Voiceprint recognition and identity authentication method and system based on public safety management and medium

Abstract

The invention discloses a voiceprint recognition and identity authentication method and system based on public safety management and a medium, and relates to the field of voiceprint recognition, and the method comprises the following steps: S1, carrying out the preprocessing of an original voice signal, extracting three types of features in parallel, and splicing the three types of features into a joint feature for representation; s2, coding the joint feature representation, outputting a high-level shared feature vector, and generating a voiceprint identification vector and a living body attribute vector through branches; s3, when high-security-level authentication is judged, dynamic challenge content is sent, response voice is received, and instant behavior characteristics are analyzed; and S4, calculating voiceprint matching similarity and living body attribute probability, fusing the voiceprint matching similarity, the living body attribute probability and the selectable dynamic evidence score, and generating a final authentication decision. By introducing the joint feature coding method based on the correlation constraint, the deep discrimination information can be better reserved, the defense capability on advanced speech synthesis and conversion attacks is enhanced, and the overall security level of the system is improved.

Description

Technical Field

[0001] This invention relates to the field of voiceprint recognition technology, and in particular to a voiceprint recognition and identity authentication method, system and medium based on public safety management. Background Technology

[0002] With the intelligent development of public safety management, voiceprint recognition technology, due to its advantages such as being contactless, easy to collect, and difficult to counterfeit, is playing an increasingly important role in scenarios such as border security checks, access control for key facilities, and remote identity verification. As one of the biometric characteristics of humans, voiceprints possess unique stability, making it possible to automatically confirm the speaker's identity by analyzing voice signals. However, the widespread application of voiceprint recognition systems has also made them targets for malicious attacks. Attackers attempt to deceive the system and impersonate legitimate individuals by playing pre-recorded voice recordings of the target (replay attacks) or by using artificial intelligence to generate forged voices (synthesis and conversion attacks), posing a serious threat to public safety.

[0003] Traditional voiceprint recognition methods primarily focus on how to more accurately extract speaker identity features from clean speech, such as the classic Mel-frequency cepstral coefficients (MFCC) and later deep learning-based voiceprint embedding vectors. However, these methods have inherent flaws in dealing with spoofing attacks: First, these methods are generally based on the assumption that the input speech is real and reliable, lacking an active liveness detection mechanism. This means they cannot distinguish whether the source of the speech signal is genuine physiological vocalization or mechanical playback or artificial synthesis, making them extremely vulnerable to simple recording and playback attacks. Second, even when independent liveness detection modules are introduced, these modules are mostly based on shallow audio features for classification. For example, they identify non-live speech by analyzing spectral distortion, coding noise, or specific frequency domain anomalies in the speech signal. These are mostly classifiers based on shallow audio features (such as spectral distortion and coding noise), which are difficult to cope with the ever-evolving high-quality, high-fidelity speech synthesis and conversion technologies, resulting in poor generalization ability and a high false positive rate. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a voiceprint recognition and identity authentication method, system and medium based on public safety management.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: First aspect: The present invention provides a voiceprint recognition and identity authentication method based on public safety management, comprising the following steps:

[0006] S1. Preprocess the input raw speech signal, extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalize the three types of features and concatenate them into a joint feature representation.

[0007] S2. The joint feature representation is encoded by a shared encoder to output a high-level shared feature vector that contains both identity and liveness information. Voiceprint identification vector and liveness attribute vector are generated by the voiceprint identification branch and the liveness attribute branch, respectively. During model training, a correlation constraint loss term is introduced into the loss function to minimize the statistical dependence between the voiceprint identification vector and the liveness attribute vector and retain key coupling information.

[0008] S3. When it is determined that a high level of security authentication is required, send dynamic challenge content to the user device and receive the user's response voice to the challenge. Analyze the real-time behavioral characteristics of the response voice to obtain a dynamic evidence score.

[0009] S4. Calculate voiceprint matching similarity based on voiceprint identifier vector, calculate liveness attribute probability based on liveness attribute vector, fuse voiceprint matching similarity, liveness attribute probability and optional dynamic evidence score, and generate final authentication decision through collaborative decision rules.

[0010] In a preferred embodiment of the present invention, in step S1, the preprocessing includes:

[0011] The original speech signal is pre-emphasized using a first-order digital filter, the transfer function of which is H(z) = 1-αz. -1 The pre-weighting coefficient α ranges from 0.95 to 0.98. -1 This is expressed as a delay of one sampling period;

[0012] The pre-emphasized signal is divided into frames with a frame length of 20-30 ms and a frame shift of 8-12 ms.

[0013] A Hamming window is added to each frame of signal to reduce spectral leakage.

[0014] In a preferred embodiment of the present invention, in step S1, the spectral structure feature extraction includes: performing a Fourier transform on each frame of speech signal to obtain the spectrum, then filtering it through a Mel filter bank and taking the logarithmic energy to obtain a Mel frequency cepstral; calculating the first-order and second-order difference coefficients of the Mel frequency cepstral to capture dynamic changes in the spectrum;

[0015] The extraction of high-frequency details and phase features includes: extracting the sub-band energy distribution of the speech signal in the frequency range of 8-20 kHz; calculating the instantaneous phase derivative through Hilbert transform, or calculating the group delay feature, to detect phase discontinuities;

[0016] The prosodic behavior feature extraction includes: estimating the fundamental frequency profile in the 50-500 Hz range; obtaining the energy envelope through short-time energy smoothing; and calculating short-time speech rate changes and pause frequencies and durations.

[0017] In a preferred embodiment of the present invention, in step S2, the voiceprint identification branch is composed of a 2- to 3-layer fully connected neural network, and each layer of the fully connected neural network uses the ReLU activation function; the output of the voiceprint identification branch is a 192-dimensional or 256-dimensional voiceprint identification vector; the training of the voiceprint identification branch uses additional corner margin loss or triplet loss as its main loss function.

[0018] In a preferred embodiment of the present invention, in step S2, the liveness attribute branch is composed of 2 to 3 layers of fully connected neural networks, each layer of which uses the ReLU activation function; the output of the liveness attribute branch is a scalar with dimension 1, which is represented as the speech liveness attribute probability after passing through the Sigmoid activation function, or the output is a 64-dimensional or 128-dimensional liveness attribute vector; the training of the liveness attribute branch uses binary cross-entropy loss as its main loss function.

[0019] In a preferred embodiment of the present invention, in step S2, the correlation constraint loss term is implemented by a regularization term, which is defined as the square of the Frobenius norm of the empirical covariance matrix of the vector after linear mapping of the voiceprint identification vector and the liveness attribute vector, or as the average mutual information estimate between the mapped vectors.

[0020] In a preferred embodiment of the present invention, in step S3, the dynamic challenge content is a randomly generated sequence of several digits or a short phrase of several words, which is converted into voice prompts by a text-to-speech engine; the real-time behavioral characteristics include response latency, speech fluency, and fundamental frequency variability, wherein the normal range of response latency is defined as 200-1500ms.

[0021] Secondly, this invention provides a system for voiceprint recognition and identity authentication based on public safety management, comprising:

[0022] The signal acquisition and processing module is used to receive raw voice signals from a microphone array or audio input device and perform preprocessing operations.

[0023] The feature extraction module is used to extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalizes the three types of features and concatenates them into a joint feature representation.

[0024] The joint feature encoding module is used to encode the joint feature representation, generate a high-level shared feature vector, and generate a voiceprint identifier vector and a liveness attribute vector through the voiceprint identifier branch and the liveness attribute branch, respectively.

[0025] The dynamic challenge generation module is used to generate random number sequences or short sentence challenge content in real time and provide voice prompts to the user.

[0026] The real-time behavior analysis module is used to receive users' voice responses to dynamic challenges and analyze the real-time behavioral characteristics of the response voice.

[0027] The collaborative decision-making module is used to collect voiceprint matching similarity and liveness attribute probability from the joint feature encoding module, as well as dynamic evidence score from the real-time behavior analysis module, and generate authentication pass or authentication rejection instructions by fusing multi-source evidence according to predefined cascade decision logic and thresholds.

[0028] Thirdly, the present invention provides an electronic device, comprising: at least one processor; and a memory communicatively connected to at least one of the processors;

[0029] The memory stores a computer program that is executed by at least one of the processors, such that the at least one processor is able to perform the voiceprint recognition and authentication method based on public safety management as described above.

[0030] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the voiceprint recognition and authentication method based on public security management as described above.

[0031] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0032] (1) This invention provides a voiceprint recognition and identity authentication method, system and medium based on public safety management. By introducing a joint feature encoding method based on correlation constraints and using a correlation constraint loss term, the voiceprint identifier vector and the liveness attribute vector are allowed to maintain necessary coupling information during training, rather than being completely independent. This relaxed decoupling mechanism can avoid excessive stripping of physiological features that are key to liveness detection, and at the same time prevent deceptive features from polluting the voiceprint vector, thereby improving the accuracy of liveness detection and the purity of voiceprint recognition. Compared with the forced decoupling method of adversarial autoencoders in the prior art, this invention can better retain deep discriminative information, enhance its defense against advanced speech synthesis and conversion attacks, and thus improve the overall security level of the system.

[0033] (2) In this invention, multi-source heterogeneous feature extraction and fusion are adopted. By extracting spectral structure features, high-frequency details and phase features and prosodic behavior features in parallel, a comprehensive speech representation can be constructed, covering different physical levels of identity authentication and liveness detection, so as to provide richer discrimination clues, thereby enhancing the sensitivity of features to deception attacks, improving the model's ability to distinguish between real and fake speech, thereby improving the adaptability and generalization performance to diverse deception methods and reducing the risk of missed detection.

[0034] (3) In this invention, by integrating dynamic challenge response and real-time behavior analysis, random challenges are triggered in high-security scenarios and user response characteristics are analyzed. By evaluating response delay, speech fluency and variability behavior indicators, real-time interactive verification is introduced. Its natural pronunciation behavior has a unique and difficult-to-imitate pattern, while deception attacks often exhibit mechanical or inconsistent behavior, thus providing strong liveness evidence for the authentication process, supplementing the shortcomings of static analysis, thereby effectively identifying replay and synthetic attacks, enabling intelligent security classification, and optimizing user experience while ensuring reliability.

[0035] (4) In this invention, multi-evidence collaborative decision-making and cascaded decision-making rules are adopted. The authentication strategy is flexibly adjusted according to the confidence level, which can reduce the misjudgment of a single indicator, improve the reliability of decision-making, reduce the false recognition rate and false rejection rate, and improve the authentication accuracy. Compared with the traditional fixed threshold or single decision-making mechanism, this invention is more adaptive, thereby enhancing the stability and practicality in complex public safety environments and laying the foundation for large-scale deployment. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the voiceprint recognition and identity authentication method based on public safety management according to Embodiment 1 of the present invention;

[0038] Figure 2 This is a flowchart of the preprocessing of the original speech signal and the extraction of three types of heterogeneous features in Embodiment 1 of the present invention;

[0039] Figure 3 This is a flowchart of the joint feature encoding training process in Embodiment 1 of the present invention;

[0040] Figure 4 A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown. Detailed Implementation

[0041] The technical solutions of the embodiments 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, and 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.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0043] Application Overview:

[0044] To overcome the limitations of traditional methods, more advanced solutions have emerged in existing technologies. One approach is an end-to-end liveness voiceprint recognition method based on multi-feature fusion and adversarial autoencoder (AAE) feature decoupling. This method simultaneously extracts voiceprint features and determines liveness attributes from a single-path speech signal using a unified deep learning model. Its core lies in utilizing the adversarial training mechanism of AAE to force the feature representation generated by the encoder to decompose into two independent subspaces: one containing only voiceprint identifier vectors related to speaker identity, and the other containing only liveness attribute vectors related to speech liveness. Through this feature decoupling, it is theoretically possible to ensure that during voiceprint matching, the decision-making process can eliminate false features introduced by spoofing attacks, performing authentication only based on pure identity information, thereby significantly improving the anti-spoofing security of the voiceprint recognition system.

[0045] However, after in-depth research, the applicant discovered that directly applying the aforementioned AAE feature decoupling model to voiceprint anti-spoofing scenarios in the public security field still reveals a significant deep-seated technical contradiction at the principle level. Because there is a profound and inherent coupling between the speaker's identity characteristics and living physiological characteristics in the speech signal—for example, an individual's unique timbre, formant structure, and subtle vibrato during pronunciation are not only unique identifiers of their identity but also direct manifestations of the physiological process of live vocalization—a forced and absolutist feature decoupling strategy leads to a dilemma under this complex relationship.

[0046] On the one hand, in order to meet the requirements of adversarial discriminators for the purity of the liveness attribute vector, the encoder is forced to strip all identity-related information from the vector. However, this process inevitably discards those deep physiological features that are crucial for identifying highly realistic deception attacks (such as subtle neuromuscular tremors and natural pronunciation instability), resulting in the loss or weakening of liveness discrimination features, which is the feature impairment effect. This effect makes the liveness attribute vector unable to capture the subtle and crucial physiological differences between deception speech and real speech, thereby reducing the model's ability to distinguish advanced synthesis or conversion attacks.

[0047] On the other hand, if these key, inherently coupled discriminative features cannot be effectively separated due to their close association with identity features, then this information will remain in the voiceprint identifier vector. This will directly pollute the voiceprint feature database, causing the identity features forged by deceptive voice to overlap with the real registered identity features in the vector space, interfering with the accuracy of voiceprint matching; this is the information residue effect. This pollution not only reduces the ability to detect high-quality deception attacks, but may also lead to an increase in the false rejection rate for legitimate users.

[0048] To address the aforementioned issues, this invention provides a voiceprint recognition and identity authentication method, system, and medium based on public safety management. It abandons the mandatory and absolute decoupling of voiceprint identity features and live physiological features, instead employing a strategy of relaxed decoupling and multi-evidence collaborative decision-making. It acknowledges and effectively utilizes the necessary inherent coupling information between the two, and solves the feature damage and information residue effects faced by existing voiceprint recognition technologies in dealing with advanced deception attacks through a combination of multi-level feature extraction, joint encoding of correlation constraints, and dynamic challenge response and real-time behavior analysis. This enhances the anti-deception security of voiceprint recognition.

[0049] Example 1: As Figure 1 As shown, a voiceprint recognition and identity authentication method based on public safety management includes the following steps:

[0050] S1. Preprocess the input raw speech signal, extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalize the three types of features and concatenate them into a joint feature representation.

[0051] S2. The joint feature representation is encoded by a shared encoder to output a high-level shared feature vector that contains both identity and liveness information. Voiceprint identification vector and liveness attribute vector are generated by the voiceprint identification branch and the liveness attribute branch, respectively. During model training, a correlation constraint loss term is introduced into the loss function to minimize the statistical dependence between the voiceprint identification vector and the liveness attribute vector and retain key coupling information.

[0052] S3. When it is determined that a high level of security authentication is required, send dynamic challenge content to the user device and receive the user's response voice to the challenge. Analyze the real-time behavioral characteristics of the response voice to obtain a dynamic evidence score.

[0053] S4. Calculate voiceprint matching similarity based on voiceprint identifier vector, calculate liveness attribute probability based on liveness attribute vector, fuse voiceprint matching similarity, liveness attribute probability and optional dynamic evidence score, and generate final authentication decision through collaborative decision rules.

[0054] like Figure 2 As shown, in some specific implementations, step S1 is used for multi-source heterogeneous feature extraction and fusion of the input speech signal. First, the digitized original speech signal is preprocessed. Specifically, the preprocessing includes pre-emphasis, framing, and windowing operations to enhance the energy of the high-frequency part of the speech signal to compensate for the natural attenuation effect of the vocal organs and vocal tract on the high-frequency components, thereby balancing the spectrum so that subsequent feature extraction can capture speech information more comprehensively.

[0055] In a preferred configuration, pre-emphasis is implemented using a first-order digital filter whose transfer function is set to H(z) = 1-αz. -1 The pre-emphasis coefficient α ranges from 0.95 to 0.98, preferably 0.97, which effectively enhances the energy of high-frequency components while maximally suppressing potential noise amplification, ensuring the signal-to-noise ratio. -1 This is expressed as a delay of one sampling period.

[0056] Furthermore, the preprocessing step divides the pre-emphasized speech signal into a series of short time frames. The framing operation is used to divide the quasi-static speech signal into sufficiently short time periods so that the speech characteristics within each frame can be considered stationary. The frame length is 20-30 ms, preferably 25 ms, and the frame shift is 8-12 ms, preferably 10 ms. This frame length and frame shift configuration ensures that while effectively capturing local speech features, it provides sufficient temporal resolution to analyze the dynamic changes of speech, and ensures appropriate overlap between frames to reduce edge effects.

[0057] Furthermore, a windowing operation is applied to each frame of the speech signal to smooth the signals at both ends of the frame and reduce spectral leakage caused by truncation. The Hamming window function is used to process each frame of the speech signal. The mathematical expression of the Hamming window is w(n) = 0.54-0.46*cos(2πn / (N-1)), where N represents the window length and n represents a discrete index variable with a value range of 0 to N-1.

[0058] In this embodiment, after preprocessing, three types of complementary speech features are extracted in parallel: spectral structure features, high-frequency details and phase features, and prosodic behavior features. These three types of features characterize the essential attributes of the speech signal from different dimensions, providing rich and complementary information for subsequent voiceprint recognition and liveness detection.

[0059] For the extraction of spectral structure features, this embodiment uses the Mel-spectrogram as the basis. Specifically, a discrete Fourier transform is performed on each windowed speech signal to obtain its spectral information. The window length of the Fourier transform is consistent with the frame length, and the number of Fourier transform points is set to 1024 to ensure a sufficiently high frequency resolution to distinguish the fine structure in the speech signal.

[0060] Subsequently, the obtained linear frequency spectrum is filtered by a set of Mel filter banks. In this embodiment, the number of Mel filter banks is set to 128, and the frequency response range covers 0-8000 Hz. This range covers the frequency band where human speech energy is most concentrated, and extends to some high-frequency regions to capture richer acoustic information. The logarithm of the energy of each Mel band after filtering is taken to obtain the Mel frequency cepstrum.

[0061] Furthermore, to capture the dynamic changes in the spectrum over time, this embodiment also calculates the first-order difference (delta) and second-order difference (delta-delta) coefficients of the Mel frequency cepstral plot. The difference coefficients are calculated within a 3-frame context window, specifically through linear regression or a simple difference formula, for example, delta... c(t) = (c(t+1)-c(t-1)) / 2, where delta c(t) c(t) represents the first-order difference coefficients calculated at time frame t, i.e., the dynamic features; c(t) represents the static MFCC (Mel frequency cepstral coefficients) coefficient values ​​at time frame t; and the second-order difference is a further difference calculation based on the first-order difference. Among these, the dynamic features are crucial for characterizing the speaker's pronunciation habits and speech rate changes.

[0062] For the extraction of high-frequency details and phase features, this embodiment uses a specially designed method to capture subtle information in the high-frequency region that is easily overlooked by traditional MFCC and other features, as well as the phase characteristics of the speech signal. Specifically, an energy distribution feature of the speech signal in the 8-20 kHz frequency range is extracted through a filter bank designed specifically for the high-frequency subband.

[0063] The filter bank is designed as a bandpass filter with bandwidth divided into octaves, such as multiple sub-bands including 8-11.3 kHz, 11.3-16 kHz, and 16-20 kHz. Each sub-band is implemented using a Butterworth filter with an order of 4. This division method can finely analyze the energy distribution in the high-frequency region. High-frequency components have important discriminative capabilities for distinguishing between natural speech and synthesized speech (especially high-fidelity synthesized speech), because synthesis technology often has difficulty completely mimicking the complexity of human voice when generating high-frequency overtones and noise characteristics.

[0064] In addition, this embodiment also extracts instantaneous phase derivative information; the instantaneous phase derivative is obtained by performing a Hilbert transform on the speech signal to obtain its analytic signal, and then the instantaneous phase of the analytic signal is calculated, and the phase is differentially processed.

[0065] Another approach is to calculate the group delay of the signal, defined as the negative derivative of the phase response with respect to angular frequency. These phase-related features can reveal phase discontinuities or anomalies that may be introduced during speech synthesis or conversion, as artificially generated speech typically possesses inherent properties in maintaining phase consistency. For example, when the fundamental frequency and overtone relationships of synthesized speech are modified, its phase structure changes significantly, even if these changes are not apparent in the frequency domain energy features.

[0066] For the extraction of prosodic behavioral features, this embodiment focuses on capturing the speaker's unique pronunciation rhythm, intonation, and speech rate patterns, which are also difficult to be perfectly replicated by deception attacks; specifically, prosodic behavioral features include fundamental frequency (F0) profile, energy envelope, and short-term speech rate variations.

[0067] Among them, the fundamental frequency profile reflects the frequency of vocal cord vibration and is an important clue to tone and speaker identity. In this embodiment, the YIN algorithm based on autocorrelation is used for F0 estimation, and its estimation range is set at 50-500 Hz, covering the normal F0 range of adult males and females. The energy envelope reflects the loudness change of speech. By calculating the short-time energy of each frame of speech signal and smoothing the short-time energy, for example, a moving average filter with a 50 ms window is used for smoothing to effectively remove high-frequency fluctuations and highlight the overall energy trend of speech. The short-time speech rate change is represented by syllable detection using the integrated speech activity detection (VAD) algorithm and the number of syllables per unit time. In this embodiment, a deep learning-based VAD model is used for accurate syllable boundary recognition. At the same time, the frequency and duration of pauses in speech are also analyzed, because the pause patterns in natural speech have highly personalized characteristics, while synthesized or converted speech often exhibits mechanical or unnatural distributions when generating pauses.

[0068] After extracting the three types of features, normalization is performed to eliminate differences in dimensionality and dynamic range between different features, thereby preventing certain features from dominating the subsequent learning process. This embodiment uses the mean-variance normalization (Z-score normalization) method to process each dimension of each type of feature independently over the entire speech segment.

[0069] Specifically, for each feature dimension x, its normalized value x norm Through formula x norm = (x-μ) / σ is calculated, where μ represents the mean of the feature in this dimension and σ represents the standard deviation of the feature in this dimension; this ensures that the mean of all features is zero and the variance is one, promoting stable training of the neural network; subsequently, a feature concatenation operation is performed, connecting the normalized spectral structure feature vector, high-frequency detail and phase feature vector, and prosodic behavior feature vector in the feature dimension to form a high-dimensional joint feature vector, which serves as the input to the subsequent shared encoder.

[0070] For example, if the spectral structure features are 128-dimensional, the high-frequency details and phase features are 64-dimensional, and the prosodic behavior features are 32-dimensional, then the dimension of the concatenated joint feature vector is 128 + 64 + 32 = 224-dimensional; its multi-source heterogeneous feature concatenation strategy can construct an input representation with rich information and moderate redundancy.

[0071] like Figure 3 As shown, in some specific implementations, step S2 is used to construct a deep learning model that can simultaneously capture identity information and liveness attributes and can loosely decouple the two; the model adopts a shared encoder-dual-branch decoder structure, where the shared encoder receives the joint feature representation formed in step S1 as input and encodes it into a high-level shared feature vector Z that simultaneously contains identity and liveness information.

[0072] In a preferred embodiment, the shared encoder is an architecture of a deep convolutional neural network (CNN) containing four convolutional modules. Each convolutional module is constructed as follows: a two-dimensional convolutional layer, followed by a batch normalization layer, a ReLU activation function, and a max pooling layer. Specifically, the kernel size of the first convolutional layer is set to 5×5, the stride is 1×1, and the number of output channels is 64; the kernel size of the second convolutional layer is set to 3×3, the stride is 1×2, and the number of output channels is 128; the kernel size of the third convolutional layer is set to 3×3, the stride is 1×2, and the number of output channels is 256; and the kernel size of the fourth convolutional layer is set to 3×3, the stride is 1×1, and the number of output channels is 512.

[0073] The batch normalization layer is used to accelerate network training and improve the model's generalization ability; the ReLU activation function f(x) = max(0,x) introduces non-linearity; the max pooling layer downsamples the feature map with a size of 2×2 and a stride of 2×2, gradually extracting higher-level abstract features; the network finally transforms the multi-level feature map into a fixed-dimensional vector through a global average pooling layer, and then connects it to a fully connected layer to generate a shared feature vector Z with a dimension of 1024.

[0074] Furthermore, two independent decoding branches are differentiated from the shared feature vector Z: the voiceprint identification branch and the liveness attribute branch; the voiceprint identification branch is used to map the high-level shared feature vector Z into a pure voiceprint identification vector S.

[0075] In this embodiment, the voiceprint identification branch consists of a three-layer fully connected neural network, with each layer using the ReLU activation function. The first fully connected layer maps the 1024-dimensional Z vector to 512 dimensions, and the second layer maps it to 256 dimensions, ultimately outputting a 256-dimensional voiceprint identification vector S. This vector S is a highly discriminative feature representation sensitive to speaker identity. The training of the voiceprint identification branch uses an additional angle margin loss (e.g., ArcFace loss) as its main loss function. ArcFace loss introduces an angle margin into the logistic values ​​and normalizes the feature vector, maximizing the inter-class distance between different speakers and minimizing the intra-class distance, thereby significantly enhancing the discriminative power and generalization ability of voiceprint features.

[0076] Meanwhile, the liveness attribute branch maps the shared feature vector Z to the liveness attribute vector L. In this embodiment, the liveness attribute branch also consists of three fully connected neural networks, each of which uses the ReLU activation function. The first two layers are similar to the voiceprint identification branch, mapping the 1024-dimensional Z vector step by step, and finally outputting a 64-dimensional liveness attribute vector L. This liveness attribute vector L is used to capture physiological details related to liveness in the speech signal, such as vocal tract resonance characteristics and vocal tremor. These details are crucial for identifying highly realistic deception attacks.

[0077] The training of the liveness attribute branch uses the binary cross-entropy loss as its main loss function. This loss function can effectively distinguish between real speech (label 1) and spoofed speech (label 0), and optimizes the discrimination ability of the liveness attribute vector by minimizing the classification error.

[0078] Furthermore, a correlation constraint loss term is introduced in this joint feature encoding stage. This loss term is integrated into the overall loss function of the system to minimize the statistical dependency between the voiceprint identification vector S and the liveness attribute vector L, while allowing the retention of inherent coupling information that is crucial for identifying spoofing attacks. Traditional decoupling methods often pursue absolute feature separation, but the identity features of human voice and the physiological features of liveness are highly coupled in terms of physical generation mechanism. Excessive decoupling will cause feature damage and loss of key information for liveness detection.

[0079] The relaxation decoupling strategy in this implementation acknowledges and utilizes this inherent coupling. Specifically, the regularization term is defined as the square of the Frobenius norm of the empirical covariance matrix between the voiceprint identification vector S and the liveness attribute vector L after being projected onto a shared dimension space through their respective linear mapping layers. Assuming the vectors after linear mapping are S' and L', the constraint term can be expressed as ||Cov(S', L')|| F 2 Where Cov represents the covariance matrix, ||.|| F This represents the Frobenius norm.

[0080] Another preferred definition is to calculate the estimated average mutual information (MI) between the projected vectors and use it as a regularization term, such as using a neural network-based mutual information estimator (MINE). By minimizing the estimated mutual information, S and L are made statistically less correlated, but not completely independent. The weight coefficient λ of the regularization term is an adjustable hyperparameter, set to 0.01-0.1. This weight is used to achieve an optimal balance between the degree of forced decoupling and the preservation of key coupling information. By carefully adjusting λ, it is possible to ensure the purity of the voiceprint identification vector while allowing the liveness attribute vector to contain enough physiological cues that are related to identity but crucial for liveness identification, thus effectively solving the problems of information residue and feature damage.

[0081] In some specific implementations, step S3 is used to perform dynamic challenge response and real-time behavior analysis. For high-security scenarios, a defense-in-depth mechanism is provided. For authentication scenarios that are preset to a high security level, a dynamic challenge is triggered, requiring the user to interact in real time.

[0082] Specifically, a high-quality text-to-speech (TTS) engine randomly generates a challenge, which is then converted into a voice prompt and played back to the user. The challenge can be a sequence of numbers or a short phrase. The randomness of the sequence or phrase is achieved using a cryptographically secure pseudo-random number generator (CSPRNG) combined with a predefined vocabulary, ensuring that the challenge is novel and unpredictable in each session. This significantly increases the difficulty of spoofing attacks, as pre-recorded attacks and replay attacks cannot handle unknown challenges generated in real-time. For example, the sequence of numbers can be generated by randomly selecting six numbers from 0-9, while the phrase is randomly drawn from a library containing thousands of common words and phrases, ensuring semantic coherence.

[0083] Preferably, after receiving the user's response voice, real-time behavioral feature analysis is performed immediately, including response latency, voice fluency, and fundamental frequency variability.

[0084] Furthermore, for response latency analysis, the time interval between the end of the voice prompt and the start of the user's response is measured. The end of the voice prompt is determined by the output timestamp of the TTS engine, while the start of the user's response is accurately detected by a deep learning-based Voice Activity Detection (VAD) model. This embodiment uses a VAD model based on a hybrid architecture of recurrent neural networks (RNN) and convolutional neural networks (CNN). This model has been trained on a large amount of real-world voice data and can detect the start point of the voice with an accuracy of ±10 ms. The normal range of response latency is defined as 200 ms-1500 ms. This range is based on statistics from a large amount of real user interaction data and reflects the average time required for humans to receive instructions, understand them, and respond. Delays exceeding this range or exhibiting abnormally fixed latency will be judged as abnormal behavior, indicating potential deception attacks.

[0085] Furthermore, for speech fluency and fundamental frequency variability, we conduct in-depth analysis of the speech rate, pronunciation pause patterns, and natural variability of fundamental frequency profile and energy envelope in user response speech; among which, speech rate is measured in syllables per second or phonemes per second, and pronunciation pause patterns include the frequency, duration, and positional distribution of pauses in sentences.

[0086] The aforementioned behavioral characteristics of user response speech constitute the speaker's unique speech fingerprint. The analysis results are statistically compared with the baseline prosodic behavior model established by the user during registration or normal use. The baseline prosodic behavior model is constructed by recording multiple segments of the user's natural speech during the user registration phase and extracting the statistical distribution of these behavioral characteristics: mean, variance, skewness, and kurtosis. The comparison method can use the Dynamic Time Warping (DTW) algorithm to align the fundamental frequency and energy profile and calculate their difference to quantify the nonlinear time warping distance between the response speech and the baseline pattern. In addition, statistical model anomaly detection methods based on Gaussian mixture models (GMM) or deep autoencoders can be used to project the prosodic features of the responding speech onto the distribution of the baseline model and calculate its anomaly score. For synthesized or converted speech, due to the limitations of its generation mechanism, its prosodic pattern often exhibits rigid, overly smooth, or unnatural features when dealing with unknown text. For example, the fundamental frequency profile lacks natural fine-tuning variations, the energy envelope is too regular, and the pause distribution does not conform to language habits. This implementation method can effectively identify such anomalies, thereby enhancing the system's resistance to advanced synthetic deception.

[0087] In some specific implementations, step S4 is used for multi-evidence collaborative decision-making, which integrates the dynamic evidence from step S3 and the static evidence obtained from step S4 to generate the final authentication decision.

[0088] Specifically, a static evidence score is calculated, which includes the voiceprint matching similarity score. S and the probability P of living attribute live Voiceprint matching similarity score S By comparing the voiceprint identifier vector S with the pre-registered voiceprint template S template The similarity is calculated between them.

[0089] This embodiment uses cosine similarity as the main similarity calculation method, and its calculation formula is S. cosine = (S·S template ) / (||S||·||S template Cosine similarity can effectively measure the closeness of two vectors in direction and has a certain robustness to changes in vector magnitude. In addition, it can be combined with probabilistic linear discriminant analysis (PLDA) scoring. The PLDA model can provide more discriminative scores by modeling the variability within and between speakers.

[0090] Probability P of living attribute liveThe probability value is either directly output from the liveness attribute branch or converted from the liveness attribute vector L to a probability value between 0 and 1 through a calibration module. For example, an isotonic regression-based calibrator is used, which learns a monotonically non-decreasing function on the training set to map the liveness attribute vector L to a calibrated probability value, making it more accurately reflect the true probability.

[0091] Furthermore, if a dynamic challenge has been triggered, a dynamic evidence score is calculated based on the results of real-time behavioral analysis. D Dynamic Evidence Score D The calculation obtains a comprehensive confidence score by fusing response delay, speech fluency, and fundamental frequency variability. The fusion method can adopt a summation model, in which the weights of each behavioral feature are optimized according to their ability to identify spoofing attacks.

[0092] For example, a typical summation model could be Score D = w delay * Score delay + w fluency *Score fluency + w variability * Score variability Among them, Score delay Indicates the fusion response latency; Score fluency Score indicates speech fluency. variability Indicates the fundamental frequency variability; w delay w fluency w variability The weights for fusion response latency, speech fluency, and fundamental frequency variability are respectively summed to 1 and obtained by optimization on the training set through cross-validation.

[0093] Furthermore, a collaborative decision engine is used to integrate static and dynamic evidence scores to generate the final authentication decision. In this embodiment, the decision engine is a rule-based expert system that employs cascaded decision logic to achieve a balance between authentication efficiency and user experience while ensuring security.

[0094] First-level decision: If the voiceprint matching similarity score is... S The probability P of the liveness attribute is higher than the first preset threshold. live If the value exceeds the second preset threshold, the system determines that the authentication is successful. This is the safest path with the highest confidence level, and no additional user interaction is required.

[0095] Second-level decision: If the voiceprint matching similarity score is... S The probability P of the liveness attribute is higher than the first preset threshold.live If the score is between the third and second preset thresholds, it indicates good voiceprint matching but some uncertainty in the liveness confidence. In this case, a dynamic challenge is activated. After the user completes the dynamic challenge and the system analyzes the real-time behavioral characteristics, if the dynamic evidence score is... D If the value exceeds the fourth preset threshold, the authentication is deemed successful; otherwise, the authentication is deemed rejected. This approach enhances the defense capability against medium-risk scenarios by introducing dynamic challenges as a second factor for verification.

[0096] Third-level decision: If the voiceprint matching similarity score is... S Below the first preset threshold, or the probability P of the living attribute live If the value is below the third preset threshold, authentication is directly rejected without dynamic challenge. This approach directly rejects authentication for low-confidence matches or clear signs of non-liveness, thereby maximizing system security.

[0097] Through the aforementioned logical rules, the decision engine can achieve an intelligent balance between authentication efficiency and user experience while ensuring security. Dynamic challenges are triggered only when necessary, reducing unnecessary disturbance to users and improving the overall user experience. At the same time, it provides strong defense-in-depth for high-risk situations.

[0098] Example 2: This invention provides a voiceprint recognition and identity authentication system based on public safety management, comprising:

[0099] The signal acquisition and processing module is used to receive raw voice signals from a microphone array or audio input device and perform preprocessing operations.

[0100] The feature extraction module is used to extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalizes the three types of features and concatenates them into a joint feature representation.

[0101] The joint feature encoding module is used to encode the joint feature representation, generate a high-level shared feature vector, and generate a voiceprint identifier vector and a liveness attribute vector through the voiceprint identifier branch and the liveness attribute branch, respectively.

[0102] The dynamic challenge generation module is used to generate random number sequences or short sentence challenge content in real time and provide voice prompts to the user.

[0103] The real-time behavior analysis module is used to receive users' voice responses to dynamic challenges and analyze the real-time behavioral characteristics of the response voice.

[0104] The collaborative decision-making module is used to collect voiceprint matching similarity and liveness attribute probability from the joint feature encoding module, as well as dynamic evidence score from the real-time behavior analysis module, and generate authentication pass or authentication rejection instructions by fusing multi-source evidence according to predefined cascade decision logic and thresholds.

[0105] It should be noted that the voiceprint recognition and identity authentication system based on public safety management can implement the steps in the voiceprint recognition and identity authentication method based on public safety management in Embodiment 1 above, and can achieve the same technical effect. Refer to the description in Embodiment 1 above, which will not be elaborated here.

[0106] Example 3: Figure 4 A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0107] like Figure 4 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0108] Furthermore, the processor 11 can perform various appropriate actions and processes based on a computer program stored in the read-only memory (ROM) 12 or a computer program loaded from the storage unit 18 into the random access memory (RAM) 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0109] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0110] Furthermore, processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as methods for resource management of a database.

[0111] In some specific embodiments, the method for managing database resources can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the method for managing database resources described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the method for managing database resources by any other suitable means (e.g., by means of firmware).

[0112] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0113] Computer programs used to implement the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0114] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for resource management of a database provided by this invention. The computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0115] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0116] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0117] Optionally, the computing system may include clients and servers. Clients and servers are generally geographically separated and typically interact via a communication network. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. This addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0118] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0119] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A voiceprint recognition and identity authentication method based on public safety management, characterized in that, Includes the following steps: S1. Preprocess the input raw speech signal, extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalize the three types of features and concatenate them into a joint feature representation. S2. The joint feature representation is encoded by a shared encoder to output a high-level shared feature vector that contains both identity and liveness information. Voiceprint identification vector and liveness attribute vector are generated by the voiceprint identification branch and the liveness attribute branch, respectively. During model training, a correlation constraint loss term is introduced into the loss function to minimize the statistical dependence between the voiceprint identification vector and the liveness attribute vector and retain key coupling information. S3. When it is determined that a high level of security authentication is required, send dynamic challenge content to the user device and receive the user's response voice to the challenge. Analyze the real-time behavioral characteristics of the response voice to obtain a dynamic evidence score. S4. Calculate voiceprint matching similarity based on voiceprint identifier vector, calculate liveness attribute probability based on liveness attribute vector, fuse voiceprint matching similarity, liveness attribute probability and optional dynamic evidence score, and generate final authentication decision through collaborative decision rules.

2. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S1, the preprocessing includes: The original speech signal is pre-emphasized using a first-order digital filter, the transfer function of which is H(z) = 1 - αz. -1 The pre-weighting coefficient α ranges from 0.95 to 0.

98. -1 This is expressed as a delay of one sampling period; The pre-emphasized signal is divided into frames with a frame length of 20-30 ms and a frame shift of 8-12 ms. A Hamming window is added to each frame of signal to reduce spectral leakage.

3. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S1, the spectral structure feature extraction includes: performing a Fourier transform on each frame of speech signal to obtain the spectrum, then filtering it through a Mel filter bank and taking the logarithmic energy to obtain a Mel frequency cepstral; calculating the first-order and second-order difference coefficients of the Mel frequency cepstral to capture dynamic changes in the spectrum; The extraction of high-frequency details and phase features includes: extracting the sub-band energy distribution of the speech signal in the frequency range of 8-20 kHz; calculating the instantaneous phase derivative through Hilbert transform, or calculating the group delay feature, to detect phase discontinuities; The prosodic behavior feature extraction includes: estimating the fundamental frequency profile in the 50-500 Hz range; obtaining the energy envelope through short-time energy smoothing; and calculating short-time speech rate changes and pause frequencies and durations.

4. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S2, the voiceprint identification branch consists of a 2- to 3-layer fully connected neural network, with each layer of the fully connected neural network using the ReLU activation function; the output of the voiceprint identification branch is a 192-dimensional or 256-dimensional voiceprint identification vector; the training of the voiceprint identification branch uses additional corner margin loss or triplet loss as its main loss function.

5. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S2, the liveness attribute branch consists of 2 to 3 layers of fully connected neural networks, each layer of which uses the ReLU activation function. The output of the liveness attribute branch is a scalar with a dimension of 1, which is represented as the speech liveness attribute probability after passing through the Sigmoid activation function, or the output is a 64-dimensional or 128-dimensional liveness attribute vector. The training of the liveness attribute branch uses binary cross-entropy loss as its main loss function.

6. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S2, the correlation constraint loss term is implemented through a regularization term, which is defined as the square of the Frobenius norm of the empirical covariance matrix of the vector after linear mapping of the voiceprint identification vector and the liveness attribute vector, or as the average mutual information estimate between the mapped vectors.

7. The voiceprint recognition and identity authentication method based on public safety management according to claim 1, characterized in that: In step S3, the dynamic challenge content is a randomly generated sequence of several digits or a short phrase, which is converted into voice prompts by a text-to-speech engine; the real-time behavioral characteristics include response latency, speech fluency, and fundamental frequency variability, wherein the normal range of response latency is defined as 200-1500 ms.

8. A system based on any one of claims 1-7 for voiceprint recognition and identity authentication in public safety management, characterized in that, include: The signal acquisition and processing module is used to receive raw voice signals from a microphone array or audio input device and perform preprocessing operations. The feature extraction module is used to extract spectral structure features, high-frequency details and phase features, and prosodic behavior features in parallel, and then normalizes the three types of features and concatenates them into a joint feature representation. The joint feature encoding module is used to encode the joint feature representation, generate a high-level shared feature vector, and generate a voiceprint identifier vector and a liveness attribute vector through the voiceprint identifier branch and the liveness attribute branch, respectively. The dynamic challenge generation module is used to generate random number sequences or short sentence challenge content in real time and provide voice prompts to the user. The real-time behavior analysis module is used to receive users' voice responses to dynamic challenges and analyze the real-time behavioral characteristics of the response voice. The collaborative decision-making module is used to collect voiceprint matching similarity and liveness attribute probability from the joint feature encoding module, as well as dynamic evidence score from the real-time behavior analysis module, and generate authentication pass or authentication rejection instructions by fusing multi-source evidence according to predefined cascade decision logic and thresholds.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores a computer program executed by at least one of the processors, which enables the at least one processor to perform the voiceprint recognition and authentication method based on public security management as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the voiceprint recognition and authentication method based on public security management as described in any one of claims 1-7.

Images