A self-learning dynamic voiceprint identity authentication method based on unknown corpus

Abstract

The present invention discloses a self-learning dynamic voiceprint identity authentication method based on unknown corpus. According to the registration corpus rule format, the general registration character string and a small amount of special registration character strings of the registrants are received and processed in each trained discriminator to build a registration voiceprint library. According to the verification corpus rule format, preprocessing is performed. After the precondition is passed, the digital string will obtain a first embedded code based on the text-related recognition technology; the special character string will obtain a second embedded code based on the text-independent recognition technology, so as to check whether it matches. The text-independent discriminator, the general character string and digital string text-related discriminator, and the special character string text-related voiceprint discriminator are updated according to factors such as the cumulative amount stored in the text-related database, the performance improvement degree, and the number of people who entered the corpus. The present invention has good performance, improves security based on unknown corpus, and automatically improves performance based on the dynamic increase of models and databases.

Description

Technical Field

[0001] The present invention belongs to the field of identity authentication using voiceprint technology, and in particular relates to a self-learning dynamic voiceprint identity authentication method based on unknown corpus. Background Art

[0002] Voiceprint recognition is a biometric recognition technology that uses human voiceprint features for identity authentication. It has the advantages of low acquisition cost, simple implementation, and good experience. More importantly, voiceprint features are the only low-sensitivity information among all biometric features. Compared with other biometric recognition technologies such as face recognition and fingerprint recognition, voiceprint recognition involves less personal privacy information and is more easily accepted by users. Voiceprint recognition can be divided into three types according to the recognition content: text-independent, text-related, and text-prompt. However, the software based on these three recognition schemes currently lacks anti-fraud functions, making it possible for imposters to forge voices close to the target speaker through various means, thereby defrauding the voiceprint recognition system. For example, there are such security risks in the voice locks launched by WeChat and Alipay. For text-related voiceprint recognition, the commonly used fixed verification code is easily defrauded by recording, that is, the imposter plays the recorded voice of the target user to pass the voiceprint verification. Although text-prompt voiceprint recognition has certain anti-fraud functions, it is greatly restricted by the text set. If the text set is too small, the security of the recognition system is not enough; if the text set is too large, the user needs to enter a large amount of audio, which reduces the experience. In addition, text-prompt voiceprint recognition lacks the function of anti-recording splicing fraud. For example, the pronunciation of users' fixed input numbers 0-9 or letters AZ is collected, and the voice of the target speaker saying a single number or letter is cut out to form a text voice library. When performing voiceprint confirmation, a computer program can be used to quickly splice the pronunciation of the target verification code based on the text voice library.

[0003] Therefore, there is an urgent need to propose a technology that can improve the security of the voiceprint recognition system without involving more personal privacy and reducing user experience, so as to effectively solve the problem of poor security in identity authentication schemes based on voiceprint recognition during various software logins or financial payments. Summary of the invention

[0004] The technical purpose of the present invention is to provide a self-learning dynamic voiceprint identity authentication method based on unknown corpus to solve the technical problem of poor security of voiceprint recognition.

[0005] To solve the above problems, the technical solution of the present invention is:

[0006] A self-learning dynamic voiceprint identity authentication method based on unknown corpus, configured with a dynamic voiceprint identity authentication system, the method comprising:

[0007] A1: According to the registration corpus rule format, receive and process the general registration string and special registration string of the registrant in the preset discriminator to build a registration voiceprint library; wherein the preset discriminator includes 1 text-independent discriminator, 1 general string number string text-related discriminator and several special string text-related voiceprint discriminators;

[0008] A2: According to the verification corpus rule format, the input verification corpus is identified and binary segmented to obtain a general verification string and a special verification string, and then the general verification string is identified and compared by their respective speech recognition engines. If it passes, the general verification string is preprocessed and sent to the general string, digital string, and text-related discriminator, and the first embedded code of the general string, digital string, and text-related discriminator is obtained based on the text-related recognition technology; the special verification string is preprocessed, and according to whether there is a corresponding special string text-related voiceprint discriminator for the special verification string and whether there is at least one registered embedded code, it is sent to the text-independent discriminator or the special string text-related voiceprint discriminator, and the second embedded code of the text-independent discriminator or the text-related voiceprint discriminator is obtained based on the text-independent recognition technology;

[0009] A3: Based on the cumulative amount of general verification strings stored in the text-related database and the performance improvement degree that meet a certain threshold at the same time, the general string, digital string, and text-related discriminator is updated;

[0010] If the number of people who have entered the corpus and the number of corpora stored in the text-related database according to the special verification string meet a certain threshold at the same time, the text-related corpus and the special string text-related voiceprint discriminator are updated;

[0011] The text-independent discriminator is updated based on the cumulative amount of special verification strings stored in the text-independent corpus and the performance improvement degree that meet a certain threshold at the same time.

[0012] Further preferably, in the verification process of step A2, words that have not formed a special character string text-related voiceprint discriminator will also be used. After the amount of data accumulation and the degree of performance improvement meet a certain threshold, after data amplification and robust processing, a text-related corpus required for the training model of the special character string text-related voiceprint discriminator is obtained. The method of obtaining the text-related corpus is to update and increase the text-related corpus when verified by some verifiers, so that the number of people entering the corpus and the number of their corpus meet a certain threshold at the same time, thereby generating a corresponding new special character string text-related voiceprint discriminator after training and performance verification.

[0013] Specifically, step A1 includes:

[0014] S11: inputting the general registration string and the special registration string into the dynamic voiceprint identity authentication system, the general registration string and the special registration string are respectively distinguished and processed by the front-end audio features formed by the general string digital string text related discriminator and the special string text related voiceprint discriminator, and then stored in the text related corpus;

[0015] S12: Input verification corpus, perform binary segmentation on the verification corpus, obtain a general verification string and a special verification string, obtain a first embedded code through text-related recognition technology, the first embedded code can be extracted by a general string digital string text-related discriminator, obtain a second embedded code based on text-independent recognition technology, the second embedded code can be extracted by a special string text-related voiceprint discriminator.

[0016] Specifically, step A2 includes:

[0017] S2: According to the dynamic verification rules of the dynamic voiceprint identity authentication system, a dynamic verification corpus given by the dynamic voiceprint identity authentication system is input, and the dynamic verification corpus is binary segmented to obtain a general dynamic verification string and a special dynamic verification string, and they are respectively identified and compared by their respective speech recognition engines. If they pass, they are judged to be general dynamic verification strings, and then they are subjected to pre-processing methods such as speech speed test, rearrangement, and completion and reconstruction in sequence, and then enter step S3 for subsequent processing. If they are judged to be special dynamic verification strings, they are first judged by the speech recognition engine whether they are homophonic. If so, they enter step S4 for subsequent processing, wherein the general dynamic verification string consists of a number of numbers, and the special dynamic verification characters are Chinese words or English letters or English phrases;

[0018] S3: Send the digital string to the preset general character string digital string text-related discriminator for discrimination with the preset threshold. If the discrimination is successful, send the digital string to both the text-independent corpus and the text-related database for storage. Otherwise, it is considered as a verification failure and stops.

[0019] S4: Determine whether there is a corresponding special string text-related voiceprint discriminator for the special dynamic verification string, and whether there is at least one registered corpus of the special dynamic verification string to be verified in the text-related corpus. If both exist, proceed to step S5. If the corresponding special string text-related voiceprint discriminator does not exist or both do not exist, proceed to step S6. If the corresponding special string text-related voiceprint discriminator exists but is not activated and at least one registered corpus exists in the text-related corpus, proceed to step S7.

[0020] S5: Send the special dynamic verification string to the corresponding special string text-related voiceprint discriminator for similarity discrimination. If the similarity exceeds the preset threshold, the judgment is successful, and the special dynamic verification string is sent to the text-independent corpus and the text-related database for storage. Otherwise, it is considered as a verification failure and stops;

[0021] S6: Send the special dynamic verification string to the text-independent discriminator for similarity judgment. If the similarity exceeds the preset threshold, if the judgment is successful, it is sent to the text-independent corpus and the text-dependent corpus for storage. Otherwise, it is considered as a verification failure and stops;

[0022] S7: Send the special dynamic verification character string to the text-independent discriminator for similarity judgment. If the similarity exceeds the preset threshold and the judgment is successful, it will be sent to the text-independent corpus for storage, and the text-related corpus will be post-processed. The corresponding special character string text-related voiceprint discriminator is activated, and at least one registration corpus that can be judged by the special character string text-related voiceprint discriminator is added, and one registration code is added for extracting the registration code. Otherwise, it is deemed as a verification failure and stops.

[0023] Further preferably, in step S2, before segmenting the dynamic verification corpus, the following steps are also included:

[0024] B1: Perform environmental speech quality detection on the dynamic verification corpus to determine the environmental noise level. If the calculated signal-to-noise ratio is compared with a preset threshold, if it is lower than the preset threshold, proceed to step B2, otherwise proceed to step B3;

[0025] B2: prompts to input dynamic verification corpus again, and returns to step B1. The number of times to enter step B2 is N=N+1, N=0. If N>3, it is considered as verification failure and stops. The judgment condition of N>3 is adjustable;

[0026] B3: Clear N to zero and convert the dynamic verification corpus into a format not limited to wav for segmentation;

[0027] Furthermore, in step B1, noise addition processing can be performed on the dynamic verification corpus before the environmental speech quality detection.

[0028] Wherein, step S2 specifically includes the following steps:

[0029] S21: The dynamic verification string voice is digitally recognized by the voice recognition engine and judged by the speech speed detector. If the recognition is successful and the speech speed is normal, the process proceeds to step S22. Otherwise, the verification is deemed to have failed and the process stops.

[0030] S22: Completely segment the dynamic verification string to obtain a number of digital segments containing only one number;

[0031] S23: rearrange the arrangement order of the plurality of digital fields based on the general registration character string;

[0032] S24: Adjust and complete the missing corpus according to the prior rules, and complete the arranged digital segments to reconstruct the digital string.

[0033] Wherein, step S3 specifically includes the following steps:

[0034] S31: Send the digital string to the general character string digital string text correlation discriminator for discrimination, generate a first verification embedded code, and solve the first verification embedded code with the first embedded code to obtain the similarity of the digital string for voiceprint matching. If the similarity of the digital string exceeds the preset threshold, it is determined to be successful and enter step S32, otherwise it is determined to be a verification failure and stop;

[0035] S32: sending the digital string to the text-dependent corpus and the text-independent corpus for storage, and calculating the first verification embedded code and the remaining first embedded codes to form a new first embedded code.

[0036] Wherein, step S5 specifically includes the following steps:

[0037] S51: Send the special dynamic verification string to the corresponding special string text-related voiceprint discriminator to generate a second verification embedded code, and solve the second verification embedded code with the second embedded code to obtain the similarity of the special dynamic verification string for voiceprint matching. If the similarity of the special dynamic verification string exceeds the preset threshold, it is determined to be successful and enter step S52, otherwise it is determined to be a verification failure and stop;

[0038] S52: sending the special dynamic verification character string to the text-independent corpus and the text-dependent corpus for storage, and calculating the second verification embedded code and the remaining second embedded codes to form a new second embedded code.

[0039] Wherein, step S6 specifically includes the following steps:

[0040] S61: sending the special dynamic verification string to the text-independent discriminator to generate a second verification embedded code, and solving the second verification embedded code and the second embedded code at a sentence level or a frame level to obtain the similarity of the special dynamic verification string;

[0041] S62: Perform speech recognition engine judgment on the special dynamic verification string. If the threshold is met, proceed to step S63; otherwise, it is considered as verification failure and the process stops. If the verification registration corpus exists, proceed directly to step S63;

[0042] S63: Perform voiceprint matching on the special dynamic verification string. If the match is successful, proceed to step S64; otherwise, it is considered as verification failure and the process stops;

[0043] S64: sending the special dynamic verification string to the text-independent corpus for storage, generating special string-related corpus and sending it to the temporary corpus. When the length of the special string-related corpus in the temporary corpus and the number of people who have entered the test corpus exceed the preset threshold, a special string-related corpus can be formed for training a new related discriminator.

[0044] S65: Send the special character string related corpus to the text related corpus for storage, and train the special character string related corpus to form a new special character string text related voiceprint discriminator.

[0045] Wherein, step S7 specifically includes the following steps:

[0046] S71: Send the special dynamic verification string to the text-independent discriminator to generate a second verification embedded code, and solve the second verification embedded code with the second embedded code to obtain the similarity of the special dynamic verification string. If the preset threshold is met, proceed to step S72, otherwise it is considered as verification failure and stop;

[0047] S72: Send the special dynamic verification string to the text-independent corpus for storage, and to the text-dependent corpus;

[0048] S73: Activate the corresponding special character string text-related voiceprint discriminator in the text-related corpus, and add at least one registered corpus that can be discriminated by the special character string text-related voiceprint discriminator.

[0049] Further preferably, step A3 also includes:

[0050] For text-related corpora, all corpora are trained with adversarial training using a specific verifier as the data organization method. The existing corpus text is used as input to form audio as the amplified corpus file, and noise signals are added to form new corpus files in the text-related corpus.

[0051] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0052] The invention combines speech recognition technology and voiceprint recognition technology, and has a lower false acceptance rate and false rejection rate.

[0053] A new registration and verification method is adopted, which is divided into general + special (for example, numbers + Chinese characters) verification code methods, adding certain visual interference, combined with voiceprint recognition technology, to improve security. The present invention gives full play to the advantages of text-related and text-independent voiceprint recognition, and has anti-recording fraud and anti-recording splicing fraud functions. In addition, the present invention is based on a large amount of unknown corpus, which greatly improves security.

[0054] The present invention visually adds noise to the input verification corpus to improve security and anti-theft performance.

[0055] The text-independent discriminator, the general string, number string, text-related discriminator, and the special string, text-related voiceprint discriminator jointly make autonomous judgments; the technical verification process for each person to be verified may be different. In addition, the models and databases are dynamically growing and updated, with automatic update and self-learning functions, which can continuously update the recognition system during the user's use, thereby improving the ability to recognize the target speaker and defend against imposters. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0057] Figure 1 A flowchart of a self-learning dynamic voiceprint identity authentication method based on unknown corpus of the present invention;

[0058] Figure 2 It is a practical flow chart of a self-learning dynamic voiceprint identity authentication method based on unknown corpus of the present invention;

[0059] Figure 3 A schematic diagram of a registration string of the present invention;

[0060] Figure 4 A schematic diagram of a verification string of the present invention;

[0061] Figure 5 Schematic diagram of the corpus relationship of the model of the present invention. DETAILED DESCRIPTION

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0063] In order to simplify the drawings, only the parts related to the present invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically drawn or marked. In this article, "one" not only means "only one", but also means "more than one".

[0064] The following is a further detailed description of a self-learning dynamic voiceprint identity authentication method based on unknown corpus proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description and claims.

[0065] Example

[0066] See also Figure 1 and Figure 2 This embodiment provides a self-learning dynamic voiceprint identity authentication method based on unknown corpus, which is configured for use in a dynamic voiceprint identity authentication system and can unlock and open the device through sound.

[0067] The method is as follows A1: according to the registration corpus rule format, the general registration character string and the special registration character string of the registrant are received and processed in a preset discriminator to build a registration voiceprint library; wherein the preset discriminator includes 1 text-independent discriminator, 1 general character string and number string text-related discriminator and several special character string text-related voiceprint discriminators.

[0068] A2: According to the verification corpus rule format, the input verification corpus is identified and binary segmented to obtain a general verification string and a special verification string, and they are recognized and compared by their respective speech recognition engines. If they pass the recognition and comparison, the general verification string is preprocessed and sent to the general string, number string, and text-related discriminator, and the first embedded code of the general string, number string, and text-related discriminator is obtained based on the text-related recognition technology; the special verification string is preprocessed and sent to the text-independent discriminator or the special string text-related voiceprint discriminator according to whether the special verification string has a corresponding special string text-related voiceprint discriminator and whether there is at least one registered embedded code, and the second embedded code of the text-independent discriminator or the text-related voiceprint discriminator is obtained based on the text-independent recognition technology.

[0069] A3: Based on the cumulative amount of general verification strings stored in the text-related database and the performance improvement degree that meet a certain threshold at the same time, the general string, digital string, and text-related discriminator is updated;

[0070] If the number of people who have entered corpus and the number of corpus stored in the text-related database according to the special verification string meet a certain threshold at the same time, the text-related corpus and the special string text-related voiceprint discriminator are updated.

[0071] The text-independent discriminator is updated based on the cumulative amount of special verification strings stored in the text-independent corpus and the performance improvement degree that meet a certain threshold at the same time.

[0072] Now, this embodiment is described according to practical application.

[0073] First, see Figure 3 and Figure 4 , before identity authentication, you need to enter relevant information, which is the registration stage.

[0074] The user needs to enter a registration string into the system first. The registration string is corpus information. The registration string includes general registration strings and special registration strings. Specifically, when a user registers, only K groups of registration strings (generally K 5) The content of each group of registered strings is randomly different, but the number of characters is the same. The first T groups are general registered strings, and KT groups are special strings. Usually, general strings are composed of numeric characters, and special strings are composed of letters or English phrases or Chinese characters. In this way, a registered voiceprint library is constructed, and each person has an independent part in the registered voiceprint library.

[0075] Before registration, several groups of discriminators that meet both FRR and FAR requirements have been trained, including 1 general string, number string, text-related discriminator, several special special string text-related voiceprint discriminators (100 in this embodiment), and 1 text-independent discriminator.

[0076] See also Figure 3 For the convenience of demonstration, a set of registered strings are shown, and their specific contents are not limited. Here, the design needs to take in K=5 groups of registered strings for extracting initial registered characters, including general registered strings and special registered strings. Each character content is repeated at least twice, and the number of characters in each string is 10. The first two groups of corpora are general strings, and the last three groups of corpora are special strings. The third group is letters, and the last 5 characters are repeated by the first 5 characters; the 4th and 5th groups are Chinese phrases, and each phrase is 2 characters.

[0077] Then, input according to the verification corpus given by the authentication system. The system will split the verification corpus into two segments to obtain a general verification string and a special verification string. Based on text-related recognition technology, a first embedding code is obtained. The first embedding code can be extracted by a general string digital string text-related discriminator. Based on text-independent recognition technology, a second embedding code is obtained. The second embedding code can be extracted by a preset special string text-related voiceprint discriminator. The first embedding code and the second embedding code are the initial voiceprint embedding codes of the user. Additionally, in subsequent implementations, there is only 1 general string digital string text-related discriminator and text-independent discriminator, while multiple special string text-related voiceprint discriminators can be preset. Each corresponds to a Chinese phrase or multiple letters or an English phrase, and can be increased through self-learning.

[0078] Suppose a verification corpus has a total of M characters, then it can be split into a general verification string of N characters and a special verification string of M - N characters. The general string is generally composed of N digits, and the M - N special characters are generally composed of letters or English phrases or Chinese characters. Such a design can effectively avoid the performance degradation problem of the algorithm for different channel discriminations.

[0079] See Figure 4 , for the convenience of demonstrating a set of verification corpora, which can also be called the corpora to be verified, and are not limited by their specific contents. Specifically, each verification corpus has a length of 8 characters, slightly shorter than the aforementioned registration string, and consists of a general string of N = 6 characters and a special string of M - N = 2 characters (letters or English phrases or Chinese characters). Verification corpus 1 contains all the contents in the initial registration library; verification corpus 2 contains unknown corpora, that is, "拼搏" belongs to the Chinese phrase in the special string and is located at the end of the sentence; verification corpus 3 contains the English letters "AX" in the special string and is located at the end of the sentence; verification corpus 4 consists of all initial registration corpora, and "烦恼" belongs to the Chinese phrase in the special string and is located in the middle of the sentence; verification corpus 5 contains unknown corpora, that is, "可爱" belongs to the Chinese phrase in the special string and is located at the beginning of the sentence.

[0080] After completing the above preparatory work, that is, the user can perform identity verification through the dynamic verification code given by the system.

[0081] First, see Figure 1 and Figure 2, when the user verifies, he / she inputs a dynamic verification corpus according to the dynamic verification provided by the system. The user can input the dynamic verification corpus through different audio recording devices. After input, the dynamic verification corpus can be visually noised to improve security and anti-theft, and then the dynamic verification corpus is subjected to environmental voice quality detection (including SNR calculation) to determine the environmental noise level. If the calculated signal-to-noise ratio is too low and lower than the preset threshold, it will prompt to adjust and re-enter the dynamic verification corpus. After more than three times, the system ends directly, and the specific number of times can be adjusted independently. The dynamic verification corpus higher than the preset threshold will be converted into wav format, but not limited to this single format, and it will be segmented. The general dynamic verification string and the special dynamic verification string are obtained, and they are identified separately, which can be divided into registered general dynamic verification strings, registered special dynamic verification strings, and unregistered special dynamic verification strings. Among them, the general dynamic verification string consists of a number of numbers, and the special dynamic verification characters are Chinese words or English letters or English phrases.

[0082] Assuming that the dynamic verification corpus required to be input this time is "81567903 teacup", the segmented number "81567903" and the Chinese phrase "teacup" are respectively sent to the speech recognition engine for recognition, that is, the number "81567903" is subjected to general dynamic verification string verification, and the "teacup" is subjected to special dynamic verification string verification.

[0083] See also Figure 1 and Figure 2 In this embodiment, the dynamic verification string is digitally recognized. If the recognition is successful, it enters the digital preprocessing module for subsequent processing. Otherwise, it is deemed to have failed verification and the verification is stopped. After that, the speech rate detector will be used for speech rate verification. If the speech rate is too fast, it will prompt to re-enter. If it is invalid after multiple prompts, it can be closed. For general strings, since the order of the corpus text generated when verifying is random, it is necessary to segment and reorder it to form the same general string order as the registered voiceprint library, and then verify it based on the end-to-end voiceprint recognition method in text-related recognition. Therefore, the dynamic verification string is segmented one by one in the digital preprocessing module to obtain 8 digital segments containing only one number. Then the 8 digital segments are rearranged according to the arrangement order of the general registered string to form "01356789" (for ease of understanding, it is assumed that the general registered string here is 0123456789). Next, enter the digital string reconstruction module and adjust and fill in the missing corpus according to the prior rules (i.e. 2 and 4, randomly fill in the corresponding numbers in the registered sentence whose signal-to-noise ratio reaches the threshold) to reconstruct the digital string. After filling and sorting, "0123456789" is formed.

[0084] See also Figure 1 and Figure 2, for general dynamic verification string verification, the reconstructed digital string is sent to the general string digital string text related discriminator for discrimination, and the first verification embedded code is generated, and the first verification embedded code is solved with the first embedded code to obtain the digital string similarity for voiceprint matching. If the digital string similarity exceeds the preset threshold, it is judged to be successful and enter the subsequent steps, otherwise it is considered as verification failure and stops. After the judgment is successful, the digital string is sent to the text-independent corpus and the text-related corpus for storage at the same time to enrich the user's voice data.

[0085] Assume that T sets of text-related corpora SL = {RC0, RS1, RS3, …, RS T-1}, where RC0 is a general string with N0 number and Lrc0 length, and RS1, RS3, …, RS T-1 are registered special strings, whose number and length are N1 and Lrc1, N2 and Lrc2, ..., N T-1 , length Lrc T-1 For RS and currently unregistered special strings (URS) that appear in the corpus, they are divided into two types of processing.

[0086] For general strings, they are preprocessed using preprocessing technology, and the segmented and reordered aligned corpus is added to the existing general registered corpus RC0 = {RC01, RC02,…, RC0 N0}, becomes the new RC0 = {RC01, RC02, ..., RC0 N0 ,RC0 N0+1}, and then re-enter the neural network encoder to generate a new embedding code, which is then calculated with the remaining embedding codes to form a new embedding code. For registered special strings, the reference processing method is similar and will not be repeated later. The first verification embedding code and the remaining first embedding codes are calculated to form a new first embedding code.

[0087] See also Figure 1 and Figure 2 , judge the special string.

[0088] The segmented special dynamic verification string will be identified. If it is a homophone, this step will be entered. Otherwise, it will be considered as verification failure and stop. Next, it is determined whether there is a corresponding special string text-related voiceprint discriminator, namely the "tea cup" discriminator, for the special dynamic verification string. Here, it is identified that the system has already determined whether there is a corresponding special string text-related voiceprint discriminator given the dynamic verification corpus, and whether there is at least one registered corpus of the special dynamic verification string to be verified in the text-related corpus.

[0089] See also Figure 1 and Figure 2 , if both exist, the special dynamic verification string "cup" is sent to the corresponding special string text-related voiceprint discriminator for discrimination, and the second verification embedded code is generated, and the second verification embedded code is solved with the second embedded code to obtain the similarity of the special dynamic verification string for voiceprint matching. If the similarity of the special dynamic verification string exceeds the preset threshold, it is determined to be successful and enter the subsequent steps, otherwise it is deemed to be a verification failure and stop. The special dynamic verification string is sent to the text-related corpus and text-independent corpus of the corresponding speaker for storage to enrich the speaker's voice data. The second verification embedded code and the remaining second embedded codes are calculated to form a new second embedded code.

[0090] See also Figure 1 , Figure 2 and Figure 5 For unregistered special dynamic verification strings, voiceprint recognition is carried out based on a text-independent method. The sentence-level or frame-level embedded code is used as a physical quantity to compare with the embedded code in the registered voiceprint library, and a second threshold with a lower threshold is set to achieve binary judgment; the third threshold with a higher threshold is set in combination with the speech recognition engine judgment; and the fourth threshold judgment of the binary unregistered string is achieved based on whether the requirements of the second and third thresholds are met at the same time. In addition, the system has automatic update and self-learning functions. After the first and fourth thresholds are met at the same time, it is determined that the entire corpus to be tested has passed the judgment of the recognition system. If it passes the judgment, it is classified and saved.

[0091] If the corresponding special string text-related voiceprint discriminator does not exist or both do not exist, the special dynamic verification string "teacup" is sent to the text-independent discriminator, and the sentence-level or frame-level embedded code is used as the physical quantity to compare with the embedded code in the registered voiceprint library for voiceprint matching. If the threshold is met, continue, otherwise it is considered as a verification failure and stop. The special dynamic verification string is judged by the speech recognition engine. If it meets the threshold, it proceeds to the next step, otherwise it is considered as a verification failure and stops; if the verifier registered corpus exists, it goes directly to the next step. Next, the special dynamic verification string is sent to the text-independent corpus for storage, and a new text-independent voiceprint recognition model is established to prepare for the subsequent establishment of the "teacup" text-related classifier in the text-related material library. Generate special string related corpus and send it to the temporary corpus. When the number of speakers and corpus traversed reaches a certain threshold, the special string related corpus is sent to the text related corpus for storage, and a new text related voiceprint discriminator can be constructed ("teacup" classifier, and numbered as the original last text related voiceprint discriminator plus 1). If it is not reached, continue to collect audio data of this corpus. In addition, during the execution of this step, at least one corpus is used as a group. A group of corpora can be one or five. Taking one corpus as a group as an example, after the "teacup" successfully establishes a text related voiceprint recognition classifier, another group of new irrelevant phrases can appear in the dynamic verification corpus. That is, for unregistered special strings, add them to the temporary SL and set them as RS T , when the length of the corpus is N T When a certain threshold is exceeded, RS T Put it into SL to form SL', and train it to form a new neural network encoder Net T , which can be used to generate the second embedding code. Thus, the dynamic verification corpus can be split and labeled, and retrained with the registered corpus set of the voiceprint database to form a new registered voiceprint database, a first embedding code, a second embedding code and a corresponding new embedding code.

[0092] If the corresponding special string text-related voiceprint discriminator exists but is not activated and there is at least one registered corpus in the text-related corpus, if the word "tea cup" is preset, the word discriminator has been formed [the system knows it], but the registered corpus has not yet been formed in the text-related corpus. The reason for this situation is that when it is just registered, for example: a total of 5 items are admitted, 3 of which are numbers, and the remaining 2 are words; that is, the 2 word registration corpus will not contain all the registration corpora required by the 100 discriminator groups, for example, 5 of them are covered, and the remaining 95 registered words need to be obtained in the subsequent verification process, which is the case with the hypothetical "tea cup". At this time, "tea cup" needs to enter the text-independent corpus first, and then post-process the text-related corpus to activate the discriminator with the "tea cup". Later, these 5 are used for statistical processing [the most common one is average, and other statistics can also be used to fuse a new registered embedded code]. There are N, and N are used for statistical processing, which is also dynamically trained to continuously enhance the registered embedded code.

[0093] See also Figure 5 Preferably, the self-update and self-learning of this embodiment are demonstrated in step A3.

[0094] For the related model of the digital string, the FAR and FRR conditions are met initially, so the model is generally not updated. One data is added for each verification. When the data accumulates to a threshold, for example, the performance increases by 10% and the test FAR FRR is better than the previous indicators at the same time, the model is updated to generate a new first embedded code.

[0095] When there is a lack of registered corpus for the word in spk among the 100 text-related discriminators, the registered corpus will be updated, and the model will also be updated. According to the number of people who have entered the corpus and the number of corpus of the special verification string stored in the text-related database, the text-related corpus and the special string text-related voiceprint discriminator will be updated. For the text-related corpus, all its corpus is trained in an adversarial training manner with a specific verifier as the data organization method; the existing corpus text is used as input to form an audio as an amplified corpus file, and a noise signal is added to form a new corpus file in the text-related corpus. An example is given below. Assuming that the newly added word "tea cup", this discriminator that has not been formed in the early stage meets certain conditions (for example, 10,000 speakers have 3 groups of repeated words), and then the 3 groups of repeated words are used for data amplification and robust improvement methods to train to form a model. After the model is formed, the registration corpus of the verifier will be updated for some subsequent verifiers who lack the registered corpus of the word. Several text-related voiceprint discriminators in the same situation as "teacup" can be added one by one, or they can be completed in batches of 5 and then accumulate the next 5.

[0096] The text-independent discriminator is updated when the data in the text-independent database accumulates to a certain proportion, such as an increase of 10% and the test FAR FRR is better than the previous indicators at the same time, to obtain a new text-independent discriminator and a new embedding code.

[0097] The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they still fall within the protection scope of the present invention.

Claims

1. A self-learning dynamic voiceprint identity authentication method based on unknown corpus, equipped with a dynamic voiceprint identity authentication system, characterized in that: The method comprises: A1: According to the registration corpus rule format, receive and process the general registration string and special registration string of the registrant in the preset discriminator to build a registration voiceprint library; wherein the preset discriminator includes 1 text-independent discriminator, 1 general string number string text-related discriminator and several special string text-related voiceprint discriminators; A2: According to the verification corpus rule format, the input verification corpus is identified and binary segmented to obtain a general verification string and a special verification string, and the general verification string is identified and compared by respective speech recognition engines. If the identification is successful, the general verification string is preprocessed and sent to the general string, number string, and text-related discriminator, and the first embedded code of the general string, number string, and text-related discriminator is obtained based on the text-related recognition technology; the special verification string is preprocessed, and according to whether the special verification string has a corresponding special string text-related voiceprint discriminator and whether there is at least one registered embedded code, the special verification string is sent to the text-independent discriminator or the special string text-related voiceprint discriminator, and the second embedded code of the text-independent discriminator or the text-related voiceprint discriminator is obtained based on the text-independent recognition technology; A3: updating the general character string, digital string, text correlation discriminator according to the accumulation amount of the general verification character string stored in the text correlation database and the performance improvement degree satisfying a certain threshold at the same time; If the number of people who input corpus and the number of corpus of the special verification string stored in the text-related database both meet a certain threshold, the text-related corpus and the special string text-related voiceprint discriminator are updated; According to the accumulated amount of the special verification string stored in the text-independent corpus and the performance improvement degree satisfying a certain threshold at the same time, the text-independent discriminator is updated; The step A2 comprises: S2: According to the dynamic verification rules of the dynamic voiceprint identity authentication system, a dynamic verification corpus given by the dynamic voiceprint identity authentication system is input, and the dynamic verification corpus is binary segmented to obtain a general dynamic verification string and a special dynamic verification string, and they are respectively identified and compared by their respective speech recognition engines. If they pass, they are judged to be the general dynamic verification string, and then the speech speed test, rearrangement and completion reconstruction preprocessing means are performed in sequence, and then step S3 is entered for subsequent processing. If it is judged to be the special dynamic verification string, it is first judged by the speech recognition engine whether it is homophonic. If so, step S4 is entered for subsequent processing, wherein the general dynamic verification string is composed of a number of numbers, and the special dynamic verification characters are composed of Chinese words or English letters or English phrases; S3: Send the general dynamic verification string to a preset general string digital string text related discriminator for discrimination with a preset threshold value. If the discrimination is successful, send the general dynamic verification string to both the text-independent corpus and the text-related database for storage; otherwise, it is deemed as a verification failure and stops; S4: Determine whether the special dynamic verification string has a corresponding special string text-related voiceprint discriminator, and whether the text-related corpus has at least one registered corpus of the special dynamic verification string to be verified. If both exist, proceed to step S5; if the corresponding special string text-related voiceprint discriminator does not exist or both do not exist, proceed to step S6; if the corresponding special string text-related voiceprint discriminator exists but is not activated and at least one registered corpus exists in the text-related corpus, proceed to step S7; S5: Send the special dynamic verification string to the corresponding special string text-related voiceprint discriminator to perform similarity discrimination with the registered embedded code of the person to be verified. If the similarity exceeds a preset threshold, the judgment is successful, and the special dynamic verification string is sent to the text-independent corpus and the text-related database for storage. Otherwise, it is considered as a verification failure and stops; S6: Send the special dynamic verification string to the text-independent discriminator to perform similarity discrimination with the registered embedded code of the person to be verified. If the similarity exceeds a preset threshold, if the judgment is successful, it is sent to the text-independent corpus and the text-dependent corpus for storage. Otherwise, it is deemed as a verification failure and stops; S7: Send the special dynamic verification character string to the text-independent discriminator for similarity judgment. If the similarity exceeds the preset threshold and the judgment is successful, it is sent to the text-independent corpus for storage, and the text-dependent corpus is post-processed. The corresponding special character string text-related voiceprint discriminator is activated, and at least one registered corpus that can be judged by the special character string text-related voiceprint discriminator is added. Otherwise, it is deemed as a verification failure and stops.

2. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: In the verification process of step A2, words that have not formed the special character string text-related voiceprint discriminator will also be used. After the data accumulation amount and the performance improvement degree meet a certain threshold, after data amplification and robust processing, the text-related corpus required for the training model of the special character string text-related voiceprint discriminator is obtained. The method of obtaining it is to update and increase the text-related corpus when verified by some verifiers, so that the number of people who enter the corpus and the number of their corpus meet a certain threshold at the same time, so as to generate a corresponding new special character string text-related voiceprint discriminator after training and performance verification.

3. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step A2 comprises: S11: inputting the general registration string and the special registration string into the dynamic voiceprint identity authentication system, the general registration string and the special registration string are respectively distinguished and processed by the front-end audio features formed by the general string digital string text related discriminator and the special string text related voiceprint discriminator, and then stored in the text related corpus; S12: Input the dynamic verification corpus, perform binary segmentation on the dynamic verification corpus to obtain the general verification string and the special verification string, obtain a first embedded code through text-related recognition technology, and the first embedded code can be extracted by the general string digital string text-related discriminator, obtain a second embedded code based on text-independent recognition technology, and the second embedded code can be extracted by the special string text-related voiceprint discriminator.

4. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: In the step S2, before segmenting the dynamic verification corpus, the following steps are also included: B1: Perform environmental speech quality detection on the dynamic verification corpus to determine the environmental noise level. If the calculated signal-to-noise ratio is compared with a preset threshold, if it is lower than the preset threshold, proceed to step B2, otherwise proceed to step B3; B2: prompting to input the dynamic verification corpus again, and returning to the step B1, entering the step B2 for N=N+1 times, the initial value of N is 0, if N>3, it is considered as verification failure and stops, wherein the judgment condition of N>3 is adjustable; B3: clearing N to zero, and converting the dynamic verification corpus into a wav format for segmentation; Furthermore, in the step B1, visual noise processing can be performed on the dynamic verification corpus before the environmental speech quality detection.

5. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S21: The voice of the general dynamic verification string is digitally recognized by the voice recognition engine and judged by the speech speed detector. If the recognition is successful and the speech speed is normal, the process proceeds to step S22; otherwise, the verification is deemed to have failed and the process stops; S22: Completely segment the general dynamic verification string to obtain a number of digital segments containing only one number; S23: rearrange the arrangement order of the plurality of digital segments based on the general registration character string; S24: adjusting and completing the missing corpus according to the priori rules, completing and reconstructing the arranged digital segments to obtain the general dynamic verification character string.

6. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step S3 specifically includes the following steps: S31: Send the general dynamic verification string to the general string digital string text correlation discriminator for discrimination, obtain a first verification embedded code, and solve the first verification embedded code with the first embedded code to obtain the similarity of the general dynamic verification string for voiceprint matching. If the similarity of the general dynamic verification string exceeds a preset threshold, it is determined to be successful and enter step S32, otherwise it is determined to be a verification failure and stop; S32: sending the general dynamic verification character string to the text-dependent corpus and the text-independent corpus for storage, and calculating the first verification embedded code and the remaining first embedded codes to form a new first embedded code.

7. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step S5 specifically includes the following steps: S51: sending the special dynamic verification string into the corresponding special string text-related voiceprint discriminator to generate a second verification embedded code, and solving the second verification embedded code with the second embedded code to obtain the similarity of the special dynamic verification string for voiceprint matching. If the similarity of the special dynamic verification string exceeds a preset threshold, it is determined to be successful and enter step S52, otherwise it is determined to be a verification failure and stop; S52: sending the special dynamic verification character string to the text-independent corpus for storage and the text-dependent corpus for storage, and calculating the second verification embedded code and the remaining second embedded codes to form a new second embedded code.

8. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step S6 specifically includes the following steps: S61: sending the special dynamic verification string to the text-independent discriminator to generate a second verification embedded code, and solving the second verification embedded code and the second embedded code at a sentence level or a frame level to obtain the similarity of the special dynamic verification string; S62: Perform speech recognition engine judgment on the special dynamic verification string. If it meets the threshold, proceed to step S63; otherwise, it is considered as verification failure and stops; if the verifier registration corpus exists, directly proceed to step S63; S63: Perform voiceprint matching on the special dynamic verification string. If the match is successful, proceed to step S64; otherwise, it is considered as verification failure and the process stops; S64: sending the special dynamic verification character string to the text-independent corpus for storage, generating special character string related corpus and sending it to the temporary corpus; when the length of the special character string related corpus in the temporary corpus and the number of people who input the corpus exceed the preset threshold, a special character string related corpus can be formed for training a new related discriminator; S65: sending the special character string related corpus to the text related corpus for storage, and training the special character string related corpus to form a new special character string text related voiceprint discriminator.

9. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step S7 specifically includes the following steps: S71: Send the special dynamic verification string to the text-independent discriminator to generate a second verification embedded code, and solve the second verification embedded code with the second embedded code to obtain the similarity of the special dynamic verification string for voiceprint matching. If the preset threshold is met, proceed to step S72, otherwise it is considered as verification failure and stop; S72: sending the special dynamic verification character string to the text-independent corpus for storage, and performing post-processing on the text-dependent corpus; S73: Activate the special character string text-related voiceprint discriminator corresponding to the text-related corpus, and add at least one registered corpus that can be discriminated by the special character string text-related voiceprint discriminator.

10. The self-learning dynamic voiceprint identity authentication method based on unknown corpus according to claim 1 is characterized in that: The step A3 also includes: For the text-related corpus, all its corpus is trained in an adversarial training manner using a specific verifier as the data organization method; the existing corpus text is used as input to form audio as an amplified corpus file, and a noise signal is added to form a new corpus file in the text-related corpus.

Images