A large-scale speech synthesis task processing method based on concurrent optimization scheduling

Abstract

The application discloses a large-scale speech synthesis task processing method based on concurrent optimization scheduling, which comprises the following steps: S1, receiving a plurality of speech synthesis task requests from an external system, and constructing a to-be-scheduled task set; S2, collecting the system running state of a current speech synthesis platform, and constructing a system state information set; S3, inputting the system state information set into a structure-improved deep Q network model to form a fusion state representation vector; S4, inputting the fusion state representation vector into a Q value estimation main network, and dispatching a target task in the to-be-scheduled task set to a specified computing resource node; S5, collecting the execution feedback information of the speech synthesis task, and constructing a task execution state transition sample set; and S6, inputting the state transition sample set into an experience replay pool, and performing parameter updating on the structure-improved deep Q network model. The improved deep Q network model is adopted, and intelligent scheduling optimization of large-scale speech synthesis tasks is realized.

Description

Technical Field

[0001] This invention relates to the field of speech processing and intelligent scheduling technology, and in particular to a method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling. Background Technology

[0002] In existing speech synthesis platforms, with the increasing demand for multi-task speech synthesis applications, effectively scheduling diverse tasks to ensure system response performance and speech quality has become an important direction for system optimization. Currently, mainstream speech synthesis systems typically employ static task allocation strategies or rule-based scheduling algorithms, such as first-in-first-out strategies based on task arrival time, fixed resource allocation mechanisms based on priority weights, or round-robin load balancing methods based on resource occupancy thresholds. These methods are feasible in small-scale task environments, but when facing large-scale scenarios with high concurrency, heterogeneous task inputs, and frequent resource state fluctuations, scheduling efficiency is low, easily leading to task congestion, decreased resource utilization, and fluctuations in speech synthesis quality.

[0003] Speech synthesis tasks exhibit high heterogeneity at the parameter level. The type of speech model, speech style settings, and response time requirements involved in different tasks directly affect the model loading path, computational load intensity, and synthesis latency. Furthermore, the resource status of each computing node within the platform is dynamically changing during operation, including GPU utilization, memory capacity, the number of currently queued tasks, and bandwidth usage, making effective adaptation through static configuration difficult. Simultaneously, historical feedback information during task execution, such as scheduling success rate, task failure rate, and resource utilization fluctuations, also contains potential behavioral evolution trend information, which has certain reference value for subsequent task scheduling.

[0004] Existing scheduling schemes generally suffer from insufficient representation in state modeling, failing to fully consider the multidimensional state interactions between task attributes, resource characteristics, and historical behavior. Most schedulers only use the current task length or node load as the scheduling basis, without incorporating long-term behavioral information modeling, making it difficult to accurately predict the impact of different scheduling strategies on system performance. Furthermore, traditional methods lack adaptive learning mechanisms, making it difficult to dynamically adjust scheduling strategies when the platform operating environment or task structure changes, leading to a decrease in system scheduling efficiency.

[0005] Deep reinforcement learning has been explored to some extent in resource scheduling, using state-action-reward models to learn scheduling strategies from environmental feedback. However, existing methods often use a uniform structure to handle all input states, failing to differentiate between tasks, resources, and historical feedback, resulting in limited state representation capabilities. Furthermore, some methods do not effectively utilize information from historical feedback sequences, leading to shortcomings in modeling long-term scheduling performance.

[0006] Therefore, how to provide a method for large-scale speech synthesis task processing based on concurrent optimized scheduling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a method for processing large-scale speech synthesis tasks based on concurrent optimization scheduling. This invention adopts a structurally improved deep Q-network model, integrates task queue state, resource node state, and historical scheduling feedback information, and describes in detail an intelligent scheduling mechanism for high-concurrency speech synthesis environments through multi-channel state encoding, feature fusion, and reinforcement learning strategy iteration. It has the advantages of high resource utilization, low response latency, and stable speech output quality.

[0008] A method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to an embodiment of the present invention includes the following steps:

[0009] S1. Receive multiple speech synthesis task requests from external systems and construct a set of tasks to be scheduled.

[0010] S2. Collect the current system operating status of the speech synthesis platform and construct a system status information set;

[0011] S3. Input the system state information set into the structurally improved deep Q-network model. The structurally improved deep Q-network model introduces a multi-channel state input module to extract features from the state information of the corresponding channels and concatenate all channel features to form a fused state representation vector.

[0012] S4. Input the fused state representation vector into the Q-value estimation main network, output the optimal scheduling policy, and according to the optimal scheduling policy, assign the target task in the set of tasks to be scheduled to the specified computing resource nodes, and call the target speech model to perform the speech synthesis task generation operation.

[0013] S5. Collect execution feedback information of the speech synthesis task and construct a sample set of task execution state transitions;

[0014] S6. Input the state transition sample set into the experience replay pool, periodically sample batches from the experience replay pool, and use the target network and loss function to perform parameter updates on the structurally improved deep Q network model to optimize the speech task scheduling strategy.

[0015] Optionally, each target speech synthesis task in the set of tasks to be scheduled includes input text, speech model type, speech style identifier, response time limit parameter, and allocation status identifier; the allocation status identifier includes unscheduled, in the process of scheduling, and completed.

[0016] Optionally, the system status information set includes task queue status information, resource node status information, and historical scheduling feedback information.

[0017] Optionally, S3 specifically includes:

[0018] S31. Input the system state information set into the structurally improved deep Q-network model, wherein the structurally improved deep Q-network model includes a multi-channel state input module, a feature fusion module, a Q-value estimation main network and a policy update module;

[0019] S32. The multi-channel status input module includes a task status channel coding sub-network, a resource status channel coding sub-network, and a historical trajectory status channel coding sub-network.

[0020] S33. Input the task queue status information into the task status channel encoding sub-network. The sub-network includes at least one fully connected neural network with an activation function, performs normalization processing and nonlinear feature mapping, and generates a task feature representation vector.

[0021] S34. Input the resource node status information into the resource status channel coding sub-network. The sub-network includes a normalization layer and a two-layer multi-layer perceptron structure. Extract the high-dimensional features of the resource status and perform dimensionality compression to generate a resource feature representation vector.

[0022] S35. Input historical scheduling feedback information into the historical trajectory status channel encoding sub-network. The sub-network includes a sliding window aggregation module and a feature transformation layer to model the historical data trend and generate historical behavior feature vectors.

[0023] S36. In the feature fusion module, the task feature representation vector, resource feature representation vector, and historical behavior feature vector are concatenated according to the feature dimension to generate a fused state representation vector.

[0024] Optionally, S33 specifically includes:

[0025] S331. Input the task queue status information into the task status channel coding sub-network, where each task structure vector consists of the input text length, speech model type identifier, speech style identifier, task response time limit parameter and task priority score;

[0026] S332. Perform linear normalization on each dimension of the structured vector of each task to generate a normalized task input matrix.

[0027] S333. Input the normalized task input matrix into the feature transformation module containing one or more fully connected neural networks, and apply a non-linear activation function after each layer. Perform nonlinear feature mapping to obtain the task feature representation matrix;

[0028] S334. Perform global average pooling on the task feature representation matrix, calculate the mean along the column in the sample dimension, and generate the task channel feature representation vector.

[0029] Optionally, S34 specifically includes:

[0030] S341. Input the resource node status information into the resource status channel coding sub-network, where each resource node structured vector consists of GPU utilization, remaining video memory capacity, number of currently queued tasks, bandwidth utilization ratio and node temperature.

[0031] S342. Perform linear normalization on each dimension of the structured vector of each resource node to generate a normalized resource input matrix.

[0032] S343. Input the normalized resource input matrix into the feature transformation module containing a two-layer multi-layer perceptron structure. The first layer outputs a high-dimensional intermediate feature matrix, and the second layer performs dimensionality compression mapping to generate a resource node feature matrix.

[0033] S344. Perform a global average pooling operation on the resource node feature matrix, calculate the average value column-wise along the node dimension, and generate a resource channel feature representation vector.

[0034] Optionally, S35 specifically includes:

[0035] S351. Input the historical scheduling feedback information into the historical trajectory status channel coding sub-network, where each historical feedback record vector consists of the previous round of task completion delay, the scheduling success rate of the corresponding node, the task execution failure rate, the average task execution time and the resource utilization fluctuation coefficient.

[0036] S352. Perform normalization processing on each dimension attribute in each historical feedback record to generate a normalized historical feedback input matrix;

[0037] S353. Input the normalized history feedback input matrix into the time series modeling structure containing the gating mechanism, use a one-layer GRU network for time dimension modeling, and output the hidden state sequence.

[0038] S354. Perform a time-weighted average operation on the hidden state sequence to generate the feature representation vector of the historical trajectory channel.

[0039] Optionally, S4 specifically includes:

[0040] S41. Input the fused state representation vector into the Q-value estimation main network. The Q-value estimation main network is a multilayer perceptron structure, which includes an input layer, a hidden layer and an output layer. The output layer dimension is the action space dimension, and the corresponding action value estimation result of the candidate scheduling action set is generated.

[0041] S42. In the action value estimation results, select the action with the largest Q value as the scheduling action corresponding to the current optimal scheduling strategy.

[0042] S43. According to the optimal scheduling strategy, perform scheduling operations on the target speech synthesis tasks to be assigned in the set of tasks to be scheduled, and assign the target speech synthesis tasks to be assigned to the optimal resource node.

[0043] S44. Call the speech model type and speech style identifier information corresponding to the target speech synthesis task, load and execute the specified speech synthesis model, generate the speech synthesis task and push it to the user or caller interface.

[0044] Optionally, S5 specifically includes:

[0045] S51. Collect feedback information during the execution of the speech synthesis task, including task completion delay, generated speech quality score, and real-time resource occupancy status of resource nodes.

[0046] S52. The task completion delay, the generated voice quality score, and the real-time resource occupancy status of resource nodes are used as feedback vectors.

[0047] S53. Extract the fusion state representation vector and the executed scheduling action corresponding to the current moment, and combine the two with the feedback vector to form a state transition sample triplet.

[0048] S54. Repeat steps S51 to S53 to continuously collect state transition samples for multiple speech synthesis tasks and summarize them to form a task execution state transition sample set.

[0049] Optionally, S6 specifically includes:

[0050] S61. Store the set of task execution state transition samples into the experience replay pool contained in the policy update module of the structurally improved deep Q-network model.

[0051] S62. Periodically sample fixed-size batches of samples from the experience replay pool, where each sample includes a fusion state representation vector, historical scheduling actions, and a feedback vector;

[0052] S63. Using the Q-value estimation of the main network in the structurally improved deep Q-network, calculate the action value estimation result of the current policy network, and combine it with the delay target value of the target network to define a loss function based on the time difference error:

[0053] ;

[0054] in, Optimize the loss function for the scheduling strategy. For sample batch size, This refers to the index variable of samples in the experience replay pool during the training process. For instant reward value, As a discount factor, For all scheduling actions in the target network, The target network Q-value, The predicted Q value is as follows. Let the fused state represent the next state of the vector. For the target Q-network parameter set, Let the fusion state representation vector be... For historical scheduling actions, Estimate the set of parameters for the main network to obtain the Q-value;

[0055] S64. The gradient descent algorithm is used to minimize the loss function L, and the parameter set of the policy update module in the structurally improved deep Q network model is updated to realize the continuous optimization and adaptive enhancement of the speech task scheduling policy.

[0056] The beneficial effects of this invention are:

[0057] This invention proposes a large-scale speech synthesis task processing method based on concurrent optimized scheduling. Addressing the problems of low task scheduling efficiency, uneven resource utilization, and unstable speech synthesis response performance in existing speech synthesis platforms under high-concurrency environments, it constructs a scheduling system that integrates multi-channel state modeling and a deep reinforcement learning policy update mechanism. By independently encoding task state information, resource node state information, and historical scheduling feedback information, and constructing a unified fused state representation vector during the feature fusion stage, the scheduling decision model can comprehensively perceive the current system state and evolution trend.

[0058] This invention employs a structurally improved deep Q-network model, incorporating task attributes, resource load, and historical behavior feedback into the scheduling strategy learning process. Through an experience replay mechanism and a target Q-network update mechanism, the strategy is continuously optimized and adaptively enhanced. During scheduling, task priority weights and resource allocation strategies can be dynamically adjusted, achieving intelligent matching between different task types and system states, thus improving the accuracy and response efficiency of task processing. Simultaneously, feedback information after task execution is continuously collected and constructed as state transition samples, helping the scheduling strategy accumulate experience over long-term operation and improving the model's responsiveness to dynamic changes in the system.

[0059] This method enables the speech synthesis platform to reduce task response latency, improve resource utilization efficiency, and decrease task failure rate when facing large-scale and diverse task requests, thereby effectively improving the platform's overall service performance and concurrent processing capabilities. The method also exhibits good system adaptability and scalability, maintaining relatively stable task scheduling performance under different scales and load environments. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a large-scale speech synthesis task processing method based on concurrent optimization scheduling proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the module structure of an improved deep Q-network model for a large-scale speech synthesis task processing method based on concurrent optimization scheduling proposed in this invention.

[0063] Figure 3 This is a data flow diagram of the multi-channel state input encoding and feature fusion process of a large-scale speech synthesis task processing method based on concurrent optimization scheduling proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figure 1-3 A method for processing large-scale speech synthesis tasks based on concurrent optimization scheduling includes the following steps:

[0066] S1. Receive multiple speech synthesis task requests from external systems, extract the input text, speech model type, speech style identifier and response time limit parameters for each task, and construct a set of tasks to be scheduled.

[0067] S2. Collect the current system operating status of the speech synthesis platform and construct a system status information set;

[0068] S3. Input the system state information set into the structurally improved deep Q-network model. The structurally improved deep Q-network model introduces a multi-channel state input module to extract features from the state information of the corresponding channels and concatenate all channel features to form a fused state representation vector.

[0069] S4. Input the fused state representation vector into the Q-value estimation main network, output the optimal scheduling policy, and according to the optimal scheduling policy, assign the target task in the set of tasks to be scheduled to the specified computing resource nodes, and call the target speech model to perform the speech synthesis task generation operation.

[0070] S5. Collect execution feedback information of the speech synthesis task and construct a sample set of task execution state transitions;

[0071] S6. Input the state transition sample set into the experience replay pool, periodically sample batches from the experience replay pool, and use the target network and loss function to perform parameter updates on the structurally improved deep Q network model to optimize the speech task scheduling strategy.

[0072] This invention achieves intelligent optimization of the speech synthesis task scheduling process by constructing a multi-channel state input and deep reinforcement learning policy update mechanism. By fusing task attributes, resource status, and historical feedback information, the accuracy and adaptability of the scheduling policy are improved. Experimental results show that this method can effectively reduce task response latency, improve GPU resource utilization, reduce task failure rate, and enhance the system's processing stability and resource allocation efficiency in high-concurrency scenarios, demonstrating good practical value and promising prospects for wider application.

[0073] In this embodiment, each target speech synthesis task in the set of tasks to be scheduled includes input text, speech model type, speech style identifier, response time limit parameter, and allocation status identifier; the allocation status identifier includes unscheduled, in the process of scheduling, and completed.

[0074] This invention constructs a set of tasks to be scheduled by structurally representing the input text, speech model type, style identifier, response time parameters, and allocation status identifiers of speech synthesis tasks, providing high-dimensional input support for scheduling strategies. This provides a reliable input foundation for subsequent scheduling strategy learning, helping to improve system response efficiency and resource matching capabilities.

[0075] In this embodiment, the system status information set includes task queue status information, resource node status information, and historical scheduling feedback information;

[0076] The task queue status information includes input text length, speech model type identifier, speech style identifier, task response time limit parameter, and task priority score; the resource node status information includes GPU utilization, remaining video memory capacity, number of currently queued tasks, bandwidth utilization ratio, and node temperature; the historical scheduling feedback information includes the previous round of task completion latency, the scheduling success rate of the corresponding node, the task execution failure rate, the average task execution time, and the resource utilization fluctuation coefficient.

[0077] This invention achieves comprehensive modeling of task attributes, resource load, and operational history by constructing a system status information set encompassing task queues, resource nodes, and historical feedback. Task status information reflects current task characteristics, resource status information quantifies node operational capabilities, and historical feedback information provides a reference for the effectiveness of scheduling strategies. This helps the scheduling system make more accurate resource allocation and task selection under different states, thereby improving the accuracy, stability, and overall system efficiency of task scheduling.

[0078] In this embodiment, S3 specifically includes:

[0079] S31. Input the system state information set into the structurally improved deep Q-network model, wherein the structurally improved deep Q-network model includes a multi-channel state input module, a feature fusion module, a Q-value estimation main network and a policy update module;

[0080] S32. The multi-channel status input module includes a task status channel coding sub-network, a resource status channel coding sub-network, and a historical trajectory status channel coding sub-network.

[0081] S33. Input the task queue status information into the task status channel encoding sub-network. The sub-network includes at least one fully connected neural network with an activation function, performs normalization processing and nonlinear feature mapping, and generates a task feature representation vector.

[0082] S34. Input the resource node status information into the resource status channel coding sub-network. The sub-network includes a normalization layer and a two-layer multi-layer perceptron structure. Extract the high-dimensional features of the resource status and perform dimensionality compression to generate a resource feature representation vector.

[0083] S35. Input historical scheduling feedback information into the historical trajectory status channel encoding sub-network. The sub-network includes a sliding window aggregation module and a feature transformation layer to model the historical data trend and generate historical behavior feature vectors.

[0084] S36. In the feature fusion module, the task feature representation vector, resource feature representation vector, and historical behavior feature vector are concatenated according to the feature dimension to generate a fused state representation vector.

[0085] This invention constructs an improved deep Q-network model, inputting task state, resource state, and historical feedback into corresponding channel encoding subnetworks to extract discriminative feature vectors. The task channel uses a fully connected network for normalization and nonlinear mapping, enhancing its ability to express task differences; the resource channel compresses high-dimensional resource features through a multilayer perceptron structure; and the historical channel introduces a sliding window and feature transformation structure to model the changing trends of scheduling behavior. The feature vectors output from each channel are concatenated in a fusion module to form a unified state representation, which serves as the input to the scheduling strategy. This structure improves the accuracy and completeness of state representation, enabling the scheduling system to make more accurate task allocation decisions under the influence of multiple factors, thereby improving the platform's scheduling performance and execution efficiency in complex environments.

[0086] In this embodiment, S33 specifically includes:

[0087] S331. Input the task queue status information into the task status channel coding sub-network, where each task structure vector consists of the input text length, speech model type identifier, speech style identifier, task response time limit parameter and task priority score;

[0088] S332. Perform linear normalization on each dimension of the structured vector of each task to generate a normalized task input matrix.

[0089] S333. Input the normalized task input matrix into the feature transformation module containing one or more fully connected neural networks, and apply a non-linear activation function after each layer. Perform nonlinear feature mapping to obtain the task feature representation matrix;

[0090] S334. Perform global average pooling on the task feature representation matrix, calculate the average value column-wise in the sample dimension, and generate the task channel feature representation vector.

[0091] This invention structures task queue state information into a multi-dimensional vector containing text length, speech model type, style identifier, response time, and priority score, and performs linear normalization on each dimension to construct a unified input matrix. This matrix is ​​then processed by a multi-layer fully connected neural network to extract non-linear features, and activation functions are used for feature transformation and enhancement to form a task feature representation matrix. Subsequently, global average pooling is used to average the sample dimensions, generating task channel feature representation vectors. This process improves the expressive power and abstract quality of task information, helps the scheduling model accurately understand task differences, and improves the accuracy of scheduling decisions and system processing efficiency.

[0092] In this embodiment, S34 specifically includes:

[0093] S341. Input the resource node status information into the resource status channel coding sub-network, where each resource node structured vector consists of GPU utilization, remaining video memory capacity, number of currently queued tasks, bandwidth utilization ratio and node temperature.

[0094] S342. Perform linear normalization on each dimension of the structured vector of each resource node to generate a normalized resource input matrix.

[0095] S343. Input the normalized resource input matrix into the feature transformation module containing a two-layer multi-layer perceptron structure. The first layer outputs a high-dimensional intermediate feature matrix, and the second layer performs dimensionality compression mapping to generate a resource node feature matrix.

[0096] S344. Perform a global average pooling operation on the resource node feature matrix, calculate the average value column-wise along the node dimension, and generate a resource channel feature representation vector.

[0097] This invention structures resource node status information into a multi-dimensional vector containing GPU utilization, remaining video memory capacity, number of queued tasks, bandwidth utilization ratio, and node temperature, and performs normalization processing to generate a standard input matrix. This matrix is ​​then input into a two-layer perceptron structure. The first layer outputs high-dimensional features, and the second layer performs dimensionality compression to form a resource node feature matrix. Subsequently, a global average pooling operation is used to average the node dimensions to obtain a unified resource channel feature representation vector. This processing method enhances the ability to represent differences in resource status, providing a stable, compressed, and fully expressive resource status input basis for scheduling strategies, and helping to improve the adaptability of the scheduling system in dynamic resource environments.

[0098] In this embodiment, S35 specifically includes:

[0099] S351. Input the historical scheduling feedback information into the historical trajectory status channel coding sub-network, where each historical feedback record vector consists of the previous round of task completion delay, the scheduling success rate of the corresponding node, the task execution failure rate, the average task execution time and the resource utilization fluctuation coefficient.

[0100] S352. Perform normalization processing on each dimension attribute in each historical feedback record to generate a normalized historical feedback input matrix;

[0101] S353. Input the normalized history feedback input matrix into the time series modeling structure containing the gating mechanism, use a one-layer GRU network for time dimension modeling, and output the hidden state sequence.

[0102] S354. Perform a time-weighted average operation on the hidden state sequence to generate the feature representation vector of the historical trajectory channel.

[0103] This invention constructs a historical scheduling feedback information set, encoding task completion delay, node scheduling success rate, task failure rate, average execution time, and resource fluctuation coefficient as multi-dimensional time series inputs, and performs normalization processing to form a standard input matrix. This matrix is ​​input into a time series modeling structure containing a gating mechanism, and a single-layer gated recurrent unit model is used to extract time-dimensional features, outputting a hidden state sequence. Subsequently, a time-weighted average is performed on this sequence to generate a stable historical channel feature representation vector. This method can effectively capture the trend of scheduling behavior evolving over time, providing dynamic feedback for the strategy model and improving the system's predictive ability and scheduling stability.

[0104] In this embodiment, S4 specifically includes:

[0105] S41. Input the fused state representation vector into the Q-value estimation main network. The Q-value estimation main network is a multilayer perceptron structure, which includes an input layer, a hidden layer and an output layer. The output layer dimension is the action space dimension, and the corresponding action value estimation result of the candidate scheduling action set is generated.

[0106] S42. In the action value estimation results, select the action with the largest Q value as the scheduling action corresponding to the current optimal scheduling strategy.

[0107] S43. According to the optimal scheduling strategy, perform scheduling operations on the target speech synthesis tasks to be assigned in the set of tasks to be scheduled, and assign the target tasks to be assigned to the optimal resource node.

[0108] S44. Call the speech model type and speech style identifier information corresponding to the target speech synthesis task, load and execute the specified speech synthesis model, generate the speech synthesis task and push it to the user or caller interface.

[0109] This invention inputs the fused state representation vector into a Q-value estimation main network of a multilayer perceptron structure to calculate the corresponding value of each scheduling action and selects the action with the highest Q-value as the current optimal scheduling strategy. Based on this strategy, the target task is assigned to the optimal resource node, ensuring that the task allocation process considers both task urgency and resource load. Subsequently, a specified synthesis model is loaded and executed according to the task's speech model type and style identifier to generate speech output. This mechanism achieves a closed-loop linkage between scheduling decisions and model invocation, improving the rationality of task allocation and the timeliness and consistency of speech generation.

[0110] In this embodiment, S5 specifically includes:

[0111] S51. Collect feedback information during the execution of the speech synthesis task, including task completion delay, generated speech quality score, and real-time resource occupancy status of resource nodes.

[0112] S52. The task completion delay, the generated voice quality score, and the real-time resource occupancy status of resource nodes are used as feedback vectors.

[0113] S53. Extract the fusion state representation vector and the executed scheduling action corresponding to the current moment, and combine the two with the feedback vector to form a state transition sample triplet.

[0114] S54. Repeat steps S51 to S53 to continuously collect state transition samples for multiple speech synthesis tasks and summarize them to form a task execution state transition sample set.

[0115] This invention collects task completion latency, speech quality scores, and resource node occupancy status as feedback information to construct a feedback vector that quantifies task execution performance. This vector is then combined with the pre-scheduling fusion state representation and executed actions to form a state transition sample triplet. This sample reflects the correlation between the scheduling strategy and the actual execution result, supporting continuous collection and aggregation to form a sample set. This mechanism enables the scheduling system to have feedback awareness, allowing it to reference real execution results in subsequent strategy optimization, thereby improving the effectiveness of model training and the accuracy of strategy adjustment.

[0116] In this embodiment, S6 specifically includes:

[0117] S61. Store the set of task execution state transition samples into the experience replay pool contained in the policy update module of the structurally improved deep Q-network model.

[0118] S62. Periodically sample fixed-size batches of samples from the experience replay pool, where each sample includes a fusion state representation vector, historical scheduling actions, and a feedback vector;

[0119] S63. Using the Q-value estimation of the main network in the structurally improved deep Q-network, calculate the action value estimation result of the current policy network, and combine it with the delay target value of the target network to define a loss function based on the time difference error:

[0120] ;

[0121] in, Optimize the loss function for the scheduling strategy. For sample batch size, This refers to the index variable of samples in the experience replay pool during the training process. For instant reward value, As a discount factor, For all scheduling actions in the target network, The target network Q-value, The predicted Q value is as follows. Let the fused state represent the next state of the vector. For the target Q-network parameter set, Let the fusion state representation vector be... For historical scheduling actions, Estimate the set of parameters for the main network to obtain the Q-value;

[0122] S64. The gradient descent algorithm is used to minimize the loss function L, and the parameter set of the policy update module in the structurally improved deep Q network model is updated to realize the continuous optimization and adaptive enhancement of the speech task scheduling policy.

[0123] This invention alleviates the temporal correlation between samples by storing task execution state transition samples in an experience replay pool, and periodically samples small batches of samples for policy updates. During policy optimization, the current policy network is used to estimate the value of actions, while a target network is introduced to calculate the target delay value, constructing a loss function based on temporal difference error. This loss function is based on the sample reward, the squared difference between the target Q-value and the current Q-value, reflecting the prediction bias of the policy network, and optimizes the network parameters through gradient descent. This mechanism achieves stable updates and continuous learning capabilities for the scheduling policy, contributing to improved long-term performance and adaptability of speech task scheduling strategies.

[0124] Example 1:

[0125] To verify the feasibility of this invention in practice, it was applied to the task scheduling module of a multilingual speech synthesis cloud platform. The system's comprehensive optimization capabilities in terms of resource utilization efficiency, speech synthesis response time, and speech output quality under large-scale concurrent request environments were tested. In this platform, users initiate speech synthesis tasks via API interfaces. These tasks include text in multiple languages, diverse speech model types, and different speech style requirements; some tasks also have explicit response time limits.

[0126] In practical applications, the original platform adopted a scheduling method based on static priority weighting. When the number of requests surged or the load on some resource nodes was uneven, problems such as high-priority tasks queuing, some nodes overloaded, increased synthesis response latency, and even timeout failures easily occurred. Especially during peak hours, the platform processed more than 120,000 tasks per day, with nearly 60% of the requests concentrated in a short period of time, putting significant pressure on the system scheduling. To solve this problem, the large-scale speech synthesis task processing method based on concurrent optimization scheduling proposed in this invention was integrated into the platform, replacing the original scheduling module, and a seven-day parallel comparative test was conducted with the original scheduling strategy.

[0127] During testing, the platform recorded the total number of tasks processed daily, the average response latency of each task, the speech quality score (MOS score), the average GPU utilization of resource nodes, and the task failure rate. This invention introduces a multi-channel state coding module through a structurally improved deep Q-network model to model the features of task state, resource state, and historical feedback. Based on an experience replay mechanism, the scheduling strategy is continuously optimized, making task allocation more aligned with the dynamic balance between resource availability and task urgency.

[0128] Test data shows that, under the same system resource configuration, the method adopted in this invention reduces the average task response latency to 78.6% of the original strategy. Response latency fluctuations are significantly reduced during peak task periods, and stability is improved by 22.4%. The task failure rate decreases by approximately 64%, and there are almost no timeouts or data loss under high concurrency conditions. In terms of resources, the average GPU utilization rate increases from 64.3% to 83.1%, significantly reducing resource waste. Simultaneously, in the user subjective rating of the output quality of mainstream speech synthesis models, the MOS score remains stable above 4.42, a slight improvement over the original strategy, and no decline in speech synthesis quality has been observed.

[0129] The following is a summary table comparing key system performance indicators during the testing period:

[0130] Table 1. Platform Task Scheduling Performance Comparison Data

[0131] Test metrics Original scheduling strategy mean The mean value of the method of the present invention Improvement range Daily task processing volume 121534 123089 +1.28% Average response time (milliseconds) 387 304 -21.45% Maximum response time during peak hours (ms) 1197 842 -29.67% Standard deviation of response delay (ms) 214.6 166.5 -22.4% Average task failure rate 3.59% 1.29% -63.79% Average GPU utilization 64.3% 83.1% +29.3% Average MOS speech quality score 4.37 4.42 +1.14%

[0132] As can be seen from the table above, this invention has significant advantages over traditional static priority scheduling strategies in several key performance indicators of large-scale speech synthesis task scheduling. Firstly, regarding average response latency, the method of this invention reduces the average response time of speech synthesis tasks from 387 milliseconds to 304 milliseconds, a reduction of 21.45%. Especially during high-concurrency tasks, the maximum response latency is reduced from 1197 milliseconds to 842 milliseconds, a reduction of nearly 30%, effectively alleviating the problems of task backlog and timeouts. Furthermore, a 22.4% reduction is also achieved in the standard deviation of response latency, indicating that the scheduling strategy has stronger stability and consistency, and can maintain a stable output of service quality under load fluctuations.

[0133] Regarding task completion rate, the reinforcement learning scheduling model constructed in this invention significantly reduces the occurrence of failed tasks. Test results show that the task failure rate decreased from 3.59% of the original strategy to 1.29%, a reduction of over 63%, significantly improving the overall reliability of task execution and user experience. Meanwhile, in terms of system resource utilization efficiency, the scheduling method of this invention can allocate computing resources more rationally, increasing the average GPU utilization from 64.3% to 83.1%, an improvement of 29.3%, effectively reducing resource idleness and uneven node load, and optimizing platform operating costs.

[0134] Regarding speech output quality, the scheduling mechanism of this invention provides a more precise match between the selection of the synthesis model and the execution load, ensuring that the task is completed within a reasonable timeframe without affecting the synthesized audio quality. The MOS score has steadily improved to 4.42, a slight increase compared to the original strategy's 4.37, indicating that the scheduling method of this invention improves efficiency without negatively impacting the user's perceived speech quality. Overall, this invention achieves comprehensive optimization in speech task processing speed, stability, resource utilization, and output quality, validating its practicality and advanced nature in large-scale speech platforms.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for processing large-scale speech synthesis tasks based on concurrent optimization scheduling, characterized in that, Includes the following steps: S1. Receive multiple speech synthesis task requests from external systems and construct a set of tasks to be scheduled. S2. Collect the current system operating status of the speech synthesis platform and construct a system status information set; S3. Input the system state information set into the structurally improved deep Q-network model. The structurally improved deep Q-network model introduces a multi-channel state input module to extract features from the state information of the corresponding channels and concatenate all channel features to form a fused state representation vector. S4. Input the fusion state representation vector into the Q-value estimation main network, output the optimal scheduling strategy, and according to the optimal scheduling strategy, assign the target task in the set of tasks to be scheduled to the specified computing resource node, and call the target speech model to perform the speech synthesis task generation operation. S5. Collect execution feedback information of the speech synthesis task and construct a sample set of task execution state transitions; S6. Input the state transition sample set into the experience replay pool, periodically sample batches from the experience replay pool, and use the target network and loss function to perform parameter updates on the structurally improved deep Q network model to optimize the speech task scheduling strategy.

2. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 1, characterized in that, Each target speech synthesis task in the set of tasks to be scheduled includes input text, speech model type, speech style identifier, response time limit parameter, and allocation status identifier; the allocation status identifier includes unscheduled, in the process of scheduling, and completed.

3. The method for large-scale speech synthesis task processing based on concurrent optimized scheduling according to claim 1, characterized in that, The system status information set includes task queue status information, resource node status information, and historical scheduling feedback information.

4. The method for processing large-scale speech synthesis tasks based on concurrent optimization scheduling according to claim 1, characterized in that, S3 specifically includes: S31. Input the system state information set into the structurally improved deep Q-network model, wherein the structurally improved deep Q-network model includes a multi-channel state input module, a feature fusion module, a Q-value estimation main network and a policy update module; S32. The multi-channel status input module includes a task status channel coding sub-network, a resource status channel coding sub-network, and a historical trajectory status channel coding sub-network. S33. Input the task queue status information into the task status channel encoding sub-network. The sub-network includes at least one fully connected neural network with an activation function, performs normalization processing and nonlinear feature mapping, and generates a task feature representation vector. S34. Input the resource node status information into the resource status channel coding sub-network. The sub-network includes a normalization layer and a two-layer multi-layer perceptron structure. Extract the high-dimensional features of the resource status and perform dimensionality compression to generate a resource feature representation vector. S35. Input historical scheduling feedback information into the historical trajectory status channel encoding sub-network. The sub-network includes a sliding window aggregation module and a feature transformation layer to model the historical data trend and generate historical behavior feature vectors. S36. In the feature fusion module, the task feature representation vector, resource feature representation vector, and historical behavior feature vector are concatenated according to the feature dimension to generate a fused state representation vector.

5. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 4, characterized in that, Specifically, S33 includes: S331. Input the task queue status information into the task status channel coding sub-network, where each task structure vector consists of the input text length, speech model type identifier, speech style identifier, task response time limit parameter and task priority score. S332. Perform linear normalization on each dimension of the structured vector of each task to generate a normalized task input matrix. S333. Input the normalized task input matrix into the feature transformation module containing one or more fully connected neural networks, and apply a non-linear activation function after each layer. Perform nonlinear feature mapping to obtain the task feature representation matrix; S334. Perform global average pooling on the task feature representation matrix, calculate the mean along the column in the sample dimension, and generate the task channel feature representation vector.

6. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 4, characterized in that, S34 specifically includes: S341. Input the resource node status information into the resource status channel coding sub-network, where each resource node structured vector consists of GPU utilization, remaining video memory capacity, number of currently queued tasks, bandwidth utilization ratio and node temperature. S342. Perform linear normalization on each dimension of the structured vector of each resource node to generate a normalized resource input matrix. S343. Input the normalized resource input matrix into the feature transformation module containing a two-layer multi-layer perceptron structure. The first layer outputs a high-dimensional intermediate feature matrix, and the second layer performs dimensionality compression mapping to generate a resource node feature matrix. S344. Perform a global average pooling operation on the resource node feature matrix, calculate the average value column-wise along the node dimension, and generate a resource channel feature representation vector.

7. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 4, characterized in that, Specifically, S35 includes: S351. Input the historical scheduling feedback information into the historical trajectory status channel coding sub-network, where each historical feedback record vector consists of the previous round of task completion delay, the scheduling success rate of the corresponding node, the task execution failure rate, the average task execution time and the resource utilization fluctuation coefficient. S352. Perform normalization processing on each dimension attribute in each historical feedback record to generate a normalized historical feedback input matrix; S353. Input the normalized history feedback input matrix into the time series modeling structure containing the gating mechanism, use a one-layer GRU network for time dimension modeling, and output the hidden state sequence. S354. Perform a time-weighted average operation on the hidden state sequence to generate the feature representation vector of the historical trajectory channel.

8. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 1, characterized in that, S4 specifically includes: S41. Input the fused state representation vector into the Q-value estimation main network. The Q-value estimation main network is a multilayer perceptron structure, which includes an input layer, a hidden layer and an output layer. The output layer dimension is the action space dimension, and the corresponding action value estimation result of the candidate scheduling action set is generated. S42. In the action value estimation results, select the action with the largest Q value as the scheduling action corresponding to the current optimal scheduling strategy. S43. According to the optimal scheduling strategy, perform scheduling operations on the target speech synthesis tasks to be assigned in the set of tasks to be scheduled, and assign the target tasks to be assigned to the optimal resource node. S44. Call the speech model type and speech style identification information corresponding to the target speech synthesis task, load and execute the specified speech synthesis model, generate the speech synthesis task and push it to the user or caller interface.

9. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 1, characterized in that, S5 specifically includes: S51. Collect feedback information during the execution of the speech synthesis task, including task completion delay, generated speech quality score, and real-time resource occupancy status of resource nodes. S52. The task completion delay, the generated voice quality score, and the real-time resource occupancy status of resource nodes are used as feedback vectors. S53. Extract the fusion state representation vector and the executed scheduling action corresponding to the current moment, and combine the two with the feedback vector to form a state transition sample triplet. S54. Repeat steps S51 to S53 to continuously collect state transition samples for multiple speech synthesis tasks and summarize them to form a task execution state transition sample set.

10. The method for processing large-scale speech synthesis tasks based on concurrent optimized scheduling according to claim 1, characterized in that, S6 specifically includes: S61. Store the set of task execution state transition samples into the experience replay pool contained in the policy update module of the structurally improved deep Q-network model. S62. Periodically sample fixed-size batches of samples from the experience replay pool, where each sample includes a fusion state representation vector, historical scheduling actions, and a feedback vector; S63. Using the Q-value estimation of the main network in the structurally improved deep Q-network, calculate the action value estimation result of the current policy network, and combine it with the delay target value of the target network to define a loss function based on the time difference error: ； in, Optimize the loss function for the scheduling strategy. For sample batch size, This refers to the index variable of samples in the experience replay pool during the training process. For instant reward value, As a discount factor, For all scheduling actions in the target network, The target network Q-value, The predicted Q value is as follows. The next state of the fusion state representation vector. For the target Q-network parameter set, Let the fusion state representation vector be... For historical scheduling actions, Estimate the set of parameters for the main network to obtain the Q-value; S64. The gradient descent algorithm is used to minimize the loss function L, and the parameter set of the policy update module in the structurally improved deep Q network model is updated to realize the continuous optimization and adaptive enhancement of the speech task scheduling policy.

Images