A deep speech feature construction and mental health screening method

Abstract

The present invention proposes a method for constructing deep speech features and screening mental health, including obtaining speech data and preprocessing the speech data; performing fast Fourier transform on the preprocessed speech data to obtain a power spectrum, applying a Mel filter group to the power spectrum and taking the logarithm to obtain a Mel spectrogram; for each frame of the logarithmic Mel spectrogram, calculating its first-order and second-order differences; constructing a three-channel image based on the obtained logarithmic Mel spectrogram and its first-order and second-order differences; processing the three-channel image to extract deep temporal features and deep spatial features; splicing the deep temporal features with the deep spatial features and using a fully connected layer for dimensionality reduction to obtain a fused speech feature; for the fused speech feature, using a prediction network containing a multi-task learning strategy to screen mental health and predict the severity. The method proposed by the present invention can be applied to psychological and mental health monitoring to achieve early detection, early intervention, and early treatment.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and pattern recognition in new generation information technology. Background Art

[0002] Depression is an emotional disorder. People with depression find it difficult to feel happy about things in their lives. This serious psychological disorder has become the fourth largest disease in the world. At present, the diagnosis process of depression mainly relies on professional physicians' interviews and patients' self-assessment scales. This process is long and highly dependent on the physician's clinical experience. The results of the study show that unlike healthy people, patients with depression tend to speak in a monotonous tone, with a low voice, slow response, and sometimes even stutter. At the same time, voice samples have outstanding advantages such as easy acquisition, low cost, and few restrictions.

[0003] Recently, with the continuous breakthroughs in big data technology, especially deep learning technology, a technical foundation has been provided for auxiliary diagnosis of depression. Automatically diagnosing mental health conditions by detecting acoustic changes in sound can not only achieve a more objective and accurate diagnosis of depression, but also promote early detection, early treatment, and early intervention of depression. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the purpose of the present invention is to propose a method for constructing deep speech features and screening mental health, which is used to automatically diagnose mental health conditions by detecting acoustic changes in sound.

[0006] To achieve the above objectives, the first embodiment of the present invention proposes a method for constructing deep speech features and screening mental health, including:

[0007] Acquire voice data, and pre-process the voice data;

[0008] Performing a fast Fourier transform on the preprocessed speech data to obtain a power spectrum, applying a Mel filter bank to the power spectrum and taking the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculating its first-order and second-order differences;

[0009] A three-channel image is constructed based on the obtained logarithmic Mel-spectrogram and its first-order and second-order differences;

[0010] Processing the three-channel image to extract deep temporal features and deep spatial features; splicing the deep temporal features with the deep spatial features and using a fully connected layer to perform dimensionality reduction, thereby obtaining fused speech features;

[0011] For the fused speech features, a prediction network including a multi-task learning strategy is used to screen mental health and predict its severity.

[0012] In addition, the method for constructing deep speech features and screening mental health according to the above embodiment of the present invention may also have the following additional technical features:

[0013] Furthermore, in one embodiment of the present invention, the preprocessing of the voice data includes:

[0014] The speech data is pre-emphasized, framed and windowed.

[0015] Further, in one embodiment of the present invention, the first-order and second-order differences of the logarithmic Mel-spectrogram of each frame are calculated, including:

[0016] Define the first-order difference at time t as d t , then the calculation formula is:

[0017]

[0018] Among them, c t+n and c t-n Represent the Mel frequency cepstrum coefficients at time t+n and tn respectively, for d t Using the above formula again we can get the second order difference.

[0019] Further, in one embodiment of the present invention, the processing of the three-channel image to extract the depth temporal features and the depth spatial features includes:

[0020] The cross-stacked long short-term memory network and the self-attention module are used to extract its deep temporal features. Specifically, the core component of the long short-term memory network is the gate. The forget gate, input gate, output gate, unit state and hidden state defined at time step t are f t 、i t , o t 、c t and h t , then the equation of the long short-term memory network can be expressed as:

[0021]

[0022]

[0023]

[0024] c t =f t ×c t-1 +i t×tanh(W c x t +R c h t-1 +b c ),

[0025] h t =o t ×tanh(c t ),

[0026] Where σ is the Sigmoid activation function, [W f ,R f ]、[W i ,R i ]、[W o ,R o ]、[W c ,R c ] and [b f ,b i ,b o ,b c ] are the weight matrices and biases of the forget gate, input gate, output gate and unit state, respectively, x t is the input at time t, h t-1 is the hidden state at the moment before moment t, and tanh is the hyperbolic tangent activation function; after obtaining the output of the long short-term memory network, it is fed to the self-attention module;

[0027] Define the input sequence of the self-attention module as X = [x1, x2, ..., x n ], respectively using the modulation matrix W q , W k and W v To process it:

[0028] Q=W q X,

[0029] K=W k X,

[0030] V=W v X,

[0031] Thus we can get the query matrix Q = [q1, q2, ..., q N ], key matrix K = [k1, k2, ..., k N ] and value moment V=[v1,v2,...,v N ], and then calculate each query vector q n The attention function output is:

[0032]

[0033] Therefore, through network learning, higher weights are given to speech features that are more relevant to depression, so as to increase the network's attention and obtain deep time series features.

[0034] Furthermore, in one embodiment of the present invention, the processing of the three-channel image to extract the depth temporal features and the depth spatial features further includes:

[0035] A convolutional network is selected and a compression-excitation module is introduced to extract its deep spatial features. Specifically, the dimension of the original feature map is defined as H×W×C, where H represents height, W represents width, and C represents the number of channels. H and W are compressed into one dimension through compression operation, that is, a vector of size 1×1×C is formed; the importance of each channel is predicted through the fully connected layer in the excitation part, and the importance of different channels is obtained and then applied to the corresponding channels of the feature map before compression to adjust the feature contribution of each channel, and finally the weighted deep spatial features are obtained.

[0036] Furthermore, in one embodiment of the present invention, before using the prediction network including a multi-task learning strategy to screen mental health and predict the severity of the fused speech features, the method further includes:

[0037] Using a convolutional layer and a dropout layer to compress the three-channel image, removing information features that are irrelevant or redundant to gender information, and then using an attention module and a fully connected network to obtain information related to gender in the input information;

[0038] The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

[0039] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a deep speech feature construction and mental health screening device, including the following modules:

[0040] An acquisition module, used for acquiring voice data and preprocessing the voice data;

[0041] A transformation module is used to perform a fast Fourier transform on the preprocessed speech data to obtain a power spectrum, apply a Mel filter bank to the power spectrum and take the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculate its first-order and second-order differences;

[0042] A construction module for constructing a three-channel image based on the obtained log-mel spectrogram and its first-order and second-order differences;

[0043] An extraction module, used for processing the three-channel image to extract deep temporal features and deep spatial features; concatenating the deep temporal features with the deep spatial features and performing dimensionality reduction using a fully connected layer, thereby obtaining fused speech features;

[0044] The prediction module is used to screen the mental health and predict the severity of the mental health using a prediction network including a multi-task learning strategy for the fused speech features.

[0045] Furthermore, in one embodiment of the present invention, the prediction module is further used for:

[0046] Using a convolutional layer and a dropout layer to compress the three-channel image, removing information features that are irrelevant or redundant to gender information, and then using an attention module and a fully connected network to obtain information related to gender in the input information;

[0047] The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

[0048] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, it implements the construction of a deep speech feature and a mental health screening method as described above.

[0049] To achieve the above-mentioned purpose, the fourth aspect of the present invention proposes a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the above-mentioned construction of a deep speech feature and a mental health screening method.

[0050] The construction of deep speech features and the mental health screening method proposed in the embodiment of the present invention adopt the following technical solutions: using three-channel representation to encode the Mel spectrogram extracted from speech and its first-order and second-order differential encoding as a whole; combining long short-term memory network and self-attention mechanism to extract deep temporal features; combining convolutional neural network and compression-excitation module to extract deep spatial features and fuse them with deep temporal features; extracting weighted gender loss through gender auxiliary network and introducing it into the overall network loss to constrain network convergence; screening mental health and predicting severity. The recognition method proposed in the present invention can be applied to psychological and mental health monitoring to achieve early detection, early intervention, and early treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0052] Figure 1 A flowchart of a method for constructing deep speech features and screening mental health provided by an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of the process of constructing a deep speech feature and a mental health screening device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0055] The construction of deep speech features and the mental health screening method of an embodiment of the present invention are described below with reference to the accompanying drawings.

[0056] Figure 1 A flowchart of a method for constructing deep speech features and screening mental health provided in an embodiment of the present invention.

[0057] like Figure 1 As shown, the construction of the deep speech feature and the mental health screening method include the following steps:

[0058] S101: Acquire voice data and pre-process the voice data;

[0059] S102: Perform fast Fourier transform on the preprocessed speech data to obtain a power spectrum, apply a Mel filter bank to the power spectrum and take the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculate its first-order and second-order differences;

[0060] S103: constructing a three-channel image according to the obtained logarithmic Mel-spectrogram and its first-order and second-order differences;

[0061] S104: Process the three-channel image to extract deep temporal features and deep spatial features; concatenate the deep temporal features with the deep spatial features and use a fully connected layer to perform dimensionality reduction, thereby obtaining fused speech features;

[0062] S105: Using the fused speech features, a prediction network including a multi-task learning strategy is used to screen mental health and predict its severity.

[0063] Furthermore, in one embodiment of the present invention, preprocessing the voice data includes:

[0064] Pre-emphasize, frame and window the voice data.

[0065] Further, in one embodiment of the present invention, for the logarithmic Mel-spectrogram of each frame, calculating its first-order and second-order differences includes:

[0066] Define the first-order difference at time t as d t , then the calculation formula is:

[0067]

[0068] Among them, c t+n and c t-n Represent the Mel frequency cepstrum coefficients at time t+n and tn respectively, for d t Using the above formula again we can get the second order difference.

[0069] Furthermore, in one embodiment of the present invention, the three-channel image is processed to extract the depth temporal features and the depth spatial features, including:

[0070] The cross-stacked long short-term memory network and self-attention module are used to extract its deep temporal features. Specifically, the core component of the long short-term memory network is the gate. The forget gate, input gate, output gate, unit state and hidden state defined at time step t are f t 、i t , o t 、c t and h t , then the equation of the long short-term memory network can be expressed as:

[0071]

[0072]

[0073]

[0074] c t =f t ×c t-1 +i t ×tanh(W c x t +R c h t-1 +b c ),

[0075] h t =o t ×tanh(c t ),

[0076] Where σ is the Sigmoid activation function, [W f ,R f ]、[W i ,Ri ]、[W o ,R o ]、[W c ,R c ] and [b f ,b i ,b o ,b c ] are the weight matrices and biases of the forget gate, input gate, output gate and unit state, respectively, x t is the input at time t, h t-1 is the hidden state at the moment before time t, and tanh is the hyperbolic tangent activation function; after obtaining the output of the long short-term memory network, it is fed into the self-attention module;

[0077] Define the input sequence of the self-attention module as X = [x1, x2, ..., x n ], respectively using the modulation matrix W q , W k and W v To process it:

[0078] Q=W q X,

[0079] K=W k X,

[0080] V=W v X,

[0081] Thus we can get the query matrix Q = [q1, q2, ..., q N ], key matrix K = [k1, k2, ..., k N ] and value moment V=[v1,v2,...,v N ], and then calculate each query vector q n The attention function output is:

[0082]

[0083] Therefore, through network learning, higher weights are given to speech features that are more relevant to depression, so as to increase the network's attention and obtain deep time series features.

[0084] Furthermore, in one embodiment of the present invention, before screening mental health and predicting the severity of mental health using a prediction network including a multi-task learning strategy for fused speech features, the method further includes:

[0085] The convolutional layer and the dropout layer are used to compress the three-channel image, and after removing the irrelevant or redundant information features of gender information, the attention module and the fully connected network are used to obtain the information related to gender in the input information;

[0086] The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

[0087] The present invention selects TE2E loss to be responsible for gender judgment, and constrains the convergence of the network by introducing weighted gender loss into the calculation of the overall network loss. The TE2E loss is divided into two stages during training: registration and verification. Each step of training data contains input x j～ and M registration sessions x km (m=1,…,M), can be expressed as {x j～ ,(x k1 ,x k2 ,...,x kM )}, where j and k represent the gender judgment object, which can be expressed as {e j～ ,(e k1 ,e k2 ,...,e kM )}, then find the centroid c k :

[0088]

[0089] The cosine similarity function is then used to find the distance between a specific gender and the centroid of that gender, where the weight w and bias b are learnable parameters:

[0090] S = w·cos(e j～ ,c k )+b,

[0091] The final TE2E loss is defined as:

[0092] L TE2E (e j～ ,c k )=δ(j,k)σ(S)+(1-δ(j,k))(1-σ(S)),

[0093] Where σ is the Sigmoid activation function, and the δ(j,k) function is as follows:

[0094]

[0095] For the acquired fusion speech features, a prediction network with a multi-task learning strategy is used for output. The prediction network includes two branches, which respectively include a fully connected layer, a pooling layer, and a softmax activation function. The hard parameter sharing method is used to reduce the risk of overfitting. During the training process of multi-task learning, the main task and the auxiliary task need to be optimized at the same time. The joint loss is defined as:

[0096]

[0097] Where L m and Denote the losses of the main task and auxiliary task, α and β respectively n Represents the weight coefficient of the loss function. During the training process of the model, cross entropy loss and mean square error loss are selected as the loss functions of the main task and auxiliary task respectively, and the final output is the depression tendency and severity score predicted by the model.

[0098] The construction of deep speech features and the mental health screening method proposed in the embodiment of the present invention adopt the following technical solutions: using three-channel representation to encode the mel spectrogram extracted from the speech and its first-order and second-order differential encoding as a whole; combining long short-term memory network and self-attention mechanism to extract deep temporal features; combining convolutional neural network and compression-excitation module to extract deep spatial features and fuse them with deep temporal features; extracting weighted gender loss through gender auxiliary network and introducing it into the overall network loss to constrain network convergence; screening mental health and predicting severity. The recognition method proposed in the present invention can be applied to psychological and mental health monitoring to achieve early detection, early intervention and early treatment.

[0099] Compared with the prior art, the advantages of the present invention are:

[0100] 1) In order to capture the gender information carried in the speech information, a gender network based on short-term speech features is constructed, which effectively promotes the convergence of loss and the generalization of the model.

[0101] 2) Long short-term memory networks and attention mechanisms are introduced into depression detection algorithms, which effectively solves the long-range dependency problem and compresses the feature dimension, prompting the deep network to learn in a targeted manner.

[0102] 3) A multi-task learning framework is proposed, which not only considers the identification of depressive tendencies but also predicts their severity, which helps to improve the performance of the model.

[0103] In order to realize the above-mentioned embodiments, the present invention also proposes the construction of deep speech features and a mental health screening device.

[0104] Figure 2 A schematic diagram of the structure of a deep speech feature construction and mental health screening device provided in an embodiment of the present invention.

[0105] like Figure 2 As shown, the deep speech feature construction and mental health screening device includes: an acquisition module 100, a transformation module 200, a construction module 300, an extraction module 400, and a prediction module 500, wherein:

[0106] An acquisition module is used to acquire voice data and pre-process the voice data;

[0107] A transformation module is used to perform fast Fourier transform on the preprocessed speech data to obtain a power spectrum, apply a Mel filter bank to the power spectrum and take the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculate its first-order and second-order differences;

[0108] A construction module for constructing a three-channel image based on the obtained log-mel spectrogram and its first-order and second-order differences;

[0109] An extraction module is used to process the three-channel image to extract deep temporal features and deep spatial features; the deep temporal features are concatenated with the deep spatial features and the dimension is reduced using a fully connected layer, thereby obtaining fused speech features;

[0110] The prediction module is used to screen and predict the severity of mental health by integrating speech features and using a prediction network that includes a multi-task learning strategy.

[0111] Furthermore, in one embodiment of the present invention, the prediction module is also used to:

[0112] The convolutional layer and the dropout layer are used to compress the three-channel image, and after removing the irrelevant or redundant information features of gender information, the attention module and the fully connected network are used to obtain the information related to gender in the input information;

[0113] The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

[0114] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, it implements the construction of deep speech features and mental health screening methods as described above.

[0115] To achieve the above-mentioned objectives, the fourth aspect of the present invention proposes a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the construction of deep speech features and the mental health screening method as described above are implemented.

[0116] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0117] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0118] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for constructing deep speech features and screening mental health, characterized in that: The following steps are involved: Acquire voice data, and pre-process the voice data; Performing a fast Fourier transform on the preprocessed speech data to obtain a power spectrum, applying a Mel filter bank to the power spectrum and taking the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculating its first-order and second-order differences; A three-channel image is constructed based on the obtained logarithmic Mel-spectrogram and its first-order and second-order differences; Processing the three-channel image to extract depth temporal features and depth spatial features; The deep temporal features are concatenated with the deep spatial features and dimension reduction is performed using a fully connected layer, thereby obtaining a fused speech feature; Using a prediction network including a multi-task learning strategy to screen mental health and predict the severity of the fused speech features; Wherein, the processing of the three-channel image to extract the depth temporal features and the depth spatial features further includes: A convolutional network is selected and a compression-excitation module is introduced to extract its deep spatial features. Specifically, the dimension of the original feature map is defined as H×W×C, where H represents height, W represents width, and C represents the number of channels. H and W are compressed into one dimension through compression operation, that is, a vector of size 1×1×C is formed; the importance of each channel is predicted through the fully connected layer in the excitation part, and the importance of different channels is obtained and then applied to the corresponding channels of the feature map before compression to adjust the feature contribution of each channel, and finally the weighted deep spatial features are obtained.

2. The method according to claim 1, characterized in that The preprocessing of the voice data comprises: The speech data is pre-emphasized, framed and windowed.

3. The method according to claim 1, characterized in that: The first-order and second-order differences of the logarithmic Mel spectrogram of each frame are calculated, including: Define the first-order difference at time t as d t , then the calculation formula is: Among them, c t+n and c t-n Represent the Mel frequency cepstrum coefficients at time t+n and tn respectively, for d t Using the above formula again we can get the second order difference.

4. The method according to claim 1, characterized in that: The processing of the three-channel image to extract the depth temporal features and the depth spatial features includes: The cross-stacked long short-term memory network and the self-attention module are used to extract its deep temporal features. Specifically, the core component of the long short-term memory network is the gate. The forget gate, input gate, output gate, unit state and hidden state defined at time step t are f t 、i t , o t 、c t and h t , then the equation of the long short-term memory network can be expressed as: c t =f t ×c t-1 +i t ×tanh(W c x t +R c h t-1 +b c ), h t =o t ×tanh(c t ), Where σ is the Sigmoid activation function, [W f ,R f ]、[W i ,R i ]、[W o ,R o ]、[W c ,R c ] and [b f ,b i ,b o ,b c ] are the weight matrices and biases of the forget gate, input gate, output gate and unit state, respectively, x t is the input at time t, h t-1 is the hidden state at the moment before moment t, and tanh is the hyperbolic tangent activation function; after obtaining the output of the long short-term memory network, it is fed to the self-attention module; Define the input sequence of the self-attention module as X = [x1, x2, ..., x n ], respectively using the modulation matrix W q , W k and W v To process it: Q=W q X, K=W k X, V=W v X, Thus we can get the query matrix Q = [q1, q2, ..., q N ], key matrix K = [k1, k2, ..., k N ] and value moment V=[v1,v2,...,v N ], and then calculate each query vector q n The attention function output is: Therefore, through network learning, higher weights are given to speech features that are more relevant to depression, so as to increase the network's attention and obtain deep time series features.

5. The method according to claim 1, characterized in that Before using the prediction network including a multi-task learning strategy to screen mental health and predict the severity of the fused speech features, the method further includes: Using a convolutional layer and a dropout layer to compress the three-channel image, removing information features that are irrelevant or redundant to gender information, and then using an attention module and a fully connected network to obtain information related to gender in the input information; The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

6. A deep speech feature construction and mental health screening device, characterized in that: Includes the following modules: An acquisition module, used for acquiring voice data and preprocessing the voice data; A transformation module is used to perform a fast Fourier transform on the preprocessed speech data to obtain a power spectrum, apply a Mel filter bank to the power spectrum and take the logarithm to obtain a Mel spectrogram; for the logarithmic Mel spectrogram of each frame, calculate its first-order and second-order differences; A construction module for constructing a three-channel image based on the obtained log-mel spectrogram and its first-order and second-order differences; An extraction module, used for processing the three-channel image to extract depth temporal features and depth spatial features; The deep temporal features and the deep spatial features are concatenated and dimension reduction is performed using a fully connected layer to obtain a fused speech feature, wherein a convolutional network is selected and a compression-excitation module is introduced to extract the deep spatial features, specifically including defining the dimension of the original feature map as H×W×C, wherein H represents height, W represents width, and C represents the number of channels, and compressing H and W into one dimension through a compression operation, that is, forming a vector of size 1×1×C; predicting the importance of each channel through an excitation part using a fully connected layer, obtaining the importance of different channels and then applying it to the corresponding channels of the feature map before compression, so as to adjust the feature contribution of each channel, and finally obtaining a weighted deep spatial feature; The prediction module is used to screen the mental health and predict the severity of the mental health using a prediction network including a multi-task learning strategy for the fused speech features.

7. The device according to claim 6, characterized in that The prediction module is further used for: Using a convolutional layer and a dropout layer to compress the three-channel image, removing information features that are irrelevant or redundant to gender information, and then using an attention module and a fully connected network to obtain information related to gender in the input information; The convergence of the network is constrained by introducing a weighted gender loss into the computation of the overall loss of the prediction network containing a multi-task learning strategy.

8. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the construction of deep speech features and the mental health screening method as described in any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for constructing deep speech features and screening mental health as described in any one of claims 1 to 5 is implemented.

Images