Feature recognition method and system based on deep learning

Abstract

The invention discloses a feature recognition method and system based on deep learning, and the method comprises the steps: calling to-be-recognized voice data from a database, and obtaining the to-be-recognized voice data; the method comprises the following steps: calling to-be-recognized voice data from a database, denoising the data, establishing a character recognition model, inputting the denoised to-be-recognized voice data, automatically recognizing the pronunciation of each character in the denoised voice data, translating the pronunciation into the character, obtaining the text data, proofreading the text data, and obtaining the proofreading voice text data. Privacy analysis and privacy scoring are carried out on the voice text data, the privacy score is compared with a preset threshold value, whether the voice text data affect enterprise development or not is judged, if the voice text data affect enterprise development, original voice data are encrypted and stored in a database, and in order to reserve evidences, who is a bulletin can be known. And the data is prevented from being damaged by external invasion, so that the system meets the development requirements of enterprises.

Description

Technical Field

[0001] The present invention relates to the field of feature recognition technology, and in particular to a feature recognition method and system based on deep learning. Background Art

[0002] With the development of mobile Internet, the device processing power and network bandwidth of mobile terminals have been greatly improved. Using the network to transmit voice or video data, Internet voice calls or video calls and mobile phone calls have become a common way of communication.

[0003] At present, in the process of enterprise development, if the voice data generated by Internet voice calls, video calls and mobile phone calls causes the rival company to learn the important confidential information of the enterprise, the development of the enterprise will be hindered and it will be difficult to meet the development needs of the enterprise. At present, there is a lack of a method that can monitor the voice data of the enterprise in real time and extract and analyze it to determine whether the voice data is related to the private data of the enterprise. Summary of the invention

[0004] In order to solve the above technical problems, a feature recognition method and system based on deep learning are provided. This technical solution solves the problem raised in the above background technology that in the process of enterprise development, if the voice data generated by network voice calls, video calls and mobile phone calls causes the rival company to learn the important confidential information of the enterprise, it will cause the development of the enterprise to be hindered.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A feature recognition method based on deep learning, comprising:

[0007] Retrieving the voice data to be recognized from the database to obtain the voice data to be recognized;

[0008] Performing denoising on the speech data to be recognized to obtain denoised speech data to be recognized;

[0009] Retrieving the pinyin letter string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin letter string information of each character to obtain initial consonant features and final consonant features, and performing feature extraction on the tone mark information of each character to obtain tone mark features;

[0010] Based on the initial consonant features, final consonant features and tone marker features, a character recognition model is constructed;

[0011] Input the denoised speech data to be recognized into the text recognition model to obtain speech text data;

[0012] Conduct privacy analysis on voice and text data to obtain privacy scores;

[0013] Based on the privacy score of the voice and text data, a judgment operation is performed on the original voice data and the voice and text data.

[0014] Preferably, the step of retrieving the voice data to be recognized from the database comprises the following steps:

[0015] The system automatically marks the recognized voice data with specific labels and obtains specific label information;

[0016] Traverse and process all speech data in the database based on specific label information;

[0017] The voice data containing the specific label information is the recognized voice data;

[0018] The speech data not containing the specific label information is the speech data to be recognized, and the speech data to be recognized is obtained.

[0019] Preferably, the denoising process is performed on the speech data to be recognized to obtain the denoised speech data to be recognized, comprising the following steps:

[0020] Extract the speech data to be recognized and obtain the speech signal;

[0021] Perform short-time Fourier transform on the speech signal to convert the speech signal into the frequency domain;

[0022] In the frequency domain, filters are used to remove high-frequency noise components and retain the main speech signal;

[0023] Acquiring parameter information of original voice data, wherein the parameter information includes sampling information, quantization information and coding information;

[0024] The main speech signal is converted using the same parameter information to obtain denoised speech data to be recognized.

[0025] Preferably, the step of retrieving the pinyin string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin string information of each character, obtaining initial consonant features and final consonant features, performing feature extraction on the tone mark information of each character, and obtaining tone mark features specifically comprises the following steps:

[0026] Get the pinyin string information of each word in the dictionary;

[0027] Based on the initial consonant determination rule and the final consonant determination rule, extract the initial consonant and final consonant features from the pinyin letter string information to obtain the initial consonant features and the final consonant features, wherein the initial consonant determination rule includes single consonant initial consonant determination and complex consonant initial consonant determination, and the final consonant determination rule includes single vowel determination and complex vowel determination;

[0028] The pinyin tone mark information of each character in the dictionary is obtained, and features of the pinyin tone mark information are extracted to obtain tone mark features.

[0029] Preferably, the character recognition model is constructed based on the initial consonant feature, the final consonant feature and the tone mark feature, and specifically comprises the following steps:

[0030] Based on a feedforward neural network, a text recognition model is constructed, wherein the input data of the text recognition model includes denoised speech data to be recognized, and the output data is speech text data;

[0031] Based on machine learning, the multiple sample initial consonant features, the multiple sample final consonant features, and the multiple sample tone marker features are labeled and divided to obtain a training set, a validation set, and a test set;

[0032] The training set, validation set and test set are used to perform supervised training, validation and testing on the text recognition model to obtain the text recognition model whose accuracy meets the preset accuracy requirement.

[0033] Preferably, the step of inputting the denoised speech data to be recognized into a text recognition model to obtain speech text data specifically comprises the following steps:

[0034] Inputting the denoised speech data to be recognized into the text recognition model that meets the preset accuracy requirement;

[0035] The text recognition model automatically recognizes the pronunciation of each word in the denoised speech data and translates it into text to obtain text data;

[0036] The word grouping information of the dictionary in the database is retrieved, and the adjacent words in the text data are analyzed and corrected to obtain the proofread voice text data.

[0037] Preferably, the privacy analysis of the voice and text data to obtain a privacy score specifically includes the following steps:

[0038] Get the private word model in the database;

[0039] Perform weight analysis on the private word data in the private word model to obtain the private word weight;

[0040] Segment the speech text data to obtain the speech text data of each segment;

[0041] The private words in the private word model are traversed through each segmented speech text data to obtain each segmented speech text data containing the private words;

[0042] Based on the private word weight, perform a private word sum operation on each segmented speech text data containing the private word to obtain the private word weight of each segmented speech text data;

[0043] The private word weights of each segmented voice text data are summed up again to obtain the total private word weight of the voice text data, and the total private word weight of the voice text data is the privacy score of the voice text data.

[0044] Preferably, the privacy scoring based on the voice and text data and the determination operation on the original voice data and the voice and text data specifically include the following steps:

[0045] Obtaining a privacy score for the voice and text data, comparing the privacy score for the voice and text data with a preset threshold, and obtaining a judgment operation;

[0046] If the privacy score of the voice and text data is less than the preset threshold, the judgment operation is to destroy both the original voice data and the voice and text data;

[0047] If the privacy score of the voice text data is greater than a preset threshold, the judgment operation is to encrypt the original voice data and destroy the voice text data.

[0048] Preferably, the encryption processing of the original voice data specifically includes the following:

[0049] Determine the data bit length of the original voice data and obtain the length of the data sequence;

[0050] Separating the data based on the length of the data sequence to obtain a plurality of subsequence data of a first preset length and a plurality of subsequence data of a second preset length, wherein the first preset length is smaller than the second preset length;

[0051] Performing multi-directional scrambling operations on a plurality of sub-sequence data of a first preset length and a plurality of sub-sequence data of a second preset length, respectively, to generate a plurality of multi-directional scrambled sub-data;

[0052] A plurality of multi-directional garbled sub-data generated from a plurality of sub-sequence data of a first preset length and a plurality of multi-directional garbled sub-data generated from a plurality of sub-sequence data of a second preset length are irregularly arranged and combined to obtain encrypted data.

[0053] The present invention also proposes a feature recognition system based on deep learning, comprising:

[0054] A module for acquiring voice data to be recognized, wherein the module is used to retrieve voice data to be recognized from a database to obtain the voice data to be recognized;

[0055] A denoising module, the denoising module is used to perform denoising processing on the speech data to be recognized, and obtain the denoised speech data to be recognized;

[0056] A feature extraction module, the feature extraction module is used to retrieve the pinyin letter string information and tone mark information of each character in the dictionary from the database, perform feature extraction on the pinyin letter string information of each character to obtain initial consonant features and final vowel features, perform feature extraction on the tone mark information of each character to obtain tone mark features;

[0057] A character recognition model construction module, wherein the character recognition model construction module is used to construct a character recognition model based on initial consonant features, final consonant features, and tone mark features;

[0058] A voice text data acquisition module, which is used to input the denoised voice data to be recognized into a text recognition model to obtain voice text data;

[0059] A privacy score acquisition module, wherein the privacy score acquisition module is used to perform privacy analysis on the voice text data to obtain a privacy score;

[0060] The judgment operation module is used to perform judgment operations on the original voice data and the voice text data based on the privacy score of the voice text data.

[0061] Compared with the prior art, the present invention provides a feature recognition method and system based on deep learning, which has the following beneficial effects:

[0062] The present invention retrieves voice data to be recognized from a database, denoises the data, establishes a text recognition model, inputs the denoised voice data to be recognized, automatically recognizes the pronunciation of each character in the denoised voice data and translates it into text, obtains text data and proofreads it, obtains the proofread voice text data, performs privacy analysis on the voice text data and performs privacy scoring, compares the privacy score with a preset threshold, determines whether the voice text data has an impact on the development of the enterprise, and if so, encrypts the original voice data and stores it in the database. In order to preserve evidence, it can be known who the informer is, and external intrusion and destruction of the data are prevented, so that the system meets the development needs of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the method flow of feature recognition in the present invention;

[0064] Figure 2 A schematic diagram of the flow of a method for obtaining speech data to be recognized in the present invention;

[0065] Figure 3A schematic diagram of a method flow for obtaining denoised speech data to be recognized in the present invention;

[0066] Figure 4 A schematic diagram of the flow of the method for obtaining tone marking features in the present invention;

[0067] Figure 5 It is a flow chart of the method for forming a text recognition model in the present invention;

[0068] Figure 6 A schematic diagram of the flow of the method for obtaining speech text data in the present invention;

[0069] Figure 7 A flowchart of a method for obtaining a privacy score in the present invention;

[0070] Figure 8 It is a flow chart of the method for performing determination operation on original voice data and voice text data in the present invention;

[0071] Fig. 9 The figure is a flow chart of the method for encrypting original voice data in the present invention. DETAILED DESCRIPTION

[0072] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0073] Please refer to Figure 1-9 As shown, a feature recognition method based on deep learning includes:

[0074] Retrieving the voice data to be recognized from the database to obtain the voice data to be recognized;

[0075] Performing denoising on the speech data to be recognized to obtain denoised speech data to be recognized;

[0076] Retrieving the pinyin letter string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin letter string information of each character to obtain initial consonant features and final consonant features, and performing feature extraction on the tone mark information of each character to obtain tone mark features;

[0077] Based on the initial consonant features, final consonant features and tone marker features, a character recognition model is constructed;

[0078] Input the denoised speech data to be recognized into the text recognition model to obtain speech text data;

[0079] Conduct privacy analysis on voice and text data to obtain privacy scores;

[0080] Based on the privacy score of the voice and text data, a judgment operation is performed on the original voice data and the voice and text data.

[0081] It can be understood by those skilled in the art that telephone conversations of internal employees of an enterprise are recorded, where the telephone conversations include mobile phone calls, voice calls, video calls, etc. All recorded data are stored in a database, and voice data to be recognized is extracted from the database. Since there is noise interference in telephone recordings, in order not to affect subsequent data processing, the voice data to be recognized needs to be denoised. Since each character has a pinyin letter string and a tone corresponding to the pinyin, there are four tones, and the pinyin letter string is composed of initial consonants and finals, the initial consonant features, final consonant features and tone mark features are analyzed in the dictionary, and a character recognition model can be established. The character recognition model can simulate the pronunciation of the character, so that the denoised voice data to be recognized is input into the character recognition model, and each character in the voice data can be translated, thereby forming corresponding voice text data; further, the voice text data is analyzed to see whether the characters therein involve confidential privacy messages of the enterprise, and then a judgment operation is performed on the data.

[0082] Retrieving the voice data to be recognized from the database, obtaining the voice data to be recognized specifically includes the following steps:

[0083] The system automatically marks the recognized voice data with specific labels and obtains specific label information;

[0084] Traverse and process all speech data in the database based on specific label information;

[0085] The voice data containing the specific label information is the recognized voice data;

[0086] The speech data not containing the specific label information is the speech data to be recognized, and the speech data to be recognized is obtained.

[0087] It can be understood by those skilled in the art that there are specific labels in the system, such as "recognized". When data is recognized, it is automatically marked with "recognized". The database stores recognized data and data to be recognized. "Recognized" is used to traverse all data in the entire database. Data containing "recognized" must be recognized voice data, and data without "recognized" must be voice data to be recognized.

[0088] Performing denoising on the speech data to be recognized and obtaining the denoised speech data to be recognized includes the following steps:

[0089] Extract the speech data to be recognized and obtain the speech signal;

[0090] Perform short-time Fourier transform on the speech signal to convert the speech signal into the frequency domain;

[0091] In the frequency domain, filters are used to remove high-frequency noise components and retain the main speech signal;

[0092] Obtaining parameter information of original voice data, the parameter information including sampling information, quantization information and coding information;

[0093] The main speech signal is converted using the same parameter information to obtain denoised speech data to be recognized.

[0094] It is understandable to those skilled in the art that denoising a speech signal can not only improve the clarity of the speech, but also reduce the size of the data and speed up the subsequent processing of the data by the system. The speech signal after noise removal also needs to be converted digitally. It is worth noting here that the parameter information of the original speech data is obtained and the same parameter information is used for equivalent conversion to ensure the authenticity of the data and avoid errors in subsequent data processing.

[0095] Retrieving the pinyin letter string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin letter string information of each character, obtaining initial consonant features and final consonant features, performing feature extraction on the tone mark information of each character, and obtaining tone mark features specifically includes the following steps:

[0096] Get the pinyin string information of each word in the dictionary;

[0097] Based on the initial consonant determination rules and the final consonant determination rules, the initial consonant and final consonant features are extracted from the pinyin letter string information to obtain the initial consonant features and the final consonant features. The initial consonant determination rules include single consonant initial consonant determination and complex consonant initial consonant determination, and the final consonant determination rules include single vowel determination and complex vowel determination;

[0098] The pinyin tone mark information of each character in the dictionary is obtained, and features of the pinyin tone mark information are extracted to obtain tone mark features.

[0099] It can be understood by those skilled in the art that the determination of single consonant initials includes: b, p, m, f, d, t, n, l, g, k, h, j, q, x, zh, ch, sh, r, z, c, s. When a syllable begins with any one of the above letters and does not belong to the finals, these letters will be regarded as initials; the determination of complex consonant initials includes: zh, ch, sh, r, z, c, s. When a syllable begins with any two or more of the above and does not belong to the finals, these letters will be regarded as initials. The determination of single vowels includes: a, o, e, i, u, ü. When a syllable begins with any one of the above vowels, these letters will be regarded as finals; the determination of complex vowels includes: ai, ei, ui, ao, ou, iu, ie, üe, er. When a syllable begins with any two or more of the above, these letters will be regarded as finals. There are four types of tone markers, including: first tone (ˉ), second tone (ˊ), third tone (ˇ) and fourth tone (ˋ).

[0100] Based on the initial consonant features, final consonant features and tone mark features, the character recognition model is constructed in the following steps:

[0101] Based on the feedforward neural network, a text recognition model is constructed, the input data of the text recognition model includes the denoised speech data to be recognized, and the output data is the speech text data;

[0102] Based on machine learning, multiple sample initial consonant features, multiple sample final consonant features, and multiple sample tone marker features are labeled and divided to obtain a training set, a validation set, and a test set;

[0103] The training set, validation set and test set are used to perform supervised training, validation and testing on the text recognition model to obtain a text recognition model whose accuracy meets the preset accuracy requirements.

[0104] It can be understood by those skilled in the art that a feedforward neural network is a type of artificial neural network that uses a unidirectional multi-layer structure, in which the 0th layer is called the input layer, the last layer is called the output layer, and the other intermediate layers are called hidden layers. The hidden layers can be one layer or multiple layers. Each hidden layer contains a number of neurons, each neuron can receive signals from the neurons in the previous layer and generate outputs to the next layer. There is no feedback in the entire network, and the signal propagates unidirectionally from the input layer to the output layer. The features of multiple sample initial consonants, multiple sample final consonants and multiple sample tone markers are used as construction data, and the construction data is randomly divided into a training set, a validation set and a test set according to a certain ratio, and the divided training set, validation set and test set are annotated. For example, the preferred ratio of this application is 8:1:1, wherein the training set is used to train the model, that is, to determine the learning parameters such as the weight and bias of the model; the validation set is used to verify each model using the validation set data after training multiple models in the training set, and record the model accuracy, and then select the best model and its corresponding parameters; the test set is used only once, that is, when evaluating the final model after the training is completed, and it is neither involved in the process of learning parameters nor in the process of selecting hyperparameters, but only used for the evaluation of the model. Based on machine learning, the model is supervised and trained through the training set, the obtained model is verified through the validation set, and the trained model is evaluated through the test set, and the hyperparameters are continuously adjusted until the accuracy meets the preset requirements, and the speech text data is output.

[0105] Inputting the denoised speech data to be recognized into the text recognition model, and obtaining the speech text data specifically includes the following steps:

[0106] Inputting the denoised speech data to be recognized into a text recognition model that meets the preset accuracy requirements;

[0107] The text recognition model automatically recognizes the pronunciation of each word in the denoised speech data and translates it into text to obtain text data;

[0108] The word grouping information of the dictionary in the database is retrieved, and the adjacent words in the text data are analyzed and corrected to obtain the proofread voice text data.

[0109] People skilled in the art can understand that the "dictionary" in this article is not only the Xinhua Dictionary commonly used in life, but also includes some popular Internet vocabulary in today's society, some professional terms within the company, etc., covering a wide range. Because the same pronunciation can be output in different fonts, and adjacent characters can form different words, it is necessary to compare and analyze the words, which improves the accuracy of the translated words and also improves the accuracy of subsequent privacy analysis.

[0110] The privacy analysis of voice and text data and obtaining the privacy score specifically include the following steps:

[0111] Get the private word model in the database;

[0112] Perform weight analysis on the private word data in the private word model to obtain the private word weight;

[0113] Segment the speech text data to obtain the speech text data of each segment;

[0114] The private words in the private word model are traversed through each segmented speech text data to obtain each segmented speech text data containing the private words;

[0115] Based on the private word weight, perform a private word sum operation on each segmented speech text data containing the private word to obtain the private word weight of each segmented speech text data;

[0116] The private word weights of each segment of the speech text data are summed up again to obtain the total private word weight of the speech text data, and the total private word weight of the speech text data is used as the privacy score of the speech text data.

[0117] It can be understood by those skilled in the art that different private word data correspond to different stages of development of an enterprise, such as project private word data at different stages, private word data of competitor companies at different stages, etc. These private word data constitute a private word model. Different private word data have different weights and different impacts on the enterprise. For example, when a competitor asks about the bidding situation of an enterprise, the weight is larger and the impact on the development of the enterprise is also larger. Therefore, the voice text data is segmented and processed, and then the private words in the private word model are traversed through each segment of voice text data. In this way, the size of each segment of voice text data becomes smaller, which improves the speed of data processing. In this way, the segmented voice text data containing private words are obtained. For example, a voice text data is divided into three segments of voice text data, where the first segment contains a secret word with a weight of 50, so the private word weight of the first segment is 50, the second segment contains two secret words with weights of 50 and 60 respectively, so the private word weight of the second segment is 50+60, which is 110, and the third segment contains three secret words with weights of 50, 60 and 70 respectively, so the private word weight of the third segment is 50+60+70, which is 180. Finally, the total weight of the private word is 50+110+180, which is 340, and 340 is the privacy score.

[0118] Based on the privacy scoring of voice and text data, the judgment operation on the original voice data and voice and text data specifically includes the following steps:

[0119] Obtaining a privacy score for the voice and text data, comparing the privacy score for the voice and text data with a preset threshold, and obtaining a judgment operation;

[0120] If the privacy score of the voice and text data is less than the preset threshold, the judgment operation is to destroy both the original voice data and the voice and text data;

[0121] If the privacy score of the voice text data is greater than the preset threshold, the judgment operation is to encrypt the original voice data and destroy the voice text data.

[0122] It can be understood by those skilled in the art that if the privacy score of the voice text data is less than a preset threshold, it means that the privacy of the voice text data is low and there is no need for it to exist in the database. The original voice data and voice text data can be directly destroyed, thereby expanding the storage space of the database. If the privacy score is greater than the preset threshold, it means that the privacy is greater and the data cannot be known to the outside world. The voice text data is destroyed, and the storage space of the database is further expanded. The original voice data is encrypted and stored in the database. If this voice data causes a rival company to learn of important secrets of the company and subsequently causes risks to the company, the original voice data is stored in the database. In order to preserve evidence and find out who the informant is, the original voice data is encrypted, which further improves the security of the data and prevents outside intrusion and destruction of the data.

[0123] The encryption process of the original voice data specifically includes the following:

[0124] Determine the data bit length of the original voice data and obtain the length of the data sequence;

[0125] Separating the data based on the length of the data sequence to obtain a plurality of subsequence data of a first preset length and a plurality of subsequence data of a second preset length, wherein the first preset length is less than the second preset length;

[0126] Performing a first multi-directional scrambling operation on a plurality of sub-sequence data of a first preset length respectively to generate a plurality of first multi-directional scrambled sub-data;

[0127] Performing a second multi-directional scrambling operation on a plurality of sub-sequence data of a second preset length respectively to generate a plurality of second multi-directional scrambled sub-data;

[0128] Based on a plurality of first multi-directional garbled sub-data and a plurality of second multi-directional garbled sub-data, irregular arrangements and combinations are performed to obtain encrypted data.

[0129] It can be understood by those skilled in the art that the data is divided into a plurality of subsequence data of a first preset length and a plurality of subsequence data of a second preset length, and the plurality of subsequence data of the first preset length and the plurality of subsequence data of the second preset length are respectively and sequentially subjected to multi-directional scrambling encryption, and then are subsequently irregularly arranged and combined, the values ​​of the initially separated first preset length and the second preset length are both random, and the multi-directional scrambling is also random, and the irregular arrangement and combination is also random, thereby encrypting the data and ensuring the security of the data, and preventing external intrusion from violently damaging the data.

[0130] The present invention also proposes a feature recognition system based on deep learning, comprising:

[0131] The module for acquiring the voice data to be recognized is used to retrieve the voice data to be recognized from the database to obtain the voice data to be recognized;

[0132] A denoising module, which is used to perform denoising on the speech data to be recognized, and obtain the denoised speech data to be recognized;

[0133] The feature extraction module is used to retrieve the pinyin letter string information and tone mark information of each character in the dictionary from the database, perform feature extraction on the pinyin letter string information of each character to obtain the initial consonant feature and the final consonant feature, perform feature extraction on the tone mark information of each character to obtain the tone mark feature;

[0134] A character recognition model constituting module, wherein the character recognition model constituting module is used to construct a character recognition model based on initial consonant features, final consonant features and tone mark features;

[0135] A voice text data acquisition module, which is used to input the denoised voice data to be recognized into the text recognition model to obtain voice text data;

[0136] A privacy score acquisition module is used to perform privacy analysis on voice and text data and obtain a privacy score;

[0137] The judgment operation module is used to perform judgment operations on the original voice data and the voice text data based on the privacy score of the voice text data.

[0138] In summary, the present invention retrieves the voice data to be recognized from the database, denoises the data, establishes a text recognition model, inputs the denoised voice data to be recognized, automatically recognizes the pronunciation of each character in the denoised voice data and translates it into text, obtains the text data and proofreads it, obtains the proofread voice text data, performs privacy analysis on the voice text data and performs a privacy score, compares the privacy score with a preset threshold, determines whether the voice text data has an impact on the development of the enterprise, and if so, encrypts the original voice data and stores it in the database. In order to preserve evidence, it can be known who the informant is, and external intrusion and destruction of the data are prevented, so that the system meets the development needs of the enterprise.

[0139] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A feature recognition method based on deep learning, characterized in that: include: Retrieving the voice data to be recognized from the database to obtain the voice data to be recognized; Performing denoising on the speech data to be recognized to obtain denoised speech data to be recognized; Retrieving the pinyin letter string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin letter string information of each character to obtain initial consonant features and final consonant features, and performing feature extraction on the tone mark information of each character to obtain tone mark features; Based on the initial consonant features, final consonant features and tone marker features, a character recognition model is constructed; Input the denoised speech data to be recognized into the text recognition model to obtain speech text data; Conduct privacy analysis on voice and text data to obtain privacy scores; Based on the privacy score of the voice and text data, a judgment operation is performed on the original voice data and the voice and text data.

2. A feature recognition method based on deep learning according to claim 1, characterized in that: The method of retrieving the voice data to be recognized from the database specifically comprises the following steps: The system automatically marks the recognized voice data with specific labels and obtains specific label information; Traverse and process all speech data in the database based on specific label information; The voice data containing the specific label information is the recognized voice data; The speech data not containing the specific label information is the speech data to be recognized, and the speech data to be recognized is obtained.

3. A feature recognition method based on deep learning according to claim 2, characterized in that: The denoising process is performed on the speech data to be recognized to obtain the denoised speech data to be recognized, and the denoised speech data to be recognized comprises the following steps: Extract the speech data to be recognized and obtain the speech signal; Perform short-time Fourier transform on the speech signal to convert the speech signal into the frequency domain; In the frequency domain, filters are used to remove high-frequency noise components and retain the main speech signal; Acquiring parameter information of original voice data, wherein the parameter information includes sampling information, quantization information and coding information; The main speech signal is converted using the same parameter information to obtain denoised speech data to be recognized.

4. A feature recognition method based on deep learning according to claim 3, characterized in that: The method of retrieving the pinyin letter string information and tone mark information of each character in the dictionary from the database, performing feature extraction on the pinyin letter string information of each character, obtaining initial consonant features and final consonant features, performing feature extraction on the tone mark information of each character, and obtaining tone mark features specifically comprises the following steps: Get the pinyin string information of each word in the dictionary; Based on the initial consonant determination rule and the final consonant determination rule, extract the initial consonant and final consonant features from the pinyin letter string information to obtain the initial consonant features and the final consonant features, wherein the initial consonant determination rule includes single consonant initial consonant determination and complex consonant initial consonant determination, and the final consonant determination rule includes single vowel determination and complex vowel determination; The pinyin tone mark information of each character in the dictionary is obtained, and features of the pinyin tone mark information are extracted to obtain tone mark features.

5. A feature recognition method based on deep learning according to claim 4, characterized in that: The character recognition model based on the initial consonant feature, the final consonant feature and the tone mark feature specifically comprises the following steps: Based on a feedforward neural network, a text recognition model is constructed, wherein the input data of the text recognition model includes denoised speech data to be recognized, and the output data is speech text data; Based on machine learning, the multiple sample initial consonant features, the multiple sample final consonant features, and the multiple sample tone marker features are labeled and divided to obtain a training set, a validation set, and a test set; The training set, validation set and test set are used to perform supervised training, validation and testing on the text recognition model to obtain the text recognition model whose accuracy meets the preset accuracy requirement.

6. A feature recognition method based on deep learning according to claim 5, characterized in that: The step of inputting the denoised speech data to be recognized into the text recognition model to obtain the speech text data specifically includes the following steps: Inputting the denoised speech data to be recognized into the text recognition model that meets the preset accuracy requirement; The text recognition model automatically recognizes the pronunciation of each word in the denoised speech data and translates it into text to obtain text data; The word grouping information of the dictionary in the database is retrieved, and the adjacent words in the text data are analyzed and corrected to obtain the proofread voice text data.

7. A feature recognition method based on deep learning according to claim 6, characterized in that: The privacy analysis of the voice and text data to obtain the privacy score specifically includes the following steps: Get the private word model in the database; Perform weight analysis on the private word data in the private word model to obtain the private word weight; Segment the speech text data to obtain the speech text data of each segment; The private words in the private word model are traversed through each segmented speech text data to obtain each segmented speech text data containing the private words; Based on the private word weight, perform a private word sum operation on each segmented speech text data containing the private word to obtain the private word weight of each segmented speech text data; The private word weights of each segmented voice text data are summed up again to obtain the total private word weight of the voice text data, and the total private word weight of the voice text data is the privacy score of the voice text data.

8. The feature recognition method based on deep learning according to claim 7, characterized in that: The privacy scoring based on the voice and text data and the determination operation on the original voice data and the voice and text data specifically include the following steps: Obtaining a privacy score for the voice and text data, comparing the privacy score for the voice and text data with a preset threshold, and obtaining a judgment operation; If the privacy score of the voice and text data is less than the preset threshold, the judgment operation is to destroy both the original voice data and the voice and text data; If the privacy score of the voice text data is greater than a preset threshold, the judgment operation is to encrypt the original voice data and destroy the voice text data.

9. A feature recognition method based on deep learning according to claim 8, characterized in that: The encryption process of the original voice data specifically includes the following steps: Determine the data bit length of the original voice data and obtain the length of the data sequence; Separating the data based on the length of the data sequence to obtain a plurality of subsequence data of a first preset length and a plurality of subsequence data of a second preset length, wherein the first preset length is smaller than the second preset length; Performing multi-directional scrambling operations on a plurality of sub-sequence data of a first preset length and a plurality of sub-sequence data of a second preset length, respectively, to generate a plurality of multi-directional scrambled sub-data; A plurality of multi-directional garbled sub-data generated from a plurality of sub-sequence data of a first preset length and a plurality of multi-directional garbled sub-data generated from a plurality of sub-sequence data of a second preset length are irregularly arranged and combined to obtain encrypted data.

10. A feature recognition system based on deep learning, used to implement a feature recognition method based on deep learning as described in claims 1-9, characterized in that: include: A module for acquiring voice data to be recognized, wherein the module is used to retrieve voice data to be recognized from a database to obtain the voice data to be recognized; A denoising module, the denoising module is used to perform denoising processing on the speech data to be recognized, and obtain the denoised speech data to be recognized; A feature extraction module, the feature extraction module is used to retrieve the pinyin letter string information and tone mark information of each character in the dictionary from the database, perform feature extraction on the pinyin letter string information of each character to obtain initial consonant features and final vowel features, perform feature extraction on the tone mark information of each character to obtain tone mark features; A character recognition model construction module, wherein the character recognition model construction module is used to construct a character recognition model based on initial consonant features, final consonant features, and tone mark features; A voice text data acquisition module, which is used to input the denoised voice data to be recognized into a text recognition model to obtain voice text data; A privacy score acquisition module, wherein the privacy score acquisition module is used to perform privacy analysis on the voice text data to obtain a privacy score; The judgment operation module is used to perform judgment operations on the original voice data and the voice text data based on the privacy score of the voice text data.

Images