A method and system for predicting the cognitive ability of the elderly based on voice tasks

Abstract

Embodiments of the present disclosure provide a method and system for predicting the cognitive ability of the elderly based on speech tasks. Applied to the field of data processing technology, it includes: collecting the speech data and cognitive ability data of the elderly and performing data perception to construct the speech cognitive ability feature samples of the elderly; constructing an integrated robust deep random configuration network model in a node increment manner, where the network model is composed of a number of base learners, and each learner contains a number of hidden layers; performing integrated robust prediction on the speech cognitive ability feature samples according to the network model to obtain the integrated prediction result of the network model. In this way, the present invention realizes the cognitive prediction of the elderly by using a cognitive assessment scale and speech data, and constructing an integrated robust deep random configuration network, solving the technical problems that the prior art cannot effectively and accurately diagnose the patient's cognition and the prediction effect is poor due to many influencing factors in the patient's speech data.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and particularly to a method and system for predicting the cognitive ability of the elderly based on speech tasks. Background Art

[0002] Alzheimer's Disease (AD) is a progressive neurological disease that mainly presents as progressive cognitive impairment. It causes the death of neurons and the disruption of connections in the brain, thereby affecting aspects such as memory, thinking, language, and behavior. Patients may experience symptoms such as memory loss, getting lost, inability to recognize relatives, and decline in language ability, ultimately leading to severe cognitive impairment and loss of the ability to live independently.

[0003] Currently, although there are a large number of assessment and diagnosis methods for Alzheimer's disease, they are all affected by factors such as the cultural level and emotional state of the subjects, and have high costs, complex operations, are not suitable as routine screening methods, and some methods have certain invasiveness and cannot adapt to the current diagnosis practice of Alzheimer's disease. With the development of artificial intelligence technology, early warning of Alzheimer's disease based on speech analysis has the advantages of non-invasiveness, convenience, low cost, and objectivity. However, there is a lack of prediction of cognitive ability scores.

[0004] Therefore, there is an urgent need to provide a method and corresponding system for diagnosing and predicting Alzheimer's disease based on speech data and the cognitive data of the elderly. Summary of the Invention

[0005] The present disclosure provides a method and system for predicting the cognitive ability of the elderly based on speech tasks. By using a cognitive assessment scale and speech data, and constructing an integrated robust deep random configuration network to achieve cognitive prediction of the elderly, the technical problems in the prior art that cannot effectively and accurately diagnose the patient's cognition and have poor prediction effects due to many influencing factors in the patient's speech data are solved.

[0006] According to the first aspect of the present disclosure, a method for predicting the cognitive ability of the elderly based on speech tasks is provided, including the following steps:

[0007] Collect the speech data and cognitive ability data of the elderly and perform data perception to construct a speech cognitive ability feature sample of the elderly;

[0008] Construct an integrated robust deep random configuration network model in a node increment manner, where the network model is composed of a plurality of base learners, and each base learner includes a plurality of hidden layers;

[0009] Construct a predictor based on the hidden layer, use the kernel density estimation method and construct a residual probability density function based on the speech cognitive ability feature samples, calculate the sample residual matrix of the predictor, and obtain the weighted matrix of the base learner according to the sample residual matrix;

[0010] Adopt the weighted negative correlation learning method and perform integrated robust prediction based on the weighted matrix to obtain the output result of the network model;

[0011] Update the weight distribution of the output result of the network model, and calculate the integrated prediction result of the network model.

[0012] For the aspects and any possible implementation manners as described above, a further implementation manner is provided. The process of collecting the speech data and cognitive ability data of the elderly and performing data perception to construct the speech cognitive ability feature samples of the elderly is as follows:

[0013] Use the elderly cognitive situation measurement scales for different cognitive abilities to evaluate the cognitive situation of the elderly, and collect the cognitive ability data of the elderly;

[0014] Respectively collect the recording answer results of the elderly for the speech databases of the calculation tasks of 100 - 7, 101 - 7, and 102 - 7, and collect the speech data of the elderly;

[0015] Construct the speech cognitive ability feature samples according to the cognitive ability data and the speech data.

[0016] For the aspects and any possible implementation manners as described above, a further implementation manner is provided. The elderly cognitive situation measurement scales include the Mini - Mental State Examination Scale, the Montreal Cognitive Assessment Scale, and the Wechsler Adult Intelligence Scale;

[0017] The recording answer results include the accuracy rate of answering questions, the average reaction time, the reaction time of the first answer to the question, the number of questions answered within 20 seconds, and the number of questions answered correctly within 20 seconds.

[0018] For the aspects and any possible implementation manners as described above, a further implementation manner is provided. The process of adopting the weighted negative correlation learning method and performing integrated robust prediction based on the weighted matrix to obtain the prediction result of the network model is as follows:

[0019] Construct the weighted output matrix of the robust deep random configuration network and the weighted cognitive ability of the robust deep random configuration network according to the weighted output matrix of the base learner, and construct a loss function based on the weighted output matrix and the weighted cognitive ability;

[0020] Define a regularization objective function, take the output of the loss function as the input of the objective function, and calculate the output weights of the network model after weighted negative correlation learning;

[0021] Calculate the integrated prediction results of each network model based on the output weights of the network model.

[0022] For the aspects and any possible implementation manners described above, a further implementation manner is provided. The process of updating the weight distribution of the output result of the network model and calculating the integrated prediction result of the network model is as follows:

[0023] Update the distribution weights of the samples, and calculate the weighted loss of the model output result according to the distribution weights of the samples;

[0024] Calculate the weights of the robust deep random configuration network according to the weighted loss, and calculate the integrated prediction result of the network model based on the weights of the robust deep random configuration network and the output result of the network model.

[0025] For the aspects and any possible implementation manners described above, a further implementation manner is provided. In the robust deep random configuration network model, if the model does not reach the set number of hidden layers and the corresponding number of hidden layer nodes or does not meet the preset error , then continue to increase the number of hidden layer nodes until the requirements for the number of hidden layers of the model are met.

[0026] For the aspects and any possible implementation manners described above, a further implementation manner is provided. The base learner includes a loss function and a regularization objective function. The loss function is specifically:

[0027] ;

[0028] Among them, represents the weighted output matrix of the robust deep random configuration network , represents the weighted cognitive ability of the robust deep random configuration network , represents the original weights of each robust deep random configuration network; represents the regularization parameter, , B is the output weight of the network model;

[0029] The regularization objective function is specifically:

[0030] .

[0031] According to the above aspects and any possible implementation, a further implementation is provided, wherein the accuracy of the network model ensemble prediction result is analyzed using mean absolute error and root mean square error, specifically:

[0032] ;

[0033] ;

[0034] in, RMSE is the root mean square error, MAE is the mean absolute error, To predict the results, For the real result, N The amount of data.

[0035] According to a second aspect of the present disclosure, there is provided a system for predicting cognitive ability of the elderly based on speech tasks, comprising: an elderly speech cognitive ability data acquisition module, a robust deep random configuration network construction module, and an elderly cognitive ability prediction module;

[0036] The elderly speech cognitive ability data acquisition module is used to collect the elderly speech data and cognitive ability data and perform data perception to construct the elderly speech cognitive ability feature samples;

[0037] The robust deep random configuration network construction module is used to construct an integrated robust deep random configuration network model using a node increment method, wherein the network model is composed of a number of base learners, and the learners include a number of hidden layers;

[0038] The elderly cognitive ability prediction module is used to perform integrated robust prediction on the speech cognitive ability feature samples according to the network model to obtain the network model integrated prediction results.

[0039] According to the above aspects and any possible implementation, there is further provided an implementation, wherein the robust deep random configuration network construction module includes a weighted matrix calculation module and a robust prediction module;

[0040] The weighting matrix calculation module is used to construct a predictor based on the hidden layer, use a kernel density estimation method and construct a residual probability density function based on the speech recognition ability feature sample, calculate and obtain a sample residual matrix of the predictor, and obtain a weighting matrix of the basis learner according to the sample residual matrix;

[0041] The robust prediction module is used to adopt a weighted negative correlation learning method and perform integrated robust prediction based on the weighted matrix to obtain a network model output result.

[0042] Compared with the prior art, the present invention has the following technical effects:

[0043] (1) The present invention relies on the construction of the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), Wechsler Adult Intelligence Scale, etc. for the elderly with different cognitive abilities, which can evaluate the cognitive function of patients, including multiple fields such as attention, memory, language ability, and executive function, and has the advantages of non-invasiveness, low cost, and convenient operation;

[0044] (2) The integrated robust deep random configuration network of the present invention consists of multiple base learners. The base learners adopt an objective function based on the loss function and regularization. The sample weighting function under each base learner is calculated based on Kernel Density Estimation (KDE), and weighted negative correlation learning (WNCL) is used to achieve integrated robust prediction, thereby effectively improving the accuracy and robustness of the model;

[0045] (3) The present invention can collect data through different terminals such as personal computers, home TVs, mobile phones, tablets, etc., analyze and process the data using a cloud server, and give real-time cognitive ability prediction results, achieving great improvements and breakthroughs in the implementation method, perception quality, friendly interaction, and health management.

[0046] It should be understood that the content described in the section of the invention content is not intended to limit the key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] With reference to the accompanying drawings and the following detailed description, the above and other features, advantages, and aspects of the embodiments of the present disclosure will become more obvious. The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. In the drawings, the same or similar reference numerals represent the same or similar elements, where:

[0048] Figure 1 shows a schematic flowchart of a method for predicting the cognitive ability of the elderly based on a voice task according to an embodiment of the present disclosure;

[0049] Figure 2 shows a schematic diagram of an integrated robust deep random configuration network model of a method for predicting the cognitive ability of the elderly based on a voice task according to an embodiment of the present disclosure;

[0050] Figure 3 shows a schematic diagram of the structure of a system for predicting the cognitive ability of the elderly based on a voice task according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts shall fall within the scope of protection of the present disclosure.

[0052] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] The early stage of Alzheimer's disease (AD) is mild cognitive impairment (MCI). Patients in this stage have normal daily living abilities but show progressive cognitive decline. From a treatment perspective, AD is irreversible and poses great difficulties in treatment. However, if patients can be treated at the MCI stage, the onset of dementia can be effectively delayed.

[0054] A random configuration network is a lightweight neural network model based on random weights, with advantages such as fast modeling speed and good generalization performance. However, based on the loss objective function is vulnerable to noise and outliers, and to improve the model robustness, therefore, for complex and diverse speech data, due to problems such as acquisition devices and human factors, the model robustness performance is affected. Therefore, the present invention proposes an integrated robust deep random configuration network for predicting the cognitive ability of the elderly. The integrated robust deep random configuration network consists of multiple base learners. The base learner adopts an objective function based on the loss function and regularization. The sample weighting function under each base learner is calculated based on kernel density estimation (KDE), and weighted negative correlation learning (WNCL) is used to achieve integrated robust prediction.

[0055] Referring to Figure 1 as shown, this embodiment provides a method for predicting the cognitive ability of the elderly based on speech tasks, including the following steps:

[0056] S101. Collect the speech data and cognitive ability data of the elderly and perform data perception to construct the elderly speech cognitive ability feature samples.

[0057] In this embodiment, it is constructed with the support of data such as the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), and Wechsler Adult Intelligence Scale for the elderly with different cognitive abilities, as well as their voice data. The voice data is the recordings of calculation tasks such as 100 - 7, 101 - 7, 102 - 7, etc. continuously performed by the patient.

[0058] Specifically, the cognitive impairment caused by AD will affect the language expression ability, which is then reflected in the process and content of language expression. Therefore, by collecting the voice of the user performing a preset description task (such as collecting voice through the microphone of an electronic device) to obtain voice information, and based on the algorithm of voice analysis to identify and test the user's language expression ability, and then evaluate the degree of cognitive impairment of the user, so as to detect AD and MCI based on the voice information.

[0059] The preset description task may include describing the content of a preset image and naming different target objects of a target type within a preset duration. It can be understood that cognitive impairment will affect the accuracy of describing the preset image. For the content in the image, the more severe the cognitive impairment, the worse the description accuracy. When performing the description task, the display screen of the electronic device can display the preset image. Obviously, the description accuracy of the preset image by normal users and users with cognitive impairment is different; similarly, cognitive impairment will also affect the fluency of description. For example, naming different target objects of a target type within a preset duration can be to say as many animal names as possible within a preset duration (such as 30 seconds, 1 minute, 2 minutes, etc.). Therefore, by collecting the voice information of the user performing the preset description task, it can be used to detect the degree of cognitive impairment, so as to detect the probability that the user has AD and MCI.

[0060] The specific process of obtaining the recording data is as follows: Log in to the system through a mobile phone or tablet, and then enter the answering interfaces of tasks 100 - 7, 101 - 7, 102 - 7 in sequence (the cognitive ability can be objectively evaluated by means of the diversity of answering questions). Use the recording function to answer the calculation results of 100 - 7, 101 - 7, 102 - 7, etc. Taking 100 - 7 as an example, the correct answering sequence is 93, 86, 79, 72, and so on. According to the accuracy rate of the user answering questions, the average response time of answering questions, the response time of the first answer to the question, the number of questions answered within 20 seconds, and the number of questions answered correctly within 20 seconds, the voice data of the elderly is finally obtained.

[0061] In this embodiment, the training data is set as , where represents the input data, , d represents the feature dimension; represents the cognitive ability, , Qualities representing different scales can be used for multi-output prediction with the cognitive results of scales in multiple dimensions; , N represents the number of training samples.

[0062] S102. Construct an integrated robust deep random configuration network model in a node increment manner, where the network model consists of several base learners, and each learner contains several hidden layers.

[0063] As Figure 2 shown, in this embodiment, a robust deep random configuration network is used and modeled in a node increment manner. Assume that it contains n hidden layers, and the n nodes in the th layer have been configured. The current network output can be expressed as , and the training error is defined as :

[0064] If the current model has not reached the set number of hidden layers and the corresponding number of hidden layer nodes , or does not meet the preset error , continue to add the th hidden layer node.

[0065] Randomly generate the weights and biases of multiple hidden layer nodes , and select the result that maximizes as the final parameter, where is calculated as shown in Equation (1):

[0066] (1)

[0067] where represents the output of the hidden layer node , and is calculated based on the randomly generated weights and biases through the activation function and the output of the previous layer network; ; ; .

[0068] S103. Perform integrated robust prediction on the speech cognitive ability feature samples according to the network model to obtain the integrated prediction result of the network model.

[0069] To improve the generalization performance of the model and the ability to handle outliers, use the loss function in the form of Equation (2) norm:

[0070] (2)

[0071] Among them, represents the base model weights, represents the cognitive ability, represents the base model n output of the hidden layer, represents the base model regularization parameter.

[0072] Secondly, to further improve the model performance, this embodiment designs a weighted negative correlation learning strategy based on kernel density estimation for model integration, where the number of predictors is K , and each predictor is represented as k , k = 1, 2, …, K .

[0073] In addition, compared with the residuals of normal samples, the residuals of vertical outliers are always far from the center of the overall residuals and have a low density. Therefore, the probability density of the sample residuals can be calculated through kernel density estimation (KDE) to determine the weighted matrix of each base learner, so that normal data obtains higher weights and noisy data obtains smaller weights.

[0074] The residual probability density function based on kernel density estimation (KDE) is as follows:

[0075] (3)

[0076] Among them, represents the estimated window width; represents the sample residual; represents the kernel function, usually the Gaussian kernel function, N is the number of samples.

[0077] Finally, the residual matrix of each sample of predictor k can be obtained:

[0078] (4)

[0079] Among them, is the sample weighted matrix of the predictor, is the sample residual.

[0080] Subsequently, to implement ensemble learning, with the goal of minimizing the loss function of k predictors, the objective function (5) is defined, where represents the output weight of the ensemble model after weighted negative correlation learning.

[0081] (5)

[0082] Among them, the loss function can be defined as Equation (6):

[0083] (6)

[0084] Among them, represents the weighted output matrix of the robust deep random configuration network ; represents the weighted cognitive ability of the robust deep random configuration network ; represents the original weights of each robust deep random configuration network; represents the regularization parameter; .

[0085] According to the solution of Equation (6), we can obtain , and then the results of each base model can be calculated . To perform model integration, calculate according to Equation (7) the weighted loss of the regression result , and update the distribution weights of each sample according to Equation (8):

[0086] (7)

[0087] (8)

[0088] (9)

[0089] Among them, represents the sample distribution weight, and the initial value is , represents the predictor for the sample predicted value, represents the sample actual cognitive ability value, represents the threshold function, used to update the distribution weight; represents the weight of the k-th robust deep random configuration network;

[0090] According to Equation (10), the final integrated model output result can be obtained:

[0091] (10)

[0092] Such as Figure 3As shown in the figure, this embodiment also provides a prediction system for the cognitive ability of the elderly based on speech tasks, including: an elderly speech cognitive ability data acquisition module 1, a robust deep random configuration network construction module 2, and an elderly cognitive ability prediction module 3;

[0093] The elderly speech cognitive ability data acquisition module 1 is used to collect the elderly speech data and cognitive ability data and perform data perception, and construct the elderly speech cognitive ability feature samples.

[0094] The robust deep random configuration network construction module 2 is used to construct an integrated robust deep random configuration network model in a node increment manner, where the network model is composed of several base learners, and each learner contains several hidden layers.

[0095] The robust deep random configuration network construction module 2 includes a weighted matrix calculation module 21 and a robust prediction module 22;

[0096] The weighted matrix calculation module 21 is used to construct a predictor based on the hidden layer, use the kernel density estimation method and construct a residual probability density function based on the speech cognitive ability feature samples, calculate the sample residual matrix of the predictor, and obtain the weighted matrix of the base learner according to the sample residual matrix;

[0097] The robust prediction module 22 is used to perform integrated robust prediction based on the weighted matrix by using the weighted negative correlation learning method, and obtain the output result of the network model.

[0098] The elderly cognitive ability prediction module 3 is used to perform integrated robust prediction on the speech cognitive ability feature samples according to the network model, and obtain the integrated prediction result of the network model.

[0099] Specifically, in this embodiment, data of the elderly with different cognitive abilities are collected based on a certain tertiary first-class hospital, the data of the Montreal Cognitive Assessment (MoCA) scale are recorded, and the speech data of the 100-7 task are collected based on speech. Through feature extraction, using the method of this invention, the prediction result shows that the cognitive ability prediction error is: the mean absolute error (MAE) is 3.06, and the root mean square error (RMSE) is 3.92, achieving a good prediction result.

[0100] During the implementation of the case, the user uses a platform such as a mobile phone to enter syllables according to the guidance. After the entered syllables are preprocessed, the probability of Alzheimer's disease under a single syllable and the probability of Alzheimer's disease under multiple syllables can be obtained through the predictor module and the multi-syllable fusion diagnosis module, and guiding opinions such as precautions can be put forward according to the probability.

[0101] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present disclosure is not limited by the described action sequence, because according to the present disclosure, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present disclosure.

[0102] It should be understood that various forms of processes shown above can be used, reordering, adding or deleting steps. For example, the steps described in the present disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and no limitation is imposed herein.

[0103] The above specific implementation manners do not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub - combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. A method for predicting cognitive ability of the elderly based on speech tasks, characterized by: The following steps are involved: Collect voice data and cognitive ability data of the elderly and perform data perception to construct characteristic samples of the elderly's voice and cognitive ability; An integrated robust deep random configuration network model is constructed using a node increment method, wherein the network model is composed of a plurality of base learners, each of which includes a plurality of hidden layers; Constructing a predictor based on the hidden layer, constructing a residual probability density function based on the speech recognition ability feature sample using a kernel density estimation method, calculating a sample residual matrix of the predictor, and obtaining a weighting matrix of the base learner according to the sample residual matrix; The weighted negative correlation learning method is used and integrated robust prediction is performed based on the weighted matrix to obtain the network model output results, including: Constructing a weighted output matrix of the robust deep random configuration network and a weighted cognitive ability of the robust deep random configuration network according to the weighted output matrix of the base learner, and constructing a loss function based on the weighted output matrix and the weighted cognitive ability; A regularized objective function is defined, the output of the loss function is used as the input of the objective function, and the output weight of the network model after weighted negative correlation learning is calculated; Obtaining integrated prediction results of each of the network models based on the output weights of the network models; The weight distribution of the network model output result is updated, and the network model integrated prediction result is calculated.

2. The method for predicting cognitive ability of the elderly based on speech tasks according to claim 1 is characterized in that: The process of collecting the elderly's voice data and cognitive ability data and performing data perception to construct the elderly's voice cognitive ability feature samples is as follows: The cognitive status of the elderly was assessed using a cognitive status measurement table for the elderly with different cognitive abilities, and the cognitive ability data of the elderly were collected; The recording responses of the elderly to the 100-7, 101-7 and 102-7 computing task voice databases were collected to obtain the elderly voice data; A speech cognitive ability feature sample is constructed based on the cognitive ability data and the speech data.

3. The method for predicting cognitive ability of the elderly based on speech tasks according to claim 2 is characterized in that: The cognitive status measurement scales for the elderly include the Mini-Mental State Examination, the Montreal Cognitive Assessment, and the Wechsler Adult Intelligence Scale; The recorded answer results include the accuracy of answering questions, average reaction time, reaction time for first answering questions, number of questions answered within 20 seconds, and number of questions answered correctly within 20 seconds.

4. The method for predicting cognitive ability of the elderly based on speech tasks according to claim 1 is characterized in that: The process of updating the weight distribution of the network model output result and calculating the network model integrated prediction result is as follows: Updating the distribution weights of the samples, and calculating the weighted loss of the model output results according to the distribution weights of the samples; The robust deep random configuration network weights are calculated according to the weighted loss, and the network model integration prediction results are calculated based on the robust deep random configuration network weights and the network model output results.

5. The method for predicting cognitive ability of the elderly based on speech tasks according to claim 1 is characterized in that: In the robust deep random configuration network model, if the model does not reach the set number of hidden layers and the corresponding number of hidden layer nodes or does not meet the preset error , then continue to add hidden layer nodes until the number of hidden layer nodes required by the model is met.

6. A system for predicting cognitive ability of the elderly based on speech tasks, implemented using the method for predicting cognitive ability of the elderly based on speech tasks according to any one of claims 1 to 5, characterized in that: include: Elderly people's speech cognitive ability data acquisition module (1), robust deep random configuration network construction module (2) and elderly people's cognitive ability prediction module (3); The elderly speech cognitive ability data acquisition module (1) is used to collect the elderly speech data and cognitive ability data and perform data perception to construct the elderly speech cognitive ability feature samples; The robust deep random configuration network construction module (2) is used to construct an integrated robust deep random configuration network model in a node increment manner, wherein the network model is composed of a plurality of base learners, and the learners include a plurality of hidden layers; The elderly cognitive ability prediction module (3) is used to perform integrated robust prediction on the speech cognitive ability feature samples according to the network model to obtain the network model integrated prediction result.

7. The system for predicting cognitive ability of the elderly based on speech tasks according to claim 6 is characterized in that: The robust deep random configuration network construction module (2) includes a weighted matrix calculation module (21) and a robust prediction module (22); The weighting matrix calculation module (21) is used to construct a predictor based on the hidden layer, adopt a kernel density estimation method and construct a residual probability density function based on the speech recognition ability feature sample, calculate and obtain a sample residual matrix of the predictor, and obtain a weighting matrix of the base learner according to the sample residual matrix; The robust prediction module (22) is used to adopt a weighted negative correlation learning method and perform integrated robust prediction based on the weighted matrix to obtain a network model output result.

Images