Model optimization method and device, equipment and storage medium

Abstract

The invention relates to the technical field of vehicle-mounted voice, and discloses a model optimization method, device and equipment and a storage medium, and the method comprises the steps: obtaining processing results fed back by a plurality of voice models based on the same preset task; selecting a reference result from the processing results; detecting whether a to-be-optimized voice model exists or not according to the deviation between each processing result and the reference result; and if yes, optimizing the to-be-optimized voice model according to the to-be-optimized processing result of the to-be-optimized voice model and the reference result. According to the invention, the reference result is selected from the processing results fed back by the plurality of voice models based on the same preset task, and after the to-be-optimized voice model is automatically detected according to the reference result, the detected to-be-optimized voice model is optimized by using the reference result. The user does not need to wait for actively feeding back the to-be-optimized voice model with the problem, and the use experience of the user is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle-mounted voice technology, and in particular to a model optimization method, device, equipment and storage medium. Background Art

[0002] The in-car assistant uses multiple voice models pre-configured by technicians to achieve conversations with users and other functions. In the process of the in-car assistant using voice models to achieve its functions, if there is a problem with a voice model, the voice assistant will be unable to achieve the function corresponding to the voice model.

[0003] At present, the solution to the above problem is for users to actively provide feedback, and then the technical staff will optimize the voice model based on the feedback to fix the problem. However, under this method, if the user does not promptly provide feedback on the problematic voice model, the problem of the voice model will continue to exist, resulting in the inability to implement the corresponding function of the voice model, affecting the user experience.

[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of the present invention is to provide a model optimization method, device, equipment and storage medium, aiming to solve the technical problem in the prior art that if the user does not promptly feedback the problematic speech model, the problem of the speech model will always exist, resulting in the inability to implement the function corresponding to the speech model, thereby affecting the user experience.

[0006] To achieve the above object, the present invention provides a model optimization method, which comprises the following steps:

[0007] Obtain processing results of multiple speech models based on the same preset task feedback;

[0008] Selecting a benchmark result from each of the processing results;

[0009] Detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the benchmark result;

[0010] If so, the speech model to be optimized is optimized according to the optimization processing result of the speech model to be optimized and the benchmark result.

[0011] Optionally, the step of selecting a benchmark result from each of the processing results includes:

[0012] Filtering each of the processing results, removing processing results whose deviations from other processing results exceed a preset deviation, and obtaining a plurality of filtered processing results;

[0013] Determining whether the post-filtering processing results are the same result;

[0014] If so, any of the filtered results is used as a benchmark result.

[0015] Optionally, after the step of determining whether the post-filtering processing results are the same result, the step further includes:

[0016] If not, determining the target speech model corresponding to each of the post-filtering processing results;

[0017] Performing performance evaluation on each of the target speech models according to the task type of the preset task;

[0018] Determine the optimal speech model whose performance evaluation results meet the preset performance requirements;

[0019] The processing result of the optimal speech model is selected from each of the post-filtering processing results as a benchmark result.

[0020] Optionally, the step of detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the benchmark result includes:

[0021] Determining result characteristics of each of the processing results and reference characteristics of the reference results;

[0022] Calculating the Pearson coefficient between each of the result features and the reference features, and constructing a correlation coefficient matrix according to each of the Pearson coefficients;

[0023] Based on the correlation coefficient matrix, it is detected whether there is a speech model to be optimized.

[0024] Optionally, the step of detecting whether there is a speech model to be optimized based on the correlation coefficient matrix includes:

[0025] Determine whether there is a target Pearson coefficient in the correlation coefficient matrix whose absolute value of the Pearson coefficient is lower than a preset threshold;

[0026] If the target Pearson coefficient exists, it is determined that there is a target speech model to be optimized;

[0027] Correspondingly, after the step of determining that there is a target speech model to be optimized if the target Pearson coefficient exists, the step further includes:

[0028] The processing result to be optimized corresponding to the target Pearson coefficient is determined, and the speech model corresponding to the processing result to be optimized is used as the speech model to be optimized.

[0029] Optionally, if present, the step of optimizing the speech model to be optimized according to the optimization processing result of the speech model to be optimized and the benchmark result comprises:

[0030] If so, determining a target Pearson coefficient between a plurality of target result features of the to-be-optimized processing result of the to-be-optimized speech model and the reference feature of the reference result;

[0031] Determine a change trend of each of the target Pearson coefficients, wherein the change trend represents a deviation between the target result feature and the reference feature;

[0032] Determine the result feature to be optimized from each of the target result features according to the change trend;

[0033] The speech model to be optimized is optimized based on the result features to be optimized.

[0034] Optionally, the step of optimizing the model to be optimized based on the characteristics of the result to be optimized includes:

[0035] Inputting the result feature to be optimized into a preset optimization model, wherein the preset optimization model is trained based on a plurality of result features and corresponding optimization information, wherein the optimization information is used to reduce the deviation between the result feature and the reference feature;

[0036] Receiving target optimization information fed back by the preset optimization model based on the characteristics of the result to be optimized;

[0037] The speech model to be optimized is optimized using the target optimization information.

[0038] In addition, to achieve the above object, the present invention also proposes a model optimization device, which includes:

[0039] A result feedback module is used to obtain the processing results of multiple speech models based on the same preset task feedback;

[0040] A result selection module, used for selecting a benchmark result from each of the processing results;

[0041] An optimization detection module, used for detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the reference result;

[0042] A model optimization module is used to optimize the target speech model according to the to-be-optimized processing result of the to-be-optimized speech model and the benchmark result, if any.

[0043] In addition, to achieve the above-mentioned purpose, the present invention also proposes a model optimization method device, which includes: a memory, a processor, and a model optimization method program stored in the memory and executable on the processor, and the model optimization method program is configured to implement the steps of the model optimization method described above.

[0044] In addition, to achieve the above-mentioned purpose, the present invention also proposes a storage medium, on which a model optimization method program is stored, and when the model optimization method program is executed by a processor, the steps of the model optimization method described above are implemented.

[0045] The present invention provides a model optimization method, device, equipment and storage medium. The method obtains the processing results of multiple speech models based on the same preset task feedback; then selects a benchmark result from each processing result; then detects whether there is a speech model to be optimized based on the deviation between each processing result and the benchmark result; if there is, the speech model to be optimized is optimized based on the processing result to be optimized of the speech model to be optimized and the benchmark result. The present invention selects a benchmark result from the processing results of multiple speech models based on the same preset task feedback, and then automatically detects the speech model to be optimized based on the benchmark result, and optimizes the detected speech model to be optimized using the benchmark result. There is no need to wait for the user to actively feedback the speech model to be optimized with problems, and it can ensure that the speech model to be optimized with problems is optimized in a timely manner, and ensure that the function of the speech model to be optimized is realized, which effectively improves the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the device structure of a model optimization method for a hardware operating environment involved in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the flow chart of the first embodiment of the model optimization method of the present invention;

[0048] Figure 3 This is a flow chart of a second embodiment of the model optimization method of the present invention;

[0049] Figure 4 This is a schematic diagram of the flow chart of the third embodiment of the model optimization method of the present invention;

[0050] Figure 5 This is a structural block diagram of the first embodiment of the model optimization device of the present invention.

[0051] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0053] Reference Figure 1 , Figure 1 The present invention is a schematic diagram of the device structure of a model optimization method for a hardware operating environment involved in an embodiment of the present invention.

[0054] like Figure 1 As shown, the model optimization method device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0055] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the model optimization method and apparatus, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0056] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a model optimization method program.

[0057] exist Figure 1 In the model optimization method device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the model optimization method device of the present invention can be set in the model optimization method device, and the model optimization method device calls the model optimization method program stored in the memory 1005 through the processor 1001, and executes the model optimization method provided by the embodiment of the present invention.

[0058] The embodiment of the present invention provides a model optimization method, referring to Figure 2 , Figure 2 Schematic diagram of the flow chart of the first embodiment of the model optimization method of the present invention.

[0059] In this embodiment, the model optimization method includes the following steps:

[0060] Step S10: Obtain processing results of multiple speech models based on the same preset task feedback.

[0061] It should be noted that the execution subject of the method of this embodiment can be a computing service device with model optimization, network communication and program running functions, such as a mobile phone, a tablet computer, a personal computer, etc., and can also be other electronic devices that achieve the same or similar functions. The following embodiment and the following embodiments are described using the above-mentioned model optimization device.

[0062] It is understandable that the above-mentioned voice model can be a model that supports the in-car assistant to implement corresponding functions based on the received user voice. Different voice models are used to implement different functions, among which the functions supported by the in-car assistant at least include entertainment system control, communication system control, weather query, and small talk dialogue, etc. For example, if the in-car assistant receives a call instruction input by the user's voice, the communication control voice model can be called to analyze the call instruction and automatically dial the phone of the relevant contact; if the in-car assistant receives a weather query instruction input by the user's voice, the query voice model can be called to analyze the weather query instruction and feedback the weather conditions.

[0063] It should be noted that the above-mentioned preset tasks may be basic voice tasks pre-configured for the in-vehicle assistant and capable of being recognized by different voice models, such as common tasks such as voice recognition, voice synthesis, and voice enhancement.

[0064] In a specific implementation, the above-mentioned model optimization device is applied to the in-vehicle assistant, and the above-mentioned preset tasks can be input into the model optimization device by the technicians in advance and stored, and a fixed period can be set. When the above-mentioned model optimization device detects that the current fixed period has been entered, it can call all the voice models of the in-vehicle assistant to process the preset tasks respectively, and obtain the processing results output by each voice model for processing the preset tasks.

[0065] It should be understood that when the above-mentioned model optimization device detects that the time from the input moment of the above-mentioned preset task to the current moment reaches the preset time, it can use the last user voice information recognition task that can be processed by all voice models before the current moment as a new preset task to update the preset task and avoid long-term use of the preset task affecting the model optimization accuracy.

[0066] Step S20: selecting a reference result from the processing results.

[0067] In a specific implementation, the above-mentioned model optimization device can convert the format and dimension of the processing results of each speech model so that each processing result is in the same dimension and has a unified format, and then fuse the processed results. During the fusion process, different weights can be assigned to different speech models according to the accuracy, performance or other evaluation indicators of each speech model, and then the weighted average of each processing result is calculated for each weight, and the processing result with the largest weight is selected from the processing results whose weights exceed the weighted average as the benchmark result. Among them, speech models with higher accuracy, performance or other evaluation indicators can be assigned higher weights to improve the accuracy of the benchmark results.

[0068] Step S30: Detect whether there is a speech model to be optimized based on the deviation between each of the processing results and the benchmark result.

[0069] It should be noted that the above-mentioned benchmark results combine the processing results of different speech models, which can combine the advantages of multiple speech models and reduce the errors existing in a single speech model. If the processing result of any speech model is close to the benchmark result, it can be determined that the processing result obtained by the speech model is of high accuracy and has good model performance. On the contrary, if the processing result of any speech model deviates greatly from the benchmark result, it can be determined that the processing result obtained by the speech model is of low accuracy and has problems.

[0070] In a specific implementation, the above-mentioned model optimization device can compare the processing results of each speech model with the benchmark results, calculate the mean square error between the processing results of each speech model and the benchmark results, and use the calculated mean square error as the deviation between each processing result and the benchmark result. Then, the deviation is compared with a preset mean square error threshold, and the speech model whose deviation reaches the preset mean square error threshold is used as the speech model to be optimized.

[0071] It should be understood that the above-mentioned preset mean square error threshold value can be a threshold value for judging whether the deviation between the processing result and the benchmark result is large. If the deviation between the processing result of any speech model and the benchmark result reaches the preset mean square error threshold value, it can be determined that the deviation between the processing result of any speech model and the benchmark result is large, and there is a problem with any speech model and optimization is required. On the contrary, if the deviation between the processing result of any speech model and the benchmark result does not reach the preset mean square error threshold value, it can be determined that the deviation between the processing result of any speech model and the benchmark result is small, and the performance of any speech model is normal and no optimization is required.

[0072] Step S40: If it exists, optimizing the speech model to be optimized according to the optimization processing result of the speech model to be optimized and the benchmark result.

[0073] In a specific implementation, when the above-mentioned model optimization device detects the existence of a speech model to be optimized, it can extract the deviation between the processing result to be optimized output by the speech model to be optimized and the benchmark result, generate corresponding prompt information based on the deviation, and optimize the speech model to be optimized through the prompt information, wherein the prompt information is used to guide the processing result to tend toward the benchmark result, so as to reduce the deviation between the processing result of the speech model to be optimized and the benchmark result.

[0074] For example, the optimization processing result of the speech model to be optimized is "select Movie C from Movie A, Movie B and Movie C", and the benchmark result is "because Movie C is mentioned many times by the user, select Movie C from Movie A, Movie B and Movie C". According to the deviation between the optimization processing result and the benchmark result, corresponding prompt information can be generated. For example, if the prompt information is "the user is interested in Movie C", the optimization processing result after integrating the prompt information is "the user is interested in Movie C, so select Movie C from Movie A, Movie B and Movie C". Since the prompt information is automatically generated based on the deviation, the optimization processing result after integrating the prompt information cannot be completely consistent with the benchmark result, but it can effectively reduce the deviation between the optimization processing result and the benchmark result.

[0075] This embodiment obtains the processing results of multiple speech models based on the same preset task feedback; then selects the benchmark results from each processing result; then detects whether there is a speech model to be optimized based on the deviation between each processing result and the benchmark result; if so, optimizes the speech model to be optimized based on the processing results to be optimized of the speech model to be optimized and the benchmark results. This embodiment selects the benchmark results from the processing results of multiple speech models based on the same preset task feedback, then automatically detects the speech model to be optimized based on the benchmark results, and optimizes the detected speech model to be optimized using the benchmark results. There is no need to wait for the user to actively feedback the speech model to be optimized that has problems, and it can ensure that the speech model to be optimized that has problems is optimized in a timely manner, ensuring that the function of the speech model to be optimized is realized, and effectively improving the user experience.

[0076] refer to Figure 3 , Figure 3 Schematic diagram of the flow chart of the second embodiment of the model optimization method of the present invention.

[0077] Based on the above first embodiment, in this embodiment, step S20 includes:

[0078] Step S201: Filter the processing results to remove the processing results whose deviations from other processing results exceed a preset deviation, and obtain a plurality of filtered processing results.

[0079] It should be noted that the above-mentioned preset deviation can be a threshold value for judging whether the deviation between different processing results is large. Since each processing result is generated by different speech models processing the same preset task, the deviation between each processing result should be relatively close. If the deviation between the processing result of any speech model and other processing results reaches the above-mentioned preset deviation, it can be determined that the processing result of the speech model may be inaccurate due to model error, outliers or noise interference, and the deviation from other processing results is large, that is, the speech model to be optimized may exist in the speech model that is filtered out.

[0080] In a specific implementation, the above-mentioned model optimization device can calculate the deviation between the processing results of each speech model, and can adopt the above-mentioned method of obtaining the mean square error, and then compare the calculated deviation with the preset deviation in turn. When selecting the benchmark result, the processing result whose deviation reaches the preset deviation can be temporarily removed, that is, the inaccurate results in each processing result are removed to ensure that accurate benchmark results can be obtained subsequently.

[0081] Step S202: Determine whether the post-filtering processing results are the same result.

[0082] Step S203: If yes, any of the filtered results is used as a benchmark result.

[0083] In a specific implementation, if multiple post-filtering processing results are the same result, any post-filtering result can be directly used as a benchmark result.

[0084] In an optional implementation, if there are multiple different post-filtering processing results, the number of occurrences of different post-filtering processing results can be counted, and then the number can be compared with the preset number to determine the post-filtering processing results whose number reaches the preset number. For example, if the preset number is N, N>0, the processing results include A, B, C, D and E, the number of occurrences of A is N+1, the number of occurrences of B is N+2, the number of occurrences of C is N-1, the number of occurrences of D is N-2, and the number of occurrences of E is N-3, then the number of occurrences of A and the number of occurrences of B are both greater than N, that is, all A and all B can be the target processing results. When the number of occurrences of any post-filtering processing result reaches the preset number, it is determined that the number of occurrences of the post-filtering processing result is large and can accurately reflect the processing results of each single speech model. Based on this, the post-filtering processing result with the largest number of occurrences can be selected from the post-filtering processing results whose number reaches the preset number as the benchmark result to improve the accuracy of the benchmark result.

[0085] In another optional implementation, after step S202, the method further includes:

[0086] Step S203': If not, determining the target speech model corresponding to each of the post-filtering processing results.

[0087] In a specific implementation, when the model optimization device determines that the post-filtering processing results are not the same result, it can also determine the target speech model for outputting different post-filtering processing results.

[0088] Step S204': performing performance evaluation on each of the target speech models according to the task type of the preset task.

[0089] In a specific implementation, the above-mentioned model optimization device can determine the task type of the selected preset task, and then calculate the performance indicators of different target speech models corresponding to the task type, and perform performance evaluation on the target speech model according to the level of the performance indicators. For example, if the task type of the preset task is a speech recognition type, the calculated performance indicators can be the accuracy, recall rate, and F1 score of the target speech model, and if the task type of the preset task is a speech synthesis type, the calculated performance indicators can be the naturalness and intelligibility of the target speech model.

[0090] It should be understood that after calculating the performance indicators of different target speech models, corresponding weights can be assigned to the target speech models according to the size of the performance indicators. Higher weights are assigned to models with better performance indicators, and lower weights are assigned to models with less favorable performance indicators. The weights of each target speech model are used as the performance evaluation results of each target speech model.

[0091] Step S205': Determine the optimal speech model whose performance evaluation result meets the preset performance requirements.

[0092] It should be noted that the above-mentioned preset performance requirements may be a judgment condition for evaluating whether the target speech model meets the performance requirements. When the above-mentioned performance evaluation result is composed of the weight of the target speech model, the performance evaluation result may be a preset weight, that is, the target speech model whose weight reaches the preset weight may be identified as the optimal speech model that meets the preset performance requirements.

[0093] In a specific implementation, the model optimization device may compare the weight of each target speech model in the performance evaluation result with a preset weight, and take the target speech model whose weight reaches the preset weight as the optimal speech model.

[0094] Step S206 ′: selecting the processing result of the optimal speech model from the post-filtering processing results as a benchmark result.

[0095] In a specific implementation, since the performance of the optimal speech model meets the requirements, that is, the performance is better, the above-mentioned model optimization device can directly use the processing results of the optimal speech model as the benchmark result after determining the optimal speech model from different target processing results.

[0096] In addition to the above-mentioned method of selecting the benchmark results, the processing results of each speech model can also be input into a pre-built deep learning model (such as a neural network), and the deep learning model can output the benchmark results using the feature information of each speech model. The deep learning model can be trained based on the historical processing results of different speech models and the optimal results pre-determined from the historical processing results. Among them, the above-mentioned optimal results can be determined from different historical processing results in accordance with the above-mentioned performance evaluation method.

[0097] This embodiment obtains multiple filtered processing results by filtering each processing result, removing the processing results whose deviations from other processing results reach a preset deviation; determining whether each of the filtered processing results is the same result; if so, taking any of the filtered processing results as a benchmark result. If not, determining the target speech model corresponding to each filtered processing result; performing a performance evaluation on each target speech model according to the task type of the preset task; determining the optimal speech model whose performance evaluation result meets the preset performance requirements; and selecting the target processing result of the optimal speech model from each filtered processing result as a benchmark result. This embodiment directly takes any of the filtered processing results as a benchmark result when each of the filtered processing results is the same result, and takes the processing result of the optimal speech model directly as a benchmark result according to the performance evaluation result of the target processing model corresponding to each filtered processing result when each of the filtered processing results is not the same result, without the need for tedious calculations, thereby effectively improving the efficiency of obtaining benchmark results.

[0098] refer to Figure 4 , Figure 4 Schematic diagram of the flow chart of the third embodiment of the model optimization method of the present invention.

[0099] Based on the above embodiments, in this embodiment, step S30 includes:

[0100] Step S301: Determine the result characteristics of each of the processing results and the benchmark characteristics of the benchmark results.

[0101] In a specific implementation, the above-mentioned model optimization device can determine the result features of each processing result and the benchmark features of the benchmark results according to the task type of the preset task. For example, when the task type of the preset task is a speech recognition type, the result features and the benchmark features can be the recognition accuracy, when the task type of the preset task is a speech synthesis type, the result features and the benchmark features can be the naturalness of the speech, and when the task type of the preset task is a sentiment analysis type, the result features and the benchmark features can be the recognition rates of different sentiments.

[0102] Step S302: Calculate the Pearson coefficient between each of the result features and the reference feature, and construct a correlation coefficient matrix according to each Pearson coefficient.

[0103] It should be noted that the above-mentioned Pearson coefficient can be a correlation coefficient used to measure the degree of linear correlation between the result feature of each of the above-mentioned processing results and the benchmark feature of the benchmark result. The value range of the Pearson coefficient is between -1 and 1. When the Pearson coefficient between the result feature and the benchmark feature is close to 1, it is determined that there is a strong positive correlation between the result feature and the benchmark feature. When the Pearson coefficient between the result feature and the benchmark feature is close to -1, it is determined that there is a strong negative correlation between the result feature and the benchmark feature. When the Pearson coefficient between the result feature and the benchmark feature is close to 0, it is determined that there is a weak correlation between the result feature and the benchmark feature, that is, the deviation between the two is large.

[0104] In a specific implementation, the above-mentioned model optimization device can calculate the average value of all features (including each result feature and a baseline feature), and then determine the difference between each result feature and the baseline feature and the average value, and then multiply the difference of each result feature with the difference of the baseline feature to obtain the product of the differences, and calculate the Pearson coefficient between each result feature and the baseline feature based on the product of the differences and the square of the differences, and then project each Pearson coefficient into a matrix to form a correlation coefficient matrix constructed with the Pearson coefficient.

[0105] Step S303: Detect whether a speech model is to be optimized based on the correlation coefficient matrix.

[0106] In a specific implementation, the above-mentioned model optimization device can detect the degree of proximity of each Pearson coefficient in the correlation coefficient matrix to 0. The closer the Pearson coefficient is to 0, the greater the deviation between the result feature corresponding to the Pearson coefficient and the benchmark feature. Therefore, when it is detected that there is a result feature with a Pearson coefficient close to 0, it can be determined that there is a speech model to be optimized, and the speech model corresponding to the result feature with the Pearson coefficient closest to 0 can be used as the speech model to be optimized.

[0107] Furthermore, in this embodiment, step S303 includes:

[0108] Step S3031: Determine whether there is a target Pearson coefficient in the correlation coefficient matrix whose absolute value of the Pearson coefficient is lower than a preset threshold.

[0109] It should be noted that the above-mentioned preset threshold value may be a threshold value for judging that the Pearson coefficient is less different from 0, that is, close to 0. If the absolute value of the Pearson coefficient is lower than the preset threshold value, it can be judged that the Pearson coefficient is closer to 0. On the contrary, if the absolute value of the Pearson coefficient is greater than the preset threshold value, it can be judged that the Pearson coefficient is greatly different from 0.

[0110] In a specific implementation, the above-mentioned model optimization device can traverse each Pearson coefficient in the correlation coefficient matrix, compare the absolute value of each Pearson coefficient with a preset threshold, and determine whether there is a target Pearson coefficient whose absolute value is lower than the preset threshold.

[0111] Step S3032: If the target Pearson coefficient exists, it is determined that there is a speech model to be optimized.

[0112] In a specific implementation, the above-mentioned model optimization device can determine that when it is determined that there is a target Pearson coefficient whose absolute value is lower than a preset threshold, it can determine that the deviation between the result feature corresponding to the target Pearson coefficient and the benchmark feature is large, that is, it is determined that there is a speech model to be optimized.

[0113] Accordingly, after step S3032, the following steps are further included:

[0114] Step S3033: Determine the processing result to be optimized corresponding to the target Pearson coefficient, and use the speech model corresponding to the processing result to be optimized as the speech model to be optimized.

[0115] In a specific implementation, the above-mentioned model optimization device can determine the target result feature corresponding to the target Pearson coefficient, and use the processing result corresponding to the target result feature as the processing result to be optimized, that is, the deviation between the processing result to be optimized and the benchmark result is large and needs to be optimized, so the speech model that outputs the processing result to be optimized can be used as the speech model to be optimized.

[0116] This embodiment determines the result features of each processing result and the benchmark features of the benchmark results; calculates the Pearson coefficient between each result feature and the benchmark feature, and constructs a correlation coefficient matrix based on each Pearson coefficient; and detects whether there is a speech model to be optimized based on the correlation coefficient matrix. This embodiment compares the absolute value of the Pearson coefficient in the correlation coefficient matrix with a preset threshold, and when there is a target Pearson coefficient whose absolute value of the Pearson coefficient is lower than the preset threshold, it is determined that there is a speech model to be optimized, thereby effectively improving the detection accuracy of the speech model to be optimized.

[0117] Furthermore, in this embodiment, step S40 includes:

[0118] Step S401: If so, determine the Pearson coefficient between a plurality of target result features of the to-be-optimized processing result of the to-be-optimized speech model and the reference reference feature of the reference result.

[0119] In a specific implementation, each of the above-mentioned processing results to be optimized may include multiple target result features, and the baseline features may also include multiple baseline features. The Pearson coefficient calculation process between each of the above-mentioned processing results to be optimized and the baseline features may use any result feature under the same dimension to participate in the calculation. After detecting the existence of a speech model to be optimized, the above-mentioned model optimization device may determine the processing result to be optimized of the speech model to be optimized, then extract all the target result features of the speech model to be optimized, then determine all the baseline features of the baseline results, and calculate the Pearson coefficient between the target result features and the baseline features under each dimension.

[0120] Step S402: determining a change trend of each of the Pearson coefficients, wherein the change trend represents a deviation between the target result feature and the reference feature.

[0121] In a specific implementation, the above-mentioned model optimization device can determine the change trend of the Pearson coefficient between each target result feature and each benchmark feature over time or data batches. For example, if the Pearson coefficient gradually decreases, it means that the deviation between the target result feature and the benchmark feature gradually increases, and the performance of the speech model to be optimized decreases.

[0122] Step S403: determining the result feature to be optimized from each of the target result features according to the change trend.

[0123] In a specific implementation, the above-mentioned model optimization device can select the result feature to be optimized whose absolute value of the Pearson coefficient in the change trend reaches the above-mentioned preset threshold from each target result feature, and determine that the deviation between the result feature to be optimized and the benchmark feature is large, and the result feature to be optimized needs to be optimized.

[0124] Step S404: Optimizing the target speech model based on the result features to be optimized.

[0125] In a specific implementation, the above-mentioned model optimization device can adjust the result feature to be optimized according to the baseline feature, and then optimize the target speech model according to the adjusted result feature to be optimized to reduce the deviation between the result feature to be optimized and the baseline feature, thereby repairing the problems existing in the target speech model.

[0126] Furthermore, in this embodiment, step S404 includes:

[0127] Step S4041: inputting the result feature to be optimized into a preset optimization model, wherein the preset optimization model is trained based on a plurality of result features and corresponding optimization information, wherein the optimization information is used to reduce the deviation between the result feature and the benchmark feature.

[0128] Step S4042: receiving target optimization information fed back by the preset optimization model based on the characteristics of the result to be optimized.

[0129] It should be noted that the above-mentioned optimization information may be pre-constructed information used to optimize the characteristics of the result to be optimized.

[0130] In a specific implementation, the above-mentioned model optimization device can input the above-mentioned result features to be optimized into a preset optimization model, and the preset optimization model outputs target optimization information based on the result features to be optimized. The above-mentioned preset optimization model can be trained based on multiple pre-constructed result features and optimization information assigned to the multiple result features. The optimization information can be the prompt information described in the above embodiment.

[0131] Step S4043: Optimize the speech model to be optimized using the target optimization information.

[0132] In a specific implementation, the model optimization device can obtain training samples for training the speech model to be optimized, and then update the target prompt information into the training samples to obtain updated training samples. Then, the speech model to be optimized is iteratively trained again using the updated training samples to optimize the speech model to be optimized, reduce the deviation between the result features to be optimized and the benchmark features, and thus fix the problems existing in the speech model to be optimized.

[0133] This embodiment determines the Pearson coefficients between multiple target result features and the reference features of the speech model to be optimized when there is a speech model to be optimized; determines the change trend of the Pearson coefficients of each target result feature; determines the result feature to be optimized from each of the target result features according to the change trend; and optimizes the speech model to be optimized based on the result feature to be optimized. This embodiment obtains the target optimization information corresponding to the result feature to be optimized, and optimizes the speech model to be optimized based on the target optimization information, thereby effectively improving the optimization accuracy of the speech model to be optimized.

[0134] In addition, an embodiment of the present invention further proposes a storage medium, on which a model optimization method program is stored. When the model optimization method program is executed by a processor, the steps of the model optimization method described above are implemented.

[0135] Reference Figure 5 , Figure 5 This is a structural block diagram of the first embodiment of the model optimization method device of the present invention.

[0136] like Figure 5 As shown, the model optimization method device proposed in the embodiment of the present invention includes:

[0137] The result feedback module 501 is used to obtain the processing results of multiple speech models based on the same preset task feedback.

[0138] The result selection module 502 is used to select a benchmark result from each of the processing results.

[0139] The optimization detection module 503 is used to detect whether there is a speech model to be optimized based on the deviation between each of the processing results and the benchmark result.

[0140] The model optimization module 504 is used to optimize the speech model to be optimized according to the processing result of the speech model to be optimized and the benchmark result, if any.

[0141] This embodiment obtains the processing results of multiple speech models based on the same preset task feedback; then selects the benchmark results from each processing result; then detects whether there is a speech model to be optimized based on the deviation between each processing result and the benchmark result; if so, optimizes the speech model to be optimized based on the processing results to be optimized of the speech model to be optimized and the benchmark results. This embodiment selects the benchmark results from the processing results of multiple speech models based on the same preset task feedback, then automatically detects the speech model to be optimized based on the benchmark results, and optimizes the detected speech model to be optimized using the benchmark results. There is no need to wait for the user to actively feedback the speech model to be optimized that has problems, and it can ensure that the speech model to be optimized that has problems is optimized in a timely manner, ensuring that the function of the speech model to be optimized is realized, and effectively improving the user experience.

[0142] Based on the above-mentioned first embodiment of the model optimization method device of the present invention, a second embodiment of the model optimization method device of the present invention is proposed.

[0143] In this embodiment, the result selection module 502 is also used to filter each of the processing results, remove the processing results whose deviation from other processing results exceeds a preset deviation, and obtain multiple filtered processing results; determine whether each of the filtered processing results is the same result; if so, use any of the filtered results as a benchmark result.

[0144] As an implementation mode, the result selection module 502 is also used to, if not, determine the target speech model corresponding to each of the post-filtering processing results; perform performance evaluation on each of the target speech models according to the task type of the preset task; determine the optimal speech model whose performance evaluation result meets the preset performance requirements; and select the processing result of the optimal speech model from each of the post-filtering processing results as the benchmark result.

[0145] Based on the above-mentioned embodiments of the model optimization method device of the present invention, a third embodiment of the model optimization method device of the present invention is proposed.

[0146] In this embodiment, the optimization detection module 503 is also used to determine the result characteristics of each of the processing results and the benchmark characteristics of the benchmark results; calculate the Pearson coefficient between each of the result characteristics and the benchmark characteristics, and construct a correlation coefficient matrix based on each Pearson coefficient; based on the correlation coefficient matrix, detect whether there is a speech model to be optimized.

[0147] As an implementation mode, the optimization detection module 503 is also used to determine whether there is a target Pearson coefficient in the correlation coefficient matrix whose absolute value of the Pearson coefficient is lower than a preset threshold; if the target Pearson coefficient exists, it is determined that there is a target speech model to be optimized; determine the processing result to be optimized corresponding to the target Pearson coefficient, and use the speech model corresponding to the processing result to be optimized as the speech model to be optimized.

[0148] As an implementation mode, the model optimization module 504 is also used to determine, if any, a target Pearson coefficient between multiple target result features of the to-be-optimized processing result of the speech model to be optimized and the baseline feature of the baseline result; determine a change trend of each of the target Pearson coefficients, wherein the change trend characterizes the deviation between the target result feature and the baseline feature; determine the result feature to be optimized from each of the target result features according to the change trend; and optimize the speech model to be optimized based on the result feature to be optimized.

[0149] As an implementation mode, the model optimization module 504 is also used to input the result features to be optimized into a preset optimization model, the preset optimization model is trained based on multiple result features and corresponding optimization information, and the optimization information is used to reduce the deviation between the result features and the baseline features; receive target optimization information fed back by the preset optimization model based on the result features to be optimized; and optimize the speech model to be optimized through the target optimization information.

[0150] Other embodiments or specific implementations of the model optimization method device of the present invention can refer to the above-mentioned method embodiments and will not be described in detail here.

[0151] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0152] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0153] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0154] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A model optimization method, characterized in that: The method comprises: Obtain processing results of multiple speech models based on the same preset task feedback; Selecting a benchmark result from each of the processing results; Detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the benchmark result; If so, the speech model to be optimized is optimized according to the optimization processing result of the speech model to be optimized and the benchmark result.

2. The model optimization method according to claim 1, characterized in that: The step of selecting a benchmark result from each of the processing results comprises: Filtering each of the processing results, removing processing results whose deviations from other processing results exceed a preset deviation, and obtaining a plurality of filtered processing results; Determining whether the post-filtering processing results are the same result; If so, any of the filtered results is used as a benchmark result.

3. The model optimization method according to claim 2, characterized in that: After the step of determining whether the post-filtering processing results are the same result, the method further includes: If not, determining the target speech model corresponding to each of the post-filtering processing results; Performing performance evaluation on each of the target speech models according to the task type of the preset task; Determine the optimal speech model whose performance evaluation results meet the preset performance requirements; The processing result of the optimal speech model is selected from each of the post-filtering processing results as a benchmark result.

4. The model optimization method according to claim 1, characterized in that: The step of detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the reference result comprises: Determining result characteristics of each of the processing results and reference characteristics of the reference results; Calculating the Pearson coefficient between each of the result features and the reference features, and constructing a correlation coefficient matrix according to each of the Pearson coefficients; Based on the correlation coefficient matrix, it is detected whether there is a speech model to be optimized.

5. The model optimization method according to claim 4, characterized in that: The step of detecting whether there is a speech model to be optimized based on the correlation coefficient matrix comprises: Determine whether there is a target Pearson coefficient in the correlation coefficient matrix whose absolute value of the Pearson coefficient is lower than a preset threshold; If the target Pearson coefficient exists, it is determined that there is a target speech model to be optimized; Correspondingly, after the step of determining that there is a target speech model to be optimized if the target Pearson coefficient exists, the step further includes: The processing result to be optimized corresponding to the target Pearson coefficient is determined, and the speech model corresponding to the processing result to be optimized is used as the speech model to be optimized.

6. The model optimization method according to claim 5, characterized in that: If present, the step of optimizing the speech model to be optimized according to the optimization processing result of the speech model to be optimized and the benchmark result comprises: If so, determining a target Pearson coefficient between a plurality of target result features of the to-be-optimized processing result of the to-be-optimized speech model and the reference feature of the reference result; Determine a change trend of each of the target Pearson coefficients, wherein the change trend represents a deviation between the target result feature and the reference feature; Determine the result feature to be optimized from each of the target result features according to the change trend; The speech model to be optimized is optimized based on the result features to be optimized.

7. The model optimization method according to claim 6, characterized in that: The step of optimizing the model to be optimized based on the characteristics of the result to be optimized includes: Inputting the result feature to be optimized into a preset optimization model, wherein the preset optimization model is trained based on a plurality of result features and corresponding optimization information, wherein the optimization information is used to reduce the deviation between the result feature and the reference feature; Receiving target optimization information fed back by the preset optimization model based on the characteristics of the result to be optimized; The speech model to be optimized is optimized using the target optimization information.

8. A model optimization device, characterized in that: The device comprises: A result feedback module is used to obtain the processing results of multiple speech models based on the same preset task feedback; A result selection module, used for selecting a benchmark result from each of the processing results; An optimization detection module, used for detecting whether there is a speech model to be optimized according to the deviation between each of the processing results and the reference result; A model optimization module is used to optimize the target speech model according to the optimization processing result of the speech model to be optimized and the benchmark result, if any.

9. A model optimization device, characterized in that: The device comprises: a memory, a processor, and a model optimization program stored in the memory and executable on the processor, wherein the model optimization program is configured to implement the steps of the model optimization method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a model optimization program, and when the model optimization program is executed by the processor, the steps of the model optimization method according to any one of claims 1 to 7 are implemented.

Images