Transaction processing method and device, equipment, storage medium and program product

Abstract

The invention provides a transaction processing method and device, equipment, a storage medium and a program product, and relates to the field of artificial intelligence and the field of financial science and technology. The method comprises the following steps: receiving a voice transaction request of a user; based on the voice transaction request, acquiring voice information of a user, and equipment fingerprint information and environment information of a client sending the voice information; extracting dynamic voiceprint features based on the voice information, the equipment fingerprint information and the environment information; the dynamic voiceprint features comprise acoustic features, language features and environment features; fusing the voice information and the dynamic voiceprint features to obtain fused features; obtaining a risk score about the voice transaction based on the fusion feature; and processing the voice transaction based on the risk score. According to the method, dynamic voiceprint features and multi-dimensional features of voice information are organically combined, meanwhile, risk scores are combined, the voice information is monitored, and whether the voice information is real and credible or not is determined; the voice identification accuracy is higher, and normal release of the transaction is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and the field of financial technology, and in particular to a transaction processing method and device, equipment, a storage medium and a program product. BACKGROUND

[0002] With the development of generative artificial intelligence, voice forgery means has evolved from early simple splicing and speed variation to high-level technologies such as deep forgery (Deepfake) and voiceprint conversion. Attackers only need a few seconds of real voice fragments to synthesize high-fidelity voice using deep learning models such as Tacotron, VITS, and GPT-4 voice expansion, or directly migrate the voiceprint of the target speaker to any text to achieve "one-word cloning".

[0003] Such forged voices are highly consistent with real people in terms of pitch, formant, rhythm, and even respiratory noise, and can easily bypass traditional voiceprint identification systems, posing a serious threat to identity verification in mobile banking voice login, voice payment, and other financial scenarios. SUMMARY

[0004] The embodiments of the present application provide a transaction processing method, device, equipment, storage medium and program product to achieve high voice identification accuracy and ensure normal transaction release.

[0005] In a first aspect, the embodiments of the present application provide a transaction processing method, comprising:

[0006] receiving a voice transaction request of a user;

[0007] based on the voice transaction request, obtaining voice information of the user, and device fingerprint information and environment information of a client sending the voice information;

[0008] extracting dynamic voiceprint features based on the voice information, the device fingerprint information and the environment information; the dynamic voiceprint features include acoustic features, language features and environment features;

[0009] fusing the voice information and the dynamic voiceprint features to obtain fused features;

[0010] obtaining a risk score for the voice transaction based on the fused features;

[0011] processing the voice transaction based on the risk score.

[0012] In one embodiment, obtaining a risk score for the voice transaction based on the fused features specifically includes:

[0013] obtaining a voice identification result based on the fused features;

[0014] Based on voice recognition results and user transaction behavior data, a relationship graph of user transaction behavior is constructed using graph neural networks; transaction behavior data includes transaction information, device fingerprints, and geographic information.

[0015] Based on the relationship graph of user transaction behavior and the transaction information corresponding to the transaction request, a risk score for voice transactions is obtained.

[0016] In one embodiment, based on the voice recognition results and the user's transaction behavior data, a graph neural network is used to construct a relationship graph of the user's transaction behavior, specifically including:

[0017] Voice recognition results, transaction information, device fingerprints, and geographic information are used as heterogeneous nodes, with the user identifier as the root node; where the user identifier is a unique identifier for the user in the system.

[0018] Edges are generated based on the co-occurrence relationship between nodes. Edges are established between nodes corresponding to the same device or the same geographical location. The weight of the edge is determined by the co-occurrence frequency and the time decay factor.

[0019] By training a heterogeneous graph using a graph neural network, a relational graph of user transaction behavior can be obtained.

[0020] In one embodiment, dynamic voiceprint features are extracted based on voice information, device fingerprint information, and environmental information; the dynamic voiceprint features include acoustic features, language features, and environmental features, specifically including:

[0021] Mel-frequency cepstral coefficient features and waveform encoder features are extracted based on speech information;

[0022] Align the Mel-frequency cepstral coefficient features and waveform encoder features in the time dimension to obtain aligned speech temporal features;

[0023] Key frames and key frequency bands are selected based on speech temporal features to generate acoustic features containing spectral and contextual information.

[0024] In one embodiment, the speech recognition result is obtained based on the fused features, specifically including:

[0025] The probability of speech anomalies is obtained based on fused features;

[0026] The fused features are compared with the registered voiceprint template, and the identity matching probability is output.

[0027] The voice identification result is obtained by weighted fusion of voice anomaly probability and identity matching probability.

[0028] In one embodiment, processing voice transactions based on risk scoring specifically includes:

[0029] If the risk score is greater than or equal to the first threshold, then voice transactions are blocked; or

[0030] If the risk score is greater than or equal to the second threshold, then re-verification is triggered; or

[0031] If the risk score is less than the second threshold, execute a voice transaction.

[0032] In one embodiment, the method further includes:

[0033] A first threshold and a second threshold are determined based on the user's historical voice data and transaction behavior data.

[0034] In one embodiment, the extraction of Mel-frequency cepstral coefficient features and waveform encoder features based on speech information further includes:

[0035] Noise reduction processing is performed on the voice information;

[0036] Endpoint detection is performed on the noise-reduced speech information to identify the start and end positions of the speech information;

[0037] Effective speech information is segmented based on the start and end positions;

[0038] Mel-frequency cepstral coefficient features and waveform encoder features are extracted based on effective speech information.

[0039] Secondly, embodiments of this application provide a transaction processing apparatus, including:

[0040] The receiving module is used to receive users' voice transaction requests;

[0041] The acquisition module is used to acquire the user's voice information, as well as the device fingerprint information and environmental information associated with the voice information, based on the voice transaction request.

[0042] The extraction module is used to extract dynamic voiceprint features based on voice information, device fingerprint information, and environmental information; dynamic voiceprint features include acoustic features, language features, and environmental features.

[0043] The fusion module is used to fuse speech information and dynamic voiceprint features to obtain fused features;

[0044] The processing module is used to obtain a risk score for voice transactions based on fused features;

[0045] The processing module is also used to process voice transactions based on risk scores.

[0046] Thirdly, embodiments of this application provide a transaction processing device, including: a memory and a processor;

[0047] The memory stores instructions that the computer executes;

[0048] The processor executes computer execution instructions stored in memory, causing the processor to perform any of the methods described above.

[0049] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement any of the methods described above.

[0050] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements any of the methods described above.

[0051] The transaction processing method, apparatus, device, storage medium, and program product provided in this application include: receiving a user's voice transaction request; acquiring the user's voice information, as well as device fingerprint information and environmental information of the client sending the voice information, based on the voice transaction request; extracting dynamic voiceprint features based on the voice information, device fingerprint information, and environmental information; the dynamic voiceprint features include acoustic features, linguistic features, and environmental features; fusing the voice information and dynamic voiceprint features to obtain fused features; obtaining a risk score for the voice transaction based on the fused features; and processing the voice transaction based on the risk score. By organically combining dynamic voiceprint features and multi-dimensional features of voice information, and simultaneously combining the risk score, the voice information is monitored to determine whether the voice information is authentic and reliable. Compared with traditional methods that rely solely on single-modal detection, this application has a higher accuracy rate in voice identification, ensuring normal transaction release and preventing fraudulent transactions. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] Figure 1 Application scenarios for the transaction processing method provided in this application;

[0054] Figure 2 A flowchart illustrating a transaction processing method provided in an embodiment of this application;

[0055] Figure 3 This is a schematic diagram of the structure of a transaction processing apparatus provided in an embodiment of this application;

[0056] Figure 4 A schematic diagram of the transaction processing equipment provided in this application.

[0057] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0059] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0060] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0061] First, let me explain the terms used in this application:

[0062] Mel-Frequency Cepstral Coefficients (MFCCs) are a type of "human-auditory-based" acoustic feature. They are generated by transforming a frame of speech waveform into a 12–40 dimensional vector. First, a short-time Fourier transform is performed on the speech to obtain the power spectrum. Then, a set of Mel-scaled triangular filter banks (simulating the human ear's sensitivity to different frequencies) is used to convolve the power spectrum, extracting the logarithmic energy. Finally, a Discrete Cosine Transform (DCT) is performed to compress the logarithmic energy output by the filter bank into low-dimensional coefficients. MFCCs retain information most sensitive to the human ear, such as formants and fundamental frequency trends, and are low-dimensional and noise-robust, but high-frequency details and phase information are discarded.

[0063] Waveform encoder characteristics: refers to operating directly on the original 16 kHz (or 8 kHz) sampled waveform without making explicit acoustic assumptions.

[0064] Existing technologies: Techniques that use artificial intelligence algorithms (such as deep learning models) to imitate, synthesize, or tamper with the speech of a target object can generate highly realistic forged speech based on short speech fragments. Common methods include speech synthesis, voiceprint conversion, and deepfake. These technologies can bypass traditional speech recognition systems, posing a threat to identity verification in the financial sector. In recent years, with the rapid development of generative AI (such as GPT-4 and deepfake speech synthesis), the realism of forged speech has greatly improved. It is even possible to generate highly realistic speech based on short speech fragments, making it difficult for traditional voiceprint recognition systems to distinguish between genuine and fake speech, resulting in security issues during voice transactions.

[0065] To address the shortcomings of existing technologies, the inventors of this solution, through creative research, designed a new approach. This solution provides a transaction processing method that addresses the problem of traditional voiceprint recognition systems' inability to distinguish genuine from counterfeit transactions during voice transactions. Figure 1 As shown, Figure 1 This application describes an application scenario for the transaction processing method provided in this application. This application scenario includes a client 11 and a server 12, wherein the number of clients 11 can be at least one. In practical applications, the server 12 receives a voice transaction request sent by a user through client 11 and executes the transaction processing method provided in this application. Voice transaction refers to the process by which a user initiates and completes a financial transaction through a mobile banking client by speaking (voice). A typical scenario example: A user speaks into their phone, "Transfer 500 yuan to Zhang San"; the system automatically converts the speech into text and parses out elements such as the recipient and amount; then, it uses voice verification (dynamic voiceprint comparison); after successful verification, the transfer is completed directly.

[0066] It should be noted that server 12 can also be replaced by a cluster of server 12 or other computing devices with a certain computing power. Client 11 can be a computer, mobile phone, laptop, or personal digital assistant (PDA), etc.

[0067] The following is combined Figure 1 Application scenarios, refer to Figure 2 This application describes a transaction processing method according to exemplary embodiments thereof. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited to those described herein. Figure 1 The limitations of the application scenarios shown.

[0068] It should be noted that the transaction processing method, apparatus, device, storage medium, and program product provided in this application can be used in the fields of artificial intelligence and fintech, as well as in any field other than artificial intelligence and fintech. The application fields of the transaction processing method, apparatus, device, storage medium, and product in this application are not limited.

[0069] This application provides a transaction processing method that organically combines dynamic voiceprint features and multi-dimensional features of voice information. Simultaneously, it incorporates risk scoring to monitor voice information and determine its authenticity and reliability. Compared to traditional methods that rely solely on single-modal detection, this application achieves higher accuracy in voice identification, ensuring normal transaction processing and preventing fraudulent transactions.

[0070] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0071] like Figure 2 As shown, Figure 2 This is a flowchart illustrating a transaction processing method provided in an embodiment of this application. The transaction processing method includes the following steps:

[0072] Step S201: Receive the user's voice transaction request.

[0073] Specifically, users can initiate transactions via voice, either actively or passively, such as by saying "Transfer 500 yuan to Xiao Li" on the interactive interface; or by long-pressing the button to automatically record the transaction. The server receives the voice transaction request and performs a series of transaction processing steps to complete the voice transaction.

[0074] Step S202: Based on the voice transaction request, obtain the user's voice information, as well as the device fingerprint information and environmental information of the client that sent the voice information.

[0075] Specifically, the client collects the user's voice information via microphone and obtains the client's device fingerprint and environmental information through integrated device fingerprint technology. This device fingerprint and environmental information includes the user's device model, operating system version, hardware configuration, etc. Combined with GPS geolocation technology, it generates voice information containing a timestamp. For example, when a user conducts a voice transaction through mobile banking, the front-end acquisition layer not only collects the audio recording but also records information such as the phone model and city, providing more information to determine the authenticity of the transaction.

[0076] Step S203: Extract dynamic voiceprint features based on voice information, device fingerprint information, and environmental information; dynamic voiceprint features include acoustic features, language features, and environmental features.

[0077] Specifically, language features are extracted from speech-transcribed text using natural language processing (NLP) technology, focusing on keyword frequency, grammatical structure, and semantic information. Environmental features include device model, operating system version, hardware configuration, GPS location, network latency, and IP range. By extracting dynamic voiceprint features and performing multi-dimensional complementarity, single points of failure are reduced; acoustic features can prevent recording / synthesis attacks; language features can prevent "semantic forgery" due to inconsistencies between text and voiceprint; and environmental features can prevent device tampering and abnormal geographical locations. Even if any single dimension is compromised, the remaining dimensions can still provide backup, significantly improving overall robustness. Furthermore, device upgrades, changes in speaker speed, and network environment drift all affect traditional static templates. These three types of features are calculated in real-time for each transaction, forming a "dynamic voiceprint." The model can follow changes in user behavior without retraining, resulting in a significant reduction in false rejection rates.

[0078] Step S204: After fusing the speech information and dynamic voiceprint features, the fused features are obtained.

[0079] Specifically, by fusing the "original speech signal" and the "refined dynamic voiceprint features," it is equivalent to simultaneously preserving the most original microscopic details (waveform phase, instantaneous noise) and the filtered high-order representations (speaker personality, semantics, device environment). This not only prevents the loss of key clues during the extraction process but also allows the model to perform end-to-end alignment in a unified space. This enables the downstream forgery detection / identity authentication network to obtain richer discriminative information, ultimately improving the accuracy of anti-fraud, reducing false positives, and significantly enhancing the sensitivity and overall accuracy of forgery detection.

[0080] Step S205: Obtain a risk score for voice transactions based on the fusion features.

[0081] Specifically, in one example, a model is trained based on a federated learning framework. Each bank node encrypts and processes user data locally, while the original voice, device fingerprint information, and environmental information remain stored locally within each bank. Only model parameter updates are uploaded, ensuring data is available but not visible. This protects user privacy while enabling cross-institutional collaborative data learning. A risk score for voice transactions is obtained based on the fused features.

[0082] Step S206: Process voice transactions based on risk scores.

[0083] This application organically combines dynamic voiceprint features and multi-dimensional features of speech information, and combines risk scoring to monitor speech information and determine whether the speech information is authentic and credible. Compared with traditional methods that rely solely on single-modal detection, this application has a higher accuracy rate in speech identification, which can ensure the normal release of transactions and prevent fraudulent transactions.

[0084] In one embodiment, step S206 specifically includes the following steps:

[0085] If the risk score is greater than or equal to the first threshold, the voice transaction is blocked; if the risk score is greater than or equal to the second threshold, re-verification is triggered; or if the risk score is less than the second threshold, the voice transaction is executed.

[0086] Specifically, if the risk score is greater than or equal to the first threshold, it indicates a high risk in the voice transaction, and the transaction system is immediately activated to intercept it in real time to prevent fraudulent transactions and protect the security of the voice transaction. When the risk score is greater than or equal to the second threshold, a secondary verification mechanism is triggered, requiring the user to further verify their identity through SMS verification code, fingerprint recognition, or other verification methods. If the risk score is less than the second threshold, the voice transaction is considered to have a low risk, and the voice transaction is executed, allowing it to proceed smoothly. In this example, the first threshold is 80% and the second threshold is 40%. In other embodiments, the first and second thresholds are adjustable, and this application does not limit them here.

[0087] In one embodiment, the method further includes the following steps:

[0088] A first threshold and a second threshold are determined based on the user's historical voice data and transaction behavior data.

[0089] Specifically, a user-specific feature template is constructed by acquiring all voice records of the user that have been manually or model-verified as "real" within the past 90 days, along with their corresponding device fingerprints and environmental information. The sliding window mean (window length 7 days, step size 1 day) is calculated for each of the three sub-features: acoustic, linguistic, and environmental, forming a three-dimensional dynamic template T_user. After each successful transaction, the template is refreshed in real-time using an exponentially weighted moving average (EMA, α=0.1) to ensure that the template follows the user's recent pronunciation habits, device upgrades, or location shifts. The fused feature extracted from the current transaction is denoted as F_curr. The Euclidean norm L2 normalized cosine similarity S=cosθ between F_curr and T_user is calculated. Simultaneously, the cosine distance D_fake to the global negative sample center (the mean vector of historical fake voices) is calculated as a supplement to the fakeness anomaly. The distributions of real voice P_real(S) and fake voice P_fake(S) are statistically analyzed using the user's historical records. Likelihood update: After each manual review or payment rejection confirmation, the new (S, label) is added to the Bayesian update queue, and P_real and P_fake are corrected in real time using the beta-binomial prior. The formulas for the first threshold τ1 and the second threshold τ2 are as follows:

[0090]

[0091]

[0092] The values ​​of k1, k2, and λ are automatically determined through Bayesian decision risk minimization based on the business false positive rate / false negative rate KPI. The first and second thresholds can also be obtained through other methods; this application is merely one example. In a backtesting of 200,000 real production data points, after introducing personalized thresholds, the overall false positive rate decreased from 1.8% to 0.9%; the false negative rate for spoofed voice decreased from 2.4% to 1.1%; the trigger rate for secondary verification by existing users decreased by 42%, resulting in a significant improvement in user experience.

[0093] In one embodiment, step S205 specifically includes the following steps:

[0094] Speech recognition results are obtained based on fusion features.

[0095] In one embodiment, obtaining a speech recognition result based on fused features specifically includes the following steps:

[0096] The probability of speech anomalies is obtained based on the fusion features.

[0097] The fused features are compared with the registered voiceprint template, and the identity matching probability is output.

[0098] The voice identification result is obtained by weighted fusion of voice anomaly probability and identity matching probability.

[0099] Specifically, the fused features are input into the forgery detection subnetwork, which outputs the speech anomaly probability Pfake, where Pfake ∈ [0,1], and the closer the speech anomaly probability is to 1, the higher the probability that the speech is forged or tampered with. The fused features are then compared with pre-stored registered voiceprint templates using cosine similarity calculation, and after being mapped by the sigmoid function, the identity matching probability Pmatch is output, where Pmatch ∈ [0,1], and the closer the identity matching probability is to 1, the more consistent the current speech is with the registered voiceprint. Based on the weight factor α set according to the business scenario, the speech anomaly probability and the identity matching probability are weighted and fused to obtain the speech identification result. The larger the value of R, the higher the risk of the voice transaction, which can be directly mapped to subsequent risk scoring and handling strategies.

[0100] Based on voice recognition results and user transaction behavior data, a relationship graph of user transaction behavior is constructed using graph neural networks; the transaction behavior data includes transaction information, device fingerprints, and geographic information.

[0101] In one embodiment, based on the voice recognition results and the user's transaction behavior data, a graph neural network is used to construct a relationship graph of the user's transaction behavior, specifically including the following steps:

[0102] Voice identification results, transaction information, device fingerprints, and geographic information are used as heterogeneous nodes, with the user identifier as the root node; where the user identifier is a unique identifier for the user in the system.

[0103] Edges are generated based on the co-occurrence relationships between nodes. Edges are established between nodes that are the same device or the same geographical location. The weight of the edge is determined by the co-occurrence frequency and the time decay factor.

[0104] By training a heterogeneous graph using a graph neural network, a relational graph of user transaction behavior can be obtained.

[0105] Specifically, four types of data from different sources (voice recognition results, transaction information, device fingerprints, and geographic information) are abstracted into four types of nodes, with the user's unique identifier as the root node, ensuring the graph revolves around the user and has a clear structure. Co-occurrence relationships are used to construct edges: whenever two records appear on the same device or in the same geographic location, an edge is drawn between the corresponding nodes. Weight formula: The more frequent and closer the co-occurrence of events between nodes, the greater their weight; historical co-occurrence decays over time to avoid interference from old data. A heterogeneous graph attention network (HGT / RGAT) is used to propagate information between nodes and learn low-dimensional embeddings; the training objective is node-level binary classification (normal / abnormal). Anomaly scores are output for subsequent risk scoring or to directly trigger risk control strategies. Connecting isolated transactions, devices, and locations into a graph allows for the discovery of cross-transaction and cross-device group attack chains. Each new transaction incrementally inserts nodes / edges, keeping the graph constantly fresh. Even if the voice itself is deeply forged, anomaly scores can still be increased if the device or geographical location is abnormal. Furthermore, the anomaly graph can be visualized with a single click, facilitating manual review.

[0106] Based on the relationship graph of user transaction behavior and the transaction information corresponding to the transaction request, a risk score for voice transactions is obtained.

[0107] Specifically, nodes aggregate neighbor information through message passing, automatically capturing abnormal patterns such as "short-term cross-regional" and "multiple voiceprints from the same device." The training objective is node-level binary classification (normal or abnormal), outputting anomaly scores. These anomaly scores are then fused with the voice recognition results to obtain a final risk score ranging from 0 to 100. By analyzing abnormal patterns in the relationship graph, the risk of misjudgment based on a single modality is further reduced, improving the accuracy of anti-fraud measures.

[0108] In one embodiment, step S203 specifically includes the following steps:

[0109] Mel-frequency cepstral coefficient features and waveform encoder features are extracted based on speech information.

[0110] Aligning the Mel-frequency cepstral coefficient features and waveform encoder features in the time dimension yields aligned speech temporal features.

[0111] Specifically, if the frame rates of the Mel-frequency cepstral coefficient (MFCC) features and the waveform encoder features differ, linear interpolation or 1×1 convolution is first used to unify them to the same time step. Differentiable dynamic time warping (Soft-DTW) or relative position coding Transformer is then employed to automatically align frames corresponding to fast / slow speech rates, eliminating differences in speaking speed. This yields aligned speech temporal features of consistent length, corresponding frame by frame. The generated acoustic features retain high-resolution spectral details (from key frequency bands) while incorporating the contextual semantics / prosodic information of the entire sentence (from key frames), and can be directly used for subsequent forgery detection or speaker recognition.

[0112] Key frames and key frequency bands are selected based on speech temporal features to generate acoustic features containing spectral and contextual information.

[0113] In one embodiment, the extraction of Mel-spectral coefficient features and waveform encoder features based on speech information further includes the following steps:

[0114] Noise reduction processing is performed on the voice information.

[0115] Specifically, wavelet transform and bandpass filtering are used to remove background noise, channel noise, and device noise from the speech information. This improves the signal-to-noise ratio, prevents noise from being mistaken for "speaker personality" by subsequent features, and reduces the false recognition rate.

[0116] Endpoint detection is performed on the noise-reduced speech information to identify the start and end positions of the speech information.

[0117] Specifically, endpoint detection is performed on the denoised speech information using a dual-threshold approach of energy and zero-crossing rate, or a lightweight CNN, to automatically identify the start and end frames of the speech segment. Silence segments and invalid signals before and after them are removed to reduce computational load and prevent silence noise from interfering with the model.

[0118] Effective speech information is segmented based on the start and end positions.

[0119] Specifically, the speech information is trimmed according to the start and end frames, with a 0.5-second buffer applied if necessary, to output clean speech segments, which are the valid speech information. This ensures that feature extraction only targets waveforms that actually contain spoken content, improving feature consistency.

[0120] Mel-frequency cepstral coefficient features and waveform encoder features are extracted based on effective speech information.

[0121] like Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of a transaction processing device provided in an embodiment of this application. The transaction processing device 300 includes: a receiving module 301, used to receive a user's voice transaction request; an acquisition module 302, used to acquire the user's voice information, as well as device fingerprint information and environmental information associated with the voice information, based on the voice transaction request; an extraction module 303, used to extract dynamic voiceprint features based on the voice information, device fingerprint information, and environmental information; the dynamic voiceprint features include acoustic features, language features, and environmental features; a fusion module 304, used to fuse the voice information and dynamic voiceprint features to obtain fused features; a processing module 305, used to obtain a risk score for the voice transaction based on the fused features; the processing module 305 is also used to process the voice transaction based on the risk score.

[0122] In one embodiment, the processing module 305 is used to obtain a voice identification result based on fused features; based on the voice identification result and the user's transaction behavior data, a graph neural network is used to construct a relationship graph of the user's transaction behavior; the transaction behavior data includes transaction information, device fingerprint, and geographic information; based on the relationship graph of the user's transaction behavior and the transaction information corresponding to the transaction request, a risk score for the voice transaction is obtained.

[0123] In one embodiment, the processing module 305 is used to use voice identification results, transaction information, device fingerprints and geographic information as heterogeneous nodes, and user identifiers as root nodes; wherein, user identifiers are unique identifiers of users in the system; generate edges according to the co-occurrence relationship between nodes, wherein, edges are established between nodes corresponding to the same device or the same geographic location, and the weight of the edges is determined by the co-occurrence frequency and time decay factor; and train the heterogeneous graph using a graph neural network to obtain a relational graph of user transaction behavior.

[0124] In one embodiment, the extraction module 303 is used to extract Mel-Cepstral Coefficient features and waveform encoder features based on speech information; align the Mel-Cepstral Coefficient features and waveform encoder features in the time dimension to obtain aligned speech temporal features; and filter key frames and key frequency bands based on the speech temporal features to generate acoustic features containing spectral information and contextual information.

[0125] In one embodiment, the processing module 305 is used to obtain the speech anomaly probability based on the fusion features; compare the fusion features with the registered voiceprint template and output the identity matching probability; and obtain the speech identification result by weighted fusion of the speech anomaly probability and the identity matching probability.

[0126] In one embodiment, the processing module 305 is configured to intercept the voice transaction if the risk score is greater than or equal to a first threshold; or trigger re-verification if the risk score is greater than or equal to a second threshold; or execute the voice transaction if the risk score is less than the second threshold.

[0127] In one embodiment, the processing module 305 is configured to determine a first threshold and a second threshold based on the user's historical voice data and transaction behavior data.

[0128] In one embodiment, the extraction module 303 performs noise reduction processing on the speech information; performs endpoint detection on the noise-reduced speech information to identify the start and end positions of the speech information; segments the effective speech information based on the start and end positions; and extracts Mel-frequency cepstral coefficient features and waveform encoder features based on the effective speech information.

[0129] The transaction processing device 300 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0130] This application provides a transaction processing device, including: a memory and a processor;

[0131] The memory stores instructions that the computer executes;

[0132] The processor executes computer execution instructions stored in memory, causing the processor to perform any of the methods described above.

[0133] Figure 4 A schematic diagram of the transaction processing equipment provided in this application. Figure 4 As shown, the transaction processing device 400 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the transaction processing device 400 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0134] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0135] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0136] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0137] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0138] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0139] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0140] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0141] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0142] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0143] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0144] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0145] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0146] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0147] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0148] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A transaction processing method, characterized in that, include: Receive user's voice transaction request; Based on the voice transaction request, obtain the user's voice information, as well as the device fingerprint information and environmental information of the client that sent the voice information; Dynamic voiceprint features are extracted based on the voice information, the device fingerprint information, and the environmental information; The dynamic voiceprint features include acoustic features, linguistic features, and environmental features; After fusing the speech information and the dynamic voiceprint features, a fused feature is obtained; A risk score for voice transactions is obtained based on the fusion features; The voice transaction is processed based on the risk score.

2. The method according to claim 1, characterized in that, The risk score for voice transactions obtained based on the fusion features specifically includes: The speech recognition result is obtained based on the fusion features; Based on the voice recognition results and the user's transaction behavior data, a relationship graph of the user's transaction behavior is constructed using a graph neural network; the transaction behavior data includes transaction information, device fingerprints, and geographic information. Based on the relationship graph of the user's transaction behavior and the transaction information corresponding to the transaction request, a risk score for the voice transaction is obtained.

3. The method according to claim 2, characterized in that, The step of constructing a relationship graph of user transaction behavior based on the voice recognition results and user transaction behavior data using a graph neural network specifically includes: The voice identification result, the transaction information, the device fingerprint, and the geographic information are used as heterogeneous nodes, with the user identifier as the root node; wherein, the user identifier is a unique identifier for the user in the system; Edges are generated based on the co-occurrence relationship between nodes. Edges are established between nodes corresponding to the same device or the same geographical location. The weight of the edge is determined by the co-occurrence frequency and the time decay factor. By training a heterogeneous graph using a graph neural network, a relational graph of the user's transaction behavior is obtained.

4. The method according to claim 1, characterized in that, The dynamic voiceprint features are extracted based on the voice information, the device fingerprint information, and the environmental information. The dynamic voiceprint features include acoustic features, linguistic features, and environmental features, specifically including: Based on the speech information, Mel-frequency cepstral coefficient features and waveform encoder features are extracted; Align the Mel-frequency cepstral coefficient features and the waveform encoder features in the time dimension to obtain aligned speech temporal features; Based on the aforementioned speech temporal features, key frames and key frequency bands are selected to generate acoustic features containing spectral and contextual information.

5. The method according to claim 2, characterized in that, The speech recognition result obtained based on the fusion features specifically includes: The probability of speech anomalies is obtained based on the fusion features; The fused features are compared with the registered voiceprint template, and the identity matching probability is output. The voice identification result is obtained by weighted fusion of the voice anomaly probability and the identity matching probability.

6. The method according to claim 1, characterized in that, The processing of the voice transaction based on the risk score specifically includes: If the risk score is greater than or equal to the first threshold, then the voice transaction is blocked; or If the risk score is greater than or equal to the second threshold, then re-verification is triggered; or If the risk score is less than the second threshold, the voice transaction is executed.

7. The method according to claim 6, characterized in that, The method further includes: The first threshold and the second threshold are determined based on the user's historical voice data and transaction behavior data.

8. The method according to claim 4, characterized in that, The extraction of Mel-spectral coefficient features and waveform encoder features based on the speech information further includes: The voice information is subjected to noise reduction processing; Endpoint detection is performed on the noise-reduced speech information to identify the start and end positions of the speech information. The effective voice information of the voice information is segmented based on the start position and the end position; The Mel-frequency cepstral coefficient features and the waveform encoder features are extracted based on the effective speech information.

9. A transaction processing apparatus, characterized in that, include: The receiving module is used to receive users' voice transaction requests; The acquisition module is used to acquire the user's voice information, as well as the device fingerprint information and environmental information associated with the voice information, based on the voice transaction request. The extraction module is used to extract dynamic voiceprint features based on the voice information, the device fingerprint information, and the environmental information; The dynamic voiceprint features include acoustic features, linguistic features, and environmental features; The fusion module is used to fuse the speech information and the dynamic voiceprint features to obtain fused features; The processing module is used to obtain a risk score for voice transactions based on the fusion features; The processing module is also used to process the voice transaction based on the risk score.

10. A transaction processing device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-8.

Images