Voice fatigue detection method and device and computer readable storage medium

Abstract

The invention provides a voice fatigue detection method and device and a computer readable storage medium. The method comprises the following steps: acquiring a target voice signal; preprocessing the target voice signal to obtain a corresponding target vector; and inputting the target vector into a preset cyclic convolutional neural network CRNN-Cap based on a capsule network to obtain a prediction probability value of the target voice signal belonging to each fatigue classification. According to the method, the device and the computer readable storage medium, the problem that the spatial relationship of voice features cannot be considered in an existing voice fatigue detection method can be solved.

Description

Technical Field

[0001] The invention relates to a voice fatigue detection method, a device and a computer-readable storage medium. Background Art

[0002] Most of the existing speech fatigue detection methods use convolutional neural networks (CNN). However, convolutional neural networks lose some features of speech signals when processing speech data and cannot consider the spatial relationship of speech features. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a voice fatigue detection method, device and computer-readable storage medium in response to the above-mentioned deficiencies in the prior art, introduce a capsule network to process input in vector form, retain the spatial relationship in speech features, and solve the problem that the spatial relationship of speech features cannot be considered in the existing voice fatigue detection method.

[0004] In a first aspect, the present invention provides a method for detecting speech fatigue, comprising:

[0005] Acquire a target speech signal;

[0006] Preprocessing the target speech signal to obtain a corresponding target vector;

[0007] The target vector is input into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0008] Preferably, the target speech signal is preprocessed to obtain the target

[0009] The amount includes:

[0010] Obtaining the Mel-frequency cepstral coefficients (MFCC) of the target speech signal and its first-order difference and second-order difference;

[0011] A target vector is obtained according to the MFCC and its first-order difference and second-order difference.

[0012] Preferably, the capsule network-based recurrent convolutional neural network CRNN-Cap comprises a convolutional layer, a bidirectional long short-term memory artificial neural network LSTM layer and a capsule layer connected in sequence;

[0013] The convolution layer includes a first convolution layer, a second convolution layer and a pooling layer;

[0014] The output of the first convolutional layer is connected to the input of the pooling layer;

[0015] The output of the pooling layer is connected to the input of the second convolutional layer.

[0016] Preferably, the first convolution layer and the second convolution layer both have a first convolution kernel and a second convolution kernel;

[0017] The activation functions of the first convolutional layer and the second convolutional layer are rectified linear units (ReLUs);

[0018] The pooling layer adopts maximum pooling sampling;

[0019] The activation function of the LSTM layer is the hyperbolic tangent function tanh.

[0020] Preferably, the step of inputting the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification specifically includes:

[0021] Inputting the target vector into a convolutional layer to perform convolution on the target vector;

[0022] The convolution result is input into a bidirectional long short-term memory artificial neural network LSTM, and the forward and reverse outputs of the LSTM are spliced;

[0023] The concatenated result is dynamically routed through the capsule layer to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0024] Preferably, dynamically routing the concatenated result through a capsule layer to obtain a probability of a target speech signal being at a target fatigue level includes:

[0025] Dynamically routing the concatenated result to obtain an output capsule;

[0026] Normalizing the output capsule using a squash function as an activation function of the capsule layer;

[0027] The predicted probability value of the target speech signal belonging to each fatigue level classification is obtained according to the result of the normalization processing.

[0028] In a second aspect, the present invention further provides a voice fatigue detection device, comprising:

[0029] An acquisition module, used for acquiring a target speech signal;

[0030] A preprocessing module, connected to the acquisition module, for preprocessing the target speech signal to obtain a corresponding target vector;

[0031] The probability module is connected to the preprocessing module and is used to input the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0032] Preferably, the preprocessing module specifically includes:

[0033] An acquisition unit, used to acquire the Mel-frequency cepstral coefficients MFCC and the first-order difference and second-order difference of the target speech signal;

[0034] A vector unit is used to obtain a target input vector according to the MFCC and its difference and second-order difference.

[0035] In a third aspect, the present invention further provides a voice fatigue detection device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the voice fatigue detection method described in the first aspect.

[0036] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the voice fatigue detection method as described in the first aspect above is implemented.

[0037] The speech fatigue detection method, device and computer-readable storage medium provided by the present invention first obtain a target speech signal, then preprocess the target speech signal to obtain a corresponding target vector, and then input the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification. Since the present invention introduces a capsule network to process input in the form of vectors, the spatial relationship in the speech features is retained, thereby solving the problem that the spatial relationship of speech features cannot be considered in the existing speech fatigue detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a method for detecting speech fatigue according to Embodiment 1 of the present invention;

[0039] Figure 2 Schematic diagram of a flow chart of another method for detecting voice fatigue according to Embodiment 1 of the present invention;

[0040] Figure 3 This is a structural schematic diagram of a voice fatigue detection device according to Embodiment 2 of the present invention;

[0041] Figure 4 This is a structural schematic diagram of a voice fatigue detection device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0042] In order to enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0044] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0045] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0046] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0047] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0048] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0049] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0050] Embodiment 1:

[0051] This embodiment provides a method for detecting speech fatigue. Figure 1 As shown, the method includes:

[0052] Step S1: Acquire a target speech signal.

[0053] In this embodiment, the target voice signal may be a plurality of voice data, each of which may be 3 seconds long, and the voice data may be processed as follows:

[0054] Each voice data is divided into a short axis with a window function length of 0.025 seconds, and the step length between consecutive windows is set to 0.01 seconds, with a total of 300 windows.

[0055] Step S2: preprocessing the target speech signal to obtain a corresponding target vector.

[0056] In this embodiment, in order to better preserve the spatial features in the speech signal, vector output is adopted, and the output is performed in the form of a vector in the subsequent capsule network.

[0057] Optionally, preprocessing the target speech signal to obtain a target vector includes:

[0058] Obtaining the Mel-frequency cepstral coefficients (MFCC) of the target speech signal and its first-order difference and second-order difference;

[0059] A target vector is obtained according to the MFCC and its first-order difference and second-order difference.

[0060] In this embodiment, the method of obtaining the MFCC of the speech signal and its first-order difference and second-order difference is not specifically limited.

[0061] Step S3: inputting the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0062] In this embodiment, a five-fold cross validation method can be used for the input vector to obtain the accuracy results of five groups of predicted probability values, wherein the loss function required in the cross validation process satisfies the following formula:

[0063] L k =T k max(0,m + ‖-v k ‖) 2 +λ(1-T k )max(0,‖v k ‖-m - ) 2

[0064] Among them, T k is the indicator function of the classification, and its value range is from 0 to 1. In this embodiment, the preferred value is 1. + is the upper boundary, m - is the lower boundary, m + and m - The value range of m is -1 to 1. + The preferred value of is 0.9, m -The preferred value of is -0.1. This embodiment uses λ=0.5. The final total loss value of the network is the sum of the loss values ​​of all classifications. k Represents the output vector. In this embodiment, the preferred values ​​of the parameters for calculating the loss function are the optimal values ​​selected by adjusting the parameters using the grid search method.

[0065] Optionally, the capsule network-based recurrent convolutional neural network CRNN-Cap includes a convolutional layer, a bidirectional long short-term memory artificial neural network LSTM layer and a capsule layer connected in sequence;

[0066] The convolution layer includes a first convolution layer, a second convolution layer and a pooling layer;

[0067] The output of the first convolutional layer is connected to the input of the pooling layer;

[0068] The output of the pooling layer is connected to the input of the second convolutional layer.

[0069] In this embodiment, the convolution layer can extract features from the data for learning. CRNN-Cap retains the convolution characteristics in the convolutional neural network and creates capsules to cover more areas, better extract features from the edge areas, and do not lose the location information in the area.

[0070] Since the pooling layer in the neural network itself has excellent performance in feature dimensionality reduction, data compression, reducing overfitting, and improving model fault tolerance, the appropriate introduction of a pooling layer with a smaller step size and size will not lead to a large loss of feature space information. Therefore, a layer of maximum pooling is added after the first layer of convolutional network.

[0071] Since LSTM has a high ability to process the previous and next information in speech, an LSTM layer is added before the capsule layer. At the same time, if the processing data set is too large, a Dropout layer can be added after the LSTM layer to improve the generalization ability.

[0072] Optionally, the first convolution layer and the second convolution layer both have a first convolution kernel and a second convolution kernel;

[0073] The activation functions of the first convolutional layer and the second convolutional layer are rectified linear units (ReLUs);

[0074] The pooling layer adopts maximum pooling sampling;

[0075] The activation function of the LSTM layer is the hyperbolic tangent function tanh.

[0076] In this embodiment, the model is composed of two convolution kernels, the sizes of which are [5, 3, 3, 32] and [5, 3, 32, 64], respectively, with a step size of 1. Features are extracted in the convolution layer, and the activation function is ReLu, which makes the output of some neurons 0, thus causing the sparsity of the network, reducing the interdependence of parameters, and alleviating the occurrence of overfitting problems. The output size is [batch_size, 300, 64], and the batch_size value is usually smaller than the training set, such as a positive integer between 100 and 200. For the result of the first convolution, maximum pooling is used for sampling, and the pooling window size is [2, 4].

[0077] Optionally, the step of inputting the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification specifically includes:

[0078] Inputting the target vector into a convolutional layer to perform convolution on the target vector;

[0079] The convolution result is input into a bidirectional long short-term memory artificial neural network LSTM, and the forward and reverse outputs of the LSTM are spliced;

[0080] The concatenated result is dynamically routed through the capsule layer to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0081] In this embodiment, after passing through two convolutional layers, the convolution result is expanded according to the time dimension and then input into the bidirectional LSTM layer. The activation function used by the LSTM network is tanh, and the hidden layer size is 16. It will output two sets of 8-dimensional vectors, forward and reverse, and then concatenate the forward and reverse outputs according to the time dimension as the input of the capsule network.

[0082] Optionally, dynamically routing the concatenated result through a capsule layer to obtain a probability of a target speech signal being at a target fatigue level includes:

[0083] Dynamically routing the concatenated result to obtain an output capsule;

[0084] Normalizing the output capsule using a squash function as an activation function of the capsule layer;

[0085] The predicted probability value of the target speech signal belonging to each fatigue level classification is obtained according to the result of the normalization processing.

[0086] In this embodiment, the capsule network is calculated in vector units. The 16-dimensional vector output by LSTM is used as a capsule. For each voice data, it is output as 300 capsules after passing through the LSTM layer. Then the classification result is also represented as a capsule of length 8. Next, the dynamic routing algorithm is used to obtain the output capsule of each voice data after three routings. Then the activation function squash is used to normalize the output capsule to obtain the modulus length of each output capsule. The 3-dimensional vector generated at last is used to represent the probability of classification into three different categories.

[0087] In a specific embodiment, the voice fatigue detection method may include the following steps:

[0088] like Figure 2 As shown in the figure, the model consists of the following five layers: CNN (Convolutional Neural Network, feedforward neural network) convolution layer, LSTM layer, Dropout layer, capsule layer and length layer, among which the CNN convolution layer includes the first convolution layer Conv1 and the second convolution layer Conv2.

[0089] 1) After obtaining the target speech signal, the MFCC of the target speech signal and its first-order and second-order differences are used as input to the first layer of the convolutional network. Since the pooling layer in the neural network itself has excellent performance in feature dimension reduction, data compression, reducing overfitting, and improving model fault tolerance, the appropriate introduction of a pooling layer with a smaller step size and size will not lead to a large amount of feature space information loss. Therefore, a maximum pooling layer is added after the first layer of the convolutional network. The duration of each input speech data is 3 seconds, and the data is divided into short frames with a window function length of 0.025 seconds. The step length between consecutive windows is 0.01 seconds, and there are 300 windows in total.

[0090] 2) Since LSTM has a high ability to process the previous and next information in speech, an LSTM layer is added before the capsule layer, and a Dropout layer can also be added to improve the generalization ability.

[0091] 3) Due to the superiority of the capsule network in processing features and feature spatial position information, the series of vectors output from the Dropout layer are passed to the next capsule layer. The features are extracted more abstractly through the capsule layer and dynamic routing, and a length layer is added to calculate the length of each output capsule vector in the capsule layer, that is, the existence probability of each class after classification.

[0092] The convolutional layer in the convolutional neural network can extract features from the data for learning. The capsule network retains the convolutional characteristics of the convolutional neural network, while creating capsules to cover more areas, which can better extract features from the edge areas without losing the location information within the area.

[0093] The model consists of two convolution kernels, the sizes of which are [5, 3, 3, 32] and [5, 3, 32, 64], respectively, with a step size of 1. In the convolution layer, features are extracted, and the activation function is Relu. Relu will make the output of some neurons 0, which will cause the sparsity of the network and reduce the interdependence of parameters, alleviating the occurrence of overfitting problems. The output size is [batch_size, 300, 64]. The batch_size value is usually smaller than the training set. For example, when the training set size is 300, the batch_size value range can be 100 to 200. For the result of the first convolution, the maximum pooling is used for sampling, and the pooling window size is [2, 4].

[0094] After passing through two layers of CNN convolutional layers, the convolution results are expanded according to the time dimension and then input into the bidirectional LSTM layer. The activation function used by the LSTM network is tanh, and the hidden layer size is 16. The forward and reverse 8-dimensional vector outputs are then concatenated according to the time dimension as the input of the capsule network.

[0095] Unlike ordinary neural networks, capsule networks are calculated in vector units. The 16-dimensional vector output by LSTM is used as a capsule. For each piece of data, it is represented as 300 capsules, and the classification result is also represented as a capsule of length 8. Next, the dynamic routing algorithm is used to get the output of each voice data after three routings, and then the activation function squash is used to normalize the capsule to get the modulus length of each output capsule. The final generated 3-dimensional vector is used to represent the probability of classification into three different categories.

[0096] Since the pooling layer in the neural network itself has excellent performance in feature dimensionality reduction, data compression, reducing overfitting, and improving model fault tolerance, the appropriate introduction of a pooling layer with a smaller step size will not lead to a large loss of feature space information.

[0097] In the input network, 5-fold cross validation was used, and each fold was trained for 5 rounds to obtain 5 sets of accuracy results. The average of these five sets of cross-validation results was taken as the accuracy of the final result.

[0098] The calculation formula of the loss function is shown in the formula:

[0099] L k =T k max(0,m + -‖v k ‖) 2 +λ(1-T k )max(0,‖v k ‖-m- ) 2

[0100] Where T k is the indicator function of the classification, T k =1,m + is the upper boundary, m + The value of m is 0.9, - It is the next session, m - The value of is -0.1, and the λ used in this paper is 0.5, v k The final total loss value of the network is the sum of the loss values ​​of all classes.

[0101] The speech fatigue detection method provided in an embodiment of the present invention first obtains a target speech signal, then preprocesses the target speech signal to obtain a corresponding target vector, and then inputs the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification. Since the present invention introduces a capsule network to process input in the form of a vector, the spatial relationship in the speech features is retained, thereby solving the problem that the spatial relationship of speech features cannot be considered in the existing speech fatigue detection method.

[0102] Embodiment 2:

[0103] like Figure 3 As shown, this embodiment provides a voice fatigue detection device, which is used to execute the above-mentioned voice fatigue detection method, including:

[0104] An acquisition module 11 is used to acquire a target speech signal;

[0105] A preprocessing module 12, connected to the acquisition module 11, for preprocessing the target speech signal to obtain a corresponding target vector;

[0106] The probability module 13 is connected to the preprocessing module 12 and is used to input the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0107] Preferably, the preprocessing module 12 specifically includes:

[0108] An acquisition unit, used to acquire the Mel-frequency cepstral coefficients MFCC and the first-order difference and second-order difference of the target speech signal;

[0109] A vector unit is used to obtain a target input vector according to the MFCC and its difference and second-order difference.

[0110] Preferably, the capsule network-based recurrent convolutional neural network CRNN-Cap comprises a convolutional layer, a bidirectional long short-term memory artificial neural network LSTM layer and a capsule layer connected in sequence;

[0111] The convolution layer includes a first convolution layer, a second convolution layer and a pooling layer;

[0112] The output of the first convolutional layer is connected to the input of the pooling layer;

[0113] The output of the pooling layer is connected to the input of the second convolutional layer.

[0114] Preferably, the first convolution layer and the second convolution layer both have a first convolution kernel and a second convolution kernel;

[0115] The activation functions of the first convolutional layer and the second convolutional layer are rectified linear units (ReLUs);

[0116] The pooling layer adopts maximum pooling sampling;

[0117] The activation function of the LSTM layer is the hyperbolic tangent function tanh.

[0118] Preferably, the probability module 13 specifically includes:

[0119] A convolution unit, configured to input the target vector into a convolution layer to perform convolution on the target vector;

[0120] A concatenation unit, used for inputting the convolution result into a bidirectional long short-term memory artificial neural network LSTM, and concatenating the forward and reverse outputs of the LSTM;

[0121] A routing unit is used to dynamically route the concatenated result through the capsule layer to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

[0122] Preferably, the routing unit specifically includes:

[0123] An output capsule unit, used for dynamically routing the concatenated result to obtain an output capsule;

[0124] A normalization unit, configured to normalize the output capsule using a squash function as an activation function of the capsule layer;

[0125] The probability value unit is used to obtain the predicted probability value of the target speech signal belonging to each fatigue level classification according to the result of the normalization processing.

[0126] Embodiment 3:

[0127] like Figure 4As shown, this embodiment provides a voice fatigue detection device for executing the above-mentioned voice fatigue detection method, including a memory 21 and a processor 22, the memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the voice fatigue detection method in Example 1.

[0128] The memory 21 is connected to the processor 22. The memory 21 may be a flash memory, a read-only memory or other memory. The processor 22 may be a central processing unit or a single-chip microcomputer.

[0129] Embodiment 4:

[0130] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the voice fatigue detection method in the above-mentioned embodiment 1 is implemented.

[0131] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable read only memory), flash memory or other memory technology, CD-ROM (Compact Disc Read-Only Memory), digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer.

[0132] The speech fatigue detection device and computer-readable storage medium provided in Examples 2 to 4 first obtain a target speech signal, then preprocess the target speech signal to obtain a corresponding target vector, and then input the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification. Since the present invention introduces a capsule network to process input in the form of vectors, the spatial relationship in the speech features is retained, thereby solving the problem that the spatial relationship of speech features cannot be considered in the existing speech fatigue detection method.

[0133] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting speech fatigue, characterized in that: include: Acquire a target speech signal; Preprocessing the target speech signal to obtain a corresponding target vector; The target vector is input into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

2. The method for detecting speech fatigue according to claim 1, characterized in that: The preprocessing of the target speech signal to obtain a target vector comprises: Obtaining the Mel-frequency cepstral coefficients (MFCC) of the target speech signal and its first-order difference and second-order difference; A target vector is obtained according to the MFCC and its first-order difference and second-order difference.

3. The method for detecting speech fatigue according to claim 1, characterized in that: The capsule network-based recurrent convolutional neural network CRNN-Cap includes a convolutional layer, a bidirectional long short-term memory artificial neural network LSTM layer and a capsule layer connected in sequence; The convolution layer includes a first convolution layer, a second convolution layer and a pooling layer; The output of the first convolutional layer is connected to the input of the pooling layer; The output of the pooling layer is connected to the input of the second convolutional layer.

4. The method for detecting speech fatigue according to claim 3, characterized in that: The first convolution layer and the second convolution layer both have a first convolution kernel and a second convolution kernel; The activation functions of the first convolutional layer and the second convolutional layer are rectified linear units (ReLUs); The pooling layer adopts maximum pooling sampling; The activation function of the LSTM layer is the hyperbolic tangent function tanh.

5. The method for detecting speech fatigue according to claim 3, characterized in that: The step of inputting the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification specifically includes: Inputting the target vector into a convolutional layer to perform convolution on the target vector; The convolution result is input into a bidirectional long short-term memory artificial neural network LSTM, and the forward and reverse outputs of the LSTM are spliced; The concatenated result is dynamically routed through the capsule layer to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

6. The method for detecting speech fatigue according to claim 5, characterized in that: The dynamically routing the concatenated result through the capsule layer to obtain the probability of the target speech signal being at the target fatigue level includes: Dynamically routing the concatenated result to obtain an output capsule; Normalizing the output capsule using a squash function as an activation function of the capsule layer; The predicted probability value of the target speech signal belonging to each fatigue level classification is obtained according to the result of the normalization processing.

7. A voice fatigue detection device, characterized in that: include: An acquisition module, used for acquiring a target speech signal; A preprocessing module, connected to the acquisition module, for preprocessing the target speech signal to obtain a corresponding target vector; The probability module is connected to the preprocessing module and is used to input the target vector into a preset capsule network-based recurrent convolutional neural network CRNN-Cap to obtain a predicted probability value of the target speech signal belonging to each fatigue classification.

8. The voice fatigue detection device according to claim 7, characterized in that: The preprocessing module specifically includes: An acquisition unit, used to acquire the Mel-frequency cepstral coefficients MFCC and the first-order difference and second-order difference of the target speech signal; A vector unit is used to obtain a target input vector according to the MFCC and its difference and second-order difference.

9. A voice fatigue detection device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the voice fatigue detection method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the voice fatigue detection method according to any one of claims 1 to 6 is implemented.

Images