Audio classification method based on causal hidden Markov model and related equipment

Abstract

The invention is suitable for the technical field of artificial intelligence, and provides an audio classification method and related equipment based on a causal hidden Markov model, and the method comprises the steps: extracting an initial feature set of each piece of audio data in a preset audio data set through Opensmile; screening the initial feature set according to a preset rule to obtain a preset target feature set; training a preset causal hidden Markov model according to the target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; obtaining a to-be-recognized feature set of to-be-recognized audio data through Opensmile, and screening the to-be-recognized feature set according to the preset rule to obtain a to-be-recognized target feature set; and inputting the to-be-identified target feature set into the target causal hidden Markov model to obtain an identification result corresponding to the to-be-identified audio data.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and in particular, relates to an audio classification method and related equipment based on a causal hidden Markov model. Background Art

[0002] In recent years, machine learning and deep learning methods have been increasingly used in Alzheimer's disease detection and early warning. Traditional machine learning methods, such as support vector machines (SVM) and random forests (RF), usually rely on classifying predefined features to achieve good Alzheimer's disease detection and early warning effects. With the advancement of technology, deep learning methods have gradually become a hot topic of research, especially convolutional neural networks (CNN) and recurrent neural networks (RNN), which have shown strong capabilities in processing multimodal data. For example, CNN can process brain imaging data, while RNN is suitable for analyzing time series data, such as acoustic and language features. However, although these methods have significantly improved accuracy, their interpretability is poor, and the training process requires high data volume and computing resources. In addition, in practical applications, deep learning models often require a large amount of labeled data for training, which to some extent limits their popularization and application in Alzheimer's disease detection and early warning.

[0003] Using basic acoustic features to detect and warn Alzheimer's disease is highly interpretable, but the results are often unsatisfactory; while using deep learning methods has better results, its interpretability is poor. Moreover, most current machine learning and deep learning methods do not take into account the impact of disease progression, that is, they do not consider timing factors, resulting in unsatisfactory results for Alzheimer's disease detection. Summary of the invention

[0004] The embodiment of the present application provides an audio classification method based on a causal hidden Markov model, which can solve the problem in the prior art of ignoring timing factors, resulting in unsatisfactory detection results for Alzheimer's disease.

[0005] In a first aspect, an embodiment of the present application provides an audio classification method based on a causal hidden Markov model, the method comprising:

[0006] Extract the initial feature set of each audio data in the preset audio data set through Opensmile;

[0007] Screening the initial feature set according to preset rules to obtain a preset target feature set;

[0008] Training a preset causal hidden Markov model according to the preset target feature set and attribute information corresponding to the target feature set to obtain a target causal hidden Markov model;

[0009] Obtaining a feature set to be identified of the audio data to be identified through Opensmile, and screening the feature set to be identified according to the preset rule to obtain a target feature set to be identified;

[0010] The target feature set to be identified is input into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be identified.

[0011] Optionally, the step of extracting an initial feature set of each audio data from a preset audio data set by using Opensmile includes:

[0012] Extract feature data of each audio data in different dimensions from the preset audio data set through Opensmile;

[0013] Through Python, a spectrogram of each audio data is generated according to the feature data, and the spectrogram is used as the initial feature set.

[0014] Optionally, the step of screening the initial feature set according to a preset rule to obtain a preset target feature set includes:

[0015] According to the t-test method of the preset rule, a feature with a threshold value P<0.01 is selected from the initial feature set as a first target feature to obtain a first target feature set;

[0016] According to the multivariate linear regression method of the preset rule, the features associated with the MMSE score and the age are selected from the first target feature set as the second target features, so as to obtain the second target feature set, wherein the features associated with the MMSE score and the age are features with a correlation degree between the MMSE score and the age greater than or equal to a threshold value;

[0017] According to the simple support vector machine with the preset rules, each second target feature in the second target feature set is classified and reviewed to obtain the review results of each second target feature, and a preset target feature set is selected from the second target feature set based on the review results.

[0018] Optionally, the step of training a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model includes:

[0019] Inputting the preset target feature set as image features into the preset causal hidden Markov model, and inputting the attribute information as auxiliary features into the preset causal hidden Markov model, to obtain a prediction result for the preset target feature set;

[0020] Acquire actual label data corresponding to the preset target feature set according to the attribute information, and calculate the loss function of the preset causal hidden Markov model according to the actual label data and the prediction result;

[0021] The model parameters of the preset causal hidden Markov model are updated according to the loss function until the preset causal hidden Markov model converges to obtain the target causal hidden Markov model.

[0022] Optionally, the preset causal hidden Markov model includes a splicing unit, a priori network and a posterior network, the splicing unit is used to obtain the hidden variables of the priori network at a previous time step and the attribute information corresponding to the target feature set, and splice the attribute information with the hidden variables of the previous time step to obtain a splicing feature, and input the splicing feature into the priori network;

[0023] The prior network is used to process the splicing features through the gated recurrent unit in the prior network to obtain a processing result, and input the processing result into a preset fully connected layer in the prior network to obtain a mean and a logarithmic variance vector;

[0024] The posterior network is used to extract the image features of the preset target feature set through the preset convolution layer in the posterior network, and to perform A t and the attribute information to obtain attribute features, the posterior network is further used to input the image features and the attribute features into a fully connected preset layer in the posterior network for processing to obtain a mean and logarithmic variance vector of the posterior hidden variables;

[0025] The posterior network is also used to input the hidden variables of the previous time step into the classifier in the posterior network to obtain the recognition result at the previous time step.

[0026] Optionally, the preset convolution layer in the posterior network includes a convolution layer consisting of at least three different convolution kernels, and the preset convolution layer is used to extract image features of a preset target feature set by using a time-frequency feature extraction technology.

[0027] Optionally, the loss function includes a reconstruction loss parameter of the image, a KL divergence loss parameter of the image, a reconstruction loss parameter of the data, a KL divergence loss parameter of the data, and a classification loss parameter.

[0028] On the other hand, the present application provides an audio classification device based on a causal hidden Markov model, characterized in that it includes:

[0029] The first acquisition module is used to extract an initial feature set of each audio data in a preset audio data set through Opensmile;

[0030] A feature screening module, used to screen the initial feature set according to preset rules to obtain a preset target feature set;

[0031] A training module, used for training a preset causal hidden Markov model according to the preset target feature set and attribute information corresponding to the target feature set to obtain a target causal hidden Markov model;

[0032] The second acquisition module is used to obtain a feature set to be identified of the audio data to be identified through Opensmile, and filter the feature set to be identified according to the preset rule to obtain a target feature set to be identified;

[0033] The recognition module is used to input the target feature set to be recognized into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be recognized.

[0034] In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the computer program.

[0035] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0036] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0037] Extract the initial feature set of each audio data from the preset audio data set through Opensmile; filter the initial feature set according to the preset rules to obtain the preset target feature set; train the preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain the target causal hidden Markov model; obtain the feature set to be identified of the audio data to be identified through Opensmile, and filter the feature set to be identified according to the preset rules to obtain the target feature set to be identified; input the target feature set to be identified into the target causal hidden Markov model to obtain the recognition result corresponding to the audio data to be identified. A deep learning model (causal hidden Markov model) is introduced to process and analyze acoustic features, and the time series information of the disease progression is combined to achieve efficient auxiliary detection and early warning of Alzheimer's disease. It not only has strong interpretability, so that the model can clearly reveal the cause of the disease change, but also shows a significant improvement in auxiliary detection and early warning performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 It is a flowchart of an audio classification method based on a causal hidden Markov model provided in an embodiment of the present application;

[0040] Figure 2 is a flowchart of an audio classification method based on a causal hidden Markov model provided by another embodiment of the present application;

[0041] Figure 3 is a schematic diagram of a causal hidden Markov model in an audio classification method based on a causal hidden Markov model provided in an embodiment of the present application;

[0042] Figure 4 It is a structural diagram of an audio classification device based on a causal hidden Markov model provided in an embodiment of the present application;

[0043] Figure 5 is a structural block diagram of an electronic device in one embodiment;

[0044] Figure 6 FIG. 4 is a structural block diagram of a computer storage medium in another embodiment. DETAILED DESCRIPTION

[0045] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0046] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0047] It should also be understood that the term “and / or” used in the specification and appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0048] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0049] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0050] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0051] Figure 1 A schematic flow chart of an audio classification method based on a causal hidden Markov model provided by the present application is shown. The embodiment of the present application provides an audio classification method based on a causal hidden Markov model, including:

[0052] S101, extracting an initial feature set of each audio data from a preset audio data set through Opensmile;

[0053] S102, screening the initial feature set according to a preset rule to obtain a preset target feature set;

[0054] S103, training a preset causal hidden Markov model according to the preset target feature set and attribute information corresponding to the target feature set to obtain a target causal hidden Markov model;

[0055] S104, obtaining a feature set to be identified of the audio data to be identified through Opensmile, and screening the feature set to be identified according to the preset rule to obtain a target feature set to be identified;

[0056] S105 , inputting the target feature set to be identified into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be identified.

[0057] In a possible implementation, the step of extracting an initial feature set of each audio data from a preset audio data set by using Opensmile includes:

[0058] Extract feature data of each audio data in different dimensions from the preset audio data set through Opensmile;

[0059] Through Python, a spectrogram of each audio data is generated according to the feature data, and the spectrogram is used as the initial feature set.

[0060] Through Opensmile, the initial feature set of each audio data is extracted from the preset audio data set; the initial feature set is screened according to preset rules to obtain a preset target feature set; the preset causal hidden Markov model is trained according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; the feature set to be identified of the audio data to be identified is obtained through Opensmile, and the feature set to be identified is screened according to the preset rules to obtain the target feature set to be identified; the target feature set to be identified is input into the target causal hidden Markov model to obtain the recognition result corresponding to the audio data to be identified. A deep learning model (causal hidden Markov model) is introduced to process and analyze acoustic features, and the time series information of the disease progression is combined to achieve efficient auxiliary detection and early warning of Alzheimer's disease. It not only has strong interpretability, enabling the model to clearly reveal the causes of changes in the disease, but also shows significant improvements in auxiliary detection and early warning performance.

[0061] In a possible implementation manner, the step of screening the initial feature set according to a preset rule to obtain a preset target feature set includes:

[0062] According to the t-test method of the preset rule, a feature with a threshold value P<0.01 is selected from the initial feature set as a first target feature to obtain a first target feature set;

[0063] According to the multivariate linear regression method of the preset rule, the features associated with the MMSE score and the age are selected from the first target feature set as the second target features, so as to obtain the second target feature set, wherein the features associated with the MMSE score and the age are features with a correlation degree between the MMSE score and the age greater than or equal to a threshold value;

[0064] According to the simple support vector machine with the preset rules, each second target feature in the second target feature set is classified and reviewed to obtain the review results of each second target feature, and a preset target feature set is selected from the second target feature set based on the review results.

[0065] Exemplarily, for the 4368-dimensional feature set (the initial feature set), when performing feature selection, the t-test is first used to select features with a threshold value of P<0.01 that are statistically significant. Subsequently, multiple linear regression is used to select a refined subset of features related to age and Mini Mental State Examination (MMSE) scores. Specifically, a least squares regression model is fitted to each feature, and the coefficients, p values, and 95% confidence intervals of MMSE scores and age are extracted and evaluated according to the least squares regression model to obtain regression results, and then features that are significantly associated with MMSE scores and age are selected, thereby obtaining a refined 15-dimensional feature set (A_features).

[0066] In a possible implementation, the step of training a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model includes:

[0067] Inputting the preset target feature set as image features into the preset causal hidden Markov model, and inputting the attribute information as auxiliary features into the preset causal hidden Markov model, to obtain a prediction result for the preset target feature set;

[0068] Acquire actual label data corresponding to the preset target feature set according to the attribute information, and calculate the loss function of the preset causal hidden Markov model according to the actual label data and the prediction result;

[0069] The model parameters of the preset causal hidden Markov model are updated according to the loss function until the preset causal hidden Markov model converges to obtain the target causal hidden Markov model.

[0070] In a possible implementation, the preset causal hidden Markov model includes a splicing unit, a priori network and a posterior network, the splicing unit is used to obtain the hidden variables of the priori network at a previous time step and the attribute information corresponding to the target feature set, and splice the attribute information with the hidden variables of the previous time step to obtain a splicing feature, and input the splicing feature into the priori network;

[0071] The prior network is used to process the splicing features through the gated recurrent unit in the prior network to obtain a processing result, and input the processing result into a preset fully connected layer in the prior network to obtain a mean and a logarithmic variance vector;

[0072] The posterior network is used to extract image features of a preset target feature set through a preset convolutional layer in the posterior network, and to process the acoustic features and the attribute information obtained by opensmile through a fully connected layer in the posterior network to obtain attribute features. The posterior network is also used to input the image features and the attribute features into the fully connected preset layer in the posterior network to obtain the mean and logarithmic variance vector of the posterior hidden variables;

[0073] The posterior network is also used to input the hidden variables of the previous time step into the classifier in the posterior network to obtain the recognition result in the previous time step. In the prior network, the hidden variables obtained by the posterior network in the previous time step and the personal attribute characteristics of the previous time step are used to predict the hidden variables of the current time step. When calculating the loss function to optimize the model, the KL divergence between s, v of the posterior network and s, v of the prior network is calculated, so that the posterior network can combine the knowledge of the prior network when predicting y.

[0074] For example, Figure 3 As shown, the causal hidden Markov model introduces a set of latent variables, namely s t 、v t and z t , used to capture the observed variable x that drives each time step t t , A t and t The key is that these hidden variables are also affected by an auxiliary variable B t-1 The latent variables are divided into three different parts: s t , represents additional disease-causing factors beyond observable attributes but that can be reflected in the image; v t , representing the components that determine clinical measurements; and z t , represents other concepts that may be related to the lesion area but do not directly belong to the lesion area;

[0075] The causal hidden Markov model includes a priori network and a posterior network. For the priori network, it takes the concatenated features as input, including the encoded personal attributes B t-1 , and the hidden variable (z t-1 、s t-1 、v t-1). These features are then processed by a gated recurrent unit (GRU), followed by two fully connected layers, which output the mean and log-variance vectors, respectively, calculated as follows:

[0076] z t =GRUz(B t ,z t-1 )

[0077] s t =GRUs(B t ,s t-1 )

[0078] v t =GRUv(B t ,v t-1 )

[0079] For the posterior distribution, calculated by the encoder, let x t and A t is the input at time t, and z t-1 、s t-1 and v t-1 is the latent variable at the previous time step. The posterior distribution is calculated as follows:

[0080] f z =ReLU(Linearz(Concatenate(x,B,z t-1 )))

[0081] μ z =Linearμ z (f z )

[0082]

[0083] f s =ReLU(Linearz(Concatenate(x,A,B,s t-1 )))

[0084] μ s =Linearμ s (f s )

[0085]

[0086] f v =ReLU(Linearz(Concatenate(x,A,B,v t-1 )))

[0087] μ v =Linearμ v (fv )

[0088]

[0089] Given an input tensor X whose shape is (C in , H in , W in ) and a convolution kernel K whose shape is (C out , C in , K h , K w ) where C in is the number of input channels, C out is the number of output channels, H in and W in is the height and width of the input, K h and K w are the height and width of the kernel, h and w are the spatial locations of the output, and the output Y after applying the convolution operation is given by:

[0090]

[0091] Each convolution kernel sets 5 convolution layers, and the formula for each convolution layer is as follows:

[0092] X i+1 =Dropout(ReLU(BN(Conv(X i ,W i ))))

[0093] Where Conv(X i ,W i ) indicates that X i Use the convolution kernel weight W i The convolution operation, BN (BatchNormalization) represents batch normalization, ReLU (Rectified Linear Unit) is the rectified linear unit activation function, and Dropout represents the activation of X i Use the discard operation.

[0094] In a possible implementation, it is characterized in that the preset convolution layer in the posterior network includes a convolution layer composed of at least three different convolution kernels, and the preset convolution layer is used to extract image features of a preset target feature set through time-frequency feature extraction technology.

[0095] Exemplarily, image features of a preset target feature set are extracted by time-frequency feature extraction technology, and three different convolution kernels are applied: a 9×1 convolution kernel is used to extract spectral dependency, a 1×11 convolution kernel is used for temporal dependency, and a 3×3 convolution kernel is used to capture spectral-temporal dependency.

[0096] In a possible implementation, the loss function includes an image reconstruction loss parameter, an image KL divergence loss parameter, a data reconstruction loss parameter, a data KL divergence loss parameter, and a classification loss parameter.

[0097] Exemplarily, the loss function includes: image reconstruction loss parameter, image reconstruction loss (RL) is measured by the mean square error between the reconstructed image and the original image; image KL divergence loss parameter, image KL divergence loss quantifies the divergence between the learned potential distribution and the prior distribution; data reconstruction loss parameter, RL of data A is calculated by the mean square error between the reconstructed data A and the original data A; data KL divergence loss parameter, KL divergence loss of data A measures the divergence between the learned distribution of A and its prior distribution. Classification loss parameter, using CrossEntropyLoss to calculate the classification loss to evaluate the performance of the model on the classification task. Wherein RL and KL divergence loss are expressed as:

[0098]

[0099] Among them, x i is the original input, is the output of model reconstruction. μ and log(var) are the mean and log variance learned from the data. prior and log(var prior ) are the prior mean and log-variance.

[0100] In a possible implementation, Figure 2 As shown, an audio classification method based on a causal hidden Markov model includes: selecting a suitable data set, including personal information, extracting basic acoustic features in the data set, performing feature screening, extracting a spectrogram of each audio (equivalent to extracting an initial feature set of each audio data in a preset audio data set through Opensmile), inputting the screened basic acoustic features, spectrogram and personal information into a causal hidden Markov model (equivalent to screening the initial feature set according to a preset rule to obtain a preset target feature set, inputting the preset target feature set as an image feature into the preset causal hidden Markov model, and inputting the attribute information as an auxiliary feature into the preset causal hidden Markov model), training the model and performing classification prediction (equivalent to obtaining a prediction result for the preset target feature set), calculating a loss function, and updating the parameters of the causal hidden Markov model according to the loss function until the model converges (equivalent to updating the model parameters of the preset causal hidden Markov model according to the loss function until the preset causal hidden Markov model converges to obtain the target causal hidden Markov model).

[0101] In a possible implementation, the present application provides an audio classification device based on a causal hidden Markov model, such as Figure 4 As shown, the device comprises:

[0102] The first acquisition module 201 is used to extract an initial feature set of each audio data from a preset audio data set through Opensmile;

[0103] The feature screening module 202 is used to screen the initial feature set according to a preset rule to obtain a preset target feature set;

[0104] A training module 203 is used to train a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model;

[0105] The second acquisition module 204 is used to obtain a feature set to be identified of the audio data to be identified through Opensmile, and filter the feature set to be identified according to the preset rule to obtain a target feature set to be identified;

[0106] The recognition module 205 is used to input the target feature set to be recognized into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be recognized.

[0107] Thirdly, Figure 5 As shown, an embodiment of the present application provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program: extracting an initial feature set of each audio data in a preset audio data set through Opensmile; filtering the initial feature set according to preset rules to obtain a preset target feature set; training a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; obtaining a feature set to be identified of the audio data to be identified through Opensmile, and filtering the feature set to be identified according to the preset rules to obtain a target feature set to be identified; inputting the target feature set to be identified into the target causal hidden Markov model to obtain the recognition result information corresponding to the audio data to be identified, and correcting the global path information to obtain the target path of the target unmanned boat in the target sea area.

[0108] Fourthly, Figure 6As shown, an embodiment of the present application provides a computer-readable storage medium 400, and the computer-readable storage medium stores a computer program 411. When the computer program is executed by a processor, the following steps are implemented: extracting an initial feature set of each audio data in a preset audio data set through Opensmile; filtering the initial feature set according to preset rules to obtain a preset target feature set; training a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; obtaining a feature set to be identified of the audio data to be identified through Opensmile, and filtering the feature set to be identified according to the preset rules to obtain a target feature set to be identified; inputting the target feature set to be identified into the target causal hidden Markov model to obtain the recognition result information corresponding to the audio data to be identified, and correcting the global path information to obtain the target path of the target unmanned boat in the target sea area.

[0109] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0110] In the embodiments provided in the present application, it should be understood that the disclosed devices / network equipment / terminal equipment and methods can be implemented in other ways. For example, the device / network equipment / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0111] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0112] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An audio classification method based on a causal hidden Markov model, characterized in that: The method comprises: Extract the initial feature set of each audio data in the preset audio data set through Opensmile; Screening the initial feature set according to preset rules to obtain a preset target feature set; Training a preset causal hidden Markov model according to the preset target feature set and attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; Obtaining a feature set to be identified of the audio data to be identified through Opensmile, and screening the feature set to be identified according to the preset rule to obtain a target feature set to be identified; The target feature set to be identified is input into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be identified.

2. The audio classification method based on the causal hidden Markov model as claimed in claim 1, characterized in that: The step of extracting the initial feature set of each audio data from the preset audio data set by Opensmile includes: Extract feature data of each audio data in different dimensions from the preset audio data set through Opensmile; Through Python, a spectrogram of each audio data is generated according to the feature data, and the spectrogram is used as the initial feature set.

3. The audio classification method based on causal hidden Markov model as claimed in claim 1, characterized in that: The step of screening the initial feature set according to a preset rule to obtain a preset target feature set includes: According to the t-test method of the preset rule, a feature with a threshold value P<0.01 is selected from the initial feature set as a first target feature to obtain a first target feature set; According to the multivariate linear regression method of the preset rule, the features associated with the MMSE score and the age are selected from the first target feature set as the second target features, so as to obtain the second target feature set, wherein the features associated with the MMSE score and the age are features with a correlation degree between the MMSE score and the age greater than or equal to a threshold value; According to the simple support vector machine with the preset rules, each second target feature in the second target feature set is classified and reviewed to obtain the review results of each second target feature, and a preset target feature set is selected from the second target feature set based on the review results.

4. The audio classification method based on causal hidden Markov model as claimed in claim 1, characterized in that: The step of training a preset causal hidden Markov model according to the preset target feature set and the attribute information corresponding to the target feature set to obtain a target causal hidden Markov model includes: Inputting the preset target feature set as image features into the preset causal hidden Markov model, and inputting the attribute information as auxiliary features into the preset causal hidden Markov model to obtain a prediction result for the preset target feature set; Acquire actual label data corresponding to the preset target feature set according to the attribute information, and calculate the loss function of the preset causal hidden Markov model according to the actual label data and the prediction result; The model parameters of the preset causal hidden Markov model are updated according to the loss function until the preset causal hidden Markov model converges to obtain the target causal hidden Markov model.

5. The audio classification method based on causal hidden Markov model as claimed in claim 1, characterized in that: The preset causal hidden Markov model includes a splicing unit, a priori network and a posterior network, wherein the splicing unit is used to obtain the hidden variables of the priori network at a previous time step and the attribute information corresponding to the target feature set, and splice the attribute information with the hidden variables of the previous time step to obtain a splicing feature, and input the splicing feature into the priori network; The prior network is used to process the splicing features through the gated recurrent unit in the prior network to obtain a processing result, and input the processing result into a preset fully connected layer in the prior network to obtain a mean and a logarithmic variance vector; The posterior network is used to extract the image features of the preset target feature set through the preset convolution layer in the posterior network, and to perform A t and the attribute information to obtain attribute features, the posterior network is further used to input the image features and the attribute features into a fully connected preset layer in the posterior network for processing to obtain a mean and logarithmic variance vector of the posterior hidden variables; The posterior network is also used to input the hidden variables of the previous time step into the classifier in the posterior network to obtain the recognition result at the previous time step.

6. The audio classification method based on causal hidden Markov model as claimed in claim 5, characterized in that: The preset convolution layer in the posterior network includes a convolution layer composed of at least three different convolution kernels, and the preset convolution layer is used to extract image features of a preset target feature set through a time-frequency feature extraction technology.

7. The audio classification method based on causal hidden Markov model as claimed in claim 4, characterized in that: The loss function includes a reconstruction loss parameter of an image, a KL divergence loss parameter of an image, a reconstruction loss parameter of data, a KL divergence loss parameter of data, and a classification loss parameter.

8. An audio classification device based on a causal hidden Markov model, characterized in that: The device comprises: The first acquisition module is used to extract an initial feature set of each audio data in a preset audio data set through Opensmile; A feature screening module, used to screen the initial feature set according to preset rules to obtain a preset target feature set; A training module, used for training a preset causal hidden Markov model according to the preset target feature set and attribute information corresponding to the target feature set to obtain a target causal hidden Markov model; The second acquisition module is used to obtain a feature set to be identified of the audio data to be identified through Opensmile, and filter the feature set to be identified according to the preset rule to obtain a target feature set to be identified; The recognition module is used to input the target feature set to be recognized into the target causal hidden Markov model to obtain a recognition result corresponding to the audio data to be recognized.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images