Data generation method and device based on non-uniform sampling path, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes such as financial science and technology and medical health, and discloses a data generation method and device based on a non-uniform sampling path, equipment and a medium. Quantizing contribution degree values of all time points based on process information, generating contribution degree distribution of the time points, screening a plurality of key time points, constructing a non-uniform sampling path, and gradually generating target data from initial data by using the path and a flow matching generation model. Through a time point contribution degree analysis mechanism, on the basis of the actual contribution of each time point to the data generation result, the key time points are screened out, and the non-uniform sampling path is constructed, so that the generation quality is ensured, the redundant sampling steps are effectively reduced, the reasoning efficiency of the flow matching generation model is improved, and the calculation resource consumption is reduced; and the method has better practical application value.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a data generation method, apparatus, device, and storage medium based on a non-uniform sampling path. Background Technology

[0002] In flow-matching-based text-to-speech (TTS) technologies, current mainstream solutions generally rely on multi-step sampling mechanisms to achieve a gradual transition from random noise to high-quality speech features. While this multi-step sampling can theoretically achieve high speech fidelity, it has exposed a series of technical problems in practical applications that restrict system efficiency and practicality, seriously affecting the widespread adoption of this type of method across various industries.

[0003] Specifically, existing Flow Matching TTS systems generally suffer from technical bottlenecks during the inference phase, including slow inference speed, high computational resource consumption, and an inability to flexibly adapt to different business scenarios. During multi-step sampling, each time step requires an independent forward propagation operation of the neural network. Especially with the 8-16 time steps required to generate high-quality speech, inference latency increases significantly, leading to overall system inefficiency. This problem not only limits the deployment of such models in general real-time speech generation but also severely impacts computationally resource-sensitive industry applications.

[0004] In the fintech sector, scenarios such as intelligent voice customer service, voice risk control prompts, and voice-synthesized reports place stringent demands on the response speed and deployment flexibility of voice generation. However, existing Flow Matching TTS models, with their multi-step sampling mechanisms, suffer from high computational overhead and poor real-time performance during the inference phase, making it difficult to meet the high-concurrency, low-latency business needs of financial systems. This impacts the practical application of voice technology in areas such as financial data interaction, voice notifications, and voice security alerts.

[0005] In the healthcare sector, scenarios such as voice-assisted diagnosis, health broadcasting, and doctor-patient interaction systems also place higher demands on the real-time performance and ease of deployment of speech synthesis technology. Especially in medical device terminals and mobile health management systems, existing multi-step sampling TTS solutions suffer from low inference efficiency and cannot adapt to the limited computing power of edge devices and the real-time interaction-sensitive medical environment, severely hindering the promotion and application of intelligent voice technology in key areas such as remote diagnosis and health monitoring.

[0006] Furthermore, current multi-step sampling processes suffer from a lack of basis for step selection. Related parameters are largely set based on experience, lacking a mechanism for dynamic optimization based on model training results or actual data. This results in some time steps contributing very little to the final speech quality. Redundant sampling steps not only waste computational resources but also further slow down inference speed, reducing overall system efficiency. This sampling strategy, lacking data-driven optimization, directly limits the practical application value of the Flow Matching TTS model in industries such as finance and healthcare that require efficient and reliable voice services. Summary of the Invention

[0007] The main objective of this invention is to provide a data generation method, apparatus, device, and storage medium based on a non-uniform sampling path, aiming to solve the technical problem that the lack of an invalid sampling step reduction mechanism based on the contribution of time points in the prior art leads to low inference efficiency and serious redundant calculations in the flow matching generation model.

[0008] To achieve the above objectives, the present invention provides a data generation method based on a non-uniform sampling path, comprising:

[0009] Construct a stream matching generation model for data transformation;

[0010] Obtain the process information of the flow matching generation model within a preset time interval;

[0011] The contribution value of each time point within the preset time interval is quantified based on the process information.

[0012] Generate a time-point contribution distribution based on the aforementioned contribution values;

[0013] Based on the contribution distribution at the aforementioned time points, multiple key time points are selected to construct a non-uniform sampling path;

[0014] Using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from initial data.

[0015] Furthermore, to achieve the above objectives, the present invention provides a data generation apparatus based on a non-uniform sampling path, comprising:

[0016] The model building module is used to build streaming matching generation models for data transformation;

[0017] The information extraction module is used to obtain the process information of the flow matching generation model within a preset time interval;

[0018] The contribution calculation module is used to quantify the contribution value of each time point within the preset time interval based on the process information.

[0019] A distribution generation module is used to generate a time-point contribution distribution based on the contribution value.

[0020] The path construction module is used to filter multiple key time points based on the contribution distribution of the time points and construct a non-uniform sampling path;

[0021] The data generation module is used to iteratively generate target data from initial data using the non-uniform sampling path and the flow matching generation model.

[0022] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a data generation program based on a non-uniform sampling path stored in the memory and executable on the processor, wherein when the data generation program based on a non-uniform sampling path is executed by the processor, it implements the steps of the data generation method based on a non-uniform sampling path as described above.

[0023] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a data generation program based on a non-uniform sampling path, wherein when the data generation program based on a non-uniform sampling path is executed by a processor, it implements the steps of the data generation method based on a non-uniform sampling path as described above.

[0024] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data generation method, apparatus, device, and medium based on a non-uniform sampling path, comprising: constructing a flow matching generation model for data transformation; acquiring process information of the flow matching generation model within a preset time interval; quantifying the contribution value of each time point within the preset time interval based on the process information; generating a time point contribution distribution based on the contribution values; selecting multiple key time points based on the time point contribution distribution; constructing a non-uniform sampling path; and using the non-uniform sampling path and the flow matching generation model to iteratively generate target data from initial data. This invention, through a time point contribution analysis mechanism, selects key time points based on the actual contribution of each time point to the data generation result, constructs a non-uniform sampling path, and effectively reduces redundant sampling steps while ensuring generation quality, improving the inference efficiency of the flow matching generation model, reducing computational resource consumption, and possessing better practical application value. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0026] Figure 1 This is a schematic diagram of an application environment for a data generation method based on a non-uniform sampling path according to an embodiment of the present invention;

[0027] Figure 2 This is a flowchart illustrating an embodiment of the data generation method based on a non-uniform sampling path according to the present invention;

[0028] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data generation device based on a non-uniform sampling path of the present invention;

[0029] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0030] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0031] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] The data generation method based on non-uniform sampling paths provided in this invention can be applied to, for example... Figure 1 In this application environment, the user terminal communicates with the server via a network. The server can construct a stream matching generation model for data transformation using the user terminal, obtain process information of the stream matching generation model within a preset time interval, quantify the contribution value of each time point within the preset time interval based on the process information, generate a time point contribution distribution based on the contribution value, filter multiple key time points based on the time point contribution distribution, construct a non-uniform sampling path, and use the non-uniform sampling path and the stream matching generation model to iteratively generate target data from initial data. This invention, through a time point contribution analysis mechanism, filters key time points based on the actual contribution of each time point to the data generation result, constructs a non-uniform sampling path, effectively reduces redundant sampling steps while ensuring generation quality, improves the inference efficiency of the stream matching generation model, reduces computational resource consumption, and has better practical application value. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster composed of multiple servers. The invention will be described in detail below through specific embodiments.

[0033] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the data generation method based on a non-uniform sampling path provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0034] like Figure 2 As shown, the data generation method based on a non-uniform sampling path proposed in this invention includes the following steps:

[0035] S10, Construct a stream matching generation model for data transformation;

[0036] In this embodiment, the process of constructing a flow matching generation model for data transformation includes three steps: defining the data transformation task, constructing the velocity field function, and setting and optimizing the neural network architecture. Data transformation refers to the process of mapping initial data to target data through continuous changes. The definition of the data transformation task depends on the mathematical expression or statistical characteristics of the target data distribution. The differences between the initial data and the target data can be reflected in data dimension, probability distribution, structural attributes, or feature expression methods, and the transformation path needs to be determined through data analysis.

[0037] In flow matching generative models, "flow" refers to the continuous evolution of data states over time or pseudo-time variables. The model needs to capture the dynamic changes of data within a continuous time interval. Matching is manifested in optimizing the generated data to be consistent with the target data in terms of statistical distribution, feature representation, and structural attributes. A generative model is a structure capable of generating output data through inference or sampling of input data. It typically exists in a differentiable and parameterized form, facilitating parameter adjustment through gradient optimization.

[0038] In practice, the first step is to define a velocity field function based on the data transformation task. The velocity field function describes the rate and direction of change of the data state within a time interval. Mathematically, it is expressed as the derivative of the data state with respect to time. The inputs to the function include the data state and time variables, and the output is the increment of the data state at that point in time. The specific form of the velocity field function depends on the data type and task requirements, and it must be able to express the dynamic characteristics of data evolution.

[0039] When constructing the velocity field function, a neural network architecture needs to be defined. The network input consists of the current data state and time information, and the output is the increment of the data state change. The neural network architecture design must consider the input data dimensionality, the data feature representation format, the model's expressive power, and computational efficiency. Commonly used structures include multilayer perceptrons, convolutional networks, graph networks, or structures combined with attention mechanisms. The network depth and width are adjusted according to the data complexity.

[0040] After setting the neural network structure, the model parameters need to be determined through an optimization process. During optimization, a flow matching loss function is defined to measure the difference between the generated data and the theoretical target data at each time point. The flow matching loss function can be constructed based on probability distribution differences, feature reconstruction errors, or path stability metrics. Using backpropagation and gradient descent algorithms, the model parameters are iteratively updated based on the training data, and the optimization process continues until the loss function converges or meets the set conditions.

[0041] The above process must ensure the continuity of data transformation, the sufficiency of model representation, and the stability of the training process, avoiding problems such as data state jumps, pattern collapse, or numerical instability during data generation. The smoothness of the data transformation path and the high fidelity of the generated data depend on the accurate expression of the velocity field function and the effective optimization of model parameters. Model training can combine multiple batches of samples to cover different data distributions and structures, improving the model's generalization ability and applicability.

[0042] The design of the velocity field function and the neural network architecture can be flexibly adjusted to suit different data types and application requirements. In speech data processing, deep convolutional networks can be used to construct the velocity field function, and the convolutional structure can be used to capture the local features and contextual information of the speech spectrum. In financial data processing, recurrent neural networks or Transformer structures can be used to model the long-term dependencies and dynamic trends of time series data. In medical data processing, multimodal data input can be combined, and a fusion network structure can be used to process multiple sources of information such as physiological signals and monitoring indicators simultaneously, ensuring that the velocity field function accurately reflects the collaborative change characteristics of multidimensional data.

[0043] The configuration of the loss function can also be optimized according to different scenarios. In speech generation tasks, the loss function can be combined with spectral distortion and speech clarity indicators to quantify the difference between the generated data and the target speech data. In financial data generation tasks, distribution deviation measures and risk exposure indicators can be designed to ensure that the statistical characteristics of the generated data are consistent with the real data. In medical data generation, clinical indicator constraints and data stability constraints can be introduced to improve the reliability and applicability of the generated data in medical diagnosis or auxiliary analysis.

[0044] Example Description: In the fintech business field, applications such as remote authentication, intelligent customer service, and speech synthesis broadcasting have high requirements for real-time performance and sound quality in speech generation. They need to quickly, clearly, and naturally output synthesized speech containing financial terminology and data information through a speech generation system. In such scenarios, a stream matching generation model is constructed based on a data conversion task. First, a continuous mapping process is defined from standard Gaussian noise data to commonly used speech feature data in the financial field. Speech feature data can include pronunciation dynamics, emotional stability, and terminology clarity in banking broadcasts. Considering the standardized requirements of speech styles in financial scenarios, an adaptive velocity field function is designed to ensure that the model can express speech smoothly while preserving clarity. Through a structurally optimized neural network architecture, the accuracy of expressing key domain vocabulary, monetary readings, and time information during the speech generation process is improved, ensuring the reliability and understandability of financial voice applications.

[0045] In the healthcare field, voice interaction functions in systems such as assisted diagnosis, telemedicine, and health monitoring need to generate highly natural voice information with a human-like, calm tone that does not hinder patient comprehension. To achieve this goal, the data conversion task is designed with a clear continuous mapping path from Gaussian noise data to medical voice feature data. This voice feature data must cover the tone, rhythm, and timbre characteristics of doctor's instructions, health advice, and medical prompts. When constructing the velocity field function, a time-dynamically smoothing velocity change mechanism is designed to address the soothing and professional requirements of medical voice features, avoiding abrupt changes, distortion, or emotional biases during voice generation. During training, the parameters of the stream matching generation model are optimized using a large-scale medical voice dataset to ensure high naturalness, comfortable timbre, and semantic clarity when patients receive the voice information, meeting the voice generation needs of healthcare scenarios.

[0046] This embodiment, through the joint design based on the velocity field function and the neural network architecture, combined with the parameter optimization process of the flow matching loss function, can effectively construct a flow matching generation model with data transformation capabilities, improve the accuracy of distribution matching and the smoothness of path evolution during the data generation process, reduce random fluctuations and abnormal deviations during the data generation process, enhance the reliability and application adaptability of data generation, and meet the actual needs of high-quality data generation in multiple fields such as voice, finance, and medicine.

[0047] S20, Obtain the process information of the flow matching generation model within a preset time interval;

[0048] In this embodiment, acquiring process information of the stream matching generation model within a preset time interval involves dynamically evaluating the model's operating state and output characteristics based on time series. The stream matching generation model refers to a neural network structure that controls the evolution of data states through time parameters. Its function is to gradually map random noise data into speech feature data with specific distribution characteristics. The preset time interval is a continuous time parameter range defined within the model during the data evolution process, typically standardized to a range of 0 to 1, used to control the entire process of the data state transitioning from initial noise to target speech features.

[0049] To obtain process information of the model within this time interval, multiple time points need to be selected first. The selection of time points can be uniformly distributed or non-uniformly distributed in combination with prior analysis. The number of time points is adjusted according to the requirements of speech generation quality and computational efficiency, with a common range of 8 to 64 time points. Each time point represents an intermediate result of the data state at a specific evolution stage.

[0050] At each time point, a forward propagation operation is performed on the flow matching generation model. The forward propagation process takes the current time point and data state as input and outputs the updated data state or the corresponding velocity field information. Through the forward propagation operation, the dynamic performance of the model at different time points can be captured, forming an intermediate result set that reflects the evolution trend of the data state.

[0051] The loss value distribution data is extracted from the forward propagation output. The loss value distribution data refers to the set of numerical values ​​that quantify the deviation between the model output and the theoretical ideal state. It is usually calculated based on Euclidean distance, KL divergence, or other statistical indicators, and reflects the model's data fitting error and convergence status at each time point.

[0052] Based on the forward propagation output, the gradient vector norm is further determined. The gradient vector norm is the magnitude of the gradient vector in the parameter space, which measures the sensitivity and optimization strength of the model to parameter updates at the current time point, reflecting the learning dynamics and stability of the model at a specific time point.

[0053] By extracting the loss value distribution data and gradient vector norm, a complete set of process information is formed. This process information covers the dynamic characteristics, performance indicators, and parameter change trends of the model at each time point within a preset time interval, providing a data foundation to support subsequent contribution quantification, key time point selection, and the construction of non-uniform sampling paths.

[0054] In the specific implementation, the preset time interval is set to the standard normalized interval from 0 to 1, and the time points are selected as an equally spaced sequence of 0.0, 0.1, 0.2 up to 1.0, for a total of 11 time points. Alternatively, an adaptive time point layout with dense coverage of phoneme boundaries or speech transition stages can be adopted based on the characteristics of speech data evolution. In the forward propagation operation, the input data state is the data feature matrix at the current time point. Time parameters are embedded in the form of sine and cosine time encoding to enhance the model's ability to express temporal dynamics. The output includes the updated data state and velocity field increment information. The loss value distribution data is obtained by calculating the Euclidean distance with the target speech features. The gradient vector norm is calculated using an automatic differentiation mechanism to calculate the gradient of all trainable parameters, solving for the 2-norm of the gradient vector to reflect the parameter update intensity at each time point. The process information set is stored in dictionary structure or tensor form, containing the loss value and gradient norm corresponding to all time points, supporting subsequent data analysis and contribution calculation.

[0055] Example description: In the field of healthcare, remote diagnosis and treatment voice systems need to generate highly natural and stable voice information. During the information acquisition process, combined with the specific needs of heart rate monitoring voice broadcast, the loss distribution and gradient changes of the model at each time point are dynamically monitored to ensure that the voice output is smooth and without abrupt changes, thus ensuring the comfort and accuracy of patients receiving information.

[0056] In the fintech business, intelligent voice customer service systems need to generate clear, high-fidelity voice during account information broadcasts and risk warnings. During the process of acquiring information, the system should focus on the dynamics of the model during the period involving numbers and amounts. Through process information evaluation and optimization, the system can improve the clarity and generation efficiency of voice during the important information broadcasting stage, thereby enhancing the voice interaction experience of financial services.

[0057] This embodiment obtains process information of the stream matching generation model within a preset time interval, which can comprehensively reflect the dynamic characteristics and performance of the model in the continuous evolution of data state. The intermediate index acquisition method based on time distribution avoids the limitations of single terminal result evaluation, improves the scientificity and data support level of model structure optimization, parameter adjustment and sampling path design, and ensures the coordinated balance between generation quality and inference efficiency in speech generation tasks.

[0058] S30, quantify the contribution value of each time point within the preset time interval based on the process information;

[0059] In this embodiment, the contribution value of each time point within a preset time interval is quantified based on process information. This involves analyzing dynamic data during the model execution process to measure the actual impact of different time points on the final speech generation effect. The process information comes from the dynamic monitoring data of the stream matching generation model as the data state evolves from initial noise to target speech features. Specifically, it includes the loss value distribution data and gradient vector norm corresponding to each time point. The loss value distribution data reflects the fitting error level of the model at that time point, and the gradient vector norm reflects the optimization strength and sensitivity of the parameter space.

[0060] The contribution value is a quantitative indicator calculated independently for each time point. The larger the value, the greater the contribution of that time point to the quality of the result in the data evolution path. The smaller the value, the more negligible the impact of that time point on the speech generation result.

[0061] During the quantization process, the loss value distribution data and gradient vector norm at each time point are first extracted from the process information. The loss value distribution data is obtained by calculating the Euclidean distance or cosine similarity between the output features and the target speech features, and the gradient vector norm is obtained by calculating the L2 norm of the gradients of all trainable parameters.

[0062] After extraction, calculate the global statistical characteristics of the loss value distribution data at all time points. The global statistical characteristics include the mean, standard deviation or other statistical descriptors, which are used to characterize the overall error level and distribution characteristics, and avoid contribution bias caused by local fluctuations.

[0063] Based on global statistical characteristics, the loss value distribution data at each time point is normalized. Common processing methods include z-score standardization or minimum-maximum scaling to generate normalized loss values. The normalization operation ensures that the error index at different time points is on a uniform numerical scale, eliminating the influence of dimensions.

[0064] In parallel, the information entropy value at each time point is determined based on the gradient vector norm. The information entropy value reflects the model's dynamic learning ability at that time point by measuring the uniformity and uncertainty of the gradient direction distribution. The higher the information entropy, the more sensitive the model is to data updates at that time point and the more fully the structural adjustment is performed.

[0065] Finally, the normalized loss value and the information entropy value are fused to generate the contribution value. The fusion method can be weighted summation, nonlinear combination, or adaptive mapping function, which comprehensively reflects the error control effect and model learning intensity at time points, ensuring that the contribution value has stability, discriminative power, and interpretability.

[0066] In practical implementation, process information is stored using a tensor structure, with data recorded independently at each time point. Mean squared error is used as the indicator for the loss value distribution data, and the gradient vector norm is calculated using an automatic differentiation framework, with the L2 norm used for solution. Global statistical features employ both mean and standard deviation as indicators, and normalization is performed using standard scores. Information entropy is calculated based on the covariance matrix of the gradient direction, and the fusion operation uses a weighted sum of the normalized loss value and the information entropy value, with weight coefficients determined experimentally or learned dynamically. The contribution value set is stored in vector form and subsequently used for contribution ranking, key time point selection, and non-uniform sampling path construction.

[0067] Example description: In the field of healthcare, voice interaction systems are used in chronic disease management reminder scenarios. During the contribution quantification process, the focus is on optimizing the timing of medication time and dosage announcements to ensure the clarity and stability of voice output and avoid information loss that could affect patient medication adherence.

[0068] In the fintech business, voice customer service systems handle high-precision needs such as account balance announcements and transaction confirmations. During the contribution quantification process, the system dynamically identifies the time points that affect key segments of the voice message, eliminates steps that contribute little to the result, improves response speed, reduces computing resource consumption, and ensures user experience and accuracy of information transmission.

[0069] This embodiment quantifies the contribution value of each time point based on process information, thereby achieving accurate identification of the influence of each time segment in the entire process of data evolution. It avoids the inefficiency waste caused by fixed-step sampling, supports dynamic optimization of time point selection based on actual data, improves the inference speed and generation quality of the speech generation system, balances system performance and computing resource consumption, and adapts to the real-time and high-fidelity requirements of different application scenarios.

[0070] S40, Generate a time-point contribution distribution based on the contribution value;

[0071] In this embodiment, the contribution distribution at each time point is generated based on the contribution values. This involves constructing a structured expression that reflects the trend and distribution of contribution changes at each time point within the overall time series, based on the contribution values ​​independently quantified for each time point in the preceding stages. The contribution values ​​are derived from process information analysis and have been quantified in terms of importance and actual impact of each time point in the data state evolution process by integrating error indicators and parameter gradient information.

[0072] The contribution distribution at each time point presents the overall distribution structure of the contribution at each time point within a preset time interval in a probabilistic statistical manner. This can help identify concentrated areas of key time points and time periods with low contribution, providing a data foundation for the subsequent construction of non-uniform sampling paths.

[0073] In the specific implementation process, the contribution values ​​of all selected time points within the preset time interval are first collected. The contribution values ​​are stored in a one-dimensional vector structure, indexed according to the order of the corresponding time points, and the values ​​reflect the independent contribution effect of a single time point.

[0074] After data collection is completed, the statistical distribution parameters of the contribution values ​​are calculated. These parameters include indicators such as mean, variance, extreme values, skewness, and kurtosis. They are used to quantify the overall level, fluctuation range, and distribution pattern of the contribution, providing basic statistical information for subsequent probability modeling.

[0075] Based on statistical distribution parameters, a probability density function is constructed. The probability density function uses Gaussian kernel density estimation, parameterized distribution fitting, or adaptive data-driven methods to accurately characterize the continuity and density changes of contribution values ​​in the time series, reflecting the probability distribution relationship of contribution levels at different time points.

[0076] The contribution distribution curve at time points is generated by a probability density function. The contribution distribution curve is expressed as a two-dimensional function, with the horizontal axis representing the time point sequence and the vertical axis representing the corresponding contribution probability density value. The curve is continuous and smooth, clearly showing the distribution position and amplitude changes of high contribution and low contribution regions.

[0077] Finally, the contribution distribution curve at each time point is used as the contribution distribution at each time point for subsequent screening of key time points and construction of non-uniform sampling paths. This ensures that the data representation structure is clear, the quantitative indicators are accurate, and the distribution information is complete, making it easier for subsequent operations to dynamically optimize the time point selection strategy based on this distribution information.

[0078] In implementation, contribution values ​​are stored in a data structure array, with index positions strictly corresponding to time points, and data types supporting high-precision floating-point representation. Statistical distribution parameters are calculated using a statistical calculation module, employing sliding window or global analysis methods to flexibly adapt to different time interval lengths. The probability density function is constructed preferably based on kernel density estimation methods, with kernel functions such as Gaussian kernels, Epanechnikov kernels, or other smooth functions. The bandwidth parameter is automatically adjusted according to the data distribution to ensure smooth curve representation and good resolution. Time-point contribution distribution curves are generated through a data visualization module or a mathematical modeling module, using continuous function curves or discrete point sequence interpolation reconstruction to adapt to different system call requirements, ensuring the completeness and practicality of the distribution information. While generating the contribution distribution curves, the system can set the persistent storage method for distribution information according to specific needs, supporting local caching, remote transmission, or memory mapping, improving overall system efficiency and data retrieval speed.

[0079] Example: In the healthcare business, voice generation systems are used for personalized health reminders and disease risk broadcasts. During the contribution distribution generation process, by accurately analyzing key changes in voice information, the system ensures that key information such as heart rate prompts and drug dosage information is expressed clearly and accurately, reducing the computational consumption of redundant time points and improving the system's reasoning speed and information transmission reliability.

[0080] In the fintech business, intelligent voice systems are applied to account information broadcasting, risk warnings, and business confirmation scenarios. During the contribution distribution generation process, the system dynamically identifies high-contribution time points corresponding to important information segments such as transaction amounts and interest rate changes, optimizes the voice generation steps, reduces processing latency, and improves the overall system response speed and the accuracy and stability of financial information transmission, meeting the needs of high-concurrency and low-latency financial services.

[0081] This embodiment generates a time-point contribution distribution based on the contribution value. The system can comprehensively and intuitively express the actual contribution level of each time point in the time series, identify high contribution concentration areas and low contribution redundant time periods, avoid the efficiency waste caused by blind uniform sampling, support subsequent dynamic optimization of time point selection and path construction, improve speech generation efficiency and quality, reduce system resource consumption, and enhance deployment flexibility and real-time performance in different application environments.

[0082] S50, based on the contribution distribution of the time points, select multiple key time points and construct a non-uniform sampling path;

[0083] In this embodiment, multiple key time points are selected based on the distribution of contribution at each time point, and a non-uniform sampling path is constructed. This involves analyzing the distribution of contribution at each time point to determine the set of time points that have a significant impact on the generation of target data during the evolution of data state, and using this set to form a non-uniformly distributed time point path within the time series. This path is used to guide the selection of key steps and the efficient use of computing resources in the subsequent data generation process.

[0084] The contribution distribution at each time point originates from the probability density expression constructed based on the contribution values ​​of the preceding stages, fully reflecting the relative importance and distribution pattern of each time point within the preset time interval. Quantitative analysis of this distribution information can effectively identify areas with concentrated high contributions, distinguish between redundant and critical time points, and avoid resource waste caused by uniform processing of all time points.

[0085] In the specific operation process, the average contribution and standard deviation are first determined based on the contribution distribution at all time points. The average contribution is obtained by averaging the values ​​under the contribution distribution curve at all time points, and the standard deviation is used to measure the degree of fluctuation of the contribution level in the time series. Together, they constitute the basic parameters reflecting the overall level and dispersion trend of the contribution.

[0086] After the contribution statistics parameters are determined, the contribution threshold parameters are determined based on the average contribution value and standard deviation, combined with the preset screening rules. The contribution threshold parameters can be set using the fixed multiple standard deviation offset method, the adaptive quantile method, or the dynamic adjustment method based on the business scenario, to ensure the rationality and flexibility of the threshold setting and adapt to different time series characteristics and system performance requirements.

[0087] After determining the contribution threshold parameter, the system iterates through the contribution value corresponding to each time point in the contribution distribution, compares its value with the contribution threshold parameter, and selects the time points with contribution values ​​greater than the contribution threshold parameter as key time points, thus forming a set of time points with actual contributions.

[0088] After the set of key time points is generated, the system sorts them according to their order within the time series. The sorting method maintains an ascending time structure, avoids introducing disordered interference during the selection of time points, and ensures the continuity and logical integrity of the path in the subsequent data generation process.

[0089] Finally, the sorted set of key time points is combined to form a non-uniform sampling path. The non-uniform sampling path exhibits both clustering and sparseness within the time interval. The time point density is higher in high contribution time periods and sparser in low contribution time periods. The overall path structure fully reflects the contribution distribution pattern within the time series, providing a precise basis for time point selection and an efficient resource allocation strategy for the subsequent data generation process.

[0090] In its implementation, the average contribution and standard deviation are calculated in real time using a mathematical statistics module, supporting dynamic updates and multiple iterations, making it suitable for situations where contribution levels vary significantly across different training cycles. The contribution threshold parameter can be determined using a standard deviation offset method based on empirical coefficients, where the threshold is set to the average contribution plus a certain multiple of the standard deviation, or by automatically adjusting the threshold based on changes in contribution distribution density using a data adaptive algorithm, improving the accuracy of the screening results and the system's adaptability. The screening and sorting process for time points is implemented using efficient data structures such as hash maps and balanced tree structures, ensuring efficient screening and sorting for large-scale time point sets. The generation and storage of non-uniform sampling paths support arrays, linked lists, or custom path structures, adapting to different system architectures and calling requirements, improving the calling efficiency of path structures and the overall system processing speed. The system can optimize the non-uniform sampling path structure according to business needs. For example, in speech generation tasks, the density of key time points can be appropriately increased for specific speech segments or areas with drastic changes in speech features to ensure speech quality and information transmission accuracy. In resource-constrained scenarios, the system dynamically adjusts the threshold parameter to flexibly control the number of key time points, achieving a balance between computational resources and generation quality.

[0091] Example Description: In the healthcare business field, voice generation systems are used for personalized health reminders, voice diagnosis assistance, and disease information broadcasting. By filtering key time points based on contribution distribution and constructing non-uniform sampling paths, the generation steps of high-value information segments such as heart rate change prompts and drug dosage broadcasts can be more refined, while reducing redundant calculations in low-value time periods, improving the system's real-time response capability and voice output quality, and ensuring the accuracy and timeliness of medical information transmission.

[0092] In the fintech business sector, intelligent voice systems are applied to scenarios such as account change announcements, transaction risk alerts, and business confirmations. Based on a non-uniform sampling path constructed using contribution distribution, it can dynamically focus on the voice expression of key information such as transaction amount and interest rate changes, optimize the generation step structure, reduce system inference latency, improve the real-time performance of voice interaction and the stability of information transmission, and meet the demand for efficient voice services in high-concurrency financial business environments.

[0093] This embodiment filters multiple key time points based on the distribution of contribution at each time point and constructs a non-uniform sampling path. This effectively identifies time point regions within the time series that contribute significantly to the generation of target data, avoids redundant calculations, improves efficiency and resource utilization in the data generation process, reduces system inference latency, and enhances the system's deployment flexibility and real-time performance in different application scenarios. It has significant advantages, especially in scenarios where computing resources are limited or where there are high requirements for generation quality.

[0094] S60, using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from the initial data.

[0095] In this embodiment, a non-uniform sampling path and a flow matching generation model are used to iteratively generate target data from initial data. This involves executing multiple rounds of data state update operations sequentially based on a selected sequence of key time points, combined with the dynamic calculation mechanism of the flow matching generation model, ultimately achieving a continuous transition from the initial data state to the target data state. The non-uniform sampling path is generated by the preceding steps and contains several time points. These time points are unevenly distributed within the time interval, exhibiting a structure characterized by dense high-contribution time periods and sparse low-contribution time periods, ensuring that the generation process focuses on time nodes that significantly impact the quality of the results.

[0096] The initial data serves as the starting input for the data generation process. Its specific form depends on the target task. For example, in a speech generation task, the initial data can be a Gaussian noise sequence or a standard distribution random data, which has good randomness and distribution balance, providing a mathematical basis and diversity guarantee for the subsequent evolution of the data state.

[0097] The current data state is a temporary output result during the data generation process. Each time, the system calculates the velocity field and updates the current data state, pushing the data state to gradually approach the target data structure. The velocity field reflects the evolution trend and rate of change of the data state at a specific point in time, and is dynamically calculated by the flow matching generation model.

[0098] The flow matching generation model has end-to-end data fitting capability. Based on the trained parameter structure, it can take the current data state as input at any specified time point and output the velocity field information of that state at that time point, ensuring that the data state update process conforms to the physical or statistical laws of the target data distribution.

[0099] In practice, the system first loads a non-uniform sampling path, extracts key time point sequences, and initializes the current data state on the initial data. Then, the system sequentially selects time points from the key time point sequences and processes each time point in chronological order. At each time point, the system inputs the current data state into the flow matching generation model, calculates the velocity field information for the corresponding time point, and then adjusts the current data state based on the velocity field. Typically, the Euler method or higher-order integral methods are used to update the state, ensuring the numerical stability and physical plausibility of the data evolution. This process terminates after all key time points have been processed, and the output current data state is the target data, possessing a statistical structure and representation consistent with the target distribution or characteristics, meeting the data generation requirements of a specific task.

[0100] In practical applications, the number and distribution structure of time points in non-uniform sampling paths can be dynamically adjusted based on contribution analysis results to adapt to different data complexities and computational resource constraints. The initial data generation method supports standard Gaussian distribution, uniform distribution, or a custom distribution structure based on task requirements, enhancing generation diversity and system adaptability. In the velocity field calculation process, in addition to the standard flow matching generation model, improved model structures such as residual connections, conditional control mechanisms, or multi-level feature fusion can be incorporated to enhance model expressiveness and fitting accuracy under complex data structures. The current data state update method can be based on the Euler method, Runge-Kutta method, or other higher-order numerical integration strategies to balance computational efficiency and update accuracy, adapting to the real-time and generation quality requirements of different business scenarios. In specific tasks, the system can also incorporate a dynamic step-size adjustment mechanism to flexibly adjust the time span and step distribution of each update round based on the magnitude of changes in the current data state and generation quality indicators, further optimizing the efficiency of the data generation process and the stability of the output results.

[0101] Example Description: In the healthcare field, speech synthesis systems are used for disease warning information broadcasting, personalized rehabilitation guidance, and health reminders. By using non-uniform sampling paths and flow matching generation models, the system can optimize the generation steps of key speech segments such as heart rate changes and medication prompts, ensuring the quality of speech output and the accuracy of information transmission. At the same time, it reduces the computational consumption of redundant time periods, improves the system's real-time response capability, and adapts to the application requirements of high reliability and low latency in the medical environment.

[0102] In the fintech field, intelligent voice systems are used for transaction confirmation, risk alerts, and account information broadcasting. Based on non-uniform sampling paths and flow matching generation models, they can dynamically focus on the expression of key information such as amount, interest rate, and transaction status, optimize data generation paths, reduce inference latency, and improve the real-time performance and interaction quality of voice services, meeting the dual requirements of high concurrency, low latency, and accurate information transmission in financial business scenarios.

[0103] This embodiment uses a non-uniform sampling path and a flow matching generation model to iteratively generate target data from initial data. This effectively reduces redundant computational steps, lowers system inference latency, and optimizes computational resource utilization while maintaining data generation quality. The non-uniform sampling path ensures that the generation process focuses on time periods that significantly impact the results, while the flow matching generation model provides high-precision control over the direction and magnitude of state evolution. Overall, this improves the system's practical application performance and deployment flexibility in high-quality data generation tasks.

[0104] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data generation method, apparatus, device, and medium based on a non-uniform sampling path, comprising: constructing a flow matching generation model for data transformation; acquiring process information of the flow matching generation model within a preset time interval; quantifying the contribution value of each time point within the preset time interval based on the process information; generating a time point contribution distribution based on the contribution values; selecting multiple key time points based on the time point contribution distribution; constructing a non-uniform sampling path; and using the non-uniform sampling path and the flow matching generation model to iteratively generate target data from initial data. This invention, through a time point contribution analysis mechanism, selects key time points based on the actual contribution of each time point to the data generation result, and constructs a non-uniform sampling path. While ensuring generation quality, it effectively reduces redundant sampling steps, improves the inference efficiency of the flow matching generation model, reduces computational resource consumption, and has better practical application value.

[0105] In one embodiment, step S10 includes:

[0106] S101 defines the data transformation task from the initial distribution to the target distribution;

[0107] S102, Construct a velocity field function based on the data conversion task, and determine a neural network architecture for fitting the velocity field function;

[0108] S103, optimize the parameters of the neural network architecture using the flow matching loss function to generate a flow matching generation model.

[0109] In this embodiment, a stream matching generation model for data transformation is constructed. This involves building a mathematical structure with continuous mapping capabilities based on the target data characteristics and generation requirements, and using a data-driven method to fit the process of smoothly transitioning from the initial data distribution to the target data distribution. This process first requires defining the data transformation task from the initial distribution to the target distribution. The initial distribution is the starting point of the data generation process, typically a Gaussian distribution, a uniform distribution, or a random distribution set according to the actual task requirements, possessing controllability and distribution balance. The target distribution is the output result of the data generation process, required to be consistent with the expected data structure, statistical characteristics, and physical properties of the task, and is widely applicable to the expression of various complex data structures such as speech, images, and text.

[0110] Once the data transformation task is clearly defined, a velocity field function is constructed based on this task. The velocity field function describes the dynamic evolution direction and rate of change of the data in the continuous time domain, and is usually expressed mathematically as a first-order ordinary differential equation or continuous-time modeling. This function reflects how the data state gradually approaches the target distribution from the initial distribution over time, exhibiting spatial location dependence and temporal parameter correlation, and can characterize the local nonlinear change trend in complex data transformation processes.

[0111] To effectively fit the velocity field function, a neural network architecture with sufficient expressive power needs to be determined. The neural network architecture includes the design of network depth, width, activation function type, and structural connection method. Common architecture forms may include multilayer perceptrons, residual networks, temporal modeling networks, or complex structures combined with conditional control mechanisms, to ensure that the model has the ability to capture high-dimensional data features, express complex mapping relationships, and adapt to different task requirements.

[0112] After the neural network architecture is determined, the system optimizes the network parameters through the flow matching loss function. The flow matching loss function measures the difference between the velocity field output by the neural network and the theoretical ideal velocity field. It often uses second-order moment matching, distribution divergence measurement, or error function derived from theoretical path to ensure that the model training process has convergence and optimality. The resulting flow matching generation model can stably and efficiently simulate the continuous evolution of data from the initial distribution to the target distribution, meeting the application needs of diverse tasks such as data generation, transformation, or structure reconstruction.

[0113] This embodiment defines a data transformation task from an initial distribution to a target distribution, constructs a velocity field function with dynamic expressive capabilities, combines it with a highly expressive neural network architecture, and optimizes parameters using a flow matching loss function. This enables the efficient and stable acquisition of a flow matching generation model that accurately describes the continuous evolution of the data state, ensuring good smoothness, stability, and output quality in the data generation process, and improving the system's practicality and adaptability in various complex data generation tasks.

[0114] In one embodiment, step S20 above includes:

[0115] S201, Select multiple time points within a preset time interval;

[0116] S202, At each time point, the forward propagation operation of the flow matching generation model is performed to generate the time point model output;

[0117] S203, Extract the loss value distribution data from the time point model output;

[0118] S204, Determine the gradient vector norm based on the time point model output;

[0119] S205, aggregate the loss value distribution data and the gradient vector norm to form process information.

[0120] In this embodiment, acquiring the process information of the flow matching generation model within a preset time interval refers to systematically analyzing the model's state evolution and output performance under different time parameter conditions to construct an overall data structure that reflects the model's dynamic characteristics. First, multiple time points are selected within the preset time interval. The preset time interval is a continuous time range set by the system, typically covering the entire process from the initial data distribution to a distribution approaching the target data distribution. Time point selection can employ equal intervals, variable intervals, or adaptive density strategies to ensure coverage of key dynamic nodes and facilitate capturing the model's specific performance at different evolutionary stages.

[0121] At each time point, the forward propagation operation of the flow matching generation model is performed. Forward propagation refers to inputting the current data state into the model, combining it with the time parameters at that time point, and calculating the model output. The forward propagation process involves calling the velocity field function, calculating neural network parameters, and generating output features. The output is the time point model output. The time point model output reflects the transformation trend and evolution direction of the data state at that specific time point and is a direct data source for measuring the dynamic behavior and performance of the model.

[0122] Loss value distribution data is extracted from the model output at time points. The loss value distribution data reflects the error structure, prediction bias and distribution characteristics of the model at the current time point. Specifically, it includes indicators such as the overall error mean, local error distribution and gradient information. The loss value distribution data can be obtained through standard error function, distribution statistical analysis or high-order feature extraction methods, and is used to reveal the performance differences and training status of the model at different time points.

[0123] The gradient vector norm is determined based on the model output at a given time point. The gradient vector norm quantifies the parameter sensitivity and optimization difficulty of the model at the current time point. It is usually calculated by taking the first derivative of the velocity field function or loss function with respect to the model parameters and taking its Euclidean norm or other normative measures. The gradient vector norm reflects the optimization trend and stability of the model under specific time conditions and is an important indicator for evaluating the dynamic stationarity of the model and the intensity of parameter updates.

[0124] By aggregating loss value distribution data and gradient vector norms, process information is formed. Process information is a data set that comprehensively reflects the dynamic performance of the model within a complete time interval. It typically includes the loss value structure, gradient characteristics, and output trends at each time point. It has multi-dimensional, hierarchical, and visual expression capabilities, and can provide complete data support for subsequent contribution analysis, path optimization, and model performance evaluation.

[0125] This embodiment selects time points within a preset time interval and systematically performs forward propagation. By combining the loss value distribution data and gradient vector norm extracted from the time point model output, it can comprehensively and accurately obtain the dynamic process information of the flow matching generation model. It reveals the performance changes, error distribution, and parameter sensitivity of the model during the data transformation process, providing reliable data basis for subsequent optimization of sampling paths and improvement of generation efficiency and accuracy. It significantly improves the problems of arbitrary step selection and lack of dynamic features under the traditional multi-step sampling mechanism.

[0126] In one embodiment, step S30 above includes:

[0127] S301, extract the loss value distribution data and gradient vector norm from the process information;

[0128] S302, determine the global statistical characteristics of the loss value distribution data of all time points selected within the preset time interval;

[0129] S303, Based on the global statistical features, normalize the loss value distribution data at each time point to generate a normalized loss value;

[0130] S304, determine the information entropy value at each time point based on the gradient vector norm;

[0131] S305, the normalized loss value and the information entropy value are combined to generate a contribution value.

[0132] In this embodiment, the contribution value of each time point within a preset time interval is quantified based on process information. This involves using the dynamic performance data of the model at different time points to evaluate the actual impact of each time point on the data generation effect. First, loss value distribution data and gradient vector norm are extracted from the process information. The process information is a dataset that comprehensively reflects the dynamic evolution of the flow matching generation model, including the output structure, error characteristics, and gradient information at multiple time points. The loss value distribution data reflects the prediction error structure, deviation distribution, and convergence trend of the model at each time point, derived from the degree of deviation between the model output and the true target. The gradient vector norm reflects the sensitivity and adjustment magnitude during the parameter update process, derived from the gradient information of the velocity field function or loss function with respect to the model parameters.

[0133] The process involves determining the global statistical characteristics of the loss value distribution data at all selected time points within a preset time interval. These characteristics include maximum, minimum, mean, standard deviation, and distribution interval, quantifying the overall error distribution range and structural patterns, and providing a numerical benchmark for normalization and comparative analysis. This process extracts structured statistical indicators by traversing the loss value distribution data at all time points, forming complete global reference data.

[0134] Based on global statistical characteristics, the loss value distribution data at each time point is normalized to generate normalized loss values. Normalization usually employs range normalization, Z-score standardization, or other normalization methods to ensure that the loss values ​​at different time points are comparable and have a unified dimension, avoiding the interference of absolute numerical differences with the contribution analysis results. The normalized loss value reflects the relative position of the error level at each time point within the overall framework.

[0135] The information entropy value at each time point is determined based on the gradient vector norm. The information entropy value is used to measure the information complexity and uncertainty in the process of adjusting the model parameters at each time point. Specifically, the entropy index is calculated by analyzing the distribution structure of the gradient vector norm. The higher the information entropy value, the richer the amount of information carried in the parameter update and data transformation process at that time point, and the more significant the impact on the overall generation quality.

[0136] The contribution value is generated by fusing the normalized loss value and the information entropy value. The contribution value comprehensively reflects the actual role of each time point in balancing the error level and the information expression ability. Through weighted fusion, nonlinear combination or other data integration methods, a unified and quantitative contribution index is formed, which provides an accurate basis for selecting key time points and optimizing sampling paths.

[0137] This embodiment combines the loss value distribution data and gradient vector norm in the process information to systematically extract statistical features and information entropy indicators. Based on normalization processing, it integrates multi-dimensional data to generate contribution values, which can comprehensively and objectively reflect the actual contribution of each time point within the preset time interval to the data generation results. This avoids the problems of lack of basis for step selection and resource waste under traditional sampling strategies, realizes reasonable time point selection based on dynamic data performance, improves sampling efficiency, reduces inference overhead, and optimizes data generation quality and real-time performance.

[0138] In one embodiment, step S40 above includes:

[0139] S401, Collect the contribution values ​​of all selected time points within the preset time interval;

[0140] S402, determine the statistical distribution parameters of the contribution value;

[0141] S403, Construct a probability density function based on the statistical distribution parameters;

[0142] S404, Generate a time point contribution distribution curve using the probability density function, and use the time point contribution distribution curve as the time point contribution distribution.

[0143] In this embodiment, the contribution distribution at each time point is generated based on the contribution values. Specifically, this involves using statistical analysis and probabilistic modeling methods based on the obtained contribution values ​​to construct a continuous data representation structure that reflects the overall pattern of contribution characteristics at each time point. First, the contribution values ​​of all selected time points within a preset time interval are collected. These contribution values ​​originate from quantitative indicators generated based on process information in previous steps, reflecting the actual strength of the impact of different time points on the data generation effect. This collection process systematically summarizes the corresponding contribution values ​​by traversing all time points within the preset time interval, forming a complete contribution data set.

[0144] After data collection is completed, the statistical distribution parameters of the contribution values ​​are determined. The statistical distribution parameters are used to characterize the overall distribution trend, concentration and dispersion structure of the contribution values. They usually include indicators such as mean, variance, skewness, kurtosis and extreme value range. In the process of determining the parameters, histogram analysis, distribution fitting or other mathematical statistical methods can be combined to accurately quantify the structural characteristics of the contribution values, and provide the necessary mathematical basis for the subsequent construction of the probability density function.

[0145] Based on statistical distribution parameters, a probability density function is constructed. The probability density function is used to describe the relative probability of occurrence and the trend of change of different contribution values ​​in the overall distribution. By selecting an appropriate distribution model (such as Gaussian distribution, beta distribution, gamma distribution or mixed distribution structure) and combining it with the determined statistical parameters, a continuous function expression that conforms to the actual distribution characteristics of contribution values ​​is established. The probability density function can clearly reveal the probability distribution law of each contribution interval and reflect the distribution density and change pattern of different contribution levels in the time dimension.

[0146] The contribution distribution curve at each time point is generated by using a probability density function. This curve serves as the contribution distribution at each time point, and is a graphical representation of the probability density function, reflecting the continuous distribution trend of contribution values ​​within the overall structure. Generated through function calculation and data visualization, this curve visually demonstrates the distribution of contributions at different time points under varying contribution levels. Using this curve as the contribution distribution at each time point establishes a unified reference standard and data foundation, providing a reliable basis for subsequent selection of key time points and optimization of sampling paths.

[0147] This embodiment collects contribution values ​​and accurately extracts statistical distribution parameters through system collection and modeling using probability density functions. It then generates a contribution distribution curve at different time points that reflects the overall trend of contribution levels. This comprehensively depicts the distribution structure of contribution at different time points during the data generation process, avoiding local misjudgment problems caused by single-point data analysis. It provides a clear basis for distribution patterns, which helps to dynamically select key time points based on data characteristics, reduce redundant calculations, improve sampling efficiency, and optimize generation quality and computing resource utilization.

[0148] In one embodiment, step S50 above includes:

[0149] S501, determine the average contribution value and standard deviation based on the contribution value distribution at time points, and determine the contribution value threshold parameter based on the average contribution value and standard deviation;

[0150] S502, compare the contribution value of each time point in the contribution distribution with the contribution threshold parameter;

[0151] S503, Select the time point where the contribution value is greater than the contribution threshold parameter as the key time point;

[0152] S504, Arrange the key time points in chronological order to generate a non-uniform sampling path.

[0153] In this embodiment, multiple key time points are selected based on the time point contribution distribution to construct a non-uniform sampling path. Specifically, this involves using the generated time point contribution distribution, combined with statistical analysis, threshold screening, and sequence organization, to dynamically determine the set of time points that significantly influence the results during data generation. These time points are then arranged sequentially to form a time series sampling path with non-uniform characteristics. First, the average contribution value and standard deviation are determined based on the time point contribution distribution. The average contribution value reflects the overall central tendency of the contribution values ​​across all time points, while the standard deviation measures the dispersion and fluctuation range of the contribution values. Together, they reflect the overall structure of the time point contribution characteristics. The average and standard deviation can be directly extracted from the contribution distribution data through statistical calculations, ensuring the objectivity and accuracy of the parameter source.

[0154] The contribution threshold parameter is determined based on the average contribution value and standard deviation. The contribution threshold parameter is used to define the screening criteria for significant contribution time points. The threshold setting can be flexibly implemented based on the linear combination of the average value and standard deviation, adjustment of the multiplier coefficient, or other mathematical function transformations, to ensure that the screening criteria can take into account both the overall distribution characteristics and the dynamic changes in the specific contribution level, thereby improving the adaptability of the screening process and the data matching effect.

[0155] By comparing the contribution value of each time point in the contribution distribution with the contribution threshold parameter, and by searching the contribution distribution structure of each time point one by one, a quantitative comparison is made between the contribution value and the contribution threshold parameter corresponding to each time point, and the time points where the contribution level is higher than the set threshold standard are identified, thus ensuring the objectivity and data-driven nature of the screening results.

[0156] The time points where the contribution value is greater than the contribution threshold parameter are selected as key time points. Key time points refer to the time nodes that have a significant impact or key role in the final result during the data generation process. The contribution-based screening method can effectively avoid the interference of redundant time points, highlight the characteristics of time intervals with high contribution, and ensure that the subsequent sampling path is targeted and efficient.

[0157] By arranging key time points in chronological order, a non-uniform sampling path is formed. Unlike the equal-interval sampling strategy, the non-uniform sampling path can dynamically adjust the sampling density and distribution structure according to the actual contribution level of different time points. The sequential arrangement of key time points ensures the consistency of the path structure with the time series. The non-uniform layout optimizes the allocation of sampling resources and improves the efficiency and quality of data generation.

[0158] This embodiment dynamically determines the screening threshold based on the contribution distribution and, combined with contribution value comparison, selects a set of time points that significantly contribute to the generation process, forming a non-uniformly distributed sampling path. Compared with traditional sampling strategies with fixed step sizes or empirical settings, it can effectively avoid invalid calculations at low-contribution time points, highlight the optimization role of high-contribution time points, and significantly reduce computational complexity and inference latency while ensuring the quality of generated data, thereby improving the model's practical application performance and resource utilization efficiency.

[0159] In one embodiment, step S60 above includes:

[0160] S601, Load the non-uniform sampling path and obtain the key time point sequence;

[0161] S602, Initialize the current data state based on the initial data;

[0162] S603, Select the starting key time point in the key time point sequence as the current time point;

[0163] S604, use the flow matching generation model to determine the velocity field of the current data state at the current time point, and update the current data state based on the velocity field;

[0164] S605, determine whether there are any unprocessed key time points in the key time point sequence;

[0165] S606, when there are unprocessed key time points, select the next key time point as the current time point, and perform the velocity field determination and current data status update operations for the current time point;

[0166] S607, when there are no unprocessed key time points, output the current data state as the target data.

[0167] In this embodiment, a non-uniform sampling path and a flow matching generation model are used to iteratively generate target data from initial data. Specifically, this involves loading a non-uniform sampling path to obtain a sequence of key time points. The non-uniform sampling path contains multiple key time points dynamically selected based on contribution levels. The time point sequence retains its temporal order to guide the phased updates of the data state. This process first initializes the current data state on the initial data. The initial data can be random noise, structured data, or initial speech feature representations in a specific format. The initialization operation maps the input data to the current data state for use in the next iterative calculation.

[0168] The starting key time point in the key time point sequence is selected as the current time point to ensure that the iteration process starts from the first time point in the contribution ranking, strictly following the time series logic and avoiding temporal disorder or information loss. Subsequently, a flow matching generation model is used to determine the velocity field of the current data state at the current time point. The velocity field reflects the evolution trend and direction of change of the data state in the time dimension. The flow matching generation model fits the velocity field function through a neural network structure. The current data state and time point information are input, and the corresponding velocity field parameters are output. The current data state is updated based on this velocity field. The update process usually adopts methods such as numerical integration, solving difference equations, or continuous-time interpolation to ensure that the data state gradually approaches the target distribution along physical consistency or statistical regularity.

[0169] The process involves determining if any unprocessed critical time points exist in the sequence. If so, the next critical time point is selected as the current time point, and the velocity field determination and data state update operations continue until all critical time points have been processed. This iterative process ensures that the update operation corresponding to each critical time point is fully executed, and the data state is dynamically adjusted at each time point, gradually evolving into a state that approximates the target distribution.

[0170] When there are no unprocessed critical time points, the current data state is output as the target data. The target data is the final generated data result that meets the requirements. It can be a speech feature sequence, audio segment, or data information under other target distributions. After the output operation is completed, the data generation process ends, forming an efficient data generation link based on non-uniform sampling path optimization.

[0171] Example Explanation: In intelligent assisted diagnosis and treatment systems, it is typically necessary to generate high-quality, natural, and fluent voice prompts in real time based on electronic medical record text, structured medical examination results, or information input by doctors, to help patients understand diagnostic plans, rehabilitation suggestions, or health education content. However, existing speech synthesis methods suffer from slow inference speed and inefficient sampling processes, especially in medical terminal devices such as intelligent follow-up terminals and mobile emergency rescue equipment, where there is an urgent need for more efficient and low-latency speech generation capabilities.

[0172] To address this need, the system first defines a data transformation task based on electronic medical record data and a medical knowledge graph, converting data from a standard random Gaussian distribution to a target medical speech feature distribution. This target speech feature distribution specifically corresponds to the high-fidelity acoustic features for interpreting Chinese or multilingual medical records in the medical field. In conjunction with this task, the system constructs a velocity field function to describe the dynamic evolution path of speech data from random noise to medical speech features. The system then fits the parameters of the velocity field function using a neural network architecture (e.g., a deep convolutional network or a temporal Transformer structure), trains the network using a flow matching loss function, and generates the final flow matching generation model.

[0173] Subsequently, after model training, the system acquires detailed process information of the flow matching generation model within a defined time interval (e.g., the evolution process within 0 to 1 second). Specifically, the system selects multiple time points within this time interval (e.g., 16 non-uniformly distributed time nodes), performs the forward propagation operation of the flow matching generation model at each time point, and generates time-point model outputs, which include information such as loss value distribution data and internal gradient changes. Further, the system extracts the loss value distribution data for each time point from these model outputs to reflect the model's fitting effect at different time points, and determines the gradient vector norm for each time point based on the output, measuring the contribution of that time point to the optimization of the overall generation path. Finally, the system aggregates the loss value distribution data and gradient vector norms from all time points to form complete process information for subsequent analysis.

[0174] Based on the aforementioned process information, the system quantifies and evaluates the contribution value at each time point within the time interval. First, it extracts the loss value distribution data and gradient vector norm for all time points. Combining these indicators, the system calculates global statistical characteristics, including the average loss value and variance level at each time point. Further, based on these global statistical characteristics, it performs normalization processing on the loss value distribution data at each time point, generating standardized normalized loss values ​​to facilitate horizontal comparisons across different time points. Simultaneously, combined with the gradient vector norm, the system calculates the information entropy value at each time point, reflecting the importance of that time point in information representation and model uncertainty. Finally, it merges the normalized loss value and the information entropy value to form a comprehensive contribution value for each time point, objectively measuring the actual role of that time point in the generation path.

[0175] To comprehensively analyze the contribution distribution, the system collects contribution values ​​at all time points. Based on statistical methods (such as maximum / minimum, mean, and standard deviation analysis), it determines the statistical distribution parameters of the contribution values. Furthermore, it constructs a probability density function reflecting the overall trend and uses this function to generate contribution distribution curves at different time points, visually displaying the contribution distribution during the data generation process. These contribution distribution curves are stored and used as the basis for subsequent key time point selection.

[0176] Based on the contribution distribution at each time point, the system identifies multiple key time points and constructs a non-uniform sampling path accordingly to improve generation efficiency. Specifically, the system calculates the average and standard deviation of the contribution based on the contribution distribution at each time point, and sets a contribution threshold parameter using these parameters. The contribution value of each time point is compared with the contribution threshold parameter, and all time points with contribution values ​​greater than the threshold are selected as key time points, while redundant and low-contribution time points are eliminated. After selection, the system arranges all key time points in chronological order and constructs an optimized non-uniform sampling path, significantly reducing unnecessary model calculation steps and lowering system resource consumption.

[0177] In the actual speech generation stage, the system loads a non-uniform sampling path to acquire a sequence of key time points. Based on the received initial data (such as random Gaussian noise or the initial state of preset medical speech features), the system initializes the current data state to ensure the accurate starting point of the generation process. Each time point in the key time point sequence is selected sequentially as the current time point. For each time point, a flow matching generation model is used to calculate the velocity field of the current data state at that time point, clarifying the evolution direction and velocity information of the current data state at that moment. The current data state is then updated based on the velocity field, gradually guiding the data state towards the target medical speech distribution. This process iterates until all key time points have been processed. Finally, the system outputs the current data state as the target data. The generated target data corresponds to clear, natural speech information that meets the needs of medical scenarios and can be directly used in patient voice reminders, doctor voice-assisted explanations, or health education scenarios, meeting the actual needs of healthcare businesses for high-efficiency, low-latency, and high-quality speech synthesis.

[0178] In the fintech business, speech synthesis is widely used in scenarios such as intelligent customer service, voice warnings, and financial assistants. Especially when facing real-time voice broadcasts, voice identity authentication, or investment risk warnings for customers, the system needs to improve generation efficiency, reduce hardware resource consumption, and ensure stable operation in high-concurrency or mobile environments while ensuring the clarity, naturalness, and accuracy of the speech.

[0179] To address this, the system, based on the multilingual and multi-style speech output requirements in the financial sector, constructs a data conversion task from a standard random Gaussian distribution to a specific financial speech style distribution. The target speech style distribution can cover Chinese, English, regional dialects, etc., adapting to various financial transactions such as customer Q&A, transaction confirmation, and risk warnings. In conjunction with this task, the system designs a velocity field function to describe the evolution path of data from a random noise state to the target financial speech features. The velocity field function is parameter-fitted based on deep neural network structures (such as multilayer perceptrons and attention mechanism networks), and the model is continuously optimized using a flow matching loss function to form the final flow matching generation model used for data conversion.

[0180] After model training, the system, addressing the real-time and accuracy requirements of financial transactions, acquires detailed process information of the stream matching generation model within a preset time interval (e.g., a continuous time period from 0 to 1 second). During operation, the system selects multiple time points within the time interval, typically setting the time point density and distribution strategy based on actual business needs. At each time point, the system performs forward propagation of the stream matching generation model, generating time-point model outputs that include metrics such as prediction error and gradient changes. Furthermore, it extracts loss value distribution data from the time-point model outputs to reflect the model's fit performance at each time point, and calculates the corresponding gradient vector norm to measure the impact of each time point on the overall speech generation path. By aggregating the loss value distribution data and gradient vector norms from each time point, the system generates complete process information, facilitating the identification of differences in contribution at different time points.

[0181] Based on process information, the system quantifies the contribution value at each time point, assisting in optimizing the efficient sampling path in financial speech generation. Specifically, this includes extracting the loss value distribution data and gradient vector norm at each time point from the process information, statistically analyzing the overall characteristics of each time point, and calculating global statistical parameters such as mean and variance. Combining the statistical results, the system performs normalization processing on the loss value distribution data at each time point, generating normalized loss values ​​under a unified standard to ensure the reliability of horizontal data comparison. Further, by combining the gradient vector norm, the system determines the information entropy value at each time point, measuring the data uncertainty and information content. By fusing the normalized loss value and the information entropy value, the system generates a comprehensive contribution value, reflecting the actual value of each time point in the financial speech generation path.

[0182] The system generates a time-point contribution distribution based on the contribution values, comprehensively displaying the distribution of effects at different time points. During operation, the system collects contribution values ​​for all time points, constructs a probability density function based on statistical distribution parameters (such as maximum, minimum, mean, and standard deviation), generates a time-point contribution distribution curve, and stores this curve as a time-point contribution distribution, serving as the basis for optimizing the sampling path.

[0183] Based on the contribution distribution at different time points, the system selects multiple key time points and constructs a non-uniform sampling path to improve speech generation efficiency. First, the system calculates the average and standard deviation of the contribution, sets a contribution threshold based on these parameters, and removes redundant time points with contribution values ​​below the threshold. High-value time points are retained as key time points and further arranged chronologically to form an optimized non-uniform sampling path. This significantly reduces unnecessary model calculations and adapts to the stringent real-time response requirements of financial services.

[0184] In the actual speech generation process, the system loads a non-uniform sampling path to acquire key time point sequences and initializes the current data state based on initial data (such as random noise or a set initialization state). Key time points are sequentially selected as the current time point, and a flow matching generation model is used to calculate the velocity field of the current data state at that time point, obtaining the trend of data state changes. The current data state is updated based on the velocity field, gradually guiding the data state to converge towards the target financial speech feature distribution. This process is repeated until all key time points are processed, and finally, the current data state is output as the target data. The generated target data corresponds to clear, accurate, and natural speech content, which can be directly applied to scenarios such as financial voice customer service, transaction voice confirmation, and intelligent risk control broadcasting, meeting the needs for high-quality, low-latency, and resource-friendly financial speech generation.

[0185] This embodiment loads a non-uniform sampling path, retaining only time points with high contribution, and combines this with a flow matching generation model to dynamically update the current data state. This avoids redundant calculations for low-contribution time points, highlights the optimization role of key time points, reduces the number of forward propagations and resource consumption, and improves the efficiency and practicality of the data generation process. Compared with traditional uniform sampling or fixed step size methods, it can effectively reduce inference latency while ensuring the quality of the generated results, and enhance the real-time performance and adaptability of the system in application scenarios such as speech generation, speech synthesis, and speech reconstruction.

[0186] In one embodiment, a data generation apparatus based on a non-uniform sampling path is provided, which corresponds one-to-one with the data generation method based on a non-uniform sampling path described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data generation device based on a non-uniform sampling path according to the present invention. The modules include a model building module 10, an information extraction module 20, a contribution calculation module 30, a distribution generation module 40, a path building module 50, and a data generation module 60. Detailed descriptions of each functional module are as follows:

[0187] Model building module 10 is used to build a stream matching generation model for data transformation;

[0188] Information extraction module 20 is used to obtain process information of the flow matching generation model within a preset time interval;

[0189] Contribution calculation module 30 is used to quantify the contribution value of each time point within the preset time interval based on the process information;

[0190] The distribution generation module 40 is used to generate a time-point contribution distribution based on the contribution value.

[0191] The path construction module 50 is used to filter multiple key time points based on the contribution distribution of the time points and construct a non-uniform sampling path.

[0192] The data generation module 60 is used to iteratively generate target data from initial data using the non-uniform sampling path and the flow matching generation model.

[0193] In one embodiment, the model building module 10 is specifically used for:

[0194] Define a data transformation task from the initial distribution to the target distribution;

[0195] Based on the data transformation task, a velocity field function is constructed, and a neural network architecture for fitting the velocity field function is determined;

[0196] The parameters of the neural network architecture are optimized using a flow matching loss function to generate a flow matching generative model.

[0197] In one embodiment, the information extraction module 20 is specifically used for:

[0198] Select multiple time points within a preset time interval;

[0199] At each time point, the forward propagation operation of the flow matching generation model is performed to generate the time-point model output;

[0200] Extract the loss value distribution data from the model output at the aforementioned time points;

[0201] The gradient vector norm is determined based on the output of the time-point model.

[0202] The loss value distribution data and the gradient vector norm are aggregated to form process information.

[0203] In one embodiment, the contribution calculation module 30 is specifically used for:

[0204] Extract the loss value distribution data and gradient vector norm from the process information;

[0205] Determine the global statistical characteristics of the loss value distribution data for all selected time points within the preset time interval;

[0206] Based on the global statistical features, the loss value distribution data at each time point is normalized to generate a normalized loss value.

[0207] The information entropy value at each time point is determined based on the norm of the gradient vector.

[0208] The contribution value is generated by combining the normalized loss value and the information entropy value.

[0209] In one embodiment, the distribution generation module 40 is specifically used for:

[0210] Collect the contribution values ​​of all selected time points within the preset time interval;

[0211] Determine the statistical distribution parameters of the contribution value;

[0212] Construct a probability density function based on the statistical distribution parameters;

[0213] The probability density function is used to generate the contribution distribution curve at each time point, and the contribution distribution curve at each time point is used as the contribution distribution at each time point.

[0214] In one embodiment, the path construction module 50 is specifically used for:

[0215] The average contribution and standard deviation are determined based on the contribution distribution at time points, and the contribution threshold parameter is determined based on the average contribution and standard deviation.

[0216] Compare the contribution value at each time point in the contribution distribution with the contribution threshold parameter;

[0217] Select the time points where the contribution value is greater than the contribution threshold parameter as the key time points;

[0218] Arrange the key time points in chronological order to generate a non-uniform sampling path.

[0219] In one embodiment, the data generation module 60 is specifically used for:

[0220] Load the non-uniform sampling path to obtain the key time point sequence;

[0221] Initialize the current data state based on the initial data;

[0222] Select the starting key time point in the key time point sequence as the current time point;

[0223] The flow matching generation model is used to determine the velocity field of the current data state at the current time point, and the current data state is updated based on the velocity field.

[0224] Determine whether there are any unprocessed key time points in the key time point sequence;

[0225] If there are unprocessed critical time points, select the next critical time point as the current time point, and perform velocity field determination and current data status update operations for the current time point.

[0226] If there are no unprocessed critical time points, the current data state will be output as the target data.

[0227] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When executed by the processor, the computer program implements the functions or steps of a data generation method based on a non-uniform sampling path on the server side.

[0228] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the user-side functions or steps of a data generation method based on a non-uniform sampling path.

[0229] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0230] Construct a stream matching generation model for data transformation;

[0231] Obtain the process information of the flow matching generation model within a preset time interval;

[0232] The contribution value of each time point within the preset time interval is quantified based on the process information.

[0233] Generate a time-point contribution distribution based on the aforementioned contribution values;

[0234] Based on the contribution distribution at the aforementioned time points, multiple key time points are selected to construct a non-uniform sampling path;

[0235] Using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from initial data.

[0236] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0237] Construct a stream matching generation model for data transformation;

[0238] Obtain the process information of the flow matching generation model within a preset time interval;

[0239] The contribution value of each time point within the preset time interval is quantified based on the process information.

[0240] Generate a time-point contribution distribution based on the aforementioned contribution values;

[0241] Based on the contribution distribution at the aforementioned time points, multiple key time points are selected to construct a non-uniform sampling path;

[0242] Using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from initial data.

[0243] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0244] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0245] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0246] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A data generation method based on a non-uniform sampling path, characterized in that, Includes the following steps: Construct a stream matching generation model for data transformation; Obtain the process information of the flow matching generation model within a preset time interval; The contribution value of each time point within the preset time interval is quantified based on the process information. Generate a time-point contribution distribution based on the aforementioned contribution values; Based on the contribution distribution at the aforementioned time points, multiple key time points are selected to construct a non-uniform sampling path; Using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from the initial data.

2. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Construct a streaming matching generation model for data transformation, including: Define a data transformation task from the initial distribution to the target distribution; Based on the data transformation task, a velocity field function is constructed, and a neural network architecture for fitting the velocity field function is determined; The parameters of the neural network architecture are optimized using a flow matching loss function to generate a flow matching generative model.

3. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Obtaining the process information of the flow matching generation model within a preset time interval includes: Select multiple time points within a preset time interval; At each time point, the forward propagation operation of the flow matching generation model is performed to generate the time-point model output; Extract the loss value distribution data from the model output at the aforementioned time points; The gradient vector norm is determined based on the output of the time-point model. The loss value distribution data and the gradient vector norm are aggregated to form process information.

4. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Based on the process information, the contribution value of each time point within the preset time interval is quantified, including: Extract the loss value distribution data and gradient vector norm from the process information; Determine the global statistical characteristics of the loss value distribution data for all selected time points within the preset time interval; Based on the global statistical features, the loss value distribution data at each time point is normalized to generate a normalized loss value. The information entropy value at each time point is determined based on the norm of the gradient vector. The contribution value is generated by combining the normalized loss value and the information entropy value.

5. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Generate a time-point contribution distribution based on the stated contribution values, including: Collect the contribution values ​​of all selected time points within the preset time interval; Determine the statistical distribution parameters of the contribution value; Construct a probability density function based on the statistical distribution parameters; The probability density function is used to generate the contribution distribution curve at each time point, and the contribution distribution curve at each time point is used as the contribution distribution at each time point.

6. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Based on the contribution distribution at each time point, multiple key time points are selected, and a non-uniform sampling path is constructed, including: The average contribution and standard deviation are determined based on the contribution distribution at time points, and the contribution threshold parameter is determined based on the average contribution and standard deviation. Compare the contribution value at each time point in the contribution distribution with the contribution threshold parameter; Select the time points where the contribution value is greater than the contribution threshold parameter as the key time points; Arrange the key time points in chronological order to generate a non-uniform sampling path.

7. The data generation method based on a non-uniform sampling path as described in claim 1, characterized in that, Using the non-uniform sampling path and the flow matching generation model, target data is generated iteratively from initial data, including: Load the non-uniform sampling path to obtain the key time point sequence; Initialize the current data state based on the initial data; Select the starting key time point in the key time point sequence as the current time point; The flow matching generation model is used to determine the velocity field of the current data state at the current time point, and the current data state is updated based on the velocity field. Determine whether there are any unprocessed key time points in the key time point sequence; If there are unprocessed critical time points, select the next critical time point as the current time point, and perform velocity field determination and current data status update operations for the current time point. If there are no unprocessed critical time points, the current data state will be output as the target data.

8. A data generation device based on a non-uniform sampling path, characterized in that, The data generation device based on a non-uniform sampling path includes: The model building module is used to build streaming matching generation models for data transformation; The information extraction module is used to obtain the process information of the flow matching generation model within a preset time interval; The contribution calculation module is used to quantify the contribution value of each time point within the preset time interval based on the process information. A distribution generation module is used to generate a time-point contribution distribution based on the contribution value. The path construction module is used to filter multiple key time points based on the contribution distribution of the time points and construct a non-uniform sampling path; The data generation module is used to iteratively generate target data from initial data using the non-uniform sampling path and the flow matching generation model.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a data generation program based on a non-uniform sampling path stored in the memory and executable on the processor, wherein the data generation program based on a non-uniform sampling path, when executed by the processor, implements the steps of the data generation method based on a non-uniform sampling path as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a data generation program based on a non-uniform sampling path, which, when executed by a processor, implements the steps of the data generation method based on a non-uniform sampling path as described in any one of claims 1-7.

Images