this disclosure should not be defined simply as names of terms, but rather based on the meanings of the terms and the overall content of this disclosure.

[0033] In this specification, expressions such as “has,” “can have,” “includes,” or “may include” indicate the presence of a feature (e.g., a number, function, operation, or component such as a part), and do not exclude the presence of additional features.

[0034] The expression "at least one of A and / or B" should be understood to mean either "A" or "B" or "A and B".

[0035] The expressions “first,” “second,” “first,” or “second,” etc., used herein can mathematically represent various components, regardless of order and / or importance, and are only used to distinguish one component from another, but do not limit the components.

[0036] When it is said that a component (e.g., a first component) is “(operatively or communicatively) coupled with / to” or “connected to” another component (e.g., a second component), it should be understood that the component may be directly coupled to the other component, or may be connected through another component (e.g., a third component).

[0037] Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, terms such as "comprise" or "consist of" are intended to indicate the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but should be understood not to preclude the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0038] In the present disclosure, a "module" or "part" performs at least one function or operation and may be implemented as hardware or software, or as a combination of hardware and software. Furthermore, multiple "modules" or multiple "parts" may be integrated into at least one module and implemented as at least one processor (not shown), excluding any "modules" or "parts" that need to be implemented as specific hardware.

[0039] Hereinafter, a device (10) for polluting voice data in real time according to one embodiment of the present invention will be abbreviated as a voice data polluting device (10).

[0040] FIG. 1 is a flowchart illustrating a method for polluting voice data in real time according to one embodiment of the present invention.

[0041] According to one embodiment of the present invention, in block 1001, the voice data contamination device (10) can preprocess voice data.

[0042] The process of preprocessing voice data involves voice signal acquisition, digital signal processing (DSP), normalization, feature extraction, postprocessing, and feature vector utilization. The detailed process is described below.

[0043] According to one embodiment of the present invention, in block 1002, the voice data contamination device (10) can add contamination noise in the process of upsampling preprocessed voice data in the frequency domain.

[0044] According to one embodiment of the present invention, the voice data contamination device (10) in block 1001 can convert voice data into a frequency domain using a short-time Fourier transform (STFT).

[0045] According to one embodiment of the present invention, a voice data contamination device (10) contaminates data in the process of converting data into a frequency domain, and then reverse-converts the contaminated data to generate voice data in the time domain.

[0046] According to one embodiment of the present invention, the step of converting into a frequency domain is characterized by including the steps of dividing time domain voice data into blocks with a window length of 256 samples, applying a Hamming window to each block, maintaining continuity of overlapping voice data between each block using an overlap of 75%, applying a short-time Fourier transform to convert each block into a frequency domain vector, and removing samples corresponding to negative frequencies from the frequency domain vector to reduce the size of the vector in the frequency domain to a window length of 129 samples.

[0047] Here, overlap refers to the overlapping section between one window block and the next window block in signal processing, and is used to ensure signal continuity and smooth transition.

[0048] More specifically, in one embodiment of the present invention, the inclusion of overlapping data between consecutive blocks using a 75% overlap by the voice data contamination device (10) means that each window shares 75% of the samples with the previous window block, and the remaining 25% of the samples are newly added. This maintains signal continuity and enhances the accuracy of spectrum analysis.

[0049] For example, if the window size is 256 samples and the overlap ratio is 75%, the starting point of each block can be determined by the following calculation:

[0050] Window size: 256 samples

[0051] Overlap ratio: 75%

[0052] Number of overlapping samples: 256 * 75% = 192 samples

[0053] In this case, the new starting point of each window block is shifted by 64 samples (256 - 192) from the starting point of the previous block. For example, the starting points for the first five consecutive blocks could be as follows:

[0054] First block: 0 to 255 samples

[0055] Second block: 64 to 319 samples

[0056] Third block: 128 to 383 samples

[0057] Fourth block: samples 192 to 447

[0058] Fifth block: 256 to 511 samples

[0059] In this way, each new block overlaps the previous block by 192 samples. That is, the last 192 samples of the current block overlap the first 192 samples of the next block.

[0060] The use of overlap preserves signal continuity and improves the quality of the transformed spectrum by retaining more frequency information. It also reduces artifacts that can occur at the edges when using window functions.

[0061] According to one embodiment of the present invention, a voice data contamination device (10) uses STFT to transform a voice signal into a frequency domain, and can transform voice data according to set parameters. At this time, a prediction input can be configured using eight consecutive STFT vectors. In addition, noise can be added in the frequency domain to contaminate data, and then the noise can be transformed through a process of transforming the data back into the time domain, thereby generating voice data with added noise. At this time, the transformation process can be carefully controlled so that the noise is not audible to the human ear.

[0062] Also, the reason for removing negative frequencies is that negative frequencies are created by the symmetry of real-valued signals, so the entire frequency component can be restored by taking only the positive frequency components.

[0063] According to one embodiment of the present invention, noise data to be added to voice data can be generated by considering the frequency range and intensity of voice data. Here, the noise data includes data in a hearing range that is inaudible to humans. For example, the noise data is characterized by not including a hearing range of 20 Hz to 20,000 Hz. In other words, the voice data contamination device (10) of the present invention prevents the indiscriminate use of voice data during the artificial intelligence learning process by adding noise data in a range that is inaudible to the voice data.

[0064] According to one embodiment of the present invention, in block 1003, the voice data contamination device (10) can input voice data with added contamination noise into a deep learning model to generate contaminated voice data.

[0065] According to one embodiment of the present invention, in block 1004, the voice data contamination device (10) can generate a voice signal by converting the contaminated voice data into a time domain using a regression neural network.

[0066] More specifically, it may include a step of converting voice data into a frequency domain using a Fourier transform (STFT), a step of calculating an output estimate using a plurality of STFT vectors including continuous noise generated in the frequency domain conversion process as input values, a step of transforming voice data by adding noise data to the frequency domain voice data, and a step of converting the transformed voice data into voice data in the time domain using an inverse short-time Fourier transform (ISTFT).

[0067] In addition, the step of generating contaminated voice data according to one embodiment of the present invention may include the step of setting the contaminated voice data as a predictive variable and setting the original voice data as a target signal, and the step of generating the contaminated voice data using a deep learning model according to the set conditions.

[0068]

[0069] According to one embodiment of the present invention, a voice data contamination device (10) can propose a voice signal upsampling algorithm for improving the efficiency of artificial intelligence learning and strengthening data security.

[0070] According to one embodiment of the present invention, a voice data contamination device (10) can perform data preprocessing including a data division process that can perform a training set, a verification set, and a test set through a data collection and processing process of normalizing a voice signal and extracting features.

[0071] More specifically, the data preprocessing process includes voice signal collection, digital signal processing (DSP), normalization, feature extraction, postprocessing, and feature vector use.

[0072] First, voice data can be collected and analog signals can be converted to digital signals.

[0073] Digital signal processing is achieved through sampling and quantization. Sampling allows for the measurement of a signal at specific time intervals to convert an analog signal into a digital signal. According to Nyquist theorem, the sampling frequency must be at least twice the maximum frequency of the original signal. Quantization allows the sampled signal to be represented using a finite number of bits, and the greater the number of quantized bits, the more precise the signal representation.

[0074] Next, because the speech signal changes over time, it is divided into short time intervals (typically 20-40 ms), called frames. Frame segmentation allows for feature extraction from each frame.

[0075] Next, a window function is applied to each frame to reduce discontinuities that occur at the beginning and end of the frame. Commonly used window functions include the Hamming window and the Hanning window.

[0076] Normalization is the process of making the signal amplitude constant, reducing fluctuations in the audio signal and facilitating signal comparison in subsequent processing steps. For example, this is achieved through loudness normalization, which adjusts the overall amplitude of the loudness signal to a constant level, and DC offset removal, which removes the DC component contained in the signal to make the signal symmetrical around zero.

[0077] The feature extraction process is a process of extracting meaningful information from a voice signal, and features can be extracted using methods such as MFCC (Mel-Frequency Cepstral Coefficients), LPC (Linear Predictive Coding), and Chroma Features.

[0078] In post-processing, the extracted feature vectors can be additionally normalized or filtered. This is done to make the distribution of the feature vectors uniform or to reduce noise.

[0079] Next, the normalized and feature-extracted vectors can be used as inputs to machine learning models to perform various tasks such as speech recognition and sentiment analysis.

[0080] This process is an essential preprocessing step for understanding and analyzing voice signals, and the algorithms and technologies applied at each step are important.

[0081] Preprocessed voice signals can be decontaminated using artificial intelligence voice learning data contamination technology.

[0082] According to one embodiment of the present invention, a voice data contamination device (10) can generate voice data containing noise by using a voice signal upsampling algorithm, a neural network output precision maximization model, and an optimized regression neural network algorithm.

[0083] Here, PyTorch and TensorFlow can be used for modeling, and Horovod and Deepspeed can be used for distributed learning.

[0084] Voice data containing this generated noise can be distributed through a creator data protection and transaction service platform applicable to various industries. This platform may include a blockchain platform.

[0085]

[0086] According to one embodiment of the present invention, a voice data contamination device (10) can collect voice data, such as an individual's voice, and process noise using a data contamination algorithm. The data contamination algorithm may include a voice signal upsampling algorithm, a neural network output precision maximization model, and an optimized regression neural network algorithm.

[0087] Contaminated voice data can be protected from unauthorized use and prevent privacy violations. Contaminated voice data can be traded through a secure voice data platform.

[0088] Tainted voice data refers to a technology that intentionally adds minor errors or modifications to voice data to prevent unauthorized use or duplication of the original data. This technique enhances data security by altering some of the characteristics of the original voice, making it difficult for unauthorized users to recognize or utilize the data. Data tampering is primarily used for copyright protection and privacy, and can also be used to prevent data misuse during AI learning.

[0089] FIG. 2 is a drawing specifically explaining a voice data contamination technology based on real-time frequency conversion and audio signal modification technology according to one embodiment of the present invention.

[0090] In particular, according to one embodiment of the present invention, the voice data contamination device (10) may include a voice contamination data AI model that creates an input value that intentionally induces misclassification by using a specific noise (Noise or Perturbation) value created by exploiting an internal vulnerability of a deep learning model.

[0091] According to one embodiment of the present invention, the voice data contamination device (10) increases the complexity of data through upsampling of the audio signal and sets the prediction variables and target signals of the neural network by utilizing the size and spectrum of the transformed and clean audio signal, thereby preventing unauthorized learning that exceeds the processing capacity of the AI ​​model.

[0092] In this process, a recurrent neural network algorithm is trained to minimize the mean squared error between the output target and the input target based on predicted input values, which can be used to transform noisy audio into the time domain. This approach is designed to disrupt learning by including variations that are not perceptible to the human ear.

[0093] In particular, according to one embodiment of the present invention, a voice data contamination device (10) utilizes voice recognition technology to collect and automatically label voice data, and maximizes the precision of neural network output to complicate data so that deep learning recognition is impossible when learning contaminated data. In addition, a voice signal upsampling technology that enables general use of original voice data can be used in fields other than artificial intelligence learning.

[0094] FIG. 3 is a diagram illustrating a voice signal upsampling algorithm for strengthening data security according to one embodiment of the present invention.

[0095] According to one embodiment of the present invention, a voice data contamination device (10) can use a voice signal upsampling algorithm to improve the processing efficiency of an artificial intelligence model and strengthen data security.

[0096] In particular, by upsampling the original and copied audio signals to 8 kHz, the computational load of the neural network can be increased, and in the process, the frequency range of the voice data can be expanded.

[0097] Upsampling increases the complexity of the data, which increases the amount of data that the AI ​​model has to process, causing overload.

[0098] FIG. 4 is a diagram for explaining a neural network output precision maximization model that integrates audio signal transformation according to one embodiment of the present invention.

[0099] According to one embodiment of the present invention, a voice data contamination device (10) can generate an audio signal deformation that exceeds the human hearing range through an integrated neural network output precision maximization model.

[0100] The diagram in Figure 4 illustrates how two attack models can output data containing noise by using the output of specific embedding layers to access and analyze data. This allows the original voice data to be trained in increasingly complex ways, step by step. "Shallow Embeddings" and "Deep Embeddings" are feature vectors extracted from different layers, each containing different levels of information and abstracted features.

[0101] This can hinder voice learning by preventing voice signals from being used for AI learning.

[0102] FIG. 5 is a diagram for explaining a regression neural network algorithm for precise conversion of audio data including noise data according to one embodiment of the present invention.

[0103] According to one embodiment of the present invention, a voice data contamination device (10) can generate an input value that intentionally induces misclassification by using a specific noise value created by exploiting an internal vulnerability of a deep learning model.

[0104] FIG. 6 is a drawing for explaining the configuration of a device for polluting voice data in real time according to one embodiment of the present invention.

[0105] According to one embodiment of the present invention, a voice data contamination device (10) may include a memory (100), a processor (200), and a communication unit (300).

[0106] According to one embodiment of the present invention, a voice data contamination device (10) may include a memory (100) that stores computer executable instructions, and a processor (200) that preprocesses voice data by executing the computer executable instructions, adds contamination noise in the process of upsampling the preprocessed voice data in the frequency domain, inputs the voice data with the contamination noise added into a deep learning model to generate contaminated voice data, and converts the contaminated voice data into a time domain using a regression neural network to generate a voice signal.

[0107] The memory (100) can store commands or data related to at least one other component of the voice data contamination device (10). In particular, the memory (100) can be implemented as a non-volatile memory, a volatile memory, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), etc. The memory (100) is accessed by the processor (200), and data reading / writing / modifying / deleting / updating, etc. can be performed by the processor (200). In the present disclosure, the term "memory" can include the memory (100), the processor (200), a ROM (not shown), a RAM (not shown), or a memory card (not shown) (e.g., a micro SD card, a memory stick) mounted on the voice data contamination device (10). In addition, the memory (100) can store programs and data for configuring various screens to be displayed on the display (120) area.

[0108] Additionally, the memory (100) can store an artificial intelligence agent for generating summary information, and, according to one embodiment of the present disclosure, can store an artificial intelligence learning model (i.e., a document summary model). Meanwhile, according to another embodiment, the artificial intelligence learning model can be stored in another electronic device.

[0109] Additionally, the memory (100) can store at least some of the various modules included in the processor.

[0110] The communication unit (300) by the start of work enables communication such as data transmission and reception between the voice data contamination device (10) and the external device.

[0111] The processor (200) by the start of work can obtain the speech recognition error rate of the contaminated data by using at least one of CER and WER.

[0112] CER measures the error rate at the character level, and is calculated by dividing the number of incorrectly recognized characters by the total number of characters. CER = (number of incorrectly recognized characters) / (total number of characters)

[0113] For example, if the word "HELLO" is recognized as "HEXYO", there are two error characters ("L" is incorrectly recognized as "X" and "O" is incorrectly recognized as "Y"), and there are five characters in total, so the CER is 0.4.

[0114] WER measures the error rate at the word level and is calculated by dividing the number of incorrectly recognized words by the total number of words. WER = (Number of incorrectly recognized words) / (Total number of words)

[0115] For example, if the sentence "Hello world" is recognized as "Hello there", there is 1 error word and 2 total words, so WER = 0.5.

[0116] Both CER and WER indicate that a higher error rate indicates lower system performance, and effective model construction and data-based learning are required to lower these values.

[0117] The processor (200) by the start of work can evaluate the performance of the voice recognition system according to the following criteria.

[0118] Test: Each test is run by testing 300 cases of 10 seconds of situation-specific second-by-second voice data.

[0119] Applying pollution techniques: Apply various pollution techniques to the original speech data and convert it into spoken text. This is where the recognition accuracy is evaluated.

[0120] Error rate: The higher the error rate, the less accurate the speech recognition, and the more successful the contamination technique is considered to be.

[0121] Furthermore, pure neural network performance evaluation is based on categorical cross-entropy of output nodes for learning and evaluation, without performing additional backend processing such as beam search. Alphabet-based character units are used as units. This means the model focuses on recognizing individual characters.

[0122] FIG. 7 is a diagram for explaining technical features of a blockchain platform that distributes voice data according to one embodiment of the present invention.

[0123] According to one embodiment of the present invention, in block 901, the voice data protection device (10) can form a blockchain network based on a smart contract between a seller and a consumer.

[0124] Among the terms used in this specification, "block" refers to small data created in a P2P manner, linked to other blocks, and executed as a single data storage environment. Here, one block contains information about the time the block was created and the previously created blocks, and is connected like a chain, so it is called a block chain.

[0125] That is, the 1000th block created contains all the information from before the 1000th block was created and after the 999th block was created, as well as the information included in blocks 1 to 999.

[0126] According to one embodiment of the present invention, in block 902, the voice data protection device (10) can create a blockchain wallet in which a secret key is set between the seller and the consumer.

[0127] According to one embodiment of the present invention, in block 903, the voice data protection device (10) can register voice data including noise data in a blockchain network.

[0128] The voice data protection device (10) is configured to include a plurality of participating nodes (Peer #1, #2, #3, #n) of a blockchain network, and a blockchain processing application equipped in each participating node creates a transaction in which data is recorded, and creates a block including the transaction according to a preset consensus algorithm and stores it in the blockchain of each participating node.

[0129] In the system according to the present invention, when information on voice data including noise data is input, the blockchain processing application equipped on any participating node generates a transaction including the data, creates and stores a file block including the transaction according to a preset consensus algorithm, and stores this stored copy in the blockchain of each peer node. The forgery and tampering of the stored files can be confirmed by locating each file in the leaf node of the Merkle Tree using the Merkle Tree, and checking only the root node of the Merkle Tree. This series of processes is managed by the Distributed Hash Table (DHT), where a unique ID is assigned to each file and information on the original file is stored. Basically, the blockchain processing application includes a File Broadcast or Block Generate function, and also includes a DATA I / O API function for receiving a request from a file inquiry application and transmitting a text file according to the request as an output.

[0130] Meanwhile, the file search application is an application installed on the user's terminal, etc., and used directly by the user. When the user requests a voice data file, it requests the necessary information from the blockchain or data storage device, and retrieves the voice data file from the blockchain or provides the original file from the data storage device.

[0131] Meanwhile, the voice data protection device (10) having the aforementioned configuration stores voice data files in the form of text files, making them very convenient for inquiry and inspection. Furthermore, the voice data protection device (10) according to the present invention records and stores information related to smart contracts on a blockchain, thereby allowing the range of noise data in voice data files to be changed and provided based on whether the information is related to the purchaser set in the user's smart contract.

[0132] According to one embodiment of the present invention, at block 904, the voice data protection device (10) can transmit voice data including noise data to the consumer as payment from the consumer is confirmed.

[0133] According to one embodiment of the present invention, when voice data is downloaded by a third party other than a consumer through a smart contract, a step of lowering the quality of the voice data by expanding the range of noise data included in the voice data may be included.

[0134] According to one embodiment of the present invention, a method for detecting errors in a smart contract involves conducting a thorough security review using various techniques and tools prior to the creation, review, and deployment of the smart contract. For example, errors in smart contracts can be detected using methods such as code review, testing, static analysis, formal verification, bug bounty programs, following best practices, using linting tools, and regular updates and audits.

[0135]

[0136] The steps of a method or algorithm described in connection with an embodiment of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in a random access memory (RAM), a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a hard disk, a removable disk, a CD-ROM, or any other form of computer-readable recording medium well known in the art to which the present invention pertains.

[0137] The components of the present invention may be implemented as programs (or applications) and stored on a medium to be executed in conjunction with a computer, which is hardware. The components of the present invention may be implemented as software programs or software elements. Similarly, the embodiments may be implemented in a programming or scripting language such as C, C++, Java, or an assembler, including various algorithms implemented as a combination of data structures, processes, routines, or other programming components. Functional aspects may be implemented as algorithms that are executed on one or more processors.

[0138] While the embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will appreciate that the present invention can be implemented in other specific forms without altering the technical concept or essential features thereof. Therefore, the embodiments described above should be understood to be illustrative in all respects and not restrictive.

Claims

1. Step of preprocessing voice data; A step of adding contamination noise in the process of upsampling the above preprocessed voice data in the frequency domain; A step of generating polluted voice data by inputting the voice data with the above polluted noise added into a deep learning model; and A method for polluting speech data in real time, comprising the step of converting the polluted speech data into a time domain using a recurrent neural network to generate a speech signal.

2. In paragraph 1, A method for polluting voice data in real time, further comprising the step of generating pollution noise to be added to the voice data, taking into account the frequency range and intensity of the voice data.

3. In paragraph 1, The above pollution noise is, A method for polluting voice data in real time, characterized in that the hearing range of 20 Hz to 20,000 Hz is not included.

4. In paragraph 1, The step of generating the above contaminated voice data is: A step of setting the above-mentioned contaminated voice data as a predictor variable and setting the original voice data as a target signal; and A method for polluting voice data in real time, comprising a step of generating the polluted voice data using the deep learning model according to the above-described conditions.

5. Memory that stores computer executable instructions; and A device for polluting voice data in real time, comprising a processor that preprocesses voice data by executing the computer-executable instructions, adds pollution noise in the process of upsampling the preprocessed voice data in the frequency domain, inputs the voice data with the added pollution noise into a deep learning model to generate polluted voice data, and converts the polluted voice data into the time domain using a recurrent neural network to generate a voice signal.

Images