Multi-model combined voiceprint identification method and device

Abstract

The invention provides a multi-model combined voiceprint identification method and device, and the method comprises the steps: obtaining a voice sample, and extracting key features in the voice sample through a deep neural network model; inputting a specified voice into the deep neural network model, and extracting a high-dimensional feature vector; converting the extracted high-dimensional feature vectors into digital features by using an i-vector model, and constructing a database; the i-vector is optimized by using a joint factor analysis method so as to improve the distinguishing ability of voiceprint recognition; the high-precision and high-stability voiceprint identification method is realized by fusing the deep neural network model, the i-vector model and the joint factor analysis method, and the method not only improves the accuracy of voiceprint identification, but also reduces the dependence of the system on hardware especially under the conditions of variable environmental noise and sound recording attack resistance, and improves the accuracy of voiceprint identification. And the voiceprint identification is more efficient and practical.

Description

Technical Field

[0001] The present invention belongs to the field of biometric identification technology, and in particular relates to a multi-model combined voiceprint identification method and device. Background Art

[0002] With the continuous advancement of science and technology, biometric technology has become one of the important means of identity authentication. Among the many biometric technologies, voiceprint recognition has gradually gained attention due to its unique advantages, such as non-contact, difficult to counterfeit and low cost. Especially for the elderly, because their fingerprints may become blurred due to age, or because of physical conditions that make it inconvenient to use contact-based identity authentication methods, voiceprint recognition technology provides a more convenient and secure solution. The invention of the voiceprint lock is based on this background, aiming to provide a more friendly and reliable security protection measure for the elderly. The invention covers the following technologies: 1. Smart door locks are devices that use modern technology to open and close doors electronically and informatically. Compared with traditional mechanical door locks, smart door locks provide more convenience and security. At present, the main ways to use smart door locks are: users enter fingerprints, portrait information, door opening commands in advance, or memorize passwords in advance so that they can use the door locks normally.

[0003] 2. Database technology is an important branch of information technology, which involves the storage, management, retrieval and analysis of data. The core of database technology is the database management system (DBMS), which is a software system used to create and manage databases. It can be used to store users' personal data.

[0004] 3. Liveness detection technology is a biometric security measure used to distinguish between real human individuals and forged identities (such as photos, videos, masks, etc.). This technology is crucial to improving the security of biometric systems, especially in application scenarios with high security requirements, such as financial services, border control, high-end security facilities, etc.

[0005] 4. Biometric technology: Biometric technology is a technology that verifies personal identity by analyzing personal physiological or behavioral characteristics. With the development of science and technology, biometric technology has been widely used in many fields such as security verification, access control, and personal identity verification. The current biometric technologies mainly include fingerprints, faces, irises, voiceprints, etc.

[0006] 5. The inventors have found that when the elderly use the above method, the following technical problems often occur: First, users need to enter their human body characteristics in the smart door lock in advance by entering fingerprints, face photos, and door unlocking commands. Visitors cannot use the smart door lock normally when they visit for the first time.

[0007] Second, as they age, the fingerprints of the elderly may become less clear due to wear, dryness or skin aging, which limits the application of traditional fingerprint recognition technology among the elderly.

[0008] Third, older adults may have difficulty performing fast and accurate physical operations, such as using keys or touching a screen, due to arthritis, muscle stiffness, or other physical illnesses.

[0009] Fourth, the elderly may suffer from memory loss and easily forget passwords or lose their keys, which brings inconvenience to their daily lives.

[0010] Fifth, older people may not be as adaptable to new technologies as younger people, so they need an identity authentication method that is simple, intuitive and easy to operate. Summary of the invention

[0011] In order to solve the above technical problems, the present invention provides a multi-model combined voiceprint identification method and device.

[0012] The technical solution adopted by the present invention is: a multi-model combined voiceprint identification method, comprising: Obtain speech samples and use a deep neural network model to extract key features from the speech samples; Input the specified speech into the deep neural network model and extract the high-dimensional feature vector; The i-vector model is used to convert the extracted high-dimensional feature vectors into digital features and build a database; The i-vector is optimized using joint factor analysis to improve the distinguishing ability of voiceprint recognition.

[0013] The acquired speech samples include speech data from different people, as well as environmental noise samples; After obtaining the speech samples, denoising, segmentation and normalization are required to improve the accuracy of subsequent feature extraction.

[0014] Also includes: Voiceprint identification is performed based on the optimized i-vector. During voiceprint identification, a new voice sample is input, and the i-vector is used to obtain the voiceprint data of the voice to be identified, which is compared with the data in the database to calculate the similarity score.

[0015] Also includes: Liveness detection: During voiceprint identification, a liveness detection mechanism is introduced to ensure the authenticity of the voiceprint sample.

[0016] Also includes: Decision and output: Based on the similarity score after voiceprint identification, a preset threshold is used for judgment. If the score exceeds the threshold, it is confirmed that the voice to be identified matches the claimed speaker. The output voiceprint identification result is a binary match decision or a score indicating confidence.

[0017] It also includes fine-tuning the deep neural network model to adapt to different environmental noise conditions to ensure that the voiceprint identification process can be carried out in real-time or near real-time environment.

[0018] The present invention also provides a multi-model combined voiceprint identification device, which includes an electronic device and a computer-readable medium.

[0019] Among them, an electronic device is configured with a processor and a memory, and the memory is configured to store a computer program, and when the program is executed by the processor, the voiceprint identification method as described in any one of claims 1-6 is implemented.

[0020] A computer-readable medium stores a computer program, which, when executed by a processor, implements the voiceprint identification method as claimed in any one of claims 1 to 6.

[0021] The beneficial effects of the present invention are: The present invention realizes a high-precision and high-stability voiceprint identification method by integrating a deep neural network model, an i-vector model and a joint factor analysis method. This method not only improves the accuracy of voiceprint recognition, especially under conditions of variable environmental noise and counter-recording attacks, but also reduces the system's dependence on hardware, making voiceprint identification more efficient and practical.

[0022] The method of the present invention enhances the security of the system through the liveness detection mechanism, simplifies the user operation process, and improves the user experience. It also has broad application prospects and provides a safer and more convenient development direction for the smart lifestyle of the elderly. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] A multi-model combined voiceprint identification method, comprising: Get voice samples and use the deep neural network model (DNN) to extract key features from the voice samples; DNN is a multi-layered neural network with sufficiently powerful function expression capabilities (exponentially related to the number of layers), and can design various flexible network structures for domain knowledge, and has efficient training methods (such as stochastic gradient descent SGD). In particular, the hierarchical structure of deep neural networks provides a very effective carrier for feature learning. Therefore, DNN can be used to learn voiceprint features or speaker discrimination factors.

[0025] This step includes two aspects: data collection preprocessing and deep neural network model training. The data collection and preprocessing method is: collect a certain number of speech samples, including speech data from different speakers, and possible environmental noise samples, and preprocess the speech samples, including denoising, segmentation and normalization, to improve the accuracy of subsequent feature extraction. Deep neural network model training is to design a deep neural network model that contains multiple convolutional layers and recurrent layers to extract the time-frequency features of speech signals; then use a large amount of speech data to train the deep neural network model until the model can accurately extract the key features in the speech signal.

[0026] Input the specified speech into the deep neural network model and extract the high-dimensional feature vector; The i-vector model is used to convert the extracted high-dimensional feature vectors into digital features and build a database. The extracted high-dimensional feature vectors are used to train the i-vector model, which can map the high-dimensional features into a low-dimensional space of a fixed length to form the i-vector representation of the speaker. Then, the i-vector calculation is performed on the speech sample of each speaker to build a voiceprint database.

[0027] The i-vector is optimized using joint factor analysis to improve the distinguishing ability of voiceprint recognition by modeling the variability of speakers and conversations.

[0028] Also includes: Voiceprint identification is performed based on the optimized i-vector. During voiceprint identification, a new voice sample is input, and the i-vector is used to obtain the voiceprint data of the voice to be identified, which is compared with the data in the database to calculate the similarity score.

[0029] Also includes: Liveness detection: When performing voiceprint identification, a liveness detection mechanism is introduced to ensure the authenticity of the voiceprint sample, such as requiring the user to make a specific sound pattern or make a random voice response.

[0030] Also includes: Decision and output: Based on the similarity score after voiceprint identification, a preset threshold is used for judgment. If the score exceeds the threshold, it is confirmed that the voice to be identified matches the claimed speaker. The output voiceprint identification result is a binary match decision or a score indicating confidence.

[0031] It also includes fine-tuning the deep neural network model to adapt to different environmental noise conditions to ensure that the voiceprint identification process can be carried out in real-time or near real-time environment.

[0032] This solution uses deep neural networks (DNN) to automatically extract high-dimensional features of speech signals. These features contain rich voiceprint information, such as pitch, rhythm, and intonation, which improves the accuracy and efficiency of feature extraction. At the same time, the i-vector model maps the high-dimensional feature vector to a low-dimensional space of fixed length, which not only reduces the dimension of the data, but also retains the key voiceprint features of the speaker, making the voiceprint representation more compact and comparable. By applying the joint factor analysis (JFA) method to optimize the i-vector, while ensuring the combination of deep learning models with traditional voiceprint recognition models, a powerful voiceprint identification framework is formed, which can make full use of their respective advantages and improve the accuracy of voiceprint identification.

[0033] The application of the i-vector model in this scheme reduces the dimension of the features, making voiceprint data easier to process and compare, while retaining sufficient information for accurate voiceprint identification.

[0034] The application of joint factor analysis in this scheme improves the generalization ability of the model to different speakers and conversation conditions, so that the voiceprint identification system can maintain a high accuracy in different environments and conditions.

[0035] An electronic device is provided with a processor and a memory, wherein the memory is configured to store a computer program, and the program implements the above-mentioned voiceprint identification method when executed by the processor.

[0036] A computer-readable medium stores a computer program, which implements the above-mentioned voiceprint identification method when executed by a processor.

[0037] Specifically, in real life, the above-mentioned voiceprint identification method can be applied to different devices, and the following usage steps of several embodiments are listed here. Example

[0038] A bank voice authentication system comprises the following steps: Step 1: During the customer registration phase, use a deep neural network model to extract voiceprint features from a text read by the customer.

[0039] Step 2: Input the extracted voiceprint features into the i-vector model, generate the customer's i-vector and store it in the bank database.

[0040] Step 3: When the customer tries to access the account, the system asks the customer to read the same text again.

[0041] Step 4: The system uses the same deep neural network model to extract voiceprint features and compares the newly generated i-vector with the i-vector in the database.

[0042] Step 5: Apply joint factor analysis to analyze the two i-vectors to eliminate the influence of inter-session variation.

[0043] Step 6: If the analysis results show that the two i-vectors are similar enough, the customer's identity is confirmed and access to the account is allowed; otherwise, access is denied. Example

[0044] An intelligent door lock system comprises the following steps: Step 1: During the system setup phase, family members each record a voice sample, and the system uses a deep neural network model to extract voiceprint features.

[0045] Step 2: Use the extracted features to train the i-vector model, generate a unique i-vector for each family member and store it.

[0046] Step 3: When family members request access to the smart door lock device, they only need to say the preset command.

[0047] Step 4: The system extracts the voiceprint features in the command and generates an i-vector.

[0048] Step 5: Use joint factor analysis to compare the real-time i-vector with the i-vector in the database to confirm the identity of the speaker.

[0049] Step 6: If the identity authentication is successful, the system performs corresponding operations, such as opening the door. Example

[0050] A border management and entry-exit control system comprises the following steps: Step 1: When a traveler applies for a visa or passport, record their voice sample and use a deep neural network model to extract voiceprint features.

[0051] Step 2: Use the extracted voiceprint features to train the i-vector model, generate the traveler's i-vector and store it in the database.

[0052] Step 3: During border inspection, travelers are asked to read a text or provide a voice sample.

[0053] Step 4: The system extracts the voiceprint features from the speech sample and generates a new i-vector.

[0054] Step 5: Apply joint factor analysis to compare the new i-vector with the i-vector in the database to verify the identity of the traveler.

[0055] Step 6: If the verification is successful, the traveler is allowed to enter or exit the country; if it fails, further review is required. Example

[0056] An enterprise employee attendance system includes the following steps: Step 1: When an employee joins the company, record their voice samples and use a deep neural network model to extract voiceprint features.

[0057] Step 2: Use the extracted features to train the i-vector model, generate a unique i-vector for each employee and store it in the attendance system.

[0058] Step 3: When employees arrive at or leave work, they only need to say their name or a preset phrase to the attendance machine.

[0059] Step 4: The attendance machine extracts the voiceprint features from the voice sample and generates an i-vector.

[0060] Step 5: Use joint factor analysis to compare the real-time i-vector with the i-vector in the database to confirm the identity of the employee.

[0061] Step 6. If identity verification is successful, the system records the attendance information; if it fails, the employee is required to use other methods to verify attendance.

[0062] By deeply combining the high-dimensional feature extraction capability of deep neural networks, the low-dimensional robust representation of the i-vector model, and the optimization technology of the joint factor analysis method, a voiceprint identification method that can maintain high accuracy and high stability under a variety of environmental conditions is realized. At the same time, this method also has the characteristics of real-time processing, easy integration and high security. It is suitable for a wide range of application scenarios, such as smart device unlocking, remote identity authentication, etc., providing users with an innovative, safe and convenient means of personal identity authentication. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-model combined voiceprint identification method, characterized in that: include: Obtain speech samples and use a deep neural network model to extract key features from the speech samples; Input the specified speech into the deep neural network model and extract the high-dimensional feature vector; The i-vector model is used to convert the extracted high-dimensional feature vectors into digital features and build a database; The i-vector is optimized using joint factor analysis to improve the distinguishing ability of voiceprint recognition.

2. The multi-model combined voiceprint identification method according to claim 1, characterized in that: The acquired speech samples include speech data from different people, as well as environmental noise samples; After obtaining the speech samples, denoising, segmentation and normalization are required to improve the accuracy of subsequent feature extraction.

3. The multi-model combined voiceprint identification method according to claim 1, characterized in that: Also includes: Voiceprint identification is performed based on the optimized i-vector. During voiceprint identification, a new voice sample is input, and the i-vector is used to obtain the voiceprint data of the voice to be identified, which is compared with the data in the database to calculate the similarity score.

4. The multi-model combined voiceprint identification method according to claim 3, characterized in that: Also includes: Liveness detection: During voiceprint identification, a liveness detection mechanism is introduced to ensure the authenticity of the voiceprint sample.

5. The multi-model combined voiceprint identification method according to claim 4, characterized in that: Also includes: Decision and output: Based on the similarity score after voiceprint identification, a preset threshold is used for judgment. If the score exceeds the threshold, it is confirmed that the voice to be identified matches the claimed speaker. The output voiceprint identification result is a binary match decision or a score indicating confidence.

6. The multi-model combined voiceprint identification method according to claim 5, characterized in that: It also includes fine-tuning the deep neural network model to adapt to different environmental noise conditions to ensure that the voiceprint identification process can be carried out in real-time or near real-time environment.

7. A multi-model combined voiceprint identification device, characterized in that: It is configured with a processor and a memory, the memory is configured to store a computer program, and when the program is executed by the processor, the voiceprint identification method according to any one of claims 1 to 6 is implemented; Alternatively, the voiceprint identification device stores a computer program, and when the program is executed by a processor, the voiceprint identification method as described in any one of claims 1 to 6 is implemented.