A voiceprint recognition-based anti-fraud system and method

Abstract

The application discloses a kind of anti-fraud systems and methods based on voiceprint identification, it is related to voiceprint identification anti-fraud technical field.The system includes time effectiveness evaluation module, time effectiveness optimization and matching quantification module and feature matching precision tuning module.The application is by obtaining the voiceprint feature extraction data in the time window of speech signal acquisition, the time efficiency of feature extraction is quantified, and whether intelligent tuning of frame length and buffer queue threshold is judged according to this;Again based on voiceprint feature matching collaborative instruction obtains voiceprint feature alignment deviation degree, judges whether the precision tuning of feature frame change interval and tolerance window width is executed, finally outputs verification result output instruction, improves anti-fraud accuracy, solves the problem that the anti-fraud accuracy is low due to the time efficiency of voiceprint feature extraction and matching precision is not coordinated in the prior art.

Description

Technical Field

[0001] This invention relates to the field of voiceprint recognition anti-fraud technology, and in particular to an anti-fraud system and method based on voiceprint recognition. Background Technology

[0002] During the voiceprint registration phase, users record a specified voice message (such as "Registering voiceprint") through a microphone. The voice message is then preprocessed, including noise reduction to eliminate ambient noise and electrical interference, and endpoint detection to identify valid voice segments and remove silent or invalid parts.

[0003] Next, voiceprint features are extracted using algorithms such as MFCC (Mel-Frequency Cepstral Coefficients) and Fbank (Mel-scale Filter Bank Energies). Then, user voiceprint models are built using methods such as GMM (Gaussian Mixture Model), DNN (Deep Neural Network), or PLDA (Probabilistic Linear Discriminant Analysis). Finally, the models are stored in the database and bound to user identity information to provide basic data for subsequent verification.

[0004] During the voiceprint verification phase, when a user performs a sensitive operation (such as logging in or transferring money), the user's voice is collected in real time. The user speaks a specified sentence or free speech as prompted, and then the same voice preprocessing and feature extraction operations as in the registration phase are performed, including noise reduction, endpoint detection, and feature extraction. Next, the current voice features are compared with the registered voiceprint model. Common algorithms include cosine similarity, Euclidean distance, PLDA score, etc., and a similarity threshold is used to judge. If the similarity score is greater than or equal to the preset threshold, the verification is passed. If the score is lower than the threshold, it is judged as "mismatch". The system finally returns the verification result (success / failure) and records the log.

[0005] During the voiceprint recognition and anti-fraud determination stage, if verification fails, the system will trigger a risk detection mechanism. It will analyze the voiceprint recognition results and optionally perform multimodal information fusion, such as combining facial recognition and behavioral features (e.g., input rhythm, device fingerprint) for multi-factor verification. Subsequently, a fraud risk score will be generated based on indicators such as voiceprint similarity, environmental noise, and speech synthesis detection. If the risk score exceeds the threshold, it will be determined as fraudulent behavior; otherwise, it will be determined as misidentification or environmental interference. The system will execute corresponding processing strategies based on the determination results, such as blocking operations (prohibiting login, freezing accounts), secondary verification (requiring users to verify via SMS, face, or other methods), or manual review (triggering manual intervention investigation).

[0006] For example, the credit anti-fraud method, system, device, and computer-readable storage medium disclosed in Chinese invention patent announcement number CN109544324B include: when a credit verification call is detected, obtaining the risk level of the current verification customer and obtaining customer voice data collected during the credit verification call; determining a target anti-fraud strategy based on the risk level, wherein the target anti-fraud strategy is one or more of the following: voiceprint anti-fraud strategy, voice anti-fraud strategy, soft voice anti-fraud strategy, background sound anti-fraud strategy, and voice emotion anti-fraud strategy; and performing anti-fraud identification operations based on the customer voice data and the target anti-fraud strategy to obtain the anti-fraud identification result of the current verification customer.

[0007] For example, Chinese invention patent CN116433359A discloses a method, system, terminal device, and storage medium for credit card application anti-fraud based on robust voiceprint features, which includes: enhancing the robustness of voiceprint features compared to existing technologies, reducing the negative impact of environmental factors such as background noise on clustering results, and making the clustering results more accurate; and reducing the number of telephone recordings clustered together due to similar voices compared to existing technologies.

[0008] The above-mentioned technology has at least the following technical problems:

[0009] In the real-time voiceprint feature extraction process, the preprocessing and feature extraction of the voice signal need to be completed within a preset time, including noise reduction, endpoint detection and MFCC feature calculation. Although this can effectively suppress noise and improve feature quality, the computational complexity is high. When running on mobile or embedded devices, it may cause delays exceeding the preset time, affecting the user experience (such as users hanging up while waiting in telephone customer service scenarios).

[0010] Secondly, voiceprint verification relies on similarity comparison (such as cosine similarity and Euclidean distance). If the volume of the verification voice differs significantly from the registered voice (e.g., the user is near / away from the microphone), the norm change of the feature vector will lead to a lower score. Even if the voiceprint pattern is consistent, Euclidean distance treats all MFCC dimensions equally, but high-frequency dimensions are easily contaminated by noise, causing noise to dominate the distance calculation. Fixed thresholds cannot cover all modalities, resulting in an imbalance between false rejection rate and false release rate.

[0011] Finally, the feature mismatch between free speech and fixed sentences stems from spectral shifts and model biases caused by differences in text content. Fixed text registration lacks pronunciation diversity, resulting in the model failing to learn the speaker's global acoustic characteristics (such as vocal range variations and prosodic habits). During free speech verification, text-related acoustic variations (such as emphasis) are mistakenly identified as non-matching speakers, leading to low anti-fraud accuracy due to the mismatch between the time efficiency of voiceprint feature extraction and matching accuracy. Summary of the Invention

[0012] To address the problem of low anti-fraud accuracy in existing technologies due to the mismatch between the time efficiency and matching accuracy of voiceprint feature extraction, this invention provides an anti-fraud system and method based on voiceprint recognition. The technical solution is as follows:

[0013] On one hand, a voiceprint recognition-based anti-fraud system is provided, including: a timeliness assessment module, a timeliness optimization and matching quantification module, and a feature matching accuracy tuning module. The timeliness assessment module acquires voiceprint feature extraction data within the speech signal acquisition time window to reflect the time efficiency during voiceprint feature extraction. Based on the acquired voiceprint feature extraction data, a timeliness result is obtained to quantify the time efficiency of the voiceprint feature extraction process. The timeliness optimization and matching quantification module determines whether to perform intelligent timeliness tuning based on the timeliness result. If so, it sends a voiceprint feature matching coordination instruction after performing intelligent timeliness tuning; otherwise, it directly sends the voiceprint feature matching coordination instruction. The matching coordination instruction is used to reflect the feature matching data of the voiceprint feature matching accuracy, and obtains the voiceprint feature alignment deviation to quantify the degree of alignment error in the voiceprint feature alignment process. The timeliness intelligent optimization means that the frame length and the frame generation buffer queue threshold are adjusted in combination with the timeliness result. The feature matching accuracy optimization module is used to determine whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation. If so, a verification result output instruction is sent after performing feature matching accuracy optimization. Otherwise, a verification result output instruction is sent directly. Feature matching accuracy optimization means that the feature frame change range and the tolerance window width are adjusted in combination with the voiceprint feature alignment deviation. The verification result output instruction includes matching success and matching failure.

[0014] On the other hand, a voiceprint recognition-based anti-fraud method is provided, comprising: acquiring voiceprint feature extraction data within a speech signal acquisition time window to reflect the time efficiency of the voiceprint feature extraction process; obtaining a timeliness result based on the acquired voiceprint feature extraction data to quantify the time efficiency of the voiceprint feature extraction process; determining whether to perform timeliness intelligent optimization based on the timeliness result; if so, sending a voiceprint feature matching collaboration instruction after performing timeliness intelligent optimization; otherwise, directly sending a voiceprint feature matching collaboration instruction, wherein the voiceprint feature matching collaboration instruction is used to reflect the feature matching data reflecting the accuracy of voiceprint feature matching. The voiceprint feature alignment deviation is used to quantify the degree of alignment error in the voiceprint feature alignment process. The timeliness intelligent optimization means adjusting the frame length and frame generation buffer queue threshold based on the timeliness result. It is determined whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation. If so, a verification result output command is sent after performing feature matching accuracy optimization. Otherwise, a verification result output command is sent directly. The feature matching accuracy optimization means adjusting the feature frame change range and tolerance window width based on the voiceprint feature alignment deviation. The verification result output command includes matching success and matching failure.

[0015] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0016] 1. By acquiring voiceprint feature extraction data within the speech signal acquisition time window, the system reflects the time efficiency of the voiceprint feature extraction process. Based on the acquired voiceprint feature extraction data, the system obtains timeliness results to quantify the time efficiency of the voiceprint feature extraction process, enabling accurate evaluation of the real-time performance of the voiceprint feature extraction process. This provides data support for subsequent dynamic optimization of frame length and buffer queue threshold, thereby improving system response speed and resource utilization efficiency. The system determines whether to perform intelligent timeliness optimization based on the timeliness results, dynamically optimizing key parameters such as frame length and buffer queue in the voiceprint feature extraction process. This significantly improves the system's real-time response capability and computational resource utilization efficiency while ensuring feature quality. The system also quantifies the alignment error degree in the voiceprint feature alignment process based on the obtained voiceprint feature alignment deviation, providing a precise error feedback mechanism for the voiceprint recognition system and improving feature alignment accuracy. Finally, the system determines whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation, enabling dynamic optimization of the matching strategy, improving system operating efficiency and resource utilization, and ultimately enhancing anti-fraud accuracy.

[0017] 2. Based on the timeliness results, determine whether to perform timeliness-based intelligent optimization. Timeliness-based intelligent optimization includes determining whether to perform speech frame length optimization for silent segments based on the relative inter-frame energy difference. For detected silent segments, monitor their relative inter-frame energy difference to reflect the stability of internal energy fluctuations. This can effectively distinguish between real silence and low-energy speech or noise interference, thereby improving the accuracy of silent segment recognition and providing a precise basis for frame length optimization, avoiding mis-adjustment or omission, and ensuring the efficiency of voiceprint feature extraction. For detected speech segments, monitor their frequency change rate to reflect how fast the speech signal frequency changes over time. This can effectively capture the changing trend of dynamic speech features and provide a basis for adaptive adjustment of frame length and feature parameters during the voiceprint feature extraction process, thereby improving the system's adaptability to complex speech scenarios and the accuracy of feature representation. Additionally, determine whether to perform speech signal sampling frequency optimization based on the total number of speech frames. This can dynamically adapt to speech inputs of different lengths or complexities, optimize the sampling frequency to balance system processing speed and feature extraction accuracy, thereby improving anti-fraud accuracy.

[0018] 3. Determine whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation. Feature matching accuracy optimization includes determining whether to perform dynamic optimization of the feature frame change range based on the voiceprint feature alignment timestamp deviation. This can adaptively adjust the selection range of feature frames and matching strategies, thereby effectively improving the temporal alignment accuracy of voiceprint feature matching and enhancing the recognition accuracy of the system in complex speech environments. It also includes determining whether to perform dynamic optimization of the tolerance window width based on feature frame changes. This can flexibly adjust the time tolerance range in the matching process, thereby further optimizing the fault tolerance capability of voiceprint feature matching and improving the accuracy of anti-fraud. Attached Figure Description

[0019] Figure 1 A schematic diagram of the structure of an anti-fraud system based on voiceprint recognition provided in an embodiment of this application;

[0020] Figure 2 A flowchart illustrating the voice signal sampling frequency optimization of an anti-fraud system based on voiceprint recognition, provided in an embodiment of this application;

[0021] Figure 3 A flowchart illustrating the dynamic optimization of the feature frame variation range of an anti-fraud system based on voiceprint recognition, provided in an embodiment of this application;

[0022] Figure 4 A flowchart illustrating an anti-fraud method based on voiceprint recognition provided in this application embodiment. Detailed Implementation

[0023] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0024] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0025] This application provides a voiceprint recognition-based anti-fraud system and method, which solves the problem of low anti-fraud accuracy caused by the mismatch between the time efficiency and matching accuracy of voiceprint feature extraction in the prior art. By acquiring voiceprint feature extraction data within the speech signal acquisition time window, the system quantifies the time efficiency of feature extraction and determines whether to perform intelligent optimization of frame length and buffer queue threshold. Then, based on the voiceprint feature matching collaborative instruction, the system obtains the voiceprint feature alignment deviation and determines whether to perform precision optimization of feature frame change interval and tolerance window width. Finally, the system outputs the verification result output instruction, thereby improving the accuracy of anti-fraud.

[0026] The technical solution in this application embodiment aims to address the problem of low anti-fraud accuracy caused by the mismatch between the time efficiency and matching accuracy of voiceprint feature extraction. The overall approach is as follows:

[0027] By acquiring voiceprint feature extraction data within the voice signal acquisition time window, the time efficiency of the voiceprint feature extraction process is reflected. Based on the acquired voiceprint feature extraction data, a timeliness result is obtained to quantify the time efficiency of the voiceprint feature extraction process. Based on the timeliness result, it is determined whether to perform timeliness intelligent optimization. Based on the voiceprint feature alignment deviation, it is determined whether to perform feature matching accuracy optimization. If so, a verification result output command is sent after performing feature matching accuracy optimization; otherwise, a verification result output command is sent directly to improve the accuracy of anti-fraud.

[0028] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0029] This invention provides an anti-fraud system based on voiceprint recognition, such as... Figure 1 The diagram shows a structural schematic of an anti-fraud system based on voiceprint recognition. The system may include: a timeliness assessment module, a timeliness optimization and matching quantification module, and a feature matching accuracy tuning module.

[0030] As the first module of a voiceprint recognition-based anti-fraud system, the timeliness evaluation module acquires voiceprint feature extraction data within the speech signal acquisition time window to reflect the time efficiency during the voiceprint feature extraction process. Based on the acquired voiceprint feature extraction data, a timeliness result is obtained to quantify the time efficiency of the voiceprint feature extraction process. Specifically, the voiceprint feature extraction data includes: inter-frame processing interval, voiceprint feature sampling rate, and voiceprint feature extraction duration.

[0031] It needs to be explained that the time interval between two adjacent frames of speech signal is recorded by a high-precision timer to calculate the inter-frame processing interval; the voiceprint feature sampling rate is obtained by MFCC (Mel-Frequency Cepstral Coefficients); and the time difference from the input speech frame to the output feature vector is recorded by a high-precision timer to calculate the voiceprint feature extraction time.

[0032] It should be noted that the timeliness result represents the result obtained by associating the first, second, and third timeliness components, where associativity means performing an addition operation. The first timeliness component represents the result obtained by interactively processing the inter-frame processing interval factor and the inter-frame processing interval score; the second timeliness component represents the result obtained by interactively processing the voiceprint feature sampling rate factor and the voiceprint feature sampling rate score; and the third timeliness component represents the result obtained by interactively processing the voiceprint feature extraction duration factor and the voiceprint feature extraction duration score. The inter-frame processing interval score represents the result of inversely proportional processing of the relative deviation between the inter-frame processing interval reference value and the inter-frame processing interval. The explanation is as follows: the result of the relative deviation inverse proportional processing refers to the ratio of the absolute value of the inter-frame processing interval reference value to the absolute value of the difference between the inter-frame processing interval and the reference value. The voiceprint feature sampling rate score represents the result of the ratio processing of the voiceprint feature sampling rate to the voiceprint feature sampling rate reference value, where the ratio processing represents a division operation. The voiceprint feature extraction time score represents the result of the relative deviation inverse proportional processing of the voiceprint feature extraction time reference value and the voiceprint feature extraction time, where the interaction processing represents a multiplication operation.

[0033] It should be noted that the inter-frame processing interval factor, voiceprint feature sampling rate factor, voiceprint feature extraction duration factor, inter-frame processing interval reference value, voiceprint feature sampling rate reference value, and voiceprint feature extraction duration reference value are all obtained from the anti-fraud database.

[0034] It is important to note that there are close interrelationships and synergies among the voiceprint feature extraction data. First, the voiceprint feature sampling rate determines the temporal discretization density of the original speech signal. A higher sampling rate means more frames can be obtained, allowing for finer inter-frame processing interval settings, enabling feature extraction to capture more subtle speech changes. However, this also means a larger data volume, placing higher demands on computing and storage resources. The inter-frame processing interval directly affects the continuity and redundancy of the feature sequence. A shorter inter-frame processing interval results in a denser feature sequence, which helps improve the ability to capture dynamic speech changes, but also increases computational complexity. The voiceprint feature extraction duration, i.e., the total speech duration required to generate a stable voiceprint feature, must be long enough to cover a sufficient number of speech frames to ensure the robustness of the feature. The voiceprint feature extraction duration and the inter-frame processing interval together determine the total number of frames used for feature extraction, thus affecting the statistical stability and recognition performance of the feature. There are clear positive and negative correlations between the inter-frame processing interval score, the voiceprint feature sampling rate score, and the voiceprint feature extraction time score and the timeliness result: the inter-frame processing interval score is negatively correlated with the timeliness result, that is, the longer the inter-frame processing interval score, the fewer speech frames are processed per unit time, the slower the system response speed, and the lower the timeliness result; the voiceprint feature sampling rate score is positively correlated with the timeliness result, that is, the higher the voiceprint feature sampling rate score, the richer the voiceprint feature information obtained per unit time, which helps to improve the accuracy and real-time performance of feature extraction, thereby improving the timeliness result; while the voiceprint feature extraction time score is negatively correlated with the timeliness result, that is, the longer the feature extraction time, the greater the overall system processing delay, and the lower the timeliness result accordingly.

[0035] As the second module of an anti-fraud system based on voiceprint recognition, the timeliness optimization and matching quantification module is used to determine whether to perform timeliness intelligent tuning based on the timeliness results. If so, a voiceprint feature matching coordination instruction is sent after performing timeliness intelligent tuning; otherwise, a voiceprint feature matching coordination instruction is sent directly. The voiceprint feature matching coordination instruction is used to reflect the feature matching data that reflects the accuracy of voiceprint feature matching, and to obtain the voiceprint feature alignment deviation to quantify the degree of alignment error in the voiceprint feature alignment process. Timeliness intelligent tuning means adjusting the frame length and the frame generation buffer queue threshold in combination with the timeliness results. The feature matching data specifically includes: formant frequency offset, speech rate matching deviation, and timeliness result to be tested. The timeliness result to be tested indicates that if intelligent timeliness optimization has been performed, the timeliness result re-acquired after intelligent timeliness optimization will be recorded as the timeliness result to be tested; otherwise, the current timeliness result will be directly recorded as the timeliness result to be tested. The formant frequency offset is calculated by aligning the formant frequency positions of the free speech and fixed sentence voiceprint features in the spectral dimension through dynamic time warping. The speech rate matching deviation is calculated by aligning the speech rate difference between the free speech and fixed sentence in the time dimension through dynamic time warping.

[0036] It needs to be explained that the voiceprint feature alignment deviation degree represents the result obtained by associating the first, second, and third components of the alignment deviation degree, where associativity refers to addition. The first component of the alignment deviation degree represents the result obtained by interactive processing of the formant frequency offset factor and the formant frequency offset score; the second component of the alignment deviation degree represents the result obtained by interactive processing of the speech rate matching deviation factor and the speech rate matching deviation score; and the third component of the alignment deviation degree represents the result obtained by interactive processing of the timeliness result factor and the timeliness result score to be tested. The formant frequency offset score represents the result of processing the relative deviation ratio between the formant frequency offset and the set formant frequency offset value, where the relative deviation ratio is the ratio of the absolute value of the difference between the formant frequency offset and the set formant frequency offset value to the set formant frequency offset value. The speech rate matching deviation score represents the result of processing the relative deviation ratio between the speech rate matching deviation and the set speech rate matching deviation value. The timeliness result score to be tested represents the result of processing the timeliness result set value as a percentage. Interactive processing refers to multiplication, and percentage processing refers to division.

[0037] It should be noted that the formant frequency offset factor, speech rate matching deviation factor, timeliness result factor, formant frequency offset setting value, speech rate matching deviation setting value, and timeliness result setting value are all obtained from the anti-fraud database.

[0038] It should be noted that there is a significant positive and negative correlation between the feature matching data. Specifically, when the formant frequency shift increases, it indicates that the formant positions of the free speech and the fixed sentence differ significantly in the spectral dimension, usually reflecting unstable pronunciation or mismatch of voiceprint features. In this case, it is often accompanied by an increase in speech rate matching deviation, that is, the difference in speech rate between the free speech and the fixed sentence in the time dimension intensifies. Both of these factors together lead to a decrease in the timeliness result of the test, that is, a decrease in the real-time processing and matching performance of the system. Conversely, when both the formant frequency shift and the speech rate matching deviation are small, it indicates that the speech signal is highly consistent with the standard template in both the spectral and time dimensions, and the timeliness result of the test is improved accordingly. The system exhibits better real-time response and matching efficiency. There are clear positive and negative correlations between the formant frequency offset score, speech rate matching deviation score, and the timeliness result score and the voiceprint feature alignment deviation: the formant frequency offset is positively correlated with the voiceprint feature alignment deviation, that is, the larger the offset, the more significant the frequency difference between the voiceprint feature and the reference template, and the alignment deviation increases accordingly; the speech rate matching deviation is positively correlated with the voiceprint feature alignment deviation, that is, the larger the speech rate deviation, the more serious the temporal structure mismatch of the speech signal, leading to an increase in the feature alignment deviation; while the timeliness result is negatively correlated with the voiceprint feature alignment deviation, that is, the higher the timeliness result, the better the system processing speed and response efficiency, which helps to reduce the cumulative error in the feature extraction and alignment process, thereby reducing the alignment deviation.

[0039] It's important to understand that the decision to perform intelligent timeliness optimization based on the timeliness results follows these steps: If the timeliness result is greater than or equal to the upper bound of the timeliness result reference, it indicates that the processing efficiency under the current configuration has reached or exceeded the expected target. In this case, intelligent timeliness optimization is not performed to avoid unnecessary system disturbances, ensure stability, and send a voiceprint feature matching coordination command. Intelligent timeliness optimization includes critical-level optimization and inefficient-level optimization. If the timeliness result is within the timeliness result reference range, it indicates a "critical state," meaning that optimal performance has not yet been achieved, but there is no significant lag. In this case, critical-level optimization is performed to improve system efficiency through fine-tuning and prevent slipping into an inefficient state. The timeliness result reference range is used to characterize the open interval formed by the lower and upper bounds of the timeliness result reference. Critical-level optimization includes determining whether to perform silence segment speech frame length optimization. If so, it is performed after silence segment speech frame length optimization, and then it is determined whether to perform speech segment speech frame length optimization. Otherwise, it is directly determined whether to perform speech segment speech frame length optimization. If the timeliness result is less than or equal to the lower bound of the timeliness result reference, it indicates that the system processing efficiency is seriously insufficient, and there may be frame accumulation, excessive latency, or resource bottleneck. In this case, inefficient-level tuning is performed. The purpose is to quickly restore system performance through deep adjustment. Inefficient-level tuning includes determining whether to perform frame generation buffer queue threshold range tuning. If so, after adjusting the frame generation buffer queue threshold, it is determined whether to perform voice signal sampling frequency tuning. If not, it is directly determined whether to perform voice signal sampling frequency tuning.

[0040] As a further explanation, the specific steps for determining whether to perform silence segment speech frame length optimization are as follows:

[0041] Speech activity detection is enabled to receive input speech signals in real time and call the speech activity detection algorithm to analyze the speech signals frame by frame. The input speech signals are monitored in real time to distinguish between silence segments and speech segments. For the detected silence segments, the relative inter-frame energy difference is monitored to reflect the stability of energy fluctuations within the silence segment.

[0042] If the relative inter-frame energy difference is lower than the first energy threshold, the current silence segment is determined to be a stable silence segment, and a first frame length adjustment instruction is generated to adjust the current speech frame length to the first target frame length. The specific adjustment steps are as follows: the timeliness result and energy difference correction amount are queried from the constructed inefficient level optimization mapping table to obtain the first frame length mapping value. Based on the obtained first frame length mapping value, the current speech frame length is incrementally reduced to avoid the impact of drastic frame length changes on system stability, ensure that the adjustment process is smooth and controllable, and at the same time realize the gradual release of computing resources to obtain the first target frame length. It should be explained that incremental reduction refers to the adjustment method of reducing the current speech frame length multiple times in a step-by-step manner based on the first frame length mapping value, with a preset fixed step size or dynamic adaptation step size, until the first target frame length is reached. Its core feature is to avoid a large-scale adjustment of the frame length all at once. Instead, it achieves a smooth transition from the current value to the target value of the frame length through a phased, gradual reduction. This adapts to the low-energy, low-dynamic characteristics of stable silence segments while ensuring the stability of the speech processing flow. Let the current speech frame length be P0, the first target frame length be Pn, and the single step size corresponding to the first mapping value of the frame length be ΔP (ΔP is greater than 0). First adjustment: Update the current frame length to P1 = P0 - ΔP. At this time, it is necessary to determine whether P1 is less than or equal to Pn. If not, proceed to the next adjustment. Repeated adjustment: Calculate P2 = P1 - ΔP, P3 = P2 - ΔP in sequence until the m-th adjustment yields Pm, which satisfies that Pm is less than or equal to Pn. At this point, Pm is determined as the final first target frame length (if Pm is less than Pn, it can be precisely equal to Pn through a final fine-tuning). The energy difference correction represents the difference between the relative inter-frame energy difference and the first energy threshold.

[0043] If the relative inter-frame energy difference is higher than the second energy threshold, the current silence segment is determined to be a sudden silence segment, and a second frame length adjustment instruction is generated to adjust the current speech frame length to the second target frame length. The specific adjustment steps are as follows: the timeliness result and energy difference offset are queried from the constructed inefficient-level optimization mapping table to obtain the second frame length mapping value. Based on the obtained second frame length mapping value, the current speech frame length is incrementally increased to balance the adjustment speed and system stability, and to avoid signal distortion or processing errors caused by sudden changes in frame length. It should be explained that incremental increase refers to the multiple, step-by-step increase of the current speech frame length based on the second frame length mapping value, with a preset fixed step size or a dynamically adapted step size, until the second target frame length is reached. Let the current speech frame length be A0, the second target frame length be At, and the single step size corresponding to the second frame length mapping value be ΔA (ΔA>0). Initial adjustment: The current frame length is updated to A1 = A0 + ΔA. At this point, it is necessary to determine whether A1 is greater than or equal to At. If not, the next adjustment is performed, calculating A2 = A1 + ΔA, A3 = A2 + ΔA, and so on, until An is obtained after the nth adjustment, satisfying that An is greater than or equal to At. At this point, An is determined as the final second target frame length (if An is greater than At, it can be precisely equal to At through a final fine-tuning). This achieves intelligent identification and frame length optimization of sudden silence segments. Furthermore, through a smooth and controllable adjustment mechanism, the system's efficient and stable operation is ensured, significantly improving the voiceprint processing system's ability to capture sudden signals and its robustness. The energy difference offset represents the difference between the relative inter-frame energy difference and the second energy threshold.

[0044] If the relative inter-frame energy difference is within the energy threshold range, it indicates that the current silence segment is neither a stable silence segment nor a sudden silence segment, but is in a state of moderate and relatively stable energy fluctuation. In this case, silence segment speech frame length optimization is not performed, and the current speech frame length is kept unchanged to avoid system overhead and performance fluctuations caused by frequent frame length adjustments. The energy threshold range represents the closed interval formed by the first energy threshold and the second energy threshold.

[0045] It should be noted that by enabling voice activity detection to perform real-time frame-by-frame analysis of the input speech signal, the system accurately distinguishes between silence segments and speech segments, and monitors the relative inter-frame energy difference for silence segments to assess the stability of internal energy fluctuations. Through refined silence segment classification, intelligent frame length adjustment strategies, and a smooth incremental optimization mechanism, the system's adaptability to various silence states, feature extraction quality, and operating efficiency are significantly improved, while effectively balancing system resource consumption and processing performance.

[0046] As a further explanation, the specific steps for determining whether to perform speech segment frame length optimization are as follows:

[0047] For the detected speech segment, its frequency change rate is monitored to reflect how fast the frequency of the speech signal changes over time.

[0048] If the frequency change rate is higher than the first change rate threshold, the current speech segment is determined to be a high dynamic range speech segment, and a third frame length adjustment instruction is generated to adjust the current speech frame length to the third target frame length. The specific adjustment steps are as follows: the timeliness result and the change rate offset are queried from the constructed inefficient level optimization mapping table to obtain the third frame length mapping value. Based on the obtained third frame length mapping value, the current speech frame length is incrementally reduced. Incremental reduction means that the current speech frame length is not directly adjusted to the final expected third target frame length all at once, but is gradually reduced according to a certain step size until the third target frame length is reached. For example, assuming the current speech frame length is L0 and the adjustment step size corresponding to the third mapping value of the frame length is L, the current speech frame length is initially adjusted to L1=L0-L, and then adjusted again to L2=L1-L, and so on. After multiple adjustments, it is adjusted to the third target frame length Ln that meets the requirements, thus obtaining the third target frame length. This achieves rapid response and accurate processing of speech segments with high dynamic characteristics. It also avoids system oscillation through a smooth adjustment mechanism, optimizes resource utilization, and enhances the robustness and adaptability of the system. The rate of change offset represents the difference between the frequency change rate and the first rate of change threshold.

[0049] If the frequency change rate is lower than the second change rate threshold, the current speech segment is determined to be a slowly changing speech segment, and a fourth frame length adjustment instruction is generated to adjust the current speech frame length to the fourth target frame length. The specific adjustment steps are as follows: the timeliness result and the change rate correction amount are queried from the constructed inefficient level optimization mapping table to obtain the fourth frame length mapping value. Based on the obtained fourth frame length mapping value, the current speech frame length is incrementally increased. This means that based on the fourth frame length mapping value, the current speech frame length is increased multiple times in a stepwise manner with a preset fixed step size or a dynamic step size determined according to the change rate correction amount, until the fourth target frame length is reached. The core idea is to avoid increasing the frame length drastically all at once. Instead, it is to gradually increase the frame length in stages to achieve a smooth transition from the current value to the target value, so as to adapt to the characteristics of slowly changing speech segments. Let the current speech frame length be K0, the fourth target frame length be Kt, and the single adjustment step size corresponding to the fourth mapping value of the frame length be ΔK (ΔK>0). Update the current frame length to K1=K0+ΔK, and determine whether K1 is greater than or equal to Kt. If not, proceed to the next adjustment. Calculate K2=K1+ΔK, K3=K2+ΔK, and so on, until the nth adjustment is completed and Kn is obtained, which satisfies that Kn is greater than or equal to Kt. At this point, Kn is determined as the fourth target frame length (if Kn is greater than Kt, Kn can be precisely equal to Kt through a final fine-tuning) to obtain the fourth target frame length, in order to improve the integrity of low-frequency features, adapt to the slow-changing characteristics, and ensure the coherence and efficiency of speech processing to cover complete features. However, excessively increasing the frame length will result in too few frames per unit time, which may lead to the loss of subtle change information. The rate difference offset represents the difference between the frequency change rate and the second change rate threshold.

[0050] If the frequency change rate is within the change rate threshold range, and its frequency change rate is at a moderate level, it does not belong to the high-frequency, rapidly changing, high-dynamic speech segment, nor to the low-frequency, stable, slowly changing speech segment. Therefore, speech segment frame length optimization is not performed, and the current speech frame length is kept unchanged to avoid invalid adjustments and reduce computational redundancy. The change rate threshold range represents the closed interval formed by the second change rate threshold and the first change rate threshold.

[0051] It should be noted that in speech segment processing, by monitoring the frequency change rate of the speech signal in real time, fine-grained classification and intelligent frame length optimization of the dynamic characteristics of speech segments are achieved. Through intelligent classification, dynamic mapping table lookup, and incremental optimization mechanisms, accurate adaptation and efficient processing of speech segments with different dynamic characteristics are achieved, significantly improving the feature extraction quality, real-time response capability, and resource utilization efficiency of the voiceprint recognition system, while also ensuring the system's stability and adaptability.

[0052] As a further explanation, the specific steps for determining whether to perform frame generation buffer queue threshold range optimization are as follows:

[0053] The decision to perform frame generation buffer queue threshold range optimization is based on the frame generation rate. The frame generation rate reflects the frame generation frequency within the speech signal acquisition time window. The specific steps are as follows: If the frame generation rate exceeds the set upper limit, the timeliness result and generation rate offset are retrieved from the constructed inefficient-level optimization mapping table to obtain the upper limit mapping value of the frame generation buffer queue. Based on the obtained upper limit mapping value, the upper limit value of the frame generation buffer queue is scaled. The scaling process involves multiplying the upper limit mapping value of the frame generation buffer queue with the upper limit value of the frame generation buffer queue using the scaling factor as a proportionality coefficient. This results in a dynamic upper limit value for the frame generation buffer queue. The proportionality coefficient refers to the frame generation buffer queue upper limit value obtained through the inefficient-level optimization mapping table. The generation buffer queue upper limit mapping value is a dimensionless parameter used to scale and adjust the upper limit value of the current frame generation buffer queue. This coefficient is jointly determined by the timeliness result and the generation rate offset. Its value reflects the adjustment intensity of the current system load on the queue upper limit. The calculation formula for the dynamic frame generation buffer queue upper limit value is: Dynamic frame generation buffer queue upper limit value = Current frame generation buffer queue upper limit value × proportional coefficient. This achieves intelligent balancing of system load and efficient optimization of resource utilization, avoids processing delays or data loss caused by frame generation overload, and enhances the real-time response capability and stability of the system. The generation rate offset represents the difference between the frame generation rate and the set upper limit of the generation rate.

[0054] If the frame generation rate is less than the set lower limit of the generation rate, the timeliness result and the generation rate correction amount are retrieved from the constructed inefficiency-level optimization mapping table to obtain the lower limit mapping value of the frame generation buffer queue. Based on the obtained lower limit mapping value, a weighted correction process is performed on the lower limit value of the frame generation buffer queue. The weighted correction process refers to the calculation method of adjusting the current lower limit value of the frame generation buffer queue based on the retrieved lower limit mapping value to obtain a dynamic lower limit value of the frame generation buffer queue. Its core is to use the lower limit mapping value of the frame generation buffer queue as a weight factor to scale and adjust the current lower limit value of the queue, thereby achieving adaptive optimization of the lower limit of the queue and obtaining a dynamic lower limit value of the frame generation buffer queue. This achieves reasonable allocation of system resources and optimization of processing efficiency, avoids idle processing resources or task delays caused by slow frame generation, and improves the system's response flexibility and operational stability. The generation rate correction amount represents the difference between the frame generation rate and the set lower limit of the generation rate.

[0055] If the frame generation rate is within the set range, the frame generation buffer queue threshold range optimization will not be performed. This avoids introducing system jitter or computational redundancy due to unnecessary dynamic adjustments, ensuring that the system maintains an efficient and stable operating state under moderate load. The set range represents the closed interval formed by the lower limit and upper limit of the set generation rate.

[0056] It should be noted that by monitoring the frame generation rate in real time and determining whether it exceeds the set range, when the rate is too high, the upper limit of the queue is dynamically increased by proportional scaling to prevent overload; when the rate is too low, the lower limit of the queue is dynamically decreased by weighted correction to avoid resource idleness; and within the normal range, the queue threshold remains unchanged, thereby achieving adaptive balance of system load, efficient utilization of resources, and stable and controllable operation.

[0057] like Figure 2 The diagram shows a flowchart of a voice signal sampling frequency optimization process for an anti-fraud system based on voiceprint recognition, provided in an embodiment of this application. The specific logic is as follows: First, determine if the total number of voice frames exceeds the upper limit of the frame generation buffer queue. If so, query the timeliness result and queue overflow deviation from the constructed inefficient-level optimization mapping table to obtain a first mapping value for the voice signal sampling frequency. Based on the obtained first mapping value, scale down the current voice signal sampling frequency to obtain an adjusted value. Second, determine if the total number of voice frames is less than the lower limit of the frame generation buffer queue. If so, query the timeliness result and queue overflow offset from the constructed inefficient-level optimization mapping table to obtain a second mapping value for the voice signal sampling frequency. Based on the obtained second mapping value, scale up the current voice signal sampling frequency to obtain a corrected value. Third, determine if the total number of voice frames is within the threshold range of the frame generation buffer queue to be compared. If so, do not perform voice signal sampling frequency optimization.

[0058] As a further explanation, the specific steps for determining whether to perform voice signal sampling frequency optimization are as follows:

[0059] If frame generation buffer queue threshold interval optimization is performed, the frame generation buffer queue threshold interval is re-acquired after the optimization and recorded as the frame generation buffer queue threshold interval to be compared. Otherwise, the current frame generation buffer queue threshold interval is directly recorded as the frame generation buffer queue threshold interval to be compared. The frame generation buffer queue threshold interval to be compared represents the closed interval formed by the lower limit and the upper limit of the frame generation buffer queue.

[0060] The decision to perform speech signal sampling frequency optimization is based on the total number of speech frames. The total number of speech frames reflects the total number of speech frames waiting to be processed in the frame generation buffer queue. The specific steps are as follows: If the total number of speech frames exceeds the upper limit set for the frame generation buffer queue, the timeliness result and queue overflow deviation are retrieved from the constructed inefficient-level optimization mapping table to obtain the first mapping value of the speech signal sampling frequency. Based on the obtained first mapping value, the current speech signal sampling frequency is proportionally scaled down. This proportional scaling down process refers to adjusting the speech signal sampling frequency based on the value retrieved from the inefficient-level optimization mapping table. The first frequency mapping value is used to perform a multiplicative scaling operation on the current speech signal sampling frequency, and this mapping value is less than 1, thereby reducing the speech signal sampling frequency and the speech frame generation rate to alleviate the congestion of the frame generation buffer queue. This process is a linear proportional adjustment with lossless, reversible, and dynamically adaptive technical characteristics. It obtains the speech signal sampling frequency adjustment value, thereby dynamically reducing the speech frame generation rate, effectively alleviating the queue congestion, and ensuring that the system can still maintain stable operation and efficient resource utilization under high load. The queue overflow deviation represents the difference between the total number of speech frames and the set upper limit of the frame generation buffer queue.

[0061] If the total number of speech frames is less than the lower limit set for the frame generation buffer queue, the timeliness result and the queue overflow offset are retrieved from the constructed inefficiency-level optimization mapping table to obtain the second mapping value of the speech signal sampling frequency. Based on the obtained second mapping value of the speech signal sampling frequency, the current speech signal sampling frequency is scaled up. This means that based on the second mapping value of the speech signal sampling frequency obtained from the inefficiency-level optimization mapping table, a multiplicative scaling operation is performed on the current speech signal sampling frequency, and the mapping value is greater than 1, thereby increasing the speech signal sampling frequency to increase the speech frame generation rate, alleviate the idle state of the frame generation buffer queue, and obtain the speech signal sampling frequency correction value, thereby dynamically improving the speech frame generation rate and effectively alleviating the queue idle state. The queue overflow offset represents the difference between the total number of speech frames and the lower limit set for the frame generation buffer queue.

[0062] If the total number of voice frames is within the threshold range of the frame generation buffer queue, voice signal sampling frequency optimization will not be performed to avoid introducing system jitter or computational redundancy due to unnecessary dynamic adjustments, and to ensure that the system maintains an efficient and stable operating state under moderate load.

[0063] It should be noted that by dynamically adjusting the threshold range of the frame generation buffer queue and the audio signal sampling frequency, precise control and adaptive balance between the audio frame generation rate and the queue load are achieved. The overall mechanism possesses lossless, reversible, linear adjustment, and dynamic adaptive characteristics, ensuring that the system maintains efficient and stable operation under different load conditions, achieving optimized resource utilization and continuous stability of system performance.

[0064] As the third module of a voiceprint recognition-based anti-fraud system, the feature matching accuracy tuning module is used to determine whether to perform feature matching accuracy tuning based on the voiceprint feature alignment deviation. If so, a verification result output instruction is sent after performing feature matching accuracy tuning; otherwise, a verification result output instruction is sent directly. Feature matching accuracy tuning means adjusting the feature frame variation range and the tolerance window width in combination with the voiceprint feature alignment deviation. The verification result output instruction includes matching success and matching failure.

[0065] It should be noted that the determination of whether to perform feature matching accuracy optimization is based on the voiceprint feature alignment deviation. The specific determination steps are as follows: If the voiceprint feature alignment deviation is lower than or equal to the voiceprint feature alignment deviation reference value, feature matching accuracy optimization is not performed. If the voiceprint feature alignment deviation is higher than the voiceprint feature alignment deviation reference value, feature matching accuracy optimization is performed. Feature matching accuracy optimization includes determining whether to perform dynamic optimization of the feature frame change range. If so, after performing dynamic optimization of the feature frame change range, it is determined whether to perform dynamic optimization of the tolerance window width; otherwise, it is directly determined whether to perform dynamic optimization of the tolerance window width.

[0066] like Figure 3 The diagram shows a flowchart of dynamic optimization of the feature frame change range for an anti-fraud system based on voiceprint recognition, provided in an embodiment of this application. The specific logic is as follows: First, determine whether the voiceprint feature alignment timestamp deviation is within the alignment time deviation range. If so, do not perform dynamic optimization of the feature frame change range. Otherwise, determine whether the voiceprint feature alignment timestamp deviation is greater than the upper limit of the alignment time deviation reference. If so, query the alignment deviation and alignment time deviation from the constructed feature frame change range mapping table to obtain the upper limit adjustment amount of the feature frame change range. Based on the obtained upper limit adjustment amount, expand the current feature frame change upper limit to obtain the dynamic feature change upper limit. Otherwise, determine whether the voiceprint feature alignment timestamp deviation is less than the lower limit of the alignment time deviation reference. If so, query the alignment deviation and alignment time offset from the constructed feature frame change range mapping table to obtain the lower limit adjustment amount of the feature frame change range. Based on the obtained lower limit adjustment amount, reduce the current feature frame change lower limit to obtain the dynamic feature change lower limit.

[0067] As a further explanation, the specific steps for determining whether to perform dynamic optimization of the feature frame change range are as follows:

[0068] Feature frame variation refers to the degree of difference in the voiceprint feature vector between each pair of aligned acoustic feature frames during the time alignment of two speech segments (e.g., a registered speech and a test speech). Feature frame variation is used to reflect the degree of difference in voiceprint feature vectors during the voiceprint feature alignment process.

[0069] The decision to perform dynamic optimization of the feature frame change range is based on the alignment timestamp deviation of the voiceprint features. The alignment timestamp deviation of the voiceprint features is used to characterize the absolute value of the difference between the alignment timestamp of the voiceprint features and its reference value, reflecting the accuracy of time synchronization during the voiceprint feature alignment process. The specific steps are as follows: if the alignment timestamp deviation of the voiceprint features is within the alignment time deviation range, then dynamic optimization of the feature frame change range is not performed. This can filter out meaningless minor deviation interference, has a higher tolerance for timestamp fluctuations, reduces the risk of incorrect optimization due to misjudgment of the degree of deviation, and improves the stability of the system in complex environments. The alignment time deviation range represents the closed interval formed by the lower limit of the alignment time deviation reference and the upper limit of the alignment time deviation reference.

[0070] If the alignment timestamp deviation of the voiceprint feature is greater than the upper limit of the alignment time deviation reference, the alignment deviation and alignment time deviation are queried from the constructed feature frame change range mapping table to obtain the upper limit adjustment of the feature frame change. Based on the obtained upper limit adjustment of the feature frame change, the current upper limit of the feature frame change is expanded. This means that the current upper limit of the feature frame change is gradually and directionally increased. According to the specific values ​​of the alignment deviation and alignment time deviation, the upper limit of the feature frame change is expanded from the current value to a larger range according to the single expansion step size corresponding to the preset upper limit adjustment of the feature frame change, so as to adapt to the larger timestamp deviation and ensure that the voiceprint features can still be effectively aligned in a larger time range. This results in a dynamic upper limit of the feature frame change, which improves the feature alignment accuracy in large deviation scenarios. By increasing the upper limit of the feature frame change through expansion processing, the time range of feature alignment can be extended to a larger range. Let the upper limit of the current feature frame change be U0, the upper limit of the dynamic feature frame change be Ut, and the single expansion step size corresponding to the adjustment amount of the upper limit of the feature frame change be ΔU (ΔU>0). In the first expansion, the upper limit of the current feature frame change is updated to U1=U0+ΔU. It is then determined whether U1 reaches or exceeds Ut. If not, the next expansion begins, and U2=U1+ΔU, U3=U2+ΔU, etc., are calculated sequentially until the nth expansion, at which point Un is obtained, satisfying that Un is greater than or equal to Ut. At this point, Un is determined as the upper limit of the dynamic feature frame change. (If Un is greater than Ut, it can be precisely equal to Ut through a final fine-tuning.) The alignment deviation represents the difference between the voiceprint feature alignment deviation degree and the voiceprint feature alignment deviation reference value, and the alignment time deviation represents the difference between the voiceprint feature alignment timestamp deviation and the upper limit of the alignment time deviation reference.

[0071] If the alignment timestamp deviation of the voiceprint features is less than the lower limit of the alignment time deviation reference, the alignment deviation and alignment time offset are retrieved from the constructed feature frame change range mapping table to obtain the feature frame change lower limit adjustment amount. Based on the obtained feature frame change lower limit adjustment amount, the current feature frame change lower limit is reduced. The reduction process refers to a gradual and targeted adjustment of the current feature frame change lower limit based on the feature frame change lower limit adjustment amount retrieved from the feature frame change range mapping table. The core principle is to reduce the feature frame change lower limit from its current value to a smaller range according to the specific values ​​of the alignment deviation and alignment time offset, following preset rules or dynamically calculated step sizes, to adapt to smaller timestamp deviations. This ensures efficient alignment of voiceprint features within a more precise time range, resulting in a dynamic feature frame change lower limit and improving feature alignment accuracy in scenarios with small deviations. Let the current feature frame change lower limit be G0, the dynamic feature frame change lower limit be Gt, and the single reduction step size corresponding to the adjustment amount of the feature frame change lower limit be ΔG (ΔG>0). For the first reduction, update the current feature frame change lower limit to G1=G0-ΔG, and check if G1 is less than or equal to Gt. If not, proceed to the next reduction, calculating G2=G1-ΔG, G3=G2-ΔG, and so on, until the nth reduction yields Gn, which satisfies the condition that Gn is less than or equal to Gt. At this point, Ln is determined as the dynamic feature frame change lower limit (if Gn is less than Gt, a final fine-tuning can make Gn precisely equal to Gt). The alignment time offset represents the difference between the voiceprint feature alignment timestamp deviation and the alignment time deviation reference lower limit.

[0072] It should be noted that performing dynamic optimization of the feature frame change range according to different deviations in the timestamp alignment of voiceprint features can comprehensively improve the performance of the voiceprint processing system. This mechanism of dynamic adjustment based on deviation achieves comprehensive optimization of feature alignment accuracy, recognition efficiency, system stability, and resource consumption, and significantly improves the overall performance of the voiceprint processing system in various scenarios.

[0073] As a further explanation, the specific steps for determining whether to perform dynamic adjustment of the tolerance window width are as follows:

[0074] If dynamic adjustment of the feature frame change interval is performed, the feature frame change interval is re-acquired after the dynamic adjustment is performed and recorded as the feature frame change interval to be tested. Otherwise, the current feature frame change interval is directly recorded as the feature frame change interval to be tested. The feature frame change interval to be tested represents the closed interval formed by the lower limit of the feature frame change setting and the upper limit of the feature frame change setting.

[0075] The voiceprint feature time offset tolerance window refers to a constraint range of time offset set for dynamic time warping during the voiceprint feature alignment process. This window defines the upper and lower limits of the maximum allowed time offset between the current frame and the reference frame in the voiceprint feature matching path, which is used to control the search space of the alignment path to balance the accuracy and robustness of voiceprint matching.

[0076] The decision to perform dynamic adjustment of the tolerance window width is based on changes in the feature frames. These changes reflect the degree of difference between voiceprint feature vectors during the voiceprint feature alignment process. The specific steps are as follows:

[0077] If the feature frame change exceeds the upper limit set for feature frame change, the alignment deviation and feature frame change are retrieved from the constructed tolerance window width mapping table to obtain the tolerance window width reduction amount. Based on the obtained tolerance window width reduction amount, the current voiceprint feature time offset tolerance window width is reduced to obtain the voiceprint feature time offset tolerance window width for the next frame voiceprint feature alignment. This effectively improves the accuracy of voiceprint feature alignment and system processing efficiency. Timely reduction of the tolerance window width avoids including too many irrelevant features or redundant information due to an excessively wide window, reduces interference in the feature matching process, and ensures that the next frame voiceprint features can be aligned within a more accurate time offset range, thereby improving alignment accuracy. At the same time, reducing the window width can narrow the search range of the feature matching algorithm, reduce the consumption of computing resources, speed up processing, and enhance the adaptability and stability of the system in scenarios with drastic feature frame changes, thereby optimizing the overall performance of the voiceprint processing system. The feature frame change amount represents the difference between the feature frame change and the upper limit set for feature frame change.

[0078] If the feature frame change is less than the lower limit set for feature frame change, the alignment deviation and feature frame change correction are retrieved from the constructed tolerance window width mapping table to obtain the tolerance window width adjustment. Based on the obtained tolerance window width adjustment, the current voiceprint feature time offset tolerance window width is expanded to obtain the voiceprint feature time offset tolerance window width for the next frame's voiceprint feature alignment. This avoids limiting the tolerance range of time offset due to an excessively narrow window, prevents the omission of potential effective matching features, and ensures that the next frame's voiceprint features can achieve accurate alignment within a more relaxed time offset range, thereby guaranteeing the integrity of feature matching. It also adapts to subtle time offsets that may occur when feature frame changes are gradual, reduces matching failures caused by window limitations, improves the system's adaptability and robustness in scenarios with stable feature frame changes, and thus optimizes the overall performance of the voiceprint processing system. The feature frame change correction represents the difference between the feature frame change and the lower limit set for feature frame change.

[0079] If the feature frame change is within the range of the feature frame change to be tested, dynamic adjustment of the tolerance window width will not be performed to avoid meaningless window adjustment operations, reduce waste of computing resources, and ensure that feature matching is performed in a stable search space, thereby improving the system's processing efficiency and stability for normal feature frame change scenarios.

[0080] It should be noted that by dynamically adjusting the width of the voiceprint feature time offset tolerance window, adaptive response and precise control to changes in feature frames are achieved. This mechanism significantly improves the accuracy of voiceprint feature alignment, system processing efficiency and scene adaptability through intelligent dynamic adjustment of window width, thereby optimizing the overall performance and reliability of the voiceprint processing system.

[0081] like Figure 4 The diagram shows a flowchart of an anti-fraud method based on voiceprint recognition provided by an embodiment of the present invention. The method includes: acquiring voiceprint feature extraction data within a speech signal acquisition time window to reflect the time efficiency during voiceprint feature extraction; obtaining a timeliness result based on the acquired voiceprint feature extraction data to quantify the time efficiency of the voiceprint feature extraction process; determining whether to perform timeliness intelligent optimization based on the timeliness result; if so, sending a voiceprint feature matching collaboration instruction after performing timeliness intelligent optimization; otherwise, directly sending the voiceprint feature matching collaboration instruction. Based on the feature matching data reflecting the accuracy of voiceprint feature matching, the voiceprint feature alignment deviation is obtained to quantify the degree of alignment error in the voiceprint feature alignment process. Timeliness intelligent optimization means that the frame length and frame generation buffer queue threshold are adjusted in combination with the timeliness result. It is determined whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation. If so, a verification result output command is sent after performing feature matching accuracy optimization. Otherwise, a verification result output command is sent directly. Feature matching accuracy optimization means that the feature frame change range and tolerance window width are adjusted in combination with the voiceprint feature alignment deviation. The verification result output command includes matching success and matching failure.

[0082] It should be explained that by constructing a dual adaptive optimization mechanism based on the timeliness of the results and the alignment deviation of the voiceprint features, intelligent optimization of the entire process of voiceprint feature extraction and matching is achieved. This technology, through a series of dual optimization mechanisms, not only ensures the efficiency of the voiceprint feature extraction and matching process, but also significantly improves the alignment accuracy and system stability, making it suitable for complex and ever-changing application scenarios.

[0083] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included 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 voiceprint recognition-based anti-fraud system, characterized in that, It includes a timeliness assessment module, a timeliness optimization and matching quantification module, and a feature matching accuracy tuning module: The timeliness evaluation module is used to acquire voiceprint feature extraction data within the voice signal acquisition time window to reflect the time efficiency in the voiceprint feature extraction process. Based on the acquired voiceprint feature extraction data, the timeliness result is obtained to quantify the time efficiency of the voiceprint feature extraction process. The timeliness optimization and matching quantification module is used to determine whether to perform timeliness intelligent tuning based on the timeliness result. If so, it sends a voiceprint feature matching coordination instruction after performing timeliness intelligent tuning; otherwise, it sends the voiceprint feature matching coordination instruction directly. The voiceprint feature matching coordination instruction is used to reflect the feature matching data of the voiceprint feature matching accuracy and obtain the voiceprint feature alignment deviation to quantify the degree of alignment error in the voiceprint feature alignment process. The timeliness intelligent tuning means adjusting the frame length and the frame generation buffer queue threshold in combination with the timeliness result. The feature matching accuracy tuning module is used to determine whether to perform feature matching accuracy tuning based on the voiceprint feature alignment deviation. If so, a verification result output instruction is sent after performing feature matching accuracy tuning; otherwise, a verification result output instruction is sent directly. The feature matching accuracy tuning means adjusting the feature frame change range and the tolerance window width in combination with the voiceprint feature alignment deviation. The verification result output instruction includes matching success and matching failure.

2. The anti-fraud system based on voiceprint recognition according to claim 1, characterized in that, The voiceprint feature extraction data specifically includes: inter-frame processing interval, voiceprint feature sampling rate, and voiceprint feature extraction duration; The timeliness result represents the result obtained by associating the first timeliness component, the second timeliness component, and the third timeliness component. The first timeliness component represents the result obtained by interactive processing of the inter-frame processing interval factor and the inter-frame processing interval score; the second timeliness component represents the result obtained by interactive processing of the voiceprint feature sampling rate factor and the voiceprint feature sampling rate score; and the third timeliness component represents the result obtained by interactive processing of the voiceprint feature extraction duration factor and the voiceprint feature extraction duration score. The inter-frame processing interval score represents the result of inversely proportional processing of the relative deviation between the inter-frame processing interval reference value and the inter-frame processing interval; the voiceprint feature sampling rate score represents the result of proportional processing of the voiceprint feature sampling rate and the voiceprint feature sampling rate reference value; and the voiceprint feature extraction duration score represents the result of inversely proportional processing of the relative deviation between the voiceprint feature extraction duration reference value and the voiceprint feature extraction duration. The specific steps for determining whether to perform intelligent timeliness optimization based on the timeliness results are as follows: If the timeliness result is greater than or equal to the upper limit of the timeliness result reference, then the timeliness intelligent optimization will not be performed, and a voiceprint feature matching collaborative instruction will be sent. The timeliness intelligent optimization includes critical-level optimization and inefficient-level optimization. If the timeliness result is within the timeliness result reference interval, then critical-level optimization is performed. The timeliness result reference interval is used to characterize the open interval formed by the lower limit of the timeliness result reference and the upper limit of the timeliness result reference. The critical-level optimization includes determining whether to perform silence segment speech frame length optimization. If so, then after performing silence segment speech frame length optimization, it is determined whether to perform speech segment speech frame length optimization. Otherwise, it is directly determined whether to perform speech segment speech frame length optimization. If the timeliness result is less than or equal to the lower bound of the timeliness result reference, then inefficient level optimization is performed. The inefficient level optimization includes determining whether to perform frame generation buffer queue threshold range optimization. If yes, then after performing frame generation buffer queue threshold adjustment, it is determined whether to perform voice signal sampling frequency optimization. If no, then it is directly determined whether to perform voice signal sampling frequency optimization.

3. The anti-fraud system based on voiceprint recognition according to claim 2, characterized in that, The specific steps for determining whether to perform silence segment speech frame length optimization are as follows: The input speech signal is monitored in real time to distinguish between silence segments and speech segments. For the detected silence segments, the relative inter-frame energy difference is monitored to reflect the stability of energy fluctuations within the silence segment. If the relative inter-frame energy difference is lower than the first energy threshold, the current silence segment is determined to be a stable silence segment, and a first frame length adjustment instruction is generated to adjust the current speech frame length to the first target frame length. The specific adjustment steps are as follows: the timeliness result and the energy difference correction amount are queried from the constructed inefficient level optimization mapping table to obtain the first frame length mapping value. Based on the obtained first frame length mapping value, the current speech frame length is incrementally reduced to obtain the first target frame length. The energy difference correction amount is used to reflect the degree of negative deviation between the relative inter-frame energy difference and the first energy threshold. If the relative inter-frame energy difference is higher than the second energy threshold, the current silence segment is determined to be a sudden silence segment, and a second frame length adjustment instruction is generated to adjust the current speech frame length to the second target frame length. The specific adjustment steps are as follows: the timeliness result and the energy difference offset are queried from the constructed inefficient level optimization mapping table to obtain the second frame length mapping value. Based on the obtained second frame length mapping value, the current speech frame length is incrementally increased to obtain the second target frame length. The energy difference offset is used to reflect the degree of positive deviation between the relative inter-frame energy difference and the second energy threshold. If the relative inter-frame energy difference is within the energy threshold range, then the silent segment speech frame length adjustment is not performed, and the current speech frame length remains unchanged. The energy threshold range is used to reflect the closed interval formed by the first energy threshold and the second energy threshold.

4. The anti-fraud system based on voiceprint recognition according to claim 2, characterized in that, The specific steps for determining whether to perform speech segment frame length optimization are as follows: For the detected speech segment, its frequency change rate is monitored to reflect how fast the frequency of the speech signal changes over time. If the frequency change rate is higher than the first change rate threshold, the current speech segment is determined to be a high dynamic range speech segment, and a third frame length adjustment instruction is generated to adjust the current speech frame length to the third target frame length. The specific adjustment steps are as follows: the timeliness result and the change rate offset are queried from the constructed inefficient level optimization mapping table to obtain the third frame length mapping value. Based on the obtained third frame length mapping value, the current speech frame length is incrementally reduced to obtain the third target frame length. The change rate offset is used to reflect the degree of positive deviation between the frequency change rate and the first change rate threshold. If the frequency change rate is lower than the second change rate threshold, the current speech segment is determined to be a slowly changing speech segment, and a fourth frame length adjustment instruction is generated to adjust the current speech frame length to the fourth target frame length. The specific adjustment steps are as follows: the timeliness result and the change rate correction amount are queried from the constructed inefficient level optimization mapping table to obtain the fourth frame length mapping value. Based on the obtained fourth frame length mapping value, the current speech frame length is incrementally increased to obtain the fourth target frame length. The rate difference offset is used to reflect the degree of negative deviation between the frequency change rate and the second change rate threshold. If the rate of frequency change is within the rate of change threshold range, then speech segment frame length optimization is not performed, and the current speech frame length remains unchanged. The rate of change threshold range refers to the closed interval formed by the second rate of change threshold and the first rate of change threshold.

5. The anti-fraud system based on voiceprint recognition according to claim 2, characterized in that, The specific steps for determining whether to perform frame generation buffer queue threshold range optimization are as follows: The frame generation rate is used to determine whether to perform frame generation buffer queue threshold range optimization. The frame generation rate is used to reflect the frame generation frequency within the speech signal acquisition time window. The specific steps are as follows: if the frame generation rate is greater than the upper limit of the generation rate, the timeliness result and the generation rate offset are queried from the constructed inefficiency level optimization mapping table to obtain the upper limit mapping value of the frame generation buffer queue. Based on the obtained upper limit mapping value of the frame generation buffer queue, the upper limit value of the frame generation buffer queue is scaled to obtain the dynamic upper limit value of the frame generation buffer queue. The generation rate offset is used to reflect the degree of positive deviation between the frame generation rate and the upper limit of the generation rate. If the frame generation rate is less than the set lower limit of the generation rate, the timeliness result and the generation rate correction amount are queried from the constructed inefficiency-level optimization mapping table to obtain the lower limit mapping value of the frame generation buffer queue. Based on the obtained lower limit mapping value of the frame generation buffer queue, the lower limit value of the frame generation buffer queue is weighted and corrected to obtain the lower limit value of the dynamic frame generation buffer queue. The generation rate correction amount is used to reflect the degree of negative deviation between the frame generation rate and the set lower limit of the generation rate. If the frame generation rate is within the set range of the generation rate, then the frame generation buffer queue threshold range optimization will not be performed. The set range of the generation rate represents the closed interval formed by the lower limit of the set generation rate and the upper limit of the set generation rate.

6. The anti-fraud system based on voiceprint recognition according to claim 2, characterized in that, The specific steps for determining whether to perform voice signal sampling frequency optimization are as follows: If frame generation buffer queue threshold interval optimization is performed, the frame generation buffer queue threshold interval is re-acquired after the frame generation buffer queue threshold interval optimization is performed and recorded as the frame generation buffer queue threshold interval to be compared; otherwise, the current frame generation buffer queue threshold interval is directly recorded as the frame generation buffer queue threshold interval to be compared. The frame generation buffer queue threshold interval to be compared represents the closed interval formed by the lower limit set by the frame generation buffer queue and the upper limit set by the frame generation buffer queue. Whether to perform voice signal sampling frequency optimization is determined based on the total number of voice frames. The total number of voice frames is used to reflect the total number of voice frames waiting to be processed in the frame generation buffer queue. The specific steps are as follows: if the total number of voice frames is greater than the upper limit set by the frame generation buffer queue, the timeliness result and the queue overflow deviation are queried from the constructed inefficiency-level optimization mapping table to obtain the first mapping value of the voice signal sampling frequency. Based on the obtained first mapping value of the voice signal sampling frequency, the current voice signal sampling frequency is proportionally scaled down to obtain the voice signal sampling frequency adjustment value. The queue overflow deviation represents the degree of positive deviation between the total number of voice frames and the upper limit set by the frame generation buffer queue. If the total number of speech frames is less than the lower limit set by the frame generation buffer queue, the timeliness result and the queue overflow offset are queried from the constructed inefficiency-level optimization mapping table to obtain the second mapping value of the speech signal sampling frequency. Based on the obtained second mapping value of the speech signal sampling frequency, the current speech signal sampling frequency is scaled up to obtain the speech signal sampling frequency correction value. The queue overflow offset represents the degree of negative deviation between the total number of speech frames and the lower limit set by the frame generation buffer queue. If the total number of speech frames is within the threshold range of the frame generation buffer queue, speech signal sampling frequency optimization will not be performed.

7. The anti-fraud system based on voiceprint recognition according to claim 1, characterized in that, The feature matching data specifically includes: formant frequency offset, speech rate matching deviation, and timeliness result to be tested; The voiceprint feature alignment deviation is characterized by the result obtained by associating the first component, the second component, and the third component of the alignment deviation. The first component of the alignment deviation represents the result obtained by interactive processing of the formant frequency offset factor and the formant frequency offset score; the second component of the alignment deviation represents the result obtained by interactive processing of the speech rate matching deviation factor and the speech rate matching deviation score; and the third component of the alignment deviation represents the result obtained by interactive processing of the timeliness result factor and the timeliness result score to be tested. The formant frequency offset score represents the result of processing the relative deviation ratio between the formant frequency offset and the set formant frequency offset value; the speech rate matching deviation score represents the result of processing the relative deviation ratio between the speech rate matching deviation and the set speech rate matching deviation value; and the timeliness result score to be tested represents the result of processing the timeliness result set value as a percentage. The specific steps for determining whether to perform feature matching accuracy optimization based on the voiceprint feature alignment deviation are as follows: If the voiceprint feature alignment deviation is lower than or equal to the voiceprint feature alignment deviation reference value, feature matching accuracy tuning will not be performed. If the voiceprint feature alignment deviation is higher than the voiceprint feature alignment deviation reference value, then feature matching accuracy optimization is performed. The feature matching accuracy optimization includes determining whether to perform dynamic optimization of the feature frame change interval. If so, then after performing dynamic optimization of the feature frame change interval, it is determined whether to perform dynamic optimization of the tolerance window width. Otherwise, it is directly determined whether to perform dynamic optimization of the tolerance window width.

8. The anti-fraud system based on voiceprint recognition according to claim 7, characterized in that, The specific steps for determining whether to perform dynamic optimization of the feature frame change range are as follows: The decision to perform dynamic optimization of the feature frame change range is based on the voiceprint feature alignment timestamp deviation. The voiceprint feature alignment timestamp deviation is used to characterize the absolute value of the difference between the voiceprint feature alignment timestamp and its reference value, reflecting the accuracy of time synchronization during the voiceprint feature alignment process. The specific steps are as follows: if the voiceprint feature alignment timestamp deviation is within the alignment time deviation range, then dynamic optimization of the feature frame change range is not performed. The alignment time deviation range represents the closed interval formed by the lower limit of the alignment time deviation reference and the upper limit of the alignment time deviation reference. If the voiceprint feature alignment timestamp deviation is greater than the upper limit of the alignment time deviation reference, the alignment deviation amount and the alignment time deviation amount are queried from the constructed feature frame change interval mapping table to obtain the upper limit adjustment amount of the feature frame change. Based on the obtained upper limit adjustment amount of the feature frame change, the current feature frame change upper limit is expanded to obtain the dynamic feature frame change upper limit. The alignment deviation amount is used to reflect the degree of positive deviation between the voiceprint feature alignment deviation degree and the voiceprint feature alignment deviation degree reference value. The alignment time deviation amount is used to reflect the degree of positive deviation between the voiceprint feature alignment timestamp deviation and the upper limit of the alignment time deviation reference. If the alignment timestamp deviation of the voiceprint feature is less than the lower limit of the alignment time deviation reference, the alignment deviation and alignment time offset are queried from the constructed feature frame change interval mapping table to obtain the lower limit adjustment of the feature frame change. Based on the obtained lower limit adjustment of the feature frame change, the current feature frame change lower limit is reduced to obtain the dynamic feature frame change lower limit. The alignment time offset is used to reflect the degree of negative deviation between the voiceprint feature alignment timestamp deviation and the lower limit of the alignment time deviation reference.

9. The anti-fraud system based on voiceprint recognition as described in claim 7, characterized in that, The specific steps for determining whether to perform dynamic adjustment of the tolerance window width are as follows: If dynamic optimization of the feature frame change interval is performed, the feature frame change interval is re-acquired after the dynamic optimization is performed and recorded as the feature frame change interval to be tested. Otherwise, the current feature frame change interval is directly recorded as the feature frame change interval to be tested. The feature frame change interval to be tested represents the closed interval formed by the lower limit of the feature frame change setting and the upper limit of the feature frame change setting. The decision to perform dynamic adjustment of the tolerance window width is based on changes in the feature frames. These changes reflect the degree of difference between voiceprint feature vectors during the voiceprint feature alignment process. The specific steps are as follows: If the feature frame change is greater than the upper limit set for feature frame change, the alignment deviation and feature frame change are queried from the constructed tolerance window width mapping table to obtain the tolerance window width reduction amount. Based on the obtained tolerance window width reduction amount, the current voiceprint feature time offset tolerance window width is reduced to obtain the voiceprint feature time offset tolerance window width for the next frame voiceprint feature alignment. The feature frame change amount is used to reflect the degree of positive deviation between the feature frame change and the upper limit set for feature frame change. If the feature frame change is less than the lower limit set for the feature frame change, the alignment deviation and feature frame change correction are queried from the constructed tolerance window width mapping table to obtain the tolerance window width adjustment. Based on the obtained tolerance window width adjustment, the current voiceprint feature time offset tolerance window width is expanded to obtain the voiceprint feature time offset tolerance window width for the next frame voiceprint feature alignment. The feature frame change correction is used to reflect the degree of negative deviation between the feature frame change and the lower limit set for the feature frame change. If the feature frame change is within the range of the feature frame change to be tested, then dynamic adjustment of the tolerance window width will not be performed.

10. A voiceprint recognition-based anti-fraud method, wherein the voiceprint recognition-based anti-fraud method applies the voiceprint recognition-based anti-fraud system as described in any one of claims 1-9, characterized in that, include: Acquire voiceprint feature extraction data within the speech signal acquisition time window to reflect the time efficiency in the voiceprint feature extraction process. Based on the acquired voiceprint feature extraction data, obtain the timeliness result to quantify the time efficiency of the voiceprint feature extraction process. Based on the timeliness results, it is determined whether to perform timeliness intelligent optimization. If so, a voiceprint feature matching coordination instruction is sent after performing timeliness intelligent optimization. If not, a voiceprint feature matching coordination instruction is sent directly. The voiceprint feature matching coordination instruction is used to reflect the feature matching data reflecting the accuracy of voiceprint feature matching, and to obtain the voiceprint feature alignment deviation to quantify the degree of alignment error in the voiceprint feature alignment process. The timeliness intelligent optimization means adjusting the frame length and the frame generation buffer queue threshold in combination with the timeliness results. The system determines whether to perform feature matching accuracy tuning based on the voiceprint feature alignment deviation. If so, a verification result output command is sent after the feature matching accuracy tuning is performed; otherwise, a verification result output command is sent directly. The feature matching accuracy tuning means adjusting the feature frame variation range and the tolerance window width in combination with the voiceprint feature alignment deviation. The verification result output command includes matching success and matching failure.

Images