Personal voice intelligent terminal for fund transaction room

Abstract

The invention discloses a personal voice intelligent terminal for a fund transaction room. The personal voice intelligent terminal comprises a voice data acquisition and classification module, a data cleaning and structuring module, a task concurrency feature extraction module, a concurrency situation prediction module, a situation classification module, a normal situation processing module and a high concurrency situation processing module. And the voice data acquisition and classification module is used for acquiring voice data of traders in real time through terminal equipment in a financial transaction scene. According to the method, the voice data is collected and classified in real time, the feature engineering and the deep learning model are combined, the voice stream processing mode is dynamically adjusted, the high task concurrency situation is effectively dealt with, and misrecognition and resource conflicts are reduced. Voice streams are segmented and fragments are identified independently under the conditions of resource saving and high concurrency in a normal situation, so that major identification errors are prevented, fund loss and strategy failure risks are reduced, the market order stability is guaranteed, and the efficiency and security of a transaction room are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of voice intelligent terminals, and in particular to a personal voice intelligent terminal used in a fund transaction room. Background Art

[0002] The personal voice intelligent terminal of the capital trading room is a highly integrated voice interaction device designed for financial trading environments and provides full voice function support. Its core functions include ordinary telephone communication, dedicated line shouting function to achieve rapid information transmission, voice message recording to save key instructions, call recording function to ensure transaction compliance and responsibility tracking, voice-to-text function to achieve real-time recording and documentation, and voice query function to support rapid retrieval of historical data and transaction information. The terminal also has high sound quality, anti-noise processing and multi-channel voice separation technology to ensure clear and efficient information transmission in noisy environments, while improving the convenience of operation through intelligent voice assistants to meet the needs of high-intensity trading rooms.

[0003] The prior art has the following deficiencies: Under existing technical conditions, when the concurrency of voice tasks in financial trading scenarios is too high, the voice terminal needs to handle multiple tasks at the same time. However, due to limited computing resources, the task priority may be reduced due to resource allocation conflicts, which in turn affects the accuracy and real-time performance of voice recognition. In particular, when traders issue large-scale buy operations through voice commands, if the terminal misidentifies "buy" as "sell" due to resource limitations or computing load problems, it may mistakenly execute large-scale sell transactions. Such errors are difficult to detect in time because of their hidden nature, often resulting in huge financial losses, the complete failure of trading strategies, and even serious impacts on market order.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide a personal voice intelligent terminal for a capital trading room, which can effectively predict and respond to high task concurrency situations by real-time collection and classification of voice data, combined with feature engineering and deep learning models, dynamically adjust the processing method of voice streams, and reduce the risk of misidentification of voice commands and resource conflicts. For normal situations, maintaining the traditional processing method not only saves resources, but also ensures the efficient operation of the system; in high-concurrency situations, by segmenting the voice stream and independently identifying each segment, it effectively prevents the major error of "buy" being misidentified as "sell", greatly reducing the risk of huge financial losses and trading strategy failures caused by voice recognition errors, so as to solve the problems in the above-mentioned background technology.

[0006] In order to achieve the above-mentioned purpose, the present invention provides the following technical solutions: a personal voice intelligent terminal for a capital trading room, comprising a voice data collection and classification module, a data cleaning and structuring module, a task concurrency feature extraction module, a concurrent situation prediction module, a situation classification module, a normal situation processing module, and a high-concurrency situation processing module; Voice data collection and classification module: In financial trading scenarios, voice data of traders is collected in real time through terminal devices, and the collected raw voice data is classified and labeled according to task type and concurrency level, while irrelevant data and noise data are eliminated; The data cleaning and structuring module converts the cleaned data into an analysis set to form a structured data set; The task concurrency feature extraction module extracts key data reflecting the concurrency of voice tasks through feature engineering technology, collects statistics and analyzes the extracted voice task concurrency-related data within a preset time window, generates indicators describing the concurrency situation, and uses the analyzed indicators to preliminarily quantify the concurrency of voice tasks in financial transaction scenarios. The concurrency prediction module inputs the quantitative indicators generated from the time window into the pre-trained deep learning model to predict the future development trend of speech task concurrency; The situation classification module divides the speech task concurrency into two categories: "high task concurrency situation" and "normal situation" according to the output results of the deep learning model; Normal situation processing module, for normal situation, continues to use the traditional voice data processing method, that is, the voice data is input as a complete voice stream, and the voice terminal relies on the normal task processing mechanism for recognition and analysis; The high-concurrency situation processing module divides the voice stream into multiple independent segments based on the voice activity detection algorithm for high-task concurrency situations. At the same time, it dynamically adjusts the number and length of the segments according to the concurrency situation predicted by the machine learning model to ensure that each segment is independently identified, reducing task overlap and resource conflicts.

[0007] Preferably, the collected original voice data is classified and labeled according to the task type and concurrency level to form a structured data set in the following specific steps: Use speech recognition technology to convert the collected raw voice data into text, and combine it with the natural language processing model to parse the instruction content, identify the task type, and match the voice data with specific transaction tasks, laying the foundation for subsequent classification and analysis; After identifying the task type, the voice data is classified according to the transaction scenario requirements. At the same time, each piece of data is labeled to establish a rich attribute description for the voice data. Convert classified and annotated speech data into structured data sets and store them uniformly in table or database format to facilitate feature extraction and modeling.

[0008] Preferably, key data reflecting the concurrency of voice tasks are extracted through feature engineering technology, wherein the extracted features include the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks simultaneously in the processing queue to the normal task processing capacity. Within a preset time window, the degree of overlap of the extracted multiple voice tasks and the ratio of the number of voice tasks simultaneously in the processing queue to the normal task processing capacity are statistically analyzed to generate voice overlap quantification values ​​and voice task accumulation quantification values, respectively. The voice task concurrency in financial transaction scenarios is preliminarily quantified through the analyzed voice overlap quantification values ​​and voice task accumulation quantification values.

[0009] Preferably, the speech overlap quantification value and the speech task accumulation quantification value generated by statistics and analysis of the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity within a preset time window are input into a pre-trained deep learning model, and a task concurrency evaluation coefficient is generated based on the deep learning model. The future development trend of the speech task concurrency is predicted through the task concurrency evaluation coefficient.

[0010] Preferably, within a preset time window, the specific steps of statistically analyzing and analyzing the overlap degree of the extracted multiple voice tasks to generate the voice overlap quantization value are as follows: In the preset time window W, the start and end times of all speech tasks are extracted, and the cumulative duration of the overlapping parts of all tasks is calculated by detecting the intersection of any two speech task time intervals. The calculation expression is:

[0011] ,in: is the cumulative duration of the overlapping parts of all tasks, N is the total number of speech tasks in the time window, i is the index of the current speech task, j is the index of another speech task compared with the current speech task i in the time window, and are the start time and end time of the i-th speech task respectively, and are the start time and end time of the jth speech task respectively; Based on the cumulative duration of all task overlaps And the time window W, define the speech task overlap density, which is used to measure the ratio of the overlapping speech duration to the overall active duration of the task. The expression for the definition of speech task overlap density is:

[0012] ,in: is the speech task overlap density, which is used to quantitatively reflect the degree of task overlap. It is the cubed cumulative value of the duration of all speech tasks, which is used to amplify the difference in overlap density under high concurrency conditions; Combined speech task overlap density The speech overlap quantization value is generated by summing up the total number of tasks N, which is used to quantify the concurrency of speech tasks. The expression for generating the speech overlap quantization value is:

[0013] ,in: Quantify the value of speech overlap and quantify the concurrency of speech tasks. It is the quadratic cumulative square root of the time window and task duration, and is used to balance the impact of the number of tasks and overlap density on the results.

[0014] Preferably, within a preset time window, the specific steps of performing statistics and analysis on the ratio of the number of voice tasks extracted and in the processing queue to the normal task processing capacity to generate the voice task accumulation quantization value are as follows: In the time window, the ratio of the total processing pressure of the voice tasks in the processing queue to the normal task processing capacity of the system is calculated. Taking into account the importance weight of the task and the task processing redundancy capacity of the system, the task load ratio is generated. The generation expression of the task load ratio is:

[0015] ,in, is the task load ratio, is the importance weight of speech task u, which is dynamically assigned according to the task type. is the maximum task processing capability of the system, Redundancy for system task processing, Indicates the total number of speech tasks in the processing queue within the time window; By analyzing the distribution of tasks in the processing queue, the task queue dispersion index is calculated to measure the distribution balance of tasks in the processing queue. The calculation expression of the task queue dispersion index is:

[0016] ,in, is the task queue dispersion index, For the task The processing demand; The task load ratio and task queue dispersion index are combined to generate the voice task accumulation quantization value through a nonlinear function. The expression for generating the voice task accumulation quantization value is:

[0017] ,in, Accumulate quantitative values ​​for speech tasks, quantify the concurrency of speech tasks and highlight extreme stress states, and To adjust the coefficient, adjust the influence weights of task load ratio and task queue dispersion index respectively. is the system baseline adjustment coefficient, which is used to correct the range under different scenarios, and e is the natural exponent, which is used to amplify the nonlinear impact of the task load ratio.

[0018] Preferably, the task concurrency evaluation coefficient generated when predicting the future development trend of the speech task concurrency situation through the pre-trained deep learning model is compared and analyzed with the pre-set task concurrency evaluation coefficient reference threshold, and the speech task concurrency situation is divided. The specific division steps are as follows: If the task concurrency evaluation coefficient is greater than the task concurrency evaluation coefficient reference threshold, the speech task concurrency situation is classified as a “high task concurrency situation”; If the task concurrency evaluation coefficient is less than or equal to the task concurrency evaluation coefficient reference threshold, the voice task concurrency situation is classified as "normal progress situation".

[0019] Preferably, for high task concurrency, the voice stream is divided into multiple independent segments based on the voice activity detection algorithm, and the number and length of the segments are dynamically adjusted according to the concurrency predicted by the machine learning model. The specific steps are as follows: The task concurrency evaluation coefficient generated by the machine learning model is compared with the preset task concurrency evaluation coefficient reference threshold to calculate the concurrency adjustment coefficient, which provides a basis for subsequent dynamic adjustment. The calculation expression of the concurrency adjustment coefficient is:

[0020] ,in, is the task concurrency evaluation coefficient, generated by the deep learning model, which quantifies the intensity of concurrency. is the reference threshold of the task concurrency evaluation coefficient, is the concurrency adjustment factor, which is used to quantify the proportion of concurrency intensity exceeding the reference threshold; In the case of high task concurrency, the number of speech segments that need to be segmented is calculated based on the concurrency adjustment coefficient. The calculation expression is:

[0021] ,in, is the split coefficient, which indicates the basic number of splits that need to be allocated per unit of task concurrency. is the number of segmented fragments, rounded up; After determining the number of segments, the length of each segment is dynamically adjusted according to the current concurrent situation. The expression for dynamic adjustment is:

[0022] ,in, is the segment length adjustment factor, which indicates the sensitivity of the reference length to changes in task concurrency. The length of the segment, which indicates the processing time range or speech duration of each voice stream.

[0023] Based on the voice activity detection algorithm, the voice stream is actually segmented according to the determined number and length of segmentation segments. The segmented segment set is represented as , , each fragment The length is , the expression of the segmentation process is:

[0024] ,in, is the input original voice stream, It is a voice activity detection algorithm that is used to identify the start and end points of a speech segment and remove the silent or noisy parts. The independent speech segments after segmentation are marked in order.

[0025] In the above technical solution, the technical effects and advantages provided by the present invention are: By collecting and classifying voice data in real time, combined with feature engineering and deep learning models, it is possible to effectively predict and respond to high task concurrency situations, dynamically adjust the processing method of voice streams, and reduce the risk of misidentification of voice commands and resource conflicts. For normal situations, maintaining traditional processing methods not only saves resources, but also ensures the efficient operation of the system; in high concurrency situations, by segmenting the voice stream and independently identifying each segment, it effectively prevents the major error of "buy" being misidentified as "sell", greatly reducing the risk of huge financial losses and trading strategy failures caused by voice recognition errors. In addition, the system's dynamic adjustment and multi-level confirmation mechanism enhance the reliability and response speed of the overall operation, ensure the accuracy of financial transactions and the stability of market order, and ultimately improve the operational efficiency and security of the trading room. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0027] Figure 1 The present invention is a module schematic diagram of a personal voice intelligent terminal used in a fund transaction room. DETAILED DESCRIPTION

[0028] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0029] The present invention provides Figure 1 A personal voice intelligent terminal for a capital trading room shown in the figure includes a voice data collection and classification module, a data cleaning and structuring module, a task concurrency feature extraction module, a concurrent situation prediction module, a situation classification module, a normal situation processing module, and a high-concurrency situation processing module; Voice data collection and classification module: In financial trading scenarios, voice data of traders is collected in real time through terminal devices, and the collected raw voice data is classified and labeled according to task type (such as buy, sell, cancel order, etc.) and concurrency level, while irrelevant data and noise data are eliminated; The data cleaning and structuring module converts the cleaned data into an analysis set, including metadata such as timestamp, task type, voice length, and command keywords, to form a structured data set; Voice data, including command voice, conversation voice, and environmental noise, etc. This step requires the establishment of a stable voice collection mechanism to ensure the integrity, continuity, and diversity of the data, while adopting high sampling rate and anti-noise technology to ensure the quality of the collected data.

[0030] The specific steps for classifying and labeling the collected raw speech data according to task type and concurrency level to form a structured data set are as follows: Use speech recognition technology to convert the collected raw voice data into text, and combine it with the natural language processing (NLP) model to parse the instruction content and identify the task type (such as "buy", "sell", "cancel order"). To improve accuracy, it is necessary to establish a vocabulary containing keywords related to trading scenarios (such as "buy 1 million shares" or "cancel order"), and use context understanding technology to determine the actual meaning of keywords in the voice. For voice content that does not meet the task type (such as idle chat or invalid instructions), mark or remove it to ensure the effectiveness of task type identification.

[0031] Function: Match voice data with specific transaction tasks to lay the foundation for subsequent classification and analysis.

[0032] After identifying the task type, the voice data is classified according to the transaction scenario requirements, such as "buy task", "sell task" or "cancel task". At the same time, each piece of data is annotated, with additional information such as task type, voice length, keywords, etc., to create a rich attribute description for the voice data. The classification and annotation process can be completed with the help of automated tools to ensure the efficiency and consistency of data processing.

[0033] Function: Classification and labeling give voice data clear attributes, which facilitates subsequent structured processing and in-depth analysis.

[0034] Convert the classified and annotated voice data into a structured data set and store it in a table or database format, including metadata fields such as timestamp, task type, voice length, and command keywords. This step requires a quality check of the data to ensure the integrity and consistency of the fields and remove duplicate or erroneous entries. Efficient storage and management can be achieved through data frameworks (such as Pandas) or database tools, providing a clear and easy-to-use data foundation for subsequent analysis.

[0035] Function: Structured data sets improve the operability and analysis efficiency of data, and facilitate feature extraction and modeling.

[0036] The task concurrency feature extraction module extracts key data reflecting the concurrency of voice tasks through feature engineering technology, and performs statistics and analysis on the extracted voice task concurrency-related data within a preset time window (for example, 5 seconds or 10 seconds), generates indicators describing the concurrency situation, and preliminarily quantifies the voice task concurrency situation in financial transaction scenarios through the analyzed indicators; Feature engineering technology refers to the process of transforming, filtering and creating new features of raw data during data analysis and modeling to improve the performance and accuracy of the model. The core goal of feature engineering is to extract high-quality features that can reflect the essence of the problem from the raw data, so that the model can better understand the patterns and relationships of the data.

[0037] Key data reflecting the concurrency of voice tasks are extracted through feature engineering technology, wherein the extracted features include the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity. Within a preset time window, the degree of overlap of multiple extracted voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity are statistically analyzed to generate voice overlap quantification values ​​and voice task accumulation quantification values, respectively. The voice task concurrency in financial transaction scenarios is preliminarily quantified through the analyzed voice overlap quantification values ​​and voice task accumulation quantification values.

[0038] In financial trading scenarios, a high degree of overlap between multiple voice tasks usually indicates a higher degree of concurrency in voice tasks. This is because voice task overlap means that multiple trading instructions or tasks are issued or processed at the same time window, resulting in a shorter time interval between tasks or even overlapping with each other. The degree of voice overlap directly reflects the workload and concurrency pressure faced by the system. When this overlap occurs frequently, the system may need to parse, recognize, and process multiple instructions at the same time, increasing the complexity of resource allocation and the difficulty of voice recognition. Therefore, by quantifying the degree of voice overlap, the concurrency of voice tasks can be effectively evaluated, providing an important basis for the prediction and optimization of high concurrency situations.

[0039] The specific steps of statistically analyzing and generating speech overlap quantification values ​​for the overlap degree of multiple extracted speech tasks within a preset time window are as follows: In the preset time window W, the start and end times of all speech tasks are extracted, and the cumulative duration of the overlapping parts of all tasks is calculated by detecting the intersection of any two speech task time intervals. The calculation expression is:

[0040] ,in: is the cumulative duration of the overlapping parts of all tasks, N is the total number of speech tasks in the time window, i is the index of the current speech task, j is the index of another speech task compared with the current speech task i in the time window, and are the start time and end time of the i-th speech task respectively, and are the start time and end time of the jth speech task respectively; express: Take the earlier end time of the two tasks to represent the possible deadline of their overlapping part, ensuring that the end time of the overlapping part does not exceed the earlier end time of the two tasks; express: Take the later start time of the two tasks to represent the possible start time of their overlapping part, ensuring that the start time of the overlapping part does not start earlier than the later start time of the two tasks; express: Subtract the maximum start time from the minimum end time of the two tasks to calculate whether their time intervals overlap (the length of the overlapping part). If the result is a positive value, it means that the two tasks actually overlap; if it is a negative value, it means that the two tasks do not overlap. By calculating the potential overlapping duration, determine whether there is an intersection; express: if , then take its value to represent the overlapping duration, if , then it takes 0, indicating that the two tasks do not overlap; Make sure the result is non-negative, the overlap duration cannot be negative; This step provides basic data for the subsequent generation of speech overlap quantization values ​​by extracting the time information of the speech task and calculating the duration of the overlapping part.

[0041] Based on the cumulative duration of all task overlaps And the time window W, define the speech task overlap density, which is used to measure the ratio of the overlapping speech duration to the overall active duration of the task. The expression for the definition of speech task overlap density is:

[0042] ,in: is the speech task overlap density, which is used to quantitatively reflect the degree of task overlap. It is the cubed cumulative value of the duration of all speech tasks, which is used to amplify the difference in overlap density under high concurrency conditions; Speech task overlap density Comprehensively considering the overlap duration and task duration distribution can more accurately characterize the complexity of the concurrency level and avoid the shortcomings of simple proportional relationships.

[0043] Combined speech task overlap density The speech overlap quantization value is generated by summing up the total number of tasks N, which is used to quantify the concurrency of speech tasks. The expression for generating the speech overlap quantization value is:

[0044] ,in: Quantify the value of speech overlap and quantify the concurrency of speech tasks. It is the quadratic cumulative square root of the time window and task duration, which is used to balance the impact of the number of tasks and overlap density on the results; The speech overlap quantification value combines the overlap density, number of tasks and time distribution characteristics, which can more comprehensively quantify the degree of speech task concurrency, and the exponential form can more sensitively capture extreme cases of high concurrency.

[0045] It can be seen from the speech overlap quantification value that within the preset time window, the greater the performance value of the speech overlap quantification value generated by statistics and analysis of the overlap degree of multiple extracted speech tasks, the higher the overlap degree between multiple speech tasks within the statistical time window, which directly reflects the higher concurrency of speech tasks in financial transaction scenarios. The speech overlap quantification value intuitively expresses the complexity of multi-task concurrency by quantifying the overlapping time and frequency of tasks. When the speech overlap quantification value is large, it means that multiple speech tasks occur at the same time or the time interval between tasks is extremely short. This high concurrency situation is usually accompanied by an increase in system resource load and an increase in processing difficulty. On the contrary, when the speech overlap quantification value is small, it means that the distribution between speech tasks is more dispersed, with less overlap, the concurrency of speech tasks in financial transaction scenarios is low, and the system load is in a lighter state. Therefore, the size of the speech overlap quantification value is an important indicator for measuring the concurrency of speech tasks.

[0046] In financial trading scenarios, when the ratio of the number of voice tasks in the processing queue at the same time to the system's normal task processing capacity is large, it indicates that the concurrency of the voice tasks is higher. This is because the number of tasks in the processing queue directly reflects the current load of the system, while the normal task processing capacity is the task capacity that the system can efficiently process in a stable state. When the ratio of the two increases significantly, it means that the system is facing high concurrency pressure that exceeds the normal load, which may slow down the task processing speed, or even queue accumulation or resource competition. This situation is particularly significant in high-frequency instruction-intensive trading scenarios, which may further cause voice recognition delays or errors, affecting the accuracy and real-time nature of transactions. Therefore, by monitoring this ratio, the concurrency of voice tasks can be evaluated in real time.

[0047] The specific steps of generating the voice task accumulation quantification value by counting and analyzing the ratio of the number of voice tasks extracted and in the processing queue to the normal task processing capacity within the preset time window are as follows: In the time window, the ratio of the total processing pressure of the voice tasks in the processing queue to the normal task processing capacity of the system is calculated. Taking into account the importance weight of the task and the task processing redundancy capacity of the system, the task load ratio is generated. The generation expression of the task load ratio is:

[0048] , where (N is the total number of speech tasks in the time window), is the task load ratio, The importance weight of the voice task u is dynamically assigned according to the task type (such as "buy" or "sell"). For example, high-priority tasks have higher weights. is the maximum task processing capability of the system, that is, the upper limit of tasks that the system can efficiently process under normal conditions. Redundancy for system task processing, used to tolerate certain overload conditions. Indicates the total number of speech tasks in the processing queue within the time window; This step calculates the task load ratio by comprehensively considering the number of voice tasks and importance weights, reflecting the intensity of the current queue pressure, especially when there are more high-priority tasks.

[0049] By analyzing the distribution of tasks in the processing queue, the task queue dispersion index is calculated to measure the distribution balance of tasks in the processing queue. The calculation expression of the task queue dispersion index is:

[0050] ,in, is the task queue dispersion index, For the task The processing demand is usually allocated according to the voice length or computational complexity of the task. When the task distribution is more balanced, the value of the task queue dispersion index is higher; on the contrary, if the tasks are concentrated on a few high-demand tasks, the value of the task queue dispersion index is lower. Task The processing demand Refers to the computing resources or processing power required to complete the task. Usually, this requirement can be determined based on the length of the speech task (such as the duration of the speech) or its computational complexity. If the speech is long or requires more complex processing (such as higher recognition accuracy or multi-tasking), the processing requirement of the task will be higher. This allocation method reflects the actual occupation of system resources by the task, which helps to reasonably allocate processing power in the task queue and optimize the resource utilization efficiency of the system.

[0051] The queue dispersion index reflects the balance of task load and reveals whether the concurrency pressure comes from individual tasks or the overall load accumulation.

[0052] The task load ratio and task queue dispersion index are combined to generate the voice task accumulation quantization value through a nonlinear function. The expression for generating the voice task accumulation quantization value is:

[0053] ,in, Accumulate quantitative values ​​for speech tasks, quantify the concurrency of speech tasks and highlight extreme stress states, and To adjust the coefficient, adjust the influence weights of task load ratio and task queue dispersion index respectively. is the system benchmark adjustment coefficient, which is used to correct the range in different scenarios. is a natural exponent used to amplify the nonlinear effect of the task load ratio; It can be seen from the voice task accumulation quantification value that within the preset time window, the voice task accumulation quantification value generated by statistics and analysis of the ratio of the number of voice tasks extracted and in the processing queue to the normal task processing capacity is larger, which means that the number of voice tasks in the processing queue is close to or exceeds the processing capacity of the system, indicating that the higher the concurrency of the current voice tasks, the greater the pressure on the system; on the contrary, when the voice task accumulation quantification value is small, it means that the voice task load is relatively low, the concurrency is weak, and the system operation is closer to the normal state. Therefore, the size of the voice task accumulation quantification value can be used as an intuitive indicator for evaluating the concurrency of voice tasks.

[0054] The concurrency prediction module inputs the quantitative indicators generated from the time window into the pre-trained deep learning model to predict the future development trend of speech task concurrency; The speech overlap quantification value and speech task accumulation quantification value generated by statistics and analysis of the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity within a preset time window are input into a pre-trained deep learning model, and a task concurrency evaluation coefficient is generated based on the deep learning model. The future development trend of the voice task concurrency is predicted through the task concurrency evaluation coefficient.

[0055] A pre-trained deep learning model refers to a deep neural network model that has been fully trained using a large amount of high-quality historical data before it is officially applied. The model learns the characteristics of historical data related to speech task concurrency and grasps the laws and development trends of speech task concurrency. In the current scenario, the training data includes key features such as speech overlap quantization values ​​and speech task accumulation quantization values, as well as their patterns of change over time. The training process adjusts the model parameters to minimize the error through iterative optimization of multi-layer neural networks, so that it can accurately capture the complex dynamic characteristics of speech task concurrency. For example, the model may adopt an LSTM (Long Short-Term Memory Network) architecture, which is very suitable for modeling time series data, remembering the historical changes of speech task concurrency, and using this information to predict future development trends. In speech task concurrency analysis, the pre-trained model can receive input real-time data (such as speech overlap quantization values ​​and speech task accumulation quantization values) and process it based on the feature relationships learned during training. It does not need to learn from scratch, but directly applies existing knowledge to quickly calculate and analyze input data. The advantages of this model are high efficiency and accuracy. Especially in a dynamic, high-concurrency speech task environment, it can generate task concurrency evaluation coefficients in real time and provide a basis for predicting future trends in concurrency. This predictive capability is of great significance for optimizing resource allocation, adjusting processing strategies, and avoiding system bottlenecks.

[0056] The working mechanism of the pre-trained deep learning model includes three main links: feature extraction, nonlinear mapping, and output prediction. When inputting features such as speech overlap quantization value and speech task stacking quantization value, the model first processes these features through its first few layers of neurons to extract deeper abstract features. For example, the model may recognize that the combination of high overlap index and high stacking index is often related to the peak of speech task concurrency. Then, these extracted features are further processed into feature relationships with stronger predictive power through nonlinear mapping layers. For example, the LSTM model can remember the long-term trend of task concurrency through the gating mechanism, while ignoring short-term fluctuations to ensure the stability of prediction. At the output layer, the deep learning model generates task concurrency evaluation coefficients based on feature relationships. The task concurrency evaluation coefficients can be regarded as a comprehensive quantitative evaluation of speech task concurrency. The prediction results of the model can indicate whether the concurrency will continue to intensify or stabilize in the future, thereby providing the speech processing system with time to respond in advance. For example, when the prediction shows that the concurrency may enter a peak period, the system can take countermeasures in advance, such as dynamically adjusting resource allocation or optimizing the speech stream segmentation strategy to reduce concurrency pressure. Through this mechanism, the pre-trained deep learning model can not only improve the efficiency of speech task processing, but also significantly reduce the risk of system performance degradation or task recognition errors caused by concurrency problems.

[0057] The deep learning model is not specifically limited here, and can achieve the quantization value of speech overlap and speech task accumulation quantization value Perform comprehensive analysis to generate task concurrency evaluation coefficients In order to realize the technical solution of the present invention, the present invention provides a specific implementation method; Task concurrency evaluation coefficient The resulting calculation formula is:

[0058] , where , are the speech overlap quantization values and speech task accumulation quantization value The preset scaling factor of , Both are greater than 0.

[0059] It can be seen from the task concurrency evaluation coefficient that within the preset time window, the larger the performance value of the speech overlap quantification value generated by statistics and analysis of the degree of overlap of multiple extracted speech tasks, the larger the performance value of the speech task accumulation quantification value generated by statistics and analysis of the ratio of the number of extracted speech tasks in the processing queue at the same time to the normal task processing capacity, that is, the larger the performance value of the task concurrency evaluation coefficient generated when predicting the future development trend of speech task concurrency through a pre-trained deep learning model, the higher the concurrency degree of the current speech task, and vice versa.

[0060] The situation classification module divides the speech task concurrency into two categories: "high task concurrency situation" and "normal situation" according to the output results of the deep learning model; The task concurrency evaluation coefficient generated by predicting the future development trend of speech task concurrency through the pre-trained deep learning model is compared and analyzed with the pre-set task concurrency evaluation coefficient reference threshold, and the speech task concurrency is divided. The specific division steps are as follows: If the task concurrency evaluation coefficient is greater than the task concurrency evaluation coefficient reference threshold, the speech task concurrency situation is classified as a “high task concurrency situation”; If the task concurrency evaluation coefficient is less than or equal to the task concurrency evaluation coefficient reference threshold, the voice task concurrency situation is classified as "normal progress situation".

[0061] A high task concurrency situation indicates that the concurrency exceeds the set threshold, which may put pressure on the processing capacity of the voice terminal; a normal situation indicates that the current task concurrency is within a controllable range.

[0062] Normal situation processing module, for normal situation, continues to use the traditional voice data processing method, that is, the voice data is input as a complete voice stream, and the voice terminal relies on the normal task processing mechanism for recognition and analysis; When the voice task concurrency is classified as "normal progress", the system does not need to perform additional complex processing operations on the voice data, but maintains the existing traditional processing method, inputs the voice data as a continuous stream into the voice terminal, and recognizes and analyzes it according to the normal task processing mechanism. In this case, the voice terminal will not split or adjust the voice stream, but directly processes the complete data through its existing voice recognition model to generate results such as task category and instruction content. Since the concurrency pressure under the normal progress is low and the system resources are sufficient, the traditional processing method is sufficient to cope with it, and can keep the simplicity and consistency of the processing flow to the greatest extent.

[0063] By continuing to use this original processing method, additional resource consumption can be avoided, such as the computing resources required for operations such as splitting the voice stream or dynamically adjusting the task queue. This can not only save the system's computing overhead, but also ensure the consistency and recognition efficiency of data processing, especially in normal situations, maintaining the efficient operation of the system. In addition, this strategy can also avoid errors and delays that may be caused by additional adjustments, ensuring the accuracy and real-time processing of voice tasks. Therefore, for normal situations, maintaining the traditional method of processing voice data is a resource-efficient and stable choice, which not only simplifies operations but also optimizes overall system performance.

[0064] High concurrency situation processing module: For high task concurrency situations, the voice stream is divided into multiple independent segments based on the voice activity detection algorithm. At the same time, the number and length of the segments are dynamically adjusted according to the concurrency situation predicted by the machine learning model to ensure that each segment is independently recognized, reducing task overlap and resource conflicts; For high task concurrency, the voice stream is segmented into multiple independent segments based on the voice activity detection algorithm, and the number and length of the segments are dynamically adjusted according to the concurrency predicted by the machine learning model. The specific steps are as follows: The task concurrency evaluation coefficient generated by the machine learning model is compared with the preset task concurrency evaluation coefficient reference threshold to calculate the concurrency adjustment coefficient, which provides a basis for subsequent dynamic adjustment. The calculation expression of the concurrency adjustment coefficient is:

[0065] ,in, is the task concurrency evaluation coefficient, generated by the deep learning model, which quantifies the intensity of concurrency. It is the reference threshold of the task concurrency evaluation coefficient, indicating the upper limit of concurrency at which the system can run stably. is the concurrency adjustment factor, which is used to quantify the proportion of concurrency intensity exceeding the reference threshold.

[0066] In the case of high task concurrency, the number of speech segments that need to be segmented is calculated based on the concurrency adjustment coefficient. The calculation expression is:

[0067] ,in, The split coefficient indicates the number of basic splits that need to be allocated per unit of task concurrency. It is usually tuned based on historical data. The number of segments is rounded up to ensure that the segments are dispersed enough to reduce the system pressure. By dynamically determining the number of segments into which the voice stream is segmented, the system can effectively disperse voice tasks and avoid resource conflicts under high concurrency conditions.

[0068] After determining the number of segments, the length of each segment is dynamically adjusted according to the current concurrent situation. The expression for dynamic adjustment is:

[0069] ,in, is the segment length adjustment factor, which indicates the sensitivity of the reference length to changes in task concurrency and is usually optimized based on task load and speech stream characteristics. The length of the segment, which indicates the processing time range or speech duration of each speech stream; By dynamically adjusting the segment length, it is ensured that the voice stream segmentation can both match the system's processing capabilities and reduce the waste of resources caused by excessive segmentation.

[0070] Based on the voice activity detection (VAD) algorithm, the voice stream is actually segmented according to the determined number and length of segments. The segmented segment set is represented as , , each fragment The length is , the expression of the segmentation process is:

[0071] ,in, is the input original voice stream, It is a voice activity detection algorithm that is used to identify the start and end points of a speech segment and remove the silent or noisy parts. The independent speech segments after segmentation are marked in order; The original speech stream is divided into multiple clear and independent segments to ensure that each segment can be recognized and processed independently, reducing the impact of speech overlap on recognition accuracy.

[0072] There are many types of voice activity detection (VAD) algorithms, mainly including those based on energy threshold, statistical model, spectral feature, machine learning and deep learning. The energy threshold-based algorithm is the simplest. It distinguishes between speech and silence by setting an energy threshold, but it is easily interfered by background noise. The statistical model-based method (such as Gaussian mixture model, GMM) achieves detection by modeling the statistical distribution of speech and noise, and is suitable for low-noise environments. The spectral feature method uses the frequency domain characteristics of speech signals (such as short-time Fourier transform or Mel-frequency cepstral coefficient, MFCC) to distinguish speech segments, and the effect is relatively stable. Machine learning algorithms (such as support vector machines, SVM) can adapt to complex scenarios by extracting speech features and classifying speech and noise. Deep learning methods (such as those based on convolutional neural networks CNN or long short-term memory networks LSTM) perform well in dealing with complex background noise and dynamic environments, and can achieve high-precision voice activity detection through large-scale data training. These methods have their own advantages and disadvantages. In practical applications, a single algorithm is often selected according to the scenario or multiple methods are combined to improve the detection effect.

[0073] By solving the problems raised by the above solution, the recognition accuracy and system stability of the voice intelligent terminal in the capital trading room can be significantly improved, and the correct execution of trading instructions can be ensured. This solution can effectively predict and respond to high task concurrency situations by collecting and classifying voice data in real time, combining feature engineering and deep learning models, dynamically adjusting the processing method of voice streams, and reducing the risk of misidentification of voice commands and resource conflicts. For normal situations, maintaining the traditional processing method not only saves resources, but also ensures the efficient operation of the system; in high concurrency situations, by segmenting the voice stream and independently identifying each segment, it effectively prevents the major error of "buy" being misidentified as "sell", greatly reducing the risk of huge capital losses and trading strategy failures caused by voice recognition errors. In addition, the dynamic adjustment and multi-level confirmation mechanism of the system enhance the reliability and response speed of the overall operation, ensure the accuracy of financial transactions and the stability of market order, and ultimately improve the operational efficiency and security of the trading room.

[0074] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0075] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0076] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A personal voice intelligent terminal for a capital trading room, characterized in that: It includes voice data collection and classification module, data cleaning and structuring module, task concurrency feature extraction module, concurrent situation prediction module, situation classification module, normal situation processing module, and high concurrency situation processing module; the voice data collection and classification module, in the financial transaction scenario, collects the voice data of traders in real time through terminal equipment, classifies and labels the collected original voice data according to task type and concurrency level, and removes irrelevant data and noise data; the data cleaning and structuring module converts the cleaned data into an analysis set to form a structured data set; the task concurrency feature extraction module extracts key data reflecting the concurrency of voice tasks through feature engineering technology, and extracts the extracted data within a preset time window. The data related to voice task concurrency are counted and analyzed to generate indicators describing the concurrency situation. The concurrency situation of voice tasks in financial transaction scenarios is preliminarily quantified through the analyzed indicators. The key data reflecting the concurrency situation of voice tasks are extracted through feature engineering technology, wherein the extracted features include the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity. Within a preset time window, the degree of overlap of multiple extracted voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity are counted and analyzed to generate voice overlap quantification values ​​and voice task accumulation quantification values ​​respectively. The analyzed voice overlap quantification values ​​and voice task concurrency quantification values ​​are used to calculate the voice task concurrency situation. The accumulation quantification value preliminarily quantifies the concurrency of voice tasks in financial transaction scenarios; the concurrency situation prediction module inputs the quantitative indicators generated from the time window into the pre-trained deep learning model to predict the future development trend of the concurrency of voice tasks; the speech overlap quantification value and the speech task accumulation quantification value generated by statistics and analysis of the degree of overlap of multiple voice tasks and the ratio of the number of voice tasks in the processing queue at the same time to the normal task processing capacity within the preset time window are input into the pre-trained deep learning model, and the task concurrency evaluation coefficient is generated based on the deep learning model. The future development trend of the voice task concurrency is predicted through the task concurrency evaluation coefficient; the situation classification module According to the output results of the deep learning model, the speech task concurrency is divided into two categories: "high task concurrency situation" and "normal progress situation"; the task concurrency evaluation coefficient generated when the future development trend of the speech task concurrency is predicted by the pre-trained deep learning model is compared and analyzed with the pre-set task concurrency evaluation coefficient reference threshold, and the speech task concurrency is divided. The specific division steps are as follows: if the task concurrency evaluation coefficient is greater than the task concurrency evaluation coefficient reference threshold, the speech task concurrency is divided into "high task concurrency situation"; if the task concurrency evaluation coefficient is less than or equal to the task concurrency evaluation coefficient reference threshold, the speech task concurrency is divided into "normal progress situation";Normal situation processing module, for normal situation, continues to use the traditional voice data processing method, that is, the voice data is input as a complete voice stream, and the voice terminal relies on the normal task processing mechanism for recognition and analysis; high concurrency situation processing module, for high task concurrency situation, based on the voice activity detection algorithm, the voice stream is divided into multiple independent segments, and the number and length of the segmented segments are dynamically adjusted according to the concurrency situation predicted by the machine learning model to ensure that each segment is independently recognized, reducing task overlap and resource conflicts. ; 2. A personal voice intelligent terminal for a capital transaction room according to claim 1, characterized in that: The specific steps for classifying and labeling the collected raw speech data according to task type and concurrency level to form a structured data set are as follows: Use speech recognition technology to convert the collected raw voice data into text, and combine it with the natural language processing model to parse the instruction content, identify the task type, and match the voice data with specific transaction tasks, laying the foundation for subsequent classification and analysis; After identifying the task type, the voice data is classified according to the transaction scenario requirements. At the same time, each piece of data is labeled to establish a rich attribute description for the voice data. Convert classified and annotated speech data into structured data sets and store them uniformly in table or database format to facilitate feature extraction and modeling.

3. A personal voice intelligent terminal for a capital transaction room according to claim 1, characterized in that: The specific steps of statistically analyzing and generating speech overlap quantification values ​​for the overlap degree of multiple extracted speech tasks within a preset time window are as follows: In the preset time window W, the start and end times of all speech tasks are extracted, and the cumulative duration of the overlapping parts of all tasks is calculated by detecting the intersection of any two speech task time intervals. The calculation expression is: in: is the cumulative duration of the overlapping parts of all tasks, N is the total number of speech tasks in the time window, i is the index of the current speech task, j is the index of another speech task compared with the current speech task i in the time window, and are the start time and end time of the i-th speech task respectively, and are the start time and end time of the jth speech task respectively; Based on the cumulative duration of all task overlaps And the time window W, define the speech task overlap density, which is used to measure the ratio of the overlapping speech duration to the overall active duration of the task. The expression for the definition of speech task overlap density is: in: is the speech task overlap density, which is used to quantitatively reflect the degree of task overlap. It is the cubed cumulative value of the duration of all speech tasks, which is used to amplify the difference in overlap density under high concurrency conditions; Combined speech task overlap density The speech overlap quantization value is generated by summing up the total number of tasks N, which is used to quantify the concurrency of speech tasks. The expression for generating the speech overlap quantization value is: in: Quantify the value of speech overlap and quantify the concurrency of speech tasks. It is the quadratic cumulative square root of the time window and task duration, and is used to balance the impact of the number of tasks and overlap density on the results.

4. A personal voice intelligent terminal for a capital transaction room according to claim 3, characterized in that: The specific steps of generating the voice task accumulation quantification value by counting and analyzing the ratio of the number of voice tasks extracted and in the processing queue to the normal task processing capacity within the preset time window are as follows: In the time window, the ratio of the total processing pressure of the voice tasks in the processing queue to the normal task processing capacity of the system is calculated. Taking into account the importance weight of the task and the task processing redundancy capacity of the system, the task load ratio is generated. The generation expression of the task load ratio is: in, is the task load ratio, is the importance weight of speech task u, which is dynamically assigned according to the task type. is the maximum task processing capability of the system, Redundancy for system task processing, Indicates the total number of speech tasks in the processing queue within the time window; By analyzing the distribution of tasks in the processing queue, the task queue dispersion index is calculated to measure the distribution balance of tasks in the processing queue. The calculation expression of the task queue dispersion index is: in, is the task queue dispersion index, is the processing demand of task u; The task load ratio and task queue dispersion index are combined to generate the voice task accumulation quantization value through a nonlinear function. The expression for generating the voice task accumulation quantization value is: in, Accumulate quantitative values ​​for speech tasks, quantify the concurrency of speech tasks and highlight extreme stress states, and To adjust the coefficient, adjust the influence weights of task load ratio and task queue dispersion index respectively. is the system benchmark adjustment coefficient, which is used to correct the range in different scenarios. is a natural exponential that amplifies the nonlinear effect of the task load ratio.

5. The personal voice intelligent terminal for a capital transaction room according to claim 1, characterized in that: For high task concurrency, the voice stream is segmented into multiple independent segments based on the voice activity detection algorithm, and the number and length of the segments are dynamically adjusted according to the concurrency predicted by the machine learning model. The specific steps are as follows: The task concurrency evaluation coefficient generated by the machine learning model is compared with the preset task concurrency evaluation coefficient reference threshold to calculate the concurrency adjustment coefficient, which provides a basis for subsequent dynamic adjustment. The calculation expression of the concurrency adjustment coefficient is: in, is the task concurrency evaluation coefficient, generated by the deep learning model, which quantifies the intensity of concurrency. is the reference threshold of the task concurrency evaluation coefficient, is the concurrency adjustment factor, which is used to quantify the proportion of concurrency intensity exceeding the reference threshold; In the case of high task concurrency, the number of speech segments that need to be segmented is calculated based on the concurrency adjustment coefficient. The calculation expression is: in, is the split coefficient, which indicates the basic number of splits that need to be allocated per unit of task concurrency. is the number of segmented fragments, rounded up; After determining the number of segments, the length of each segment is dynamically adjusted according to the current concurrent situation. The expression for dynamic adjustment is: in, is the segment length adjustment factor, which indicates the sensitivity of the reference length to changes in task concurrency. The length of the segment, which indicates the processing time range or speech duration of each voice stream.

6. A personal voice intelligent terminal for a fund transaction room according to claim 5, characterized in that: Based on the voice activity detection algorithm, the voice stream is actually segmented according to the determined number and length of segmentation segments. The segmented segment set is represented as , , each fragment The length is , the expression of the segmentation process is: in, is the input original voice stream, It is a voice activity detection algorithm that is used to identify the start and end points of a speech segment and remove the silent or noisy parts. The independent speech segments after segmentation are marked in order.

Images