An Adaptive Method and System for User Voiceprint Information Based on a Large Voiceprint Model

Abstract

This invention discloses an adaptive method and system for user voiceprint information based on a large voiceprint model, relating to the fields of speech recognition and biometric authentication. The method includes a registration phase and a verification phase: in the registration phase, dual registration templates are generated and stored together using both a small and a large voiceprint model; in the verification phase, the small voiceprint model first performs initial screening and scoring, and when the confidence level falls within a preset fuzzy range, the large voiceprint model is activated in the background for secondary precise verification. After successful verification, the speech data is added to the speaker's speech database, and the dual registration templates are updated synchronously after clustering by the large voiceprint model. The system includes a registration module, an initial screening and verification module, a fuzzy judgment and secondary verification module, and a template dynamic update module. This invention achieves a balance between low latency and high accuracy through the collaboration of large and small models, and the dynamic template update adapts to voiceprint features and environmental changes, effectively reducing false call rates and improving long-term recognition accuracy. It is suitable for real-time interactive scenarios such as smart terminal wake-up and identity authentication.

Description

Technical Field

[0001] This invention relates to the field of speech recognition and biometric authentication technology, and in particular, to an adaptive method and system for voiceprint information that combines large and small model collaborative screening and dynamic updating of voiceprint registration templates; more specifically, it relates to an adaptive method and system for user voiceprint information based on a large voiceprint model. Background Technology

[0002] Voiceprint recognition, as a non-contact biometric identification technology, is widely used in scenarios such as smart device wake-up, financial identity verification, smart home control, security access control, and in-vehicle voice interaction due to its advantages such as ease of operation, unique features, difficulty in counterfeiting, and strong versatility of collection devices. With the rapid popularization of artificial intelligence technology and IoT devices, users have placed higher demands on the recognition accuracy, response speed, stability, and long-term reliability of voiceprint recognition systems.

[0003] A typical voiceprint detection system mainly consists of a registration phase and a verification phase. In the registration phase, the user is guided to record a specified amount of voice data with specified content or free text. After the system preprocesses and extracts features from the voice data, it generates a fixed voiceprint feature registration template. In the verification phase, the user's real-time audio is collected, and a verification template is generated through the same process. The verification template is then matched with the registration template for similarity. The speaker's identity is determined based on the matching score, thereby completing the device wake-up or identity authentication operation.

[0004] In practical engineering applications and large-scale deployment, traditional voiceprint recognition systems suffer from numerous insurmountable technical challenges, which have become key bottlenecks restricting the further promotion of voiceprint technology. These challenges are specifically manifested as follows:

[0005] First, the registration template is a static template; once generated, it is permanently fixed and cannot be updated or iterated. The environmental conditions, data collection devices, and pronunciation during user registration differ significantly from actual scenarios during long-term use. Factors such as ambient noise, reverberation, distance, microphone model, user age, colds, excessive smoking or drinking, and emotional changes can all cause physiological or environmental alterations in the speaker's voiceprint characteristics. These differences directly lead to a continuous decline in the matching degree between the real-time verification template and the registration template, not only causing a gradual decrease in recognition accuracy but also significantly increasing the false call rate of smart devices, severely impacting the user experience.

[0006] Second, to improve recognition accuracy and reduce false call rate, existing technologies typically optimize voiceprint models by increasing the number of model parameters, deepening the number of network layers, and expanding the training dataset. However, deploying large models directly on the front end or running them in full will lead to problems such as a surge in computing power consumption, a significant increase in verification latency, and an increase in hardware costs, which cannot meet the rigid requirements of low latency and low power consumption in real-time interactive scenarios such as smart terminals, in-vehicle devices, and mobile applications.

[0007] Third, traditional voiceprint recognition systems lack a layered and graded mechanism for precise screening of voice signals, relying solely on a single model to complete the entire verification process. When the voiceprint model's judgment of the voice signal falls within a fuzzy range, it is highly prone to misjudgment and missed judgment, leading to security issues such as false wake-up by non-target users and unauthorized access by unauthorized identities, thus failing to meet the requirements of high-security scenarios.

[0008] Fourth, most existing speech feature-based recognition and classification technologies focus on joint detection tasks such as speaker gender, age, accent, and emotion, or optimize model training accuracy through specific loss functions. They do not offer complete solutions to core issues such as static template defects in voiceprint recognition, collaborative scheduling of large and small models, dynamic updates of registered templates, and long-term adaptive optimization. Therefore, they cannot directly meet the real-time, accuracy, stability, and long-term availability requirements of voiceprint recognition. Summary of the Invention

[0009] Therefore, the purpose of this invention is to propose a user voiceprint information adaptive method and system based on a large voiceprint model. This method employs a voiceprint recognition approach that dynamically updates the registration template based on a large voiceprint model, adapting to environmental changes and temporal changes in speaker voiceprint features. It increases the number of voiceprint model parameters to meet the increased computational power consumption and verification latency requirements brought about by improving screening accuracy. It balances real-time verification, accurate screening, and template dynamics in voiceprint recognition, optimizing operational efficiency. The large model performs high-precision secondary screening of sound signals in the fuzzy range of the small model, avoiding non-target user wake-ups caused by misjudgments in the small model. It relies on the dynamic updating of the registration template to adapt to voiceprint and environmental changes, while utilizing the collaborative architecture of the large and small models to ensure low verification latency, achieving voiceprint information adaptive functionality. This improves voiceprint recognition accuracy and reduces false call rates, making it suitable for real-time voice interaction scenarios such as smart terminal wake-up and identity authentication.

[0010] This invention provides an adaptive method for user voiceprint information based on a large voiceprint model, comprising: a registration stage and a verification stage; the registration stage includes: acquiring the user's registration voice data, preprocessing the registration voice data and inputting it into a small voiceprint model and a large voiceprint model respectively, generating a small voiceprint model registration template through the small voiceprint model, generating a large voiceprint model registration template through the large voiceprint model, and storing the small voiceprint model registration template and the large voiceprint model registration template together;

[0011] Specifically, the voiceprint registration template is a set of voiceprint features generated from the user's registered voice. It serves as the matching benchmark for voiceprint verification and is divided into small voiceprint model registration templates and large voiceprint model registration templates, which can be dynamically updated using the latest features.

[0012] The verification phase includes:

[0013] S1. Collect the user's real-time voice signal, preprocess the real-time voice signal and input it into the voiceprint mini-model to generate a real-time verification template, and perform matching calculation between the real-time verification template and the voiceprint mini-model registration template to obtain the matching confidence.

[0014] Match confidence is a similarity metric score obtained by the voiceprint model after matching the real-time verification template with the registered template, representing the degree of certainty of the model in the recognition result;

[0015] S2. Determine whether the matching confidence level is within a preset fuzzy range. If not, directly output the verification result of the small voiceprint model. If yes, start the large voiceprint model to perform secondary verification on the real-time sound signal within the fuzzy range and output the verification result.

[0016] Specifically, the fuzzy interval is the preset confidence interval of the voiceprint small model. When it is in this interval, it means that the small model cannot accurately determine whether the sound signal was sent by the target user, and the large model needs to be activated for secondary screening.

[0017] The real-time sound signal is feature extracted by the large voiceprint model and similarity is calculated with the registered template of the large voiceprint model. If the calculation result is greater than a preset threshold, the verification is deemed successful; otherwise, the verification is deemed rejected.

[0018] S3. When the verification is successful, the real-time sound signal is added to the corresponding speaker's speech database. The speech data in the speaker's speech database is clustered based on the large voiceprint model. The small voiceprint model registration template and the large voiceprint model registration template are updated according to the clustering results.

[0019] Furthermore, the small voiceprint model is a lightweight, low-computing-power voiceprint recognition model used for real-time initial screening at the front end; the large voiceprint model is a high-precision voiceprint recognition model that is only activated in the background when the matching confidence level is within a preset fuzzy range.

[0020] Specifically, the voiceprint mini-model is a voiceprint recognition model with a small number of parameters, fast voiceprint feature extraction and matching speed, and low computing power consumption. It is responsible for the real-time initial screening and verification of most of the sound signals at the front end of the system.

[0021] The large voiceprint model is a voiceprint recognition model with a large number of parameters and high accuracy in voiceprint feature extraction and matching. It is only activated in the background when the small model makes a fuzzy judgment, so as to complete the accurate secondary screening of the sound signal and reduce false calls.

[0022] Furthermore, the matching calculation in step S1 uses cosine similarity or Euclidean distance;

[0023] Specifically, cosine similarity is a vector similarity calculation method that uses the cosine of the angle between two voiceprint feature vectors to represent the degree of similarity; the larger the value, the higher the matching degree.

[0024] Euclidean distance is a vector similarity calculation method. It represents the degree of similarity by calculating the Euclidean spatial distance between two voiceprint feature vectors. The smaller the distance, the higher the matching degree.

[0025] The expression for the cosine similarity is:

[0026] ;

[0027] Where A is the registration template vector, representing the voiceprint registration template feature vector pre-stored in the system database. During the verification phase, it may be the vector generated when the user initially registers, or it may be the latest template updated through subsequent large-scale model clustering iterations;

[0028] B is the verification template vector, which represents the real-time verification template feature vector generated by the system after real-time collection of user voice (such as wake word) and preprocessing and feature extraction.

[0029] n is the dimension of the voiceprint feature vector, that is, the number of feature values ​​extracted by the model;

[0030] These represent the specific feature values ​​of the registration template and the verification template in the i-th dimension, respectively. The closer the calculation result is to 1, the smaller the angle between the two vectors, the higher the voiceprint similarity, and the greater the probability of a successful match.

[0031] The expression for the Euclidean distance is:

[0032] ;

[0033] Where A is the registration template vector. Due to the use of clustering algorithm, the system may have multiple registration templates for the same user (such as normal status, cold status, etc.). In this case, A is the corresponding baseline template vector.

[0034] B is the verification template vector, which is the real-time feature vector extracted from the user's current voice signal by the voiceprint model (small model or large model);

[0035] n is the spatial dimension of the feature vector;

[0036] The square of the numerical difference between two vectors in each corresponding dimension is used to measure the absolute distance between their spatial locations. The smaller the calculated distance value, the closer the two voiceprint features are in the multidimensional space, and the higher the matching degree.

[0037] Furthermore, the preset fuzzy range is [0.4, 0.6]; when the matching confidence is greater than 0.6, the match is considered successful, and when it is less than 0.4, the match is considered rejected.

[0038] Furthermore, in step S3, multiple voiceprint sub-templates are generated after clustering. These multiple voiceprint sub-templates are adapted to at least one of the following: normal pronunciation state, disease pronunciation state, different environmental noise state, and different acquisition device state.

[0039] Furthermore, the feature calculation, clustering operation, and template update of the large voiceprint model are all executed asynchronously in the background, and the front-end verification is completed only by the small voiceprint model.

[0040] This invention also provides a user voiceprint information adaptive system based on a large voiceprint model, used to implement the user voiceprint information adaptive method based on a large voiceprint model as described above, including:

[0041] The registration module is used to obtain the user's registration voice data, preprocess the registration voice data and input it into the small voiceprint model and the large voiceprint model respectively, generate a small voiceprint model registration template through the small voiceprint model, generate a large voiceprint model registration template through the large voiceprint model, and store the small voiceprint model registration template and the large voiceprint model registration template together.

[0042] The initial screening and verification module is used to collect the user's real-time voice signal, preprocess the real-time voice signal and input it into the voiceprint mini-model to generate a real-time verification template, and perform matching calculation between the real-time verification template and the registered template of the voiceprint mini-model to obtain the matching confidence.

[0043] The fuzzy judgment and secondary verification module is used to determine whether the matching confidence is within a preset fuzzy range. If not, the verification result of the small voiceprint model is directly output. If yes, the large voiceprint model is activated to perform secondary verification on the real-time sound signal within the fuzzy range, and the verification result is output. The large voiceprint model extracts features from the real-time sound signal and calculates similarity with the registered template of the large voiceprint model. When the calculation result is greater than a preset threshold, the verification is deemed successful; otherwise, the verification is deemed rejected.

[0044] Template dynamic update module: When verification is passed, the real-time sound signal is added to the corresponding speaker's speech library, the speech data in the speaker's speech library is clustered based on the large voiceprint model, and the small voiceprint model registration template and the large voiceprint model registration template are updated according to the clustering results.

[0045] The results output module is used to output the final verification pass or rejection instruction based on the initial screening results of small models or the secondary verification results of large models.

[0046] Furthermore, the user voiceprint information adaptive system based on the large voiceprint model also includes:

[0047] Forward Iteration Module: Used to improve the initial screening accuracy of small voiceprint models and reduce the startup frequency of large voiceprint models as the number of user interactions increases;

[0048] Storage module: used to store registration templates, voice library, fuzzy interval parameters and verification thresholds; Result output module: used to output the final verification command;

[0049] The results output module is used to output the final verification pass or rejection instruction based on the initial screening results of small models or the secondary verification results of large models.

[0050] The user voiceprint information adaptive system based on a large voiceprint model of the present invention is applicable to various real-time voice interaction scenarios such as smart terminal wake-up, financial identity authentication, in-vehicle voice interaction, and security access control.

[0051] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the user voiceprint information adaptive method based on the large voiceprint model described above.

[0052] The present invention also provides a computer device, the computer device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of the user voiceprint information adaptive method based on the large voiceprint model described above.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] The user voiceprint information adaptive method and system based on a large voiceprint model provided by this invention performs high-precision secondary screening of sound signals in the fuzzy range of the small model using a large model. This effectively avoids non-target user wake-ups caused by misjudgments of the small model, significantly improves the accuracy of voiceprint recognition, and greatly reduces the false call rate (the "false call rate" is the probability that a non-target user's voice signal triggers device wake-up or successful identity authentication, and is an important indicator of the accuracy of voiceprint recognition), thus optimizing the user interaction experience. A dynamic update mechanism for the registration template is adopted, allowing the template to continuously integrate recent user voiceprint features with environmental features, adapting to physiological changes in voiceprint and environmental differences. This solves the problem of static template matching attenuation, improves long-term recognition accuracy, and ensures long-term stable operation of the system. A lightweight small model with few parameters and fast running speed is used as the front-end main verification model, undertaking most of the initial screening tasks. This results in low computational consumption and fast response speed, ensuring low computational consumption and low verification latency for the system. It meets the real-time interaction needs of scenarios such as smart terminals and real-time identity authentication; the large voiceprint model is only activated in the background when the small model is ambiguous (triggered only in the ambiguous range), avoiding the waste of computing power caused by the full-scale operation of the large model (continuous operation without distinction), greatly reducing the overall computing power consumption of the system, realizing fine allocation and efficient utilization of computing power, and improving the utilization rate of computing power resources; as the number of times users use the system increases, the registration template continuously integrates the latest voiceprint features. The more frequently users use the system, the more comprehensive the scenarios covered by the registration template, the higher the initial screening accuracy of the small model (continuously improving), and the lower the activation frequency of the large model (gradually decreasing). The system's recognition efficiency and accuracy form a positive iteration, requiring no manual intervention for optimization, and the overall system performance continues to improve; the system adopts a front-end and back-end collaborative architecture, and the template updates are completed automatically. The front-end does not require hardware upgrades, making deployment convenient and widely applicable. It can be widely adapted to various scenarios such as smart speakers, mobile phones, in-vehicle devices, financial terminals, and security access control. Attached Figure Description

[0055] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0056] In the attached diagram:

[0057] Figure 1 is an overall flowchart of the voiceprint registration and verification stages according to an embodiment of the present invention;

[0058] Figure 2 is a flowchart of the voiceprint recognition process based on dynamic updating of the registration template using a large voiceprint model according to an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of the configuration of a computer device according to an embodiment of the present invention. Detailed Implementation

[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of systems and products consistent with some aspects of this disclosure as detailed in the appended claims.

[0061] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0062] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0063] The embodiments of the present invention will be described in further detail below.

[0064] Example

[0065] This invention provides an adaptive method for user voiceprint information based on a large voiceprint model, including a registration phase and a verification phase;

[0066] The registration phase includes: the user records a specified amount of voice data according to the system instructions; the system preprocesses the voice data and inputs it into the small voiceprint model and the large voiceprint model respectively; each model extracts features from the voice data through its feature extraction module, generating a small voiceprint model registration template and a large voiceprint model registration template for the corresponding user, and storing the two templates together in the system database to complete the user's voiceprint registration; the small voiceprint model is a lightweight voiceprint recognition model with a small number of parameters, low computational consumption, and fast feature extraction and matching speed, used for initial real-time screening at the front end; the large voiceprint model is a high-precision voiceprint recognition model with a large number of parameters and high feature extraction and matching accuracy, and is only activated in the background when the small model determines ambiguity.

[0067] The verification phase includes:

[0068] S1. The system collects the user's voice signal in real time and performs preprocessing such as noise reduction and frame segmentation; the processed audio data is input into the voiceprint mini-model, which extracts voiceprint features and generates a real-time verification template; the verification template is matched and scored with the voiceprint mini-model registration template generated in the registration stage to obtain the matching confidence of the voiceprint mini-model for the voice signal.

[0069] Matching scores are calculated using either cosine similarity or Euclidean distance algorithms.

[0070] The expression for the cosine similarity is:

[0071] ;

[0072] Where A is the registration template vector, representing the voiceprint registration template feature vector pre-stored in the system database. During the verification phase, it may be the vector generated when the user initially registers, or it may be the latest template updated through subsequent large-scale model clustering iterations;

[0073] B is the verification template vector, which represents the real-time verification template feature vector generated by the system after real-time collection of user voice (such as wake word) and preprocessing and feature extraction.

[0074] n is the dimension of the voiceprint feature vector, that is, the number of feature values ​​extracted by the model;

[0075] These represent the specific feature values ​​of the registration template and the verification template in the i-th dimension, respectively. The closer the calculation result is to 1, the smaller the angle between the two vectors, the higher the voiceprint similarity, and the greater the probability of a successful match.

[0076] The expression for the Euclidean distance is:

[0077] ;

[0078] Where A is the registration template vector. Due to the use of clustering algorithm, the system may have multiple registration templates for the same user (such as normal status, cold status, etc.). In this case, A is the corresponding baseline template vector.

[0079] B is the verification template vector, which is the real-time feature vector extracted from the user's current voice signal by the voiceprint model (small model or large model);

[0080] n is the spatial dimension of the feature vector;

[0081] The square of the numerical difference between two vectors in each corresponding dimension is used to measure the absolute distance between their spatial locations. The smaller the calculated distance value, the closer the two voiceprint features are in the multidimensional space, and the higher the matching degree.

[0082] S2. Preset the confidence fuzzy range of the voiceprint mini-model and judge the confidence obtained in step S1. If the confidence exceeds the fuzzy range, the voiceprint mini-model directly outputs the verification result. If the confidence is within the fuzzy range, the background silently starts the voiceprint large model, inputs the preprocessed sound signal into the voiceprint large model to extract voiceprint features, and generates the verification template of the voiceprint large model.

[0083] The confidence level is in the fuzzy range of [0.4, 0.6]; a confidence level higher than 0.6 is considered a successful match, a confidence level lower than 0.4 is considered a rejected match, and a confidence level in the range of [0.4, 0.6] is considered a fuzzy state.

[0084] The similarity between the large model verification template and the voiceprint large model registration template generated during the registration phase is calculated. If the calculation result exceeds the preset large model verification threshold, the voice signal is determined to be emitted by the target user and the verification passes; if it does not exceed the threshold, it is determined to be the voice of a non-target user and the verification rejection result is output.

[0085] S3. If the second screening verification of the large voiceprint model passes, the system will add the verified sound signal to the speaker's speech database. Based on the large voiceprint model, the speech signals in the speaker's speech database will be re-clustered. The small voiceprint model and the large voiceprint model will update the corresponding registration templates using the newly clustered user speech data, thus completing the dynamic update of the dual templates and feature iteration.

[0086] Specifically, after clustering, multiple user voiceprint registration sub-templates are generated, covering various states such as normal pronunciation, cold, noisy environment, different distances, and different devices, adapting to the dynamic changes of voiceprint features and environment.

[0087] Figure 2 shows how the confidence level is calculated by the small voiceprint model after the voice input, and whether it is in the fuzzy interval. If it is not in the fuzzy interval, the result is output directly. If it is in the fuzzy interval, the large voiceprint model is activated for secondary verification. After the verification is passed, the voice is added to the speaker's voice database. After clustering, the registration template of the small voiceprint model is updated to complete the adaptive iteration.

[0088] The startup, feature calculation, clustering, and template update of the large voiceprint model are all performed in the background. The front end only uses the small voiceprint model to perform real-time verification, without increasing the front end verification latency.

[0089] Figure 1 illustrates the overall process of the voiceprint registration and verification stages in an embodiment of the present invention. In the registration stage, sound acquisition, voiceprint feature extraction, and voiceprint template generation are performed sequentially, and voiceprint template 1 to voiceprint template N are output. In the verification stage, sound acquisition, voiceprint feature extraction, and similarity calculation are performed sequentially, and it is determined whether the confidence level is greater than the threshold. If it is greater, the verification is passed; if it is less, the verification is not passed, and the voiceprint result is finally output.

[0090] This invention also provides a user voiceprint information adaptive system based on a large voiceprint model, comprising:

[0091] The registration module is used to obtain the user's registration voice data, preprocess the registration voice data and input it into the small voiceprint model and the large voiceprint model respectively, generate a small voiceprint model registration template through the small voiceprint model, generate a large voiceprint model registration template through the large voiceprint model, and store the small voiceprint model registration template and the large voiceprint model registration template together.

[0092] The initial screening and verification module is used to collect the user's real-time voice signal, preprocess the real-time voice signal and input it into the voiceprint mini-model to generate a real-time verification template, and perform matching calculation between the real-time verification template and the registered template of the voiceprint mini-model to obtain the matching confidence.

[0093] The fuzzy judgment and secondary verification module is used to determine whether the matching confidence is within a preset fuzzy range. If not, the verification result of the small voiceprint model is directly output. If yes, the large voiceprint model is activated to perform secondary verification on the real-time sound signal within the fuzzy range, and the verification result is output. The large voiceprint model extracts features from the real-time sound signal and calculates similarity with the registered template of the large voiceprint model. When the calculation result is greater than a preset threshold, the verification is deemed successful; otherwise, the verification is deemed rejected.

[0094] Template dynamic update module: When verification is passed, the real-time sound signal is added to the corresponding speaker's speech library, the speech data in the speaker's speech library is clustered based on the large voiceprint model, and the small voiceprint model registration template and the large voiceprint model registration template are updated according to the clustering results.

[0095] Forward Iteration Module: Used to improve the initial screening accuracy of small voiceprint models and reduce the startup frequency of large voiceprint models as the number of user interactions increases;

[0096] Storage module: used to store registration templates, voice library, fuzzy interval parameters and verification thresholds; Result output module: used to output the final verification command;

[0097] The results output module is used to output the final verification pass or rejection instruction based on the initial screening results of small models or the secondary verification results of large models.

[0098] The system of this invention adopts an integrated architecture design, with large and small models working together. The dynamic updates of the template are completed automatically in the background, without the need for additional front-end deployment costs. Only the computing resources of the large model need to be reserved in the background, making it easy to be implemented and promoted on various smart devices.

[0099] This invention employs a collaborative screening mechanism using large and small models. The small model handles the rapid initial screening of most common sound signals, ensuring low computational power and low latency. The large model only starts in the background when the small model determines ambiguity, completing accurate secondary screening and reducing false alarms from the source. Furthermore, the large model's calculation process is unnoticed by the front-end. A dual-template dynamic update mechanism is used. After the large model passes verification, the speaker's voice database is updated promptly, and the latest voiceprint features are re-clustered and then integrated into the registration templates of both the large and small models. This allows the templates to dynamically change with the environment and user voiceprint features, fundamentally solving the matching defects of static templates. A low-latency architecture is designed: all calculations of the large model and dynamic updates of the templates are completed in the background, with the front-end only requiring real-time verification by the small model. This does not increase front-end verification latency and meets the needs of real-time interactive scenarios such as smart terminals. The more times a user uses the system, the richer the latest voiceprint features integrated into the registration template become, resulting in higher initial screening accuracy of the small model and lower startup frequency of the large model. This creates a positive iteration between the system's recognition efficiency and accuracy, continuously improving the overall system performance.

[0100] Application examples

[0101] The present invention can be applied to a variety of practical scenarios. The following are specific application examples of the present invention in two different scenarios:

[0102] Example 1: Smart speaker wake-up scenario:

[0103] The user voiceprint information adaptive method based on a large voiceprint model of the present invention is applied to the wake-up of a home smart speaker. The specific steps are as follows:

[0104] 1. During the registration phase, users record five wake-up phrases in voice data following the instructions of the smart speaker. The system performs preprocessing on the voice data, including noise reduction, pre-emphasis, frame segmentation, windowing, and Mel-spectral feature extraction. The processed voice data is then input into the small voiceprint model and the large voiceprint model, respectively. The small model generates a lightweight small voiceprint model registration template, and the large model generates a high-precision large voiceprint model registration template. The two templates are associated with the user ID and stored in the local and cloud databases, completing the registration.

[0105] 2. Verification phase, as detailed below:

[0106] (1) The user who is initially screened and scored says the wake-up word. The speaker collects the sound signal and completes the preprocessing. The voiceprint mini-model is input to extract real-time features and generate a verification template. The cosine similarity is used to calculate the similarity with the mini-model registration template, and the confidence level is 0.52.

[0107] (2) Fuzzy interval determination: The preset fuzzy interval [0.4, 0.6] with a confidence level of 0.52 is within the interval. The small model cannot accurately determine this, so the large voiceprint model is silently started in the background.

[0108] (3) Secondary verification of the large model: The large model performs high-precision feature extraction on the same sound signal and calculates the similarity with the large model registration template. The result is 0.78, which is higher than the preset threshold of 0.7. It is determined to be the target user and the verification is passed.

[0109] (4) The template is dynamically updated by adding the current speech signal to the user's speech library; the large model performs K-Means clustering on all valid speech in the speech library to generate three sub-templates: normal state, cold state, and living room noise state; and updates the small voiceprint model registration template and the large voiceprint model registration template respectively.

[0110] When a user wakes up while in a cold state, the small model directly matches the cold sub-template, increasing the confidence level to 0.83. This eliminates the need to start the large model, significantly improving wake-up speed and accuracy.

[0111] Example 2: Financial Identity Authentication Scenario:

[0112] The user voiceprint information adaptive system based on a large voiceprint model of the present invention is applied to identity verification for bank APP transfers. It includes a registration module, an initial screening and verification module, a fuzzy judgment and secondary verification module, a template dynamic update module, a storage module, a forward iterative optimization module, and a result output module.

[0113] The workflow is as follows:

[0114] 1. Users complete voiceprint registration in the bank's app, generating a dual-mode version with both large and small models;

[0115] 2. When a user initiates a transfer in a noisy outdoor environment, the APP collects the voice and performs a preliminary screening using a small model. The confidence level is 0.48, indicating that the transfer has entered the fuzzy range.

[0116] 3. The system initiates a second verification using a large-scale model in the background. Once the verification is successful, the system allows the transfer.

[0117] 4. The template dynamic update module adds the noisy environment voice message to the user library, clusters it to generate noisy environment sub-templates, and updates the registered template;

[0118] 5. With repeated use by users, the registration template covers various scenarios such as quiet, outdoor, public transportation, and nighttime. The initial screening accuracy of small models remains stable at over 98.5%, while the startup frequency of large models is less than 4%.

[0119] This invention achieves adaptive optimization of voiceprint information through collaborative screening of large and small models and dynamic updating of dual templates. While ensuring low latency and low computing power, it significantly improves recognition accuracy and long-term stability, making it suitable for various scenarios that require real-time voice interaction and identity authentication.

[0120] This invention also provides a computer device. Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention; see the accompanying drawings. Figure 3 As shown, the computer device includes: an input device 23, an output device 24, a memory 22, and a processor 21; the memory 22 is used to store one or more programs; when the one or more programs are executed by the one or more processors 21, the one or more processors 21 implement the user voiceprint information adaptive method based on the large voiceprint model provided in the above embodiments; wherein the input device 23, the output device 24, the memory 22, and the processor 21 can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0121] The memory 22, as a read / write storage medium for a computing device, can be used to store software programs and computer-executable programs, such as the program instructions corresponding to the user voiceprint information adaptive method based on a large voiceprint model described in this embodiment of the invention. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the device, etc. Furthermore, the memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include memory remotely located relative to the processor 21, and these remote memories can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0122] The input device 23 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device; the output device 24 may include display devices such as a display screen.

[0123] The processor 21 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory 22, thereby realizing the above-mentioned user voiceprint information adaptive method based on the large voiceprint model.

[0124] The computer equipment provided above can be used to execute the user voiceprint information adaptive method based on the large voiceprint model provided in the above embodiments, and has corresponding functions and beneficial effects.

[0125] This invention also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the user voiceprint information adaptive method based on a large voiceprint model as provided in the above embodiments. The storage medium can be any type of memory device or storage device, including: mounting media such as CD-ROM, floppy disk, or magnetic tape; computer system memory or random access memory such as DRAM, DDRRAM, SRAM, EDORAM, Rambus RAM, etc.; non-volatile memory such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements; the storage medium may also include other types of memory or combinations thereof; furthermore, the storage medium may reside in a first computer system in which the program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet); the second computer system can provide program instructions to the first computer for execution. The storage medium includes two or more storage media that may reside in different locations (e.g., in different computer systems connected via a network). The storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0126] Of course, the computer-executable instructions provided in the embodiments of the present invention are not limited to the user voiceprint information adaptive method based on the large voiceprint model as described in the above embodiments, but can also execute related operations in the user voiceprint information adaptive method based on the large voiceprint model provided in any embodiment of the present invention.

[0127] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive method for user voiceprint information based on a large voiceprint model, characterized in that, The process includes a registration phase and a verification phase. The registration phase includes: acquiring the user's registration voice data; preprocessing the voice data and inputting it into a small voiceprint model and a large voiceprint model respectively; generating a small voiceprint model registration template using the small voiceprint model; generating a large voiceprint model registration template using the large voiceprint model; and storing the small voiceprint model registration template and the large voiceprint model registration template together. The verification phase includes: S1. Collect the user's real-time voice signal, preprocess the real-time voice signal and input it into the voiceprint mini-model to generate a real-time verification template, and perform matching calculation between the real-time verification template and the voiceprint mini-model registration template to obtain the matching confidence. S2. Determine whether the matching confidence level is within a preset fuzzy range. If not, directly output the verification result of the small voiceprint model. If yes, start the large voiceprint model to perform secondary verification on the real-time sound signal within the fuzzy range and output the verification result. The real-time sound signal is feature extracted by the large voiceprint model and similarity is calculated with the registered template of the large voiceprint model. If the calculation result is greater than a preset threshold, the verification is deemed successful; otherwise, the verification is deemed rejected. S3. When the verification is successful, the real-time sound signal is added to the corresponding speaker's speech database. The speech data in the speaker's speech database is clustered based on the large voiceprint model. The small voiceprint model registration template and the large voiceprint model registration template are updated according to the clustering results.

2. The adaptive method for user voiceprint information based on a large voiceprint model according to claim 1, characterized in that, The small voiceprint model is a lightweight, low-computing-power voiceprint recognition model used for real-time initial screening at the front end; the large voiceprint model is a high-precision voiceprint recognition model that is only activated in the background when the matching confidence level is within a preset fuzzy range.

3. The user voiceprint information adaptive method based on a large voiceprint model according to claim 1, characterized in that, The matching calculation in step S1 uses cosine similarity or Euclidean distance; the cosine similarity is: ； Where A is the registration template vector; B is the verification template vector; and n is the dimension of the voiceprint feature vector. These represent the specific feature values ​​of the registration template and the verification template in the i-th dimension, respectively. The closer the calculation result is to 1, the smaller the angle between the two vectors, the higher the voiceprint similarity, and the greater the probability of a successful match. The Euclidean distance is: ； Where A is the registration template vector. Due to the use of clustering algorithm, the system may have multiple registration templates for the same user. In this case, A is the corresponding baseline template vector; B is the verification template vector; and n is the spatial dimension of the feature vector. The square of the numerical difference between two vectors in each corresponding dimension is used to measure the absolute distance between their spatial locations. The smaller the calculated distance value, the closer the two voiceprint features are in the multidimensional space, and the higher the matching degree.

4. The adaptive method for user voiceprint information based on a large voiceprint model according to claim 1, characterized in that, The preset fuzzy range is [0.4, 0.6]; when the matching confidence is greater than 0.6, the match is considered successful, and when it is less than 0.4, the match is considered rejected.

5. The adaptive method for user voiceprint information based on a large voiceprint model according to claim 1, characterized in that, In step S3, multiple voiceprint sub-templates are generated after clustering. These multiple voiceprint sub-templates are adapted to at least one of the following: normal pronunciation state, disease pronunciation state, different environmental noise state, and different acquisition device state.

6. The adaptive method for user voiceprint information based on a large voiceprint model according to claim 1, characterized in that, The feature calculation, clustering operation, and template update of the large voiceprint model are all executed asynchronously in the background, while the front-end verification is completed only by the small voiceprint model.

7. A user voiceprint information adaptive system based on a large voiceprint model, used to implement the user voiceprint information adaptive method based on a large voiceprint model as described in any one of claims 1-6, characterized in that, include: The registration module is used to obtain the user's registration voice data, preprocess the registration voice data and input it into the small voiceprint model and the large voiceprint model respectively, generate a small voiceprint model registration template through the small voiceprint model, generate a large voiceprint model registration template through the large voiceprint model, and store the small voiceprint model registration template and the large voiceprint model registration template together. The initial screening and verification module is used to collect the user's real-time voice signal, preprocess the real-time voice signal and input it into the voiceprint mini-model to generate a real-time verification template, and perform matching calculation between the real-time verification template and the registered template of the voiceprint mini-model to obtain the matching confidence. The fuzzy judgment and secondary verification module is used to determine whether the matching confidence is within a preset fuzzy range. If not, the verification result of the small voiceprint model is directly output. If yes, the large voiceprint model is activated to perform secondary verification on the real-time sound signal within the fuzzy range, and the verification result is output. The large voiceprint model extracts features from the real-time sound signal and calculates similarity with the registered template of the large voiceprint model. When the calculation result is greater than a preset threshold, the verification is deemed successful; otherwise, the verification is deemed rejected. Template dynamic update module: When verification is successful, the real-time sound signal is added to the corresponding speaker's speech library, the speech data in the speaker's speech library is clustered based on the large voiceprint model, and the small voiceprint model registration template and the large voiceprint model registration template are updated according to the clustering results.

8. The user voiceprint information adaptive system based on a large voiceprint model according to claim 7, characterized in that, Also includes: Forward Iteration Module: Used to improve the initial screening accuracy of small voiceprint models and reduce the startup frequency of large voiceprint models as the number of user interactions increases; Storage module: Used to store registration templates, voice libraries, fuzzy interval parameters, and verification thresholds; The results output module is used to output the final verification command; The results output module is used to output the final verification pass or rejection instruction based on the initial screening results of small models or the secondary verification results of large models.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the user voiceprint information adaptive method based on a large voiceprint model as described in any one of claims 1-6.

10. A computer device, the computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the user voiceprint information adaptive method based on a large voiceprint model as described in any one of claims 1-6.

Images