Dynamic voiceprint confirmation method and computer program product

Abstract

The invention relates to an intelligent household electrical appliance, and especially relates to a dynamic voiceprint confirmation method and a computer program product, and the method comprises the steps: collecting an audio in real time, extracting voiceprint features, automatically building an initial clustering center based on the current or initial conversation voice when a target user voiceprint feature library is empty, and achieving the learning-and-use experience without pre-recording; when the voiceprint feature library is not empty, matching calculation is carried out on new voiceprint features and a historical clustering center, the feature library is dynamically updated, and meanwhile, sounder recognition is carried out in combination with a time decay strategy, so that a system can continuously track progressive changes of voice of a user and automatically adjust the gravity center of a model; the problem of recognition rate attenuation caused by the fact that a traditional static voiceprint model cannot adapt to long-term voice change is effectively solved, and recognition accuracy and robustness in complex scenes such as multi-person dialogues and noisy environments are remarkably improved through fusion of dynamic learning and timeliness judgment.

Description

Technical Field

[0001] This invention relates to the field of smart home technology, and in particular to a dynamic voiceprint verification method and computer program product. Background Technology

[0002] With the maturity of artificial intelligence technology, intelligent voice assistants have been widely used in smart glasses, smart speakers, in-vehicle systems, and smart home appliances, with full-duplex real-time voice dialogue becoming the mainstream human-computer interaction mode. This trend places higher demands on the device's ability to confirm the speaker in real-world environments, especially in complex acoustic scenarios with high noise levels and multiple speakers, such as outdoors, restaurants, and conference rooms.

[0003] Currently, to ensure voice assistants respond only to specific users, the industry generally employs speaker recognition technology based on static voiceprints as its core solution. This technology typically requires users to pre-record specified text (such as wake words or digit sequences) in a quiet environment to build a baseline voiceprint database. In practical use, the system compares the voiceprint features of the speech to be recognized with the pre-recorded static voiceprint, determining whether it belongs to the target user based on a similarity threshold. Existing technologies involve a cumbersome pre-recording process, demanding environmental requirements, negative user experience initially, and once recording is complete, the voiceprint model remains fixed and cannot be updated.

[0004] However, a speaker's voice changes dynamically with time, age, health, and emotional state. Static voiceprint databases cannot keep up with these changes, leading to a significant increase in the system's rejection rate of legitimate users after long-term use. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a dynamic voiceprint verification method and computer program product to improve the user experience of voiceprint recording and achieve speaker identification with long-term stability and high accuracy.

[0006] In a first aspect, embodiments of the present invention provide a dynamic voiceprint verification method, the method comprising: Acquire real-time audio streams; Extracting voiceprint feature vectors from real-time audio streams based on a pre-defined voice extraction model; Determine if the target user's voiceprint feature database is empty; If so, based on the voiceprint feature vectors extracted in the current or subsequent preset time period, construct initial cluster centers to initialize the target user voiceprint feature library; If not, the current voiceprint feature vector is matched with the historical cluster centers in the database. The target user's voiceprint feature database is updated based on the matching results, and the target speaker is identified based on the time decay strategy and the matching results.

[0007] In conjunction with the first aspect, the steps for extracting voiceprint feature vectors from a real-time audio stream based on a preset voice extraction model include: Speech activity detection is performed on real-time audio streams to filter out target speech segments from audio data containing ambient noise and silence; The voiceprint feature extractor uses a time-delay neural network architecture that has been converted and quantized in ONNX format to extract voiceprint feature vectors from the target speech segment.

[0008] In conjunction with the first aspect, the steps for constructing initial cluster centers to initialize the target user voiceprint feature library include: Determine whether the current scene is one with a preset wake word; If so, take the audio corresponding to the successful wake-up of the device as the initial trust sample, extract the first voiceprint feature vector, and take the first voiceprint feature vector or the mean of the first voiceprint feature vector as the initial cluster center. If not, in multi-turn dialogue scenarios, collect target speech segments from the first or several rounds of dialogue after the system starts, extract multiple corresponding voiceprint feature vectors, and determine the initial cluster centers through an unsupervised clustering algorithm.

[0009] In conjunction with the first aspect, the steps of matching the current voiceprint feature vector with historical cluster centers in the database and updating the target user's voiceprint feature database based on the matching results include: Calculate the similarity between the current voiceprint feature vector and each historical cluster center; Determine if there are any comparison results with a similarity greater than a preset threshold; If so, the current voiceprint feature vector is assigned to the target cluster center with the highest similarity, and the position of the target cluster center is updated according to the current voiceprint feature vector using the exponential moving average algorithm. If not, create new cluster centers using the current voiceprint feature vector.

[0010] In conjunction with the first aspect, the steps of updating the position of the target cluster center based on the current voiceprint feature vector using the exponential moving average algorithm include: Attenuation factor for obtaining historical information; Based on the attenuation factor of historical information and the historical cluster center position of the target cluster (i.e., the current voiceprint feature vector), the updated position of the target cluster center is calculated.

[0011] In conjunction with the first aspect, the steps for identifying the target speaker based on the time decay strategy and matching results include: Obtain the time decay weight of each historical cluster center; The weighted confidence score is calculated by combining the similarity obtained from the matching calculation with the time decay weight of the corresponding cluster centers. The weighted confidence level is compared with a preset dynamic decision threshold to output the final target speaker identification result.

[0012] In conjunction with the first aspect, the method also includes: When the weight of any cluster center in the target user's voiceprint feature library is lower than the preset cleanup threshold, the cluster center will be automatically removed from the target user's voiceprint feature library.

[0013] In conjunction with the first aspect, the steps for extracting speaker feature vectors from target speech segments using a time-delay neural network architecture that has undergone ONNX format conversion and quantization include: An optimized model is obtained by performing ONNX format conversion and quantization on the voiceprint feature extraction model of the time-delay neural network architecture. Based on the optimization model, voiceprint feature vectors are extracted from speech segments.

[0014] In conjunction with the first aspect, the steps for converting and quantizing the voiceprint feature extraction model of the time-delay neural network architecture to ONNX format to obtain an optimized model include: Convert the pre-trained voiceprint feature extraction model into an intermediate representation in ONNX format; Static fixed-point quantization is performed on the converted ONNX format model to convert the weight parameters and activation values ​​from floating-point precision to low-bit integer precision.

[0015] Secondly, embodiments of this application also provide a computer program product, wherein the computer program product storage medium stores computer program instructions, and the computer program instructions are read and executed by a processor to perform the above-described method.

[0016] The embodiments of the present invention bring the following beneficial effects: The present application provides a dynamic voiceprint verification method and computer program product, the method comprising: acquiring real-time audio stream; extracting voiceprint feature vectors from real-time audio stream based on a preset sound extraction model; determining whether the target user's voiceprint feature library is empty; if so, constructing an initial clustering center based on the voiceprint feature vectors extracted in the current or subsequent preset time period to initialize the target user's voiceprint feature library; if not, performing matching calculations between the current voiceprint feature vector and the historical clustering centers in the library, updating the target user's voiceprint feature library according to the matching results, and identifying the target speaker based on a time decay strategy and the matching results.

[0017] This invention acquires audio and extracts voiceprint features in real time. When the target user's voiceprint feature library is empty, it automatically constructs initial cluster centers based on the current or initial dialogue voice, achieving an on-demand learning experience without the need for pre-recording. When the voiceprint feature library is not empty, it dynamically updates the feature library by matching new voiceprint features with historical cluster centers and combining them with a time decay strategy for speaker identification. This enables the system to continuously track the gradual changes in the user's voice and automatically adjust the model's center of gravity. This not only effectively solves the problem of recognition rate decay caused by the inability of traditional static voiceprint models to adapt to long-term voice changes, but also significantly improves the recognition accuracy and robustness in complex scenarios such as multi-person dialogues and noisy environments through the fusion of dynamic learning and timeliness judgment.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the dynamic voiceprint verification method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the entire process of the dynamic voiceprint verification method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention.

[0022] Figure label: 130 - Processor, 131 - Memory, 132 - Bus, 133 - Communication interface. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To facilitate understanding of this embodiment, the application scenarios and design concepts of this application embodiment will be briefly introduced below.

[0025] In existing technical solutions, the pre-recording of voiceprints is highly demanding in terms of environmental requirements and cumbersome in operation, thus reducing the user's initial experience. Once recorded, the voiceprint model is usually static and cannot adapt to subsequent changes. A person's voice is not static; it undergoes subtle or even significant changes due to factors such as the passage of time, age, emotional fluctuations, and physical conditions (e.g., colds, fatigue). Static voiceprint databases cannot capture these dynamic changes, leading to a significant drop in recognition accuracy after a period of use, resulting in the serious problem of not responding to legitimate users (high rejection rate).

[0026] Based on this, this application provides a dynamic voiceprint verification method and a computer program product.

[0027] Example 1 This application provides a dynamic voiceprint verification method, combined with Figure 1 As shown, the method includes: S110, captures real-time audio streams.

[0028] S120 extracts voiceprint feature vectors from real-time audio streams based on a preset voice extraction model.

[0029] S130, determine whether the target user's voiceprint feature database is empty.

[0030] If yes, proceed to step S140; otherwise, proceed to step S150.

[0031] S140, Based on the voiceprint feature vectors extracted in the current or subsequent preset time period, construct the initial cluster center to initialize the target user voiceprint feature library.

[0032] S150: Match the current voiceprint feature vector with the historical cluster centers in the database, update the target user's voiceprint feature database based on the matching results, and identify the target speaker based on the time decay strategy and the matching results.

[0033] This application acquires continuous and complete acoustic input through real-time audio stream acquisition, ensuring that the speaker recognition process covers any speech segment in full-duplex dialogue. Subsequently, based on a lightweight deep neural network-based voice extraction model, the speech signal is converted into a high-dimensional voiceprint feature vector. This effectively suppresses environmental noise and channel interference while preserving the speaker's individual characteristics, providing robust feature representations for subsequent recognition. Then, by determining whether the voiceprint database is empty, the system automatically distinguishes between initial use and normal operation. During the initialization phase, initial cluster centers are constructed using wake words or early conversational speech, achieving zero pre-recording and cold start self-registration, completely eliminating the burden of user manual configuration and greatly improving system usability. During the non-initialization phase, online clustering and incremental learning of voiceprint patterns are achieved by matching new voiceprint features with historical cluster centers, continuously tracking the gradual change process of the user's voice and avoiding recognition rate decay caused by voice changes. By applying time decay weights to historical voiceprint data, the model ensures that it pays more attention to recently effective voiceprint features. The speaker is confirmed by combining the current matching results with the time decay weights, achieving highly robust and timely identity judgment in complex scenarios and significantly reducing the risk of false recognition and rejection.

[0034] In step S110, audio data can be collected in real time from the microphone array of the smart home appliance through the audio capture module.

[0035] In conjunction with the first aspect, step S120 includes: S121, Perform speech activity detection on the real-time audio stream to filter out target speech segments from audio data containing ambient noise and silence.

[0036] This step, serving as a pre-filtering stage, allows the Voice Activity Detection Module (VAD Module) to accurately distinguish between human voice segments and noise / silence segments, retaining only the target speech segments containing human voices. This avoids redundant calculations of invalid audio data by subsequent modules, significantly reducing the overall computational overhead of the system. It also effectively suppresses environmental noise contamination of voiceprint features from the data source, ensuring the purity of the speech signal input to the voiceprint feature extraction module. This lays a reliable data foundation for subsequent high-precision recognition and directly improves the system's robustness in noisy environments.

[0037] S122 extracts voiceprint feature vectors from the real-time audio stream of the target speech segment using an optimized time-delay neural network architecture that has undergone ONNX format conversion and quantization processing via a voiceprint feature extractor.

[0038] The Voiceprint Extraction Module employs a lightweight deep neural network and an optimized time-delay neural network architecture. This architecture efficiently captures deep, stable features representing the speaker's identity from short-duration speech signals, generating highly discriminative voiceprint feature vectors that ensure accurate clustering and recognition. In this embodiment, the lightweight deep neural network and optimized time-delay neural network architecture drastically compress the time consumed in a single feature extraction, thus fully meeting the low latency (e.g., within 20 milliseconds) performance requirements of full-duplex real-time voice dialogue scenarios, achieving an optimal balance between accuracy and speed.

[0039] Subsequently, the system checks whether the target user's voiceprint database is empty to automatically distinguish between initial use and normal operation. In the initial use scenario, the system uses the wake word or early conversational voice to build initial cluster centers, achieving zero pre-recording and cold start self-registration, completely eliminating the burden of user manual configuration and greatly improving system usability.

[0040] In conjunction with the first aspect, step S122 includes: S1221, the voiceprint feature extraction model of the time-delay neural network architecture is converted to ONNX format and quantized to obtain an optimized model.

[0041] S1222, based on an optimized model, extracts voiceprint feature vectors from speech segments.

[0042] In conjunction with the first aspect, step S1221 includes: S1221a converts the pre-trained voiceprint feature extraction model into an intermediate representation in ONNX format.

[0043] S1221b performs static fixed-point quantization on the converted ONNX format model, converting the weight parameters and activation values ​​from floating-point precision to low-bit integer precision.

[0044] Step S122 aims to efficiently and accurately convert the detected raw speech signal into high-dimensional mathematical features (i.e., voiceprint feature vectors) that can characterize the speaker's identity. Specifically, in step S1221a, the original Time Delay Neural Network (TDNN) model obtained using training frameworks such as PyTorch and TensorFlow is converted into ONNX (Open Neural Network Exchange), an open and neutral model representation format. This decouples the model from the training framework and subsequent inference hardware, ensuring that the model can be seamlessly deployed and run efficiently on various edge computing devices (such as processors in smart glasses). Step S1221b, based on the ONNX conversion, performs static fixed-point quantization (usually INT8 quantization) on the model. This process converts the model's internal weight parameters and activation values ​​from 32-bit floating-point (FP32) precision to 8-bit integer (INT8) precision, significantly reducing the model size (by approximately 75%) and memory usage. It also allows subsequent calculations to utilize hardware acceleration instructions for integer operations, thereby significantly reducing computational complexity and power consumption.

[0045] Step S1222 loads the optimized model obtained in step S1221 into memory. When the speech activity detection module outputs a speech segment, the speech data (or its features, such as MFCC) is immediately input into this optimized model. The optimized model performs forward propagation calculations on the input data. Thanks to the optimized operating environment of the ONNX format and the efficient integer computation after quantization, this inference process is extremely fast, outputting a fixed-dimensional, dense floating-point vector (e.g., 192-dimensional). This vector is the speaker feature vector, which highly summarizes the unique physiological and behavioral characteristics of the speaker contained in the speech segment and filters out the influence of speech content and environmental noise. It is the sole basis for subsequent speaker clustering, confirmation, and recognition.

[0046] The offline optimization in step S1221 ensures that the online extraction in step S1222 meets the real-time constraints. Together, they solve the problems of static scheme rigidity and wake-word scheme limited information mentioned in the background technology, enabling the system to dynamically and efficiently learn the user's voiceprint using the speech throughout the entire dialogue process, thereby achieving high-precision, low-latency speaker confirmation in complex environments.

[0047] In conjunction with the first aspect, step S140, which involves constructing initial cluster centers to initialize the target user voiceprint feature library, includes: S141, Determine whether the current scene is a scene with a preset wake word.

[0048] If yes, proceed to step S142; otherwise, proceed to step S143.

[0049] S142, take the audio corresponding to the successful wake-up of the device as the initial trust sample, extract the first voiceprint feature vector, and take the first voiceprint feature vector or the mean of the first voiceprint feature vector as the initial cluster center.

[0050] S143. In a multi-turn dialogue scenario, collect target speech segments from the first or several rounds of dialogue after the system starts, extract multiple corresponding voiceprint feature vectors, and determine the initial cluster center through an unsupervised clustering algorithm.

[0051] Specifically, step S141 determines whether the current scene is a scene with a preset wake word. By automatically identifying the current interaction mode (whether it contains a preset wake word), the most suitable voiceprint modeling strategy can be dynamically selected, ensuring that the initialization process is seamlessly integrated with the user's actual usage habits, and providing key contextual information for building a high-quality initial voiceprint library.

[0052] In scenarios where a wake word exists, step S142 automatically authenticates the audio of successfully waking up the device as an "initial trusted sample." The system can then directly extract high-value voiceprint feature vectors from this sample to construct initial cluster centers. This method requires no additional user intervention, instantly completing voiceprint registration without the user's awareness, greatly enhancing the convenience of the user experience while ensuring that the initial voiceprint originates from a high-confidence target user command.

[0053] In multi-turn dialogue scenarios, step S143 involves collecting and analyzing the multi-turn dialogue speech in the initial stage after system startup. The system can then use an unsupervised clustering algorithm to automatically discover and determine the most representative voiceprint pattern as the initial center from these samples. This approach does not rely on a specific wake word, has wider applicability, and effectively improves the stability and generalization ability of the initial voiceprint model by aggregating multiple speech segments, avoiding model bias caused by the randomness of a single sample (such as brief noise interference).

[0054] In this way, through the collaboration of S141 to S143, the cumbersome pre-recording requirements of traditional solutions are completely eliminated, enabling the system to intelligently select the optimal path and automatically complete the "cold start" of the voiceprint model, whether in a strict wake-word interaction mode or a free open dialogue mode, thus providing a technical foundation for achieving an excellent user experience of "ready to use immediately".

[0055] In conjunction with the first aspect, step S150, which involves matching the current voiceprint feature vector with historical cluster centers in the database and updating the target user's voiceprint feature database based on the matching results, includes: S151, calculate the similarity between the current voiceprint feature vector and each historical cluster center.

[0056] S152, determine whether there is a comparison result greater than the preset similarity threshold.

[0057] If yes, proceed to step S153; otherwise, proceed to step S154.

[0058] S153, assign the current voiceprint feature vector to the target cluster center with the highest similarity, and use the exponential moving average algorithm to update the position of the target cluster center according to the current voiceprint feature vector.

[0059] S154, create new cluster centers using the current voiceprint feature vector.

[0060] Step S151 quantifies the similarity between the new input features and all existing voiceprint patterns, providing an objective and quantifiable basis for subsequent clustering decisions and ensuring the scientific nature and accuracy of the update process. Step S152, based on a preset confidence threshold, determines whether the new voiceprint features are considered "tracking" existing patterns or "discovering" new patterns. This judgment mechanism enables the system to effectively distinguish between the voiceprint evolution of the target user and interference from non-target users.

[0061] For the "tracking" mode, step S153 categorizes the new features into the most similar existing patterns, achieving continuous tracking of the user's voiceprint. An exponential moving average algorithm is used to balance historical cluster centers with current new features, achieving a gradual and smooth evolution of cluster center positions. This approach not only absorbs new information to adapt to the slow changes in the user's voice but also effectively resists model mutations or jitters caused by occasional fluctuations in speech (such as brief coughs or emotional fluctuations), greatly improving the stability and robustness of the voiceprint model update process and ensuring a smooth learning process.

[0062] For the "discovery" mode, step S154 initializes the new feature as a new cluster center. This gives the Dynamic Clustering Module the ability to recognize and learn new user voiceprint states (such as changes in a user's voice after catching a cold) or to distinguish new users, dynamically expanding the flexibility of the model and avoiding performance bottlenecks caused by fixed model capacity.

[0063] In this embodiment, a stable, flexible, and adaptive online learning loop is established through the collaboration of matching, judgment, smooth updates, and dynamic creation. It not only smoothly tracks the gradual changes in user voiceprints, maintaining the model's timeliness, but also flexibly expands to accommodate new voiceprint patterns, thereby significantly improving the system's accuracy and adaptability in long-term use and fundamentally solving the problem of static model rigidity.

[0064] In conjunction with the first aspect, step S153, which employs the exponential moving average algorithm to update the position of the target cluster center based on the current voiceprint feature vector, includes: S1531, the attenuation factor for obtaining historical information.

[0065] S1532, based on the attenuation factor of historical information and the historical cluster center position of the target cluster (i.e., the current voiceprint feature vector), calculate the updated position of the target cluster center.

[0066] The attenuation factor, an adjustable parameter between 0 and 1, essentially defines the trade-off between "trust history" and "learning new knowledge." A higher attenuation factor (e.g., 0.9) means the voiceprint recognition model updates more conservatively, relying more on accumulated historical information, thus ensuring model stability; a lower attenuation factor means the model updates more aggressively, adapting to new voice changes more quickly. In step S1532, new cluster centers are calculated by weighting and fusing the historical cluster center positions with the current voiceprint feature vector according to the attenuation factor. This method is essentially a low-pass filter, effectively smoothing out feature jitter caused by accidental noise and emotional fluctuations in a single voice input, preventing unstable jumps in cluster centers, thereby improving the robustness of the model. Through progressive learning, it ensures that each model update is small and gradual, achieving smooth tracking of long-term, slow changes in user voiceprints. This avoids the model oscillations that may occur with traditional direct replacement and update methods, enabling the system to continuously and stably absorb new information without destroying the existing knowledge structure, and ultimately ensuring the reliability and accuracy of the voiceprint model in long-term use.

[0067] In conjunction with the first aspect, step S150, which involves identifying the target speaker based on the time decay strategy and the matching results, includes: S154, obtain the time decay weight of each historical cluster center.

[0068] S155 combines the similarity calculated from the matching with the time decay weight of the corresponding cluster center to calculate the weighted confidence score.

[0069] S156 compares the weighted confidence level with the preset dynamic decision threshold and outputs the final target speaker identification result.

[0070] In this embodiment, each cluster center maintains a weight that decays exponentially over time. This weight directly reflects the "freshness" of its voiceprint features and their current relevance. Recently active cluster centers have higher weights, while older, long-unupdated centers have progressively lower weights. This ensures that the decision-making system prioritizes recent and more representative user voiceprint status.

[0071] In step S155, the matching similarity is combined with the time decay weight to calculate the weighted confidence score. The "weighted confidence score," which reflects the "similarity" of the feature similarity, is combined with the time decay weight, which reflects "outdatedness," through a combination (e.g., multiplication or weighted averaging) to generate a comprehensive weighted confidence score. Thus, a template that is highly similar to the user's historical voiceprint but is outdated will have a lower final confidence score than a template with slightly lower similarity but is very recent. This makes the output decision more closely reflect the dynamic changes in the user's voice, significantly improving recognition accuracy over long-term use.

[0072] Ultimately, it compares the calculated weighted confidence score with a preset dynamic decision threshold to make a final "confirm" or "reject" judgment. Based on a comprehensive confidence score that incorporates timeliness, it not only assesses the static similarity of voice features but also evaluates the dynamic effectiveness of the voiceprint template, thus outputting highly robust and reliable recognition results in complex scenarios and long-term use. This fundamentally solves the problem of recognition rate decay caused by long-term changes in the user's voice and demonstrates excellent stability and accuracy in complex scenarios such as multi-person and noisy environments.

[0073] In conjunction with the first aspect, the method also includes: S210: When the weight of any cluster center in the target user voiceprint feature library is lower than the preset cleanup threshold, the cluster center is automatically removed from the target user voiceprint feature library.

[0074] In this embodiment, by automatically cleaning up outdated cluster centers with low weights (i.e., those that have not been used or updated for a long time), storage and computing resources can be actively released, preventing the unlimited expansion of the voiceprint feature database and ensuring the long-term high efficiency of the system. It can also ensure that the voiceprint feature database retains only recently active and high-value voiceprint patterns, which effectively purifies the model data and avoids interference from outdated and invalid patterns during matching calculations, thereby improving the speed and accuracy of subsequent recognition decisions.

[0075] Combination Figure 2The flowchart clearly illustrates the complete workflow and control logic. It sequentially reveals the closed-loop process from raw audio input to final identity decision, demonstrating the collaborative work and data flow relationships among the core modules. Specifically, the process begins with the audio acquisition module, which continuously captures raw real-time audio streams from the microphone array of smart devices (such as smart glasses), forming the signal source for all subsequent processing. Then, the continuously input audio stream undergoes voice activity detection and real-time analysis. Through its internal deep neural network decision mechanism, it accurately distinguishes target speech segments containing human voices from non-speech segments containing only environmental noise or silence. This step is crucial, effectively filtering background interference and ensuring that the system only performs deep processing on valid speech that may contain user commands, thereby significantly improving system efficiency and anti-interference capabilities.

[0076] For each identified speech segment, the system immediately initiates the feature extraction process. The speaker feature extraction module receives the speech segment and uses its internal lightweight time-delay neural network model, optimized with ONNX transformation and quantization, to convert the variable speech waveform into a fixed-dimensional high-dimensional speaker feature vector. This vector is the unique and core data representation for subsequent speaker identification.

[0077] Next, it is determined whether the target user's voiceprint feature library has been initialized, i.e., it is the first time using the system. If it is the first time using the system, the process will switch to the initialization branch to establish initial cluster centers by using the wake word or previous dialogue voice. If it is not the first time using the system, i.e., historical cluster centers already exist, the dynamic clustering learning module will calculate the similarity (such as cosine similarity) between the newly extracted voiceprint feature vector and all historical cluster centers in the library.

[0078] Then, a comparison is made based on similarity and a threshold: if the similarity is greater than the threshold, it means that the match is successful and the voice is finally confirmed to be that of the target user. The vector is assigned to the cluster with the highest similarity, and the position of the cluster center is smoothly updated using the exponential moving average algorithm to realize the progressive learning of the voiceprint model; the subsequent instruction understanding and execution module is triggered to respond to the user's voice commands. If the similarity is less than or equal to the threshold, it indicates that the match has failed and the intelligent assistant cannot respond. After recording the user as a non-target user, the clustering parameters are updated with this vector to create a new cluster center to deal with possible new users or significant changes in the user's voice. Ignoring the current voice command, the process returns to the initial audio acquisition stage and continues listening.

[0079] Secondly, embodiments of this application provide a computer program product, wherein computer program instructions are stored in a storage medium, and the computer program instructions are read and executed by a processor to perform the above-described method.

[0080] Thirdly, embodiments of this application provide an electronic device, combined with Figure 3 As shown, the electronic device includes a memory 131 and a processor 130. The memory 131 stores a computer program, and the processor 130 runs the computer program to make the electronic device perform the above-described method.

[0081] Furthermore, combined Figure 3 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0082] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0083] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0084] Fourthly, embodiments of this application provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0086] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0087] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0089] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic voiceprint verification method, characterized in that, The method includes: Acquire real-time audio streams; Based on a preset sound extraction model, voiceprint feature vectors are extracted from the real-time audio stream; Determine if the target user's voiceprint feature database is empty; If so, based on the voiceprint feature vectors extracted within the current or subsequent preset time period, an initial cluster center is constructed to initialize the target user voiceprint feature library; If not, the current voiceprint feature vector is matched with the historical cluster centers in the database. The target user's voiceprint feature database is updated based on the matching result, and the target speaker is identified based on the time decay strategy and the matching result.

2. The method according to claim 1, characterized in that, The steps of extracting voiceprint feature vectors from the real-time audio stream based on a preset voice extraction model include: Speech activity detection is performed on the real-time audio stream to filter out target speech segments from audio data containing ambient noise and silence; The voiceprint feature extractor uses a time-delay neural network architecture that has been converted and quantized in ONNX format to extract voiceprint feature vectors from the target speech segment.

3. The method according to claim 1, characterized in that, The step of constructing initial cluster centers to initialize the target user voiceprint feature library includes: Determine whether the current scene is one with a preset wake word; If so, take the audio corresponding to the successful wake-up of the device as the initial trust sample, extract the first voiceprint feature vector, and take the first voiceprint feature vector or the mean of the first voiceprint feature vector as the initial cluster center. If not, in multi-turn dialogue scenarios, collect target speech segments from the first or several rounds of dialogue after the system starts, extract multiple corresponding voiceprint feature vectors, and determine the initial cluster centers through an unsupervised clustering algorithm.

4. The method according to claim 1, characterized in that, The steps of matching the current voiceprint feature vector with historical cluster centers in the database and updating the target user's voiceprint feature database based on the matching results include: Calculate the similarity between the current voiceprint feature vector and each of the historical cluster centers; Determine if there are any comparison results with a similarity greater than a preset threshold; If so, the current voiceprint feature vector is assigned to the target cluster center with the highest similarity, and the position of the target cluster center is updated according to the current voiceprint feature vector using the exponential moving average algorithm. If not, create a new cluster center using the current voiceprint feature vector.

5. The method according to claim 4, characterized in that, The step of updating the position of the target cluster center based on the current voiceprint feature vector using the exponential moving average algorithm includes: Attenuation factor for obtaining historical information; Based on the attenuation factor of the historical information and the historical cluster center position of the target cluster, i.e. the current voiceprint feature vector, the updated position of the target cluster center is calculated.

6. The method according to claim 1, characterized in that, The steps for identifying the target speaker based on the time decay strategy and the matching results include: Obtain the time decay weight of each of the historical cluster centers; The similarity calculated by the matching is combined with the time decay weight of the corresponding cluster center to calculate the weighted confidence score. The weighted confidence level is compared with a preset dynamic decision threshold to output the final target speaker identification result.

7. The method according to claim 6, characterized in that, The method further includes: When the weight of any cluster center in the target user voiceprint feature library is lower than a preset cleanup threshold, the cluster center is automatically removed from the target user voiceprint feature library.

8. The method according to claim 2, characterized in that, The steps for extracting speaker feature vectors from the target speech segment using a time-delay neural network architecture that has undergone ONNX format conversion and quantization include: An optimized model is obtained by performing ONNX format conversion and quantization on the voiceprint feature extraction model of the time-delay neural network architecture. Based on the optimization model, voiceprint feature vectors are extracted from the speech segment.

9. The method according to claim 8, characterized in that, The steps for converting and quantizing the voiceprint feature extraction model of the time-delay neural network architecture to ONNX format to obtain an optimized model include: The pre-trained voiceprint feature extraction model is converted into an intermediate representation in ONNX format; Static fixed-point quantization is performed on the converted ONNX format model to convert the weight parameters and activation values ​​from floating-point precision to low-bit integer precision.

10. A computer program product, characterized in that, The computer program product stores computer program instructions, which are read and executed by a processor to perform the method described in any one of claims 1 to 9.

Images