A method and device for determining threshold value of wake-up model

Abstract

The present invention provides a method and apparatus for determining a threshold value for a wakeup model, wherein the method comprises: obtaining at least two types of indicator data for the wakeup model when different candidate thresholds are set for the wakeup model; constructing a fitting curve based on the indicator data corresponding to each candidate threshold; determining a target point corresponding to the gradient of the fitting curve with the smallest distance from a preset gradient; and determining a target threshold based on the candidate threshold corresponding to the target point. Because the fitting curve can reflect how the indicator data changes with changes in the threshold, the gradient of the fitting curve will also change continuously as the threshold changes. Different gradient values ​​can represent different performance of the wakeup model. Therefore, determining the target threshold based on the gradient of the fitting curve has a certain theoretical basis. The target threshold determined by the present invention can ensure that the wakeup model has a higher accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and device for determining a threshold of a wake-up model. Background Art

[0002] With the continuous development of intelligent technology, a variety of smart devices have emerged. The voice recognition function in smart devices enables non-gesture interaction, allowing users to send control commands to smart devices through voice, increasing the convenience of smart device operation. However, before voice interaction can begin, the device must first be awakened from its dormant state to function properly. When a user enters a voice wake-up word command, the voice wake-up model determines the confidence level of the command response. This confidence level is then compared with a pre-set threshold for the wake-up word. If the confidence level exceeds the threshold, the device wakes up; otherwise, no response occurs. To provide timely feedback, smart devices must maintain real-time monitoring. Each speech segment is analyzed by the voice wake-up model and a confidence level is determined. Different thresholds result in different recall rates and false alarm metrics. The choice of the confidence threshold reflects the model's performance and directly impacts the user experience.

[0003] Currently, the threshold selection method is manual selection, which selects the threshold according to different scenarios or users. The manual selection of the threshold method relies on the experience of relevant personnel and has no theoretical basis. It cannot guarantee that the selected threshold can make the model have better performance. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the defect in the prior art that the selected threshold cannot be guaranteed to enable the model to have better performance, thereby providing a threshold determination method and device for a wake-up model.

[0005] A first aspect of the present invention provides a method for determining a threshold of a wake-up model, comprising: obtaining at least two types of indicator data of the wake-up model when setting different candidate thresholds for the wake-up model; constructing a fitting curve based on the indicator data corresponding to each candidate threshold; determining a target point in the fitting curve corresponding to a gradient with the smallest distance from a preset gradient; and determining the candidate threshold corresponding to the target point as the target threshold.

[0006] Optionally, in the threshold determination method of the wake-up model provided by the present invention, the step of constructing a fitting curve based on the indicator data corresponding to each candidate threshold includes: constructing a ROC curve based on the indicator data corresponding to each candidate threshold; and fitting the ROC curve to obtain a fitting curve.

[0007] Optionally, in the threshold determination method of the wake-up model provided by the present invention, when fitting the ROC curve, the fitting function form is: f a，b，c，d (x)=ab*e(-b*(x+c)) , where (a, b, c, d) are the parameters to be fitted, x is the second indicator data, and f is the fitted first indicator data.

[0008] Optionally, in the threshold determination method of the wake-up model provided by the present invention, the gradient corresponding to each second indicator data in the fitting curve is calculated by the following formula: f′=a*b*e (-b*(x+c)) , where x is the second indicator data and f′ is its corresponding gradient.

[0009] The second aspect of the present invention provides a method for determining the threshold of a wake-up model, wherein the wake-up model includes multiple serial sub-models, and the method includes: obtaining various indicator data of each sub-model from top to bottom when setting different candidate thresholds for each sub-model; constructing a fitting curve according to the indicator data corresponding to each candidate threshold of each sub-model, wherein one candidate threshold of each sub-model corresponds to a fitting curve of an adjacent lower-level sub-model, and the number of fitting curves of the first-level sub-model is 1; respectively determining the candidate points corresponding to the gradients with the smallest distance from the preset gradients in each fitting curve of the last-level sub-model; calculating the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determining the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model; and determining the target threshold of the adjacent upper-level sub-model from bottom to top according to the intersection of the fitting curves used when determining the target threshold of each sub-model and the fitting curves of the adjacent upper-level sub-model.

[0010] Optionally, in the threshold determination method of the wake-up model provided by the present invention, the step of determining the target threshold of the adjacent upper-level sub-model based on the intersection of the fitting curve used when determining the target threshold of the sub-model and the fitting curve of the adjacent upper-level sub-model includes: if there are multiple fitting curves for the adjacent upper-level sub-model, determining a second fitting curve that intersects with the first fitting curve, the first fitting curve is the fitting curve used when determining the target threshold of the sub-model, and the second fitting curve is the fitting curve of the adjacent upper-level sub-model; and determining the candidate threshold corresponding to the intersection of the first fitting curve and the second fitting curve in the second fitting curve as the target threshold of the adjacent upper-level sub-model.

[0011] The third aspect of the present invention provides a threshold determination device for a wake-up model, comprising: a first data acquisition module, used to obtain various indicator data of the wake-up model when different candidate thresholds are set for the wake-up model; a first fitting data construction module, used to construct a fitting curve based on the indicator data corresponding to each candidate threshold; a target point determination module, used to determine the target point corresponding to the gradient with the smallest distance from the preset gradient in the fitting curve; and a first target threshold determination module, used to determine the candidate threshold corresponding to the target point as the target threshold.

[0012] According to a fourth aspect of the present invention, a device for determining a threshold of a wake-up model is provided, wherein the wake-up model includes a plurality of serial sub-models, and the device includes: a second data acquisition module for sequentially acquiring, from top to bottom, various types of indicator data of each sub-model when different candidate thresholds are set for each sub-model; a second fitting data construction module for constructing a fitting curve based on the indicator data corresponding to each candidate threshold of each sub-model, wherein one candidate threshold of each sub-model corresponds to one fitting curve of an adjacent lower-level sub-model, and the number of fitting curves of a first-level sub-model is 1; a candidate point determination module for respectively determining the candidate point corresponding to the gradient with the smallest distance from a preset gradient in each fitting curve of the last-level sub-model; a second target threshold determination module for calculating the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determining the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model; and a third target threshold determination module for sequentially determining, from bottom to top, the target threshold of the adjacent upper-level sub-model based on the intersection of the fitting curve used when determining the target threshold of each sub-model and the fitting curve of the adjacent upper-level sub-model.

[0013] The fifth aspect of the present invention provides a computer device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to perform the threshold determination method for the wake-up model provided in the first aspect of the present invention or the second aspect of the present invention.

[0014] A sixth aspect of the present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the threshold determination method of the wake-up model provided in the first aspect of the present invention or the second aspect of the present invention.

[0015] The technical solution of the present invention has the following advantages:

[0016] 1. The present invention provides a method and device for determining a threshold value of a wake-up model. After setting different candidate threshold values ​​for the wake-up model, the index data corresponding to each candidate threshold value is obtained, and a fitting curve is established based on the index data. The candidate threshold value is determined based on the gradient of the fitting curve. Since the fitting curve can reflect the change of the index data as the threshold value changes, the gradient of the fitting curve will also change continuously as the threshold value continues to change. Different gradient values ​​can represent different performance of the wake-up model. Therefore, determining the target threshold value based on the gradient of the fitting curve has a certain theoretical basis. The target threshold value determined by the present invention can make the wake-up model have higher accuracy.

[0017] 2. The present invention provides a method and apparatus for determining a threshold value for a wakeup model. When setting target thresholds for sequential submodels in the wakeup model, different candidate thresholds are set for each submodel from top to bottom. Then, indicator data corresponding to each candidate threshold is obtained and a fitting curve for each submodel is established based on the indicator data. First, candidate points corresponding to the gradients with the smallest distance from a preset gradient are determined in each fitting curve of the final submodel. Then, the target threshold for the final submodel is determined based on the area of ​​the upper left corner of the coordinates corresponding to each candidate point. Because the fitting curve reflects how the indicator data changes with threshold values, the gradient of the fitting curve also changes with changes in the threshold. Different gradient values ​​can represent different performance characteristics of the wakeup model. Furthermore, because the area under the fitting curve measures the quality of the fitting curve, a larger area indicates a better fitting curve. The smaller the area of ​​the corresponding upper left corner, the larger the area under the curve indicates a better curve. By combining the optimal candidate points in the optimal fitting curve, the target threshold for the final submodel is determined, resulting in better performance for the final submodel. Furthermore, considering that the target threshold selection of the upper-level submodel will affect the indicators of the lower-level submodels, in the present invention, after first determining the target threshold of the last-level submodel, a step-by-step tracing method is used to sequentially obtain the target threshold of the upper-level submodel, so that each submodel can achieve better performance. Based on the above analysis, the present invention can achieve the determination of the target threshold of each submodel in a multi-level wake-up model, and by implementing the target threshold of each submodel determined by the present invention, each submodel can achieve better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is a flowchart of a specific example of a method for determining a threshold value of a wake-up model in an embodiment of the present invention;

[0020] Figure 2 Schematic diagram of an ROC curve and a fitting curve constructed with the awakening rate as the vertical axis and the false awakening index as the horizontal axis in an embodiment of the present invention;

[0021] Figure 3 This is a flowchart of another specific example of a method for determining a threshold value of a wake-up model in an embodiment of the present invention;

[0022] Figure 4Schematic diagram of the ROC fitting curve of the two-level model in an embodiment of the present invention;

[0023] Figure 5 This is a principle block diagram of a specific example of a device for determining a threshold value of a wake-up model in an embodiment of the present invention;

[0024] Figure 6 is a principle block diagram of another specific example of a device for determining a threshold value of a wake-up model in an embodiment of the present invention;

[0025] Figure 7 This is a principle block diagram of a specific example of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] In the description of the present invention, it should be noted that the terms "first", "second" and "third" are only used for descriptive purposes and should not be understood as indicating or implying relative importance.

[0028] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] The embodiment of the present invention provides a method for determining a threshold value of a wake-up model, such as Figure 1 As shown, the following steps are included:

[0030] Step S11: obtaining at least two types of indicator data of the wake-up model when different candidate thresholds are set for the wake-up model.

[0031] In an optional embodiment, the test samples are input into the wake-up model to obtain the wake-up results corresponding to each test sample. When different wake-up thresholds are set for the wake-up model, the wake-up results of the same test sample may be different. Therefore, when different wake-up thresholds are set for the wake-up model, the values ​​of the same type of indicator data are also different.

[0032] In an optional embodiment, the wake-up model may be a voice wake-up model.

[0033] In an optional embodiment, the indicator data is indicator data that can form a ROC curve, such as a wake-up rate and a false wake-up indicator.

[0034] In an optional embodiment, the wake-up rate is the ratio of the number of samples correctly recognized by the model to the total number of samples input into the model.

[0035] In an optional embodiment, the wake-up indicator may be the ratio of false wake-up samples to the total number of tests, or the number of false wake-ups within a fixed time.

[0036] In an optional embodiment, the candidate threshold value range is [0, 1].

[0037] Step S12: constructing a fitting curve according to the indicator data corresponding to each candidate threshold value.

[0038] In an optional embodiment, two types of indicator data construct a fitting curve. When only two types of indicator data are obtained in step S11, one fitting curve is constructed. When more than two types of indicator data are obtained in step S11, multiple fitting curves can be constructed.

[0039] In an optional embodiment, when the indicator data is the awakening rate and the false awakening index, a fitting curve is constructed with the awakening rate as the vertical axis and the false awakening index as the horizontal axis as shown in FIG. Figure 2 As shown, the points on the fitting curve are composed of candidate thresholds, awakening rates, and false awakening indicators.

[0040] Step S13: determining a target point corresponding to the gradient with the smallest distance from the preset gradient in the fitting curve.

[0041] In an optional embodiment, the fitting curves constructed by different indicator data correspond to respective preset gradients.

[0042] For example, when the indicator data used to construct the fitting curve are the awakening rate and the false awakening index, generally speaking, the higher the awakening rate, the higher the false awakening index. The point with a gradient value of 1 is in the upper left corner of the fitting curve. The awakening rate and false awakening corresponding to this point are more appropriate. When the value is less than 1, the awakening rate decreases more, while the false awakening index does not decrease much. On the contrary, when the value is greater than 1, although the awakening rate increases, the false awakening index also increases more. Therefore, for the fitting curve constructed by the awakening rate and the false awakening index, its preset gradient setting can be determined to be 1. For fitting curves constructed by other types of indicator data, if the relationship between the indicator data is the same as the relationship between the awakening rate and the false awakening index, and the change trend of the fitting curve still conforms to the above rule, the preset gradient can still be set to 1.

[0043] Step S14: Determine the target threshold value based on the candidate threshold values ​​corresponding to the target point.

[0044] In an optional embodiment, if only two types of indicator data are obtained and only one fitting curve is constructed based on the two types of indicator data, the target threshold is determined based on the candidate threshold corresponding to the target point determined by the fitting curve.

[0045] In an optional embodiment, if multiple categories of indicator data are obtained and two or more fitting curves are constructed based on the multiple categories of indicator data, after determining each target point in each fitting curve, the candidate threshold corresponding to each target point is determined, and the final target threshold is selected from each candidate threshold.

[0046] For a model, the fitting curve constructed by the indicator data expresses its performance, and the value of the indicator data used to evaluate the model needs to be determined according to the threshold. Different thresholds have different corresponding indicator data values. The fitting curve can reflect the changes in the indicator data as the threshold changes. As the threshold changes, the gradient of the fitting curve will also change. Different gradient values ​​can represent different performance of the awakening model. Therefore, the target threshold of the model can be determined by determining the gradient of the curve.

[0047] For example, as the threshold continues to increase, the awakening rate will become lower and lower, and the false awakening index will also be lower. From the perspective of the curve, as the threshold continues to increase, the gradient of the curve also continues to increase. Therefore, the problem can be transformed into determining the gradient of the curve, thereby determining the threshold of the model.

[0048] The threshold determination method for the wake-up model provided in an embodiment of the present invention sets different candidate thresholds for the wake-up model, obtains indicator data corresponding to each candidate threshold, establishes a fitting curve based on the indicator data, and determines the candidate threshold based on the gradient of the fitting curve. Since the fitting curve can reflect the change of indicator data as the threshold changes, the gradient of the fitting curve will also change continuously as the threshold continues to change. Different gradient values ​​can represent different performance of the wake-up model. Therefore, determining the target threshold based on the gradient of the fitting curve has a certain theoretical basis. The target threshold determined by this method can make the wake-up model have higher accuracy.

[0049] In an optional embodiment, in the method for determining the threshold of the wake-up model provided in the embodiment of the present invention, after executing step S11 and before executing step S12, the method further includes:

[0050] The data of each indicator are standardized separately. The data standardization method is adopted to subtract the mean and divide the standard deviation of each indicator data to obtain data under the same dimension. The calculation method is:

[0051]

[0052] Among them, x is the indicator data, mean(.) means finding the mean, and std(.) means finding the standard deviation.

[0053] In an optional embodiment, when the indicator data includes first indicator data and second indicator data, the step of constructing a fitting curve according to the first indicator data and the second indicator data specifically includes:

[0054] First, a ROC curve is constructed based on the first indicator data and the second indicator data corresponding to each candidate threshold.

[0055] Then, the ROC curve is fitted to obtain a fitting curve.

[0056] In the embodiment of the present invention, the constructed ROC curve is fitted by using a function fitting method, the ROC curve is fitted into a continuous curve, and the relevant threshold is determined according to the slope of the curve.

[0057] In an optional embodiment, when fitting the ROC curve, the fitting function form is:

[0058] f a，b，c，d (x)=ab*e (-d(x+c)) ,

[0059] Among them, (a, b, c, d) are the parameters that need to be fitted, x is the second indicator data, and f is the fitted first indicator data.

[0060] The fitting function is fitted according to the provided first indicator data and the false awakening indicator. The fitting method is the least squares method, and the loss function constructed is:

[0061]

[0062] Among them, y is the real first indicator data, f is the fitted first indicator data, and x is the second indicator data.

[0063] In a specific embodiment, any fitting method can be used to fit the ROC curve, and the present invention does not limit the fitting method.

[0064] In an optional embodiment, the gradient corresponding to each false awakening indicator in the fitting curve is calculated by the following formula:

[0065] f′=a*b*e (-b*(x+c)) ,

[0066] Wherein, x is the second indicator data, and f′ is its corresponding gradient.

[0067] The threshold determination method of the wake-up model provided by the embodiment of the present invention uses the first indicator data as the vertical coordinate and the second indicator data as the horizontal coordinate to construct a fitting curve. Among all the gradients, the point closest to the preset gradient is selected, and its corresponding The optimal second indicator data, the candidate threshold on the line is the optimal threshold. Specifically, it is necessary to fit the curve composed of two sets of data into a smooth curve, such as two sets of data x and y. According to the two sets of data, the function f is fitted, and the function value f is obtained by inputting x. x , we can get its gradient calculation formula f′ by fitting the function, input x, and get the gradient f′ x , calculate f′ x The point closest to the preset gradient, the corresponding midpoint of x is the optimal point, the value in x is the second indicator data, and the corresponding value in y is the first indicator data.

[0068] Illustratively, in the above embodiment, the first indicator data is the wake-up rate, and the second indicator data is the false wake-up indicator.

[0069] If the wake-up model is a multi-level model, comprising multiple sub-models, each sub-model focuses on different metrics, resulting in correspondingly different threshold selections. Multi-level models can be divided into parallel and serial models. In a parallel multi-level model, since the confidence and metric data of each sub-model are independent of each other, thresholds can be recommended separately. Specifically, target thresholds for each sub-model can be recommended using the method provided in any of the above embodiments.

[0070] For serial multi-level models, it is necessary to jointly recommend target thresholds. The serial method means that the next level operation is performed after the previous level sub-model is awakened. Therefore, after the previous level target threshold is selected, the next level will be meaningful and it will also obtain corresponding indicator data.

[0071] The embodiment of the present invention provides a threshold determination method for a wake-up model, which is used to determine a target threshold for a wake-up model including at least two serial sub-models, such as Figure 3 As shown, the method includes:

[0072] Step S21: Obtain various indicator data of each sub-model from top to bottom when setting different candidate thresholds for each sub-model.

[0073] Since the target threshold selection of the previous-level sub-model will affect the index data of the next-level sub-model, if different thresholds are selected in the previous-level sub-model, even if two identical thresholds are selected in the next-level sub-model, the values ​​of the corresponding index data will be different. Therefore, in an embodiment of the present invention, different first candidate thresholds are first set for the first-level sub-model, and the values ​​of various types of index data corresponding to each first candidate threshold are determined. Then, corresponding to each first candidate threshold, different second candidate thresholds are set for each second-level sub-module, and the values ​​of various types of index data corresponding to each second candidate threshold are determined. One first candidate threshold corresponds to multiple second candidate thresholds, and so on, until various types of index data of each sub-model are obtained when different candidate thresholds are set for each sub-model.

[0074] Step S22: constructing a fitting curve based on the index data corresponding to each candidate threshold of each sub-model. One candidate threshold of each sub-model corresponds to one fitting curve of the adjacent lower-level sub-model. The number of fitting curves of the first-level sub-model is 1.

[0075] In the embodiment of the present invention, the number of fitting curves of a sub-model is the same as the number of candidate thresholds of an adjacent upper-level sub-model.

[0076] Step S23: determining the candidate points corresponding to the gradients with the smallest distances from the preset gradients in the fitting curves of the last-level sub-model respectively.

[0077] Step S24: Calculate the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determine the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model.

[0078] The area under the fitted curve constructed based on the indicator data corresponding to each candidate threshold value of the sub-model can measure the quality of the fitted curve. The larger the area, the better the fitted curve. The smaller the area of ​​the corresponding upper left corner, the larger the area under the curve, and the better the curve. Therefore, in this embodiment of the present invention, after selecting the optimal candidate point from each of the multiple fitted curves corresponding to the last-level sub-model, the optimal fitting curve is selected based on the area of ​​the upper left corner of each candidate point. The optimal candidate point in the optimal fitting curve is combined to comprehensively obtain the target threshold of the last-level sub-model.

[0079] Step S25: determining the target threshold of the adjacent upper-level sub-model from bottom to top based on the intersection of the fitting curve used when determining the target threshold of each sub-model and the fitting curve of the adjacent upper-level sub-model.

[0080] Since the selection of the target threshold of the upper-level sub-model will affect the indicator data of the lower-level sub-model, after the optimal target threshold is determined in the last-level sub-model, the target threshold of the upper level is determined step by step according to the step-by-step backtracking method, so that the sub-models at all levels can have better performance.

[0081] In an embodiment of the present invention, a method for determining a threshold for a wakeup model is provided. When setting target thresholds for sequential submodels in the wakeup model, different candidate thresholds are set for each submodel from top to bottom. Then, various indicator data corresponding to each candidate threshold is obtained, and fitting curves for each submodel are established based on the indicator data. First, candidate points corresponding to the gradients with the smallest distance to a preset gradient are determined in each fitting curve of the final submodel. Then, the target threshold for the final submodel is determined based on the area of ​​the upper left corner of the coordinates corresponding to each candidate point. Because the fitting curve can reflect how the indicator data changes with changes in the threshold, the gradient of the fitting curve also changes with changes in the threshold. Different gradient values ​​can represent different performance of the wakeup model. Furthermore, because the area under the fitting curve can measure the quality of the fitting curve, the larger the area, the better the fitting curve. The smaller the area of ​​the corresponding upper left corner, the larger the area under the curve, the better the curve. By combining the optimal candidate points in the optimal fitting curve, the target threshold for the final submodel is determined, resulting in better performance for the final submodel. Furthermore, considering that the target threshold selection of the upper-level sub-model will affect the indicators of the lower-level sub-models, in an embodiment of the present invention, after first determining the target threshold of the last-level sub-model, a step-by-step tracing method is used to sequentially obtain the target threshold of the upper-level sub-model, so that each sub-model can have better performance. Based on the above analysis, the method provided by the embodiment of the present invention can achieve the determination of the target threshold of each sub-model in the multi-level wake-up model, and by implementing the target threshold of each sub-model determined by the present invention, each sub-model can have better performance.

[0082] In an optional embodiment, in the above step S25, the step of determining the target threshold of the adjacent upper-level submodel according to the intersection of the fitting curve used when determining the target threshold of the submodel and the fitting curve of the adjacent upper-level submodel includes:

[0083] If there are multiple fitting curves for adjacent upper-level sub-models, perform the following steps:

[0084] First, a second fitting curve intersecting the first fitting curve is determined. The first fitting curve is the fitting curve used when determining the target threshold of the sub-model, and the second fitting curve is the fitting curve of the adjacent upper-level sub-model.

[0085] When there are multiple fitting curves in the upper-level sub-model, if the first fitting curve is extended infinitely, the multiple curves in the upper-level sub-model will intersect with the first fitting curve, thereby obtaining multiple intersection points. In an embodiment of the present invention, the first fitting curve in the upper-level sub-model that intersects with the first fitting curve is used as the second fitting curve.

[0086] Then, a candidate threshold corresponding to the intersection of the first fitting curve and the second fitting curve in the second fitting curve is determined as a target threshold of the adjacent upper-level sub-model.

[0087] In a specific embodiment, if Figure 4 As shown in the figure, it is a ROC fitting curve diagram of the two-level model. The fitting curve is established based on the awakening rate and false awakening index, where the horizontal axis represents the false awakening index and the vertical axis represents the awakening rate. The target threshold selection of the first-level model directly affects the awakening rate and false awakening index of the second-level model. For the second-level model, its target threshold is the candidate threshold corresponding to the point whose gradient is closest to the preset gradient. For different first-level target thresholds, there are different second-level models. Under each first-level target threshold, the area of ​​the upper left corner can be calculated for each point whose ROC curve gradient of the corresponding second-level model is closest to the preset gradient. The candidate threshold corresponding to the point with the smallest area (i.e., point B in the figure) is selected as the target threshold of the second-level model. Figure 4 Point A in is the target of the first-level model.

[0088] The embodiment of the present invention provides a threshold determination device for a wake-up model, such as Figure 5 As shown, including:

[0089] The first data acquisition module 11 is used to obtain various indicator data of the wake-up model when different candidate thresholds are set for the wake-up model. For details, please refer to the description of step S11 in the above embodiment and will not be repeated here.

[0090] The first fitting data construction module 12 is used to construct a fitting curve according to the indicator data corresponding to each candidate threshold and the false awakening indicator. For details, please refer to the description of step S12 in the above embodiment and will not be repeated here.

[0091] The target point determination module 13 is used to determine the target point corresponding to the gradient with the smallest distance from the preset gradient in the fitting curve. For details, please refer to the description of step S13 in the above embodiment and will not be repeated here.

[0092] The first target threshold determination module 14 is used to determine the candidate threshold corresponding to the target point as the target threshold. For details, please refer to the description of step S14 in the above embodiment, which will not be repeated here.

[0093] The embodiment of the present invention provides a threshold determination device for a wake-up model, wherein the wake-up model includes a plurality of serial sub-models, such as Figure 6 As shown, the device includes:

[0094] The second data acquisition module 21 is used to sequentially acquire various indicator data of each sub-model from top to bottom when different candidate thresholds are set for each sub-model. For details, please refer to the description of step S21 in the above embodiment and will not be repeated here.

[0095] The second fitting data construction module 22 is used to construct a fitting curve based on the indicator data corresponding to each candidate threshold of each sub-model. One candidate threshold of each sub-model corresponds to a fitting curve of the adjacent lower-level sub-model. The number of fitting curves of the first-level sub-model is 1. For details, please refer to the description of step S22 in the above embodiment and will not be repeated here.

[0096] The candidate point determination module 23 is used to respectively determine the candidate points corresponding to the gradients with the smallest distances from the preset gradients in the fitting curves of the last level sub-model. For details, please refer to the description of step S23 in the above embodiment and will not be repeated here.

[0097] The second target threshold determination module 24 is used to calculate the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determine the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model. For details, please refer to the description of step S24 in the above embodiment and will not be repeated here.

[0098] The third target threshold determination module 25 is used to determine the target threshold of the adjacent upper-level sub-model from bottom to top based on the intersection of the fitting curve used when determining the target threshold of each sub-model and the fitting curve of the adjacent upper-level sub-model. For details, please refer to the description of step S25 in the above embodiment and will not be repeated here.

[0099] An embodiment of the present invention provides a computer device, such as Figure 7 As shown, the computer device mainly includes one or more processors 31 and a memory 32. Figure 7 A processor 31 is taken as an example.

[0100] The computer device may further include an input device 33 and an output device 34 .

[0101] The processor 31, memory 32, input device 33 and output device 34 may be connected via a bus or other means. Figure 7 The bus connection is taken as an example.

[0102] The processor 31 may be a central processing unit (CPU). The processor 31 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips. The general-purpose processor may be a microprocessor or the processor may be any conventional processor. The memory 32 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; and the data storage area may store data created based on the use of the threshold determination device of the wake-up model. In addition, the memory 32 may include a high-speed random access memory and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 32 may optionally include a memory remotely located relative to the processor 31, and these remote memories may be connected to the threshold determination device of the wake-up model via a network. The input device 33 can receive a calculation request (or other digital or character information) input by the user and generate a key signal input related to the threshold determination device of the wake-up model. The output device 34 can include a display device such as a display screen to output the calculation results.

[0103] An embodiment of the present invention provides a computer-readable storage medium that stores computer instructions. The computer storage medium stores computer-executable instructions that can execute the threshold determination method for the wake-up model in any of the above-mentioned method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); the storage medium can also include a combination of the above-mentioned types of memory.

[0104] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for determining a threshold of a wake-up model, characterized in that: The wake-up model includes a plurality of serial sub-models, and the method includes: Obtain at least two types of indicator data for each sub-model from top to bottom when setting different candidate thresholds for each sub-model; A fitting curve is constructed based on the indicator data corresponding to each candidate threshold of each sub-model. One candidate threshold of each sub-model corresponds to a fitting curve of the adjacent lower-level sub-model. The number of fitting curves of the first-level sub-model is 1. Determine the candidate points corresponding to the gradients with the smallest distance from the preset gradients in each fitting curve of the last level sub-model respectively; Calculate the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determine the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model; From bottom to top, the target threshold of the adjacent upper-level sub-model is determined according to the intersection of the fitting curve used when determining the target threshold of each sub-model and the fitting curve of the adjacent upper-level sub-model.

2. The method for determining the threshold of the wake-up model according to claim 1, characterized in that: The step of determining the target threshold of the adjacent upper-level submodel according to the intersection of the fitting curve used when determining the target threshold of the submodel and the fitting curve of the adjacent upper-level submodel includes: If there are multiple fitting curves in the adjacent upper-level sub-model, Determine a second fitting curve that intersects the first fitting curve, wherein the first fitting curve is a fitting curve used when determining a target threshold of a sub-model, and the second fitting curve is a fitting curve of an adjacent upper-level sub-model; A candidate threshold corresponding to an intersection of the first fitting curve and the second fitting curve in the second fitting curve is determined as a target threshold of the adjacent upper-level sub-model.

3. A device for determining a threshold value of a wake-up model, characterized in that: The wake-up model includes a plurality of serial sub-models, and the apparatus includes: The second data acquisition module is used to sequentially acquire various indicator data of each sub-model from top to bottom when different candidate thresholds are set for each sub-model; The second fitting data construction module is used to construct a fitting curve based on the indicator data corresponding to each candidate threshold of each sub-model. One candidate threshold of each sub-model corresponds to one fitting curve of the adjacent lower-level sub-model. The number of fitting curves of the first-level sub-model is 1. A candidate point determination module is used to determine the candidate point corresponding to the gradient with the smallest distance from the preset gradient in each fitting curve of the last level sub-model; The second target threshold determination module is used to calculate the area of ​​the upper left corner corresponding to each candidate point in the same coordinate axis, and determine the candidate threshold corresponding to the candidate point with the smallest area as the target threshold of the last-level sub-model; The third target threshold determination module is used to determine the target threshold of the adjacent upper-level submodel from bottom to top according to the intersection of the fitting curve used when determining the target threshold of each submodel and the fitting curve of the adjacent upper-level submodel.

4. A computer device, characterized in that: include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to perform the threshold determination method of the wake-up model as claimed in claim 1 or 2.

5. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the threshold determination method for the wake-up model according to claim 1 or 2.

Images