Signal processing detection method based on attention guidance learnable time domain filter bank

Abstract

The invention discloses a signal processing detection method based on an attention guidance learnable time domain filter bank. Comprising the following steps: S1, establishing a joint learning framework of a learnable time domain filter bank with attention guidance; s2, spectral features related to tasks are generated through the learning module, and acoustic features with biological significance are generated by using a learnable time domain filter; s3, guiding the learnable time domain filter to reserve useful frequency sub-bands through the multi-scale spectrum attention module; and S4, establishing a classification module, and generating the output of the segment-level label by taking the learned spectral features as the input through the classification module. The Mel spectral features in the existing deep SDD framework can be easily replaced by the arranged learning module, and the central frequency of the time domain filter concentrated in the specific formant range of Chinese vowels, especially in the frequency range of 600-700Hz can be learned, so that the time domain filter is a biomarker which is beneficial to distinguishing depression patients.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing detection, and in particular relates to a signal processing detection method based on an attention-guided learnable time-domain filter group. Background Art

[0002] Depression is a common mental illness worldwide and a major factor in the global burden of disease in young people. Depression usually occurs initially in adolescence and may continue or relapse in adulthood. It often becomes a lifelong chronic mental illness. At present, the screening and detection methods of depression mainly rely on questionnaires and interviews, supplemented by clinical assessments by psychiatrists, such as the Patient Health Questionnaire (PHQ) (Klenk & Spitzer, 2002) and the Hamilton Depression Rating Scale (HAMD) (Hamilton, 1986). However, these assessments are easily biased by the experience of the interviewer, the quality of the question scheme, and the patient's willingness (Cummins et al., 2015). Therefore, an objective and convenient method for detecting depression is desirable. Speech is a non-invasive and easily accessible biological information, and different emotional states have different voice characteristics (De Villers, Vidrascu, & Lamel, 2005). Previous studies have reported that the speech of patients with depression is usually characterized by whispering, monotony, slowness, slurring, long pauses and pauses.

[0003] Depression is a global mental health problem, but there is a lack of effective screening methods to help early detection and treatment. This paper aims to promote large-scale screening of depression by focusing on the speech depression detection (SDD) task. Currently, direct modeling of the raw signal will produce a large number of parameters. Existing deep learning-based SDD models mainly use fixed Mel-scale spectral features as input. However, these features are not designed for depression detection, and manual settings limit the exploration of fine-grained feature representation.

[0004] There are still some problems in the existing detection methods:

[0005] (1) In terms of feature learning, direct modeling of raw signals without constraints generates a large number of parameters, which makes it impossible for the model to effectively utilize identifiable features at a fine-grained level (Tukuljac, Ricaud, Aspert, & Colbois, 2022);

[0006] (2) Existing MFbanks and MFCC features are designed to match the human perceptual system and contain rich prior factors related to speech recognition tasks. However, these general acoustic features are not tailored for assessing depression and are therefore unsuitable for SDD tasks; Mel filter banks hinder the extraction of key narrowband features in speech, such as formants associated with depression (Cummins et al., 2015; Ravanelli & Bengio, 2018);

[0007] (3) Depression can cause changes in the vocal tract, shifting speech energy from 500 Hz to the 500-1000 Hz sub-band as the severity of depression increases (Moore, Clements, Peifer, & Weiser, 2008; Ozdas et al., 2004). However, standard mel features usually use fixed hyperparameters such as the number of filters and SFFT sampling points, which is insufficient to capture the fine-grained frequency information relevant to the SDD task. To this end, we propose a signal processing detection method based on attention-guided learnable time-domain filter banks. Summary of the invention

[0008] The purpose of the present invention is to provide a signal processing detection method based on an attention-guided learnable time-domain filter group to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: a signal processing detection method based on attention-guided learnable time-domain filter bank, comprising the following steps:

[0010] S1. Establish a joint learning framework of learnable time-domain filter banks with attention guidance to directly process the original speech signal while exploring key frequency regions in an end-to-end manner, wherein the joint learning framework includes a learning module, a multi-scale spectral attention learning module, and a speech classification module;

[0011] S2. generating task-related spectral features through the learning module and generating biologically meaningful acoustic features by using a learnable time-domain filter;

[0012] S3. guiding the learnable time-domain filter to retain useful frequency sub-bands through the multi-scale spectral attention module;

[0013] S4. Establishing a classification module, through which the learned spectral features are used as input to generate segment-level label outputs;

[0014] S5. Collect a dataset based on an audio corpus of neutral readings and evaluate the performance of learnable time-domain filter banks on the audio corpus and the public DAIC-woz dataset.

[0015] Preferably, the learning module includes a learnable time domain filtering unit, an adaptive scale map pooling unit and a parameterized nonlinear transformation unit, wherein the learnable time domain filtering unit is used to perform time domain filtering on the original signal through a learnable time domain filter, the adaptive scale map pooling unit is used to downsample the output generated by the time domain filter using an adaptive scale map pool, and the parameterized nonlinear transformation unit is used to activate the output of the time domain filter group using a parameterized nonlinear transformation.

[0016] Preferably, the specific steps of S2 include:

[0017] S201. Training a set of learnable time domain filters, performing time domain filtering on the original signal through the learnable time domain filters, to obtain the time-frequency features of the original speech signal, and to capture the frequency components related to depression in the speech;

[0018] S202. downsampling the output generated by the time domain filter using an adaptive scaling pool;

[0019] S203. Use parameterized nonlinear transformation to activate the output of the learning filter group, compress the dynamic range of the learning features while maintaining the local mode, and obtain the learned filter group features.

[0020] Preferably, the output of the time domain filter is obtained by convolving the impulse response φn∈Cκ with the input segment waveform x ij ∈R1×l ij , including the following equation:

[0021] X ij [n,:]=x ij *φn,n=1,...,N;

[0022] In the formula, * represents the convolution operation, n is the filter, and X ij ∈R, N×V is a two-dimensional time-frequency feature, where V = l ij -κ+1, κ is the length of the filter;

[0023] The impulse response φn is defined as the multiplication of a Gaussian kernel and a complex sine kernel, including the following formula:

[0024] n=1,...,N;

[0025] Where t is the time step, σ n is the standard deviation, c n is the center frequency; N = 64, κ = 401, cn∈[0,0.5];

[0026] More specifically, the nth filter By a real (t) and a fictitious Partial composition;

[0027]

[0028] and Keep the filter as follows:

[0029]

[0030]

[0031] Preferably, the multi-scale spectral attention module includes a multi-scale feature extraction block and a frequency-aware attention block. The multi-scale feature extraction block is used to extract information under different receptive fields and obtain multi-scale features along the axis of the time domain filter. The frequency-aware attention block is used to obtain channel attention through convolutions at different time scales and assign different weights to the filters to extract discriminative acoustic features.

[0032] Preferably, the multi-scale feature extraction block further comprises: extracting features at the following locations using an unfolded convolution method:

[0033] Output characteristics of filter banks of various sizes Applied to each channel separately, it allows to extract multi-scale contextual information related to depression in each sub-band Yu and Koltun. Given the nth sequence X n ∈R1×M filter function and a kernel ξ,0,...,κ1, the d-dilated convolution is defined as:

[0034]

[0035] Where d is the dilation factor, t is the time step, κ is the kernel length, and (td·μ) refers to the past time step;

[0036] K dilated units are stacked in the multi-scale spectral attention module, each of which has a 2K-1 receptive field. Specifically, each dilated unit contains a dilated convolution operation, an activation function, and two 1×1 convolutions. By stacking dilated units, the model allows the receptive field to be increased. In addition, the dilated unit has two outputs, one output Rk is passed through subsequent units along the backbone network, and the other is a skip connection output Slovak, allowing the block to explicitly capture multi-scale features at multiple hierarchical levels. The formula for these two outputs is written as:

[0037]

[0038]

[0039] in represents the dilated convolution weight at layer k, relu(·) is the rectified linear unit (ReLU), and is a trainable 1×1 convolution weight.

[0040] Preferably, the frequency-aware attention block is used to simulate the importance of different channels and capture the dynamic behavior of the filter. Specifically, different learnable time-domain filters are regarded as different channels and assigned different attention weights. First, an element-level addition strategy is used to fuse multi-scale features to enhance basic features and weaken irrelevant features, including the following formula:

[0041]

[0042] in, To fuzzify the features, three parallel 1D convolutional attention modules (CAMs) with kernels of size 3, 5, and 7 are used to effectively capture the important features of X at different time scales. i f j u Then, a fully connected (FC) layer and a softmax function are used to learn the important weights aa t From the time scale characteristics of the connection;

[0043]

[0044] Where T represents transposition, a n represents the attention weight corresponding to the nth filter,

[0045] Finally, we get the weighted Use the learned weight aa t , which is written as:

[0046]

[0047] Among them, · is the product operation of Hadamard, so the frequency-aware attention block will prioritize the frequency sub-bands that have more significant impact on the SDD task. By stacking several MSSA layers, the learnable time-domain filter is guided to capture key information from the fused multi-scale features and improve the sensitivity of its depressive features.

[0048] Preferably, the classification module comprises several residual blocks, each of which consists of three one-dimensional convolutional layers with kernel sizes of 7, 5 and 3, respectively. Batch normalization and ReLU function are performed after each convolutional layer, the block output features are flattened and input into the FC layer, and the output is mapped to the probability of belonging to each class using the sigmod function. The learned spectral features are used as input to generate the output of segment-level labels, and the topic-level labels are determined by the majority vote of the part-level predictions.

[0049] Preferably, the dataset collected in S5 based on the neutral reading audio corpus includes participant recruitment, clinical observer scoring, self-assessment and speech data collection, which is used to systematically investigate the differences in speech differences between individuals diagnosed with depression and a normal control group.

[0050] Preferably, the performance of the audio corpus data set is evaluated in S5 by using PHQ-9 and HAMD17 scales to comprehensively verify the audio corpus data set.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) The present invention proposes a learning module for depression features, which adaptively constructs unbiased filter group output features through a learnable method. Different from traditional millimeter spectrum features, the learnable time domain filter group can automatically modulate the corresponding frequency region related to the classification task. By analyzing the cumulative frequency response of the learnable time domain filter, the frequency region of interest can be quantified to enhance the interpretability of the model. In addition, the learning module can easily replace the existing Mel spectrum features in the deep SDD framework, and the center frequency of the learnable time domain filter is concentrated in the specific resonance peak range of Chinese vowels, especially in the frequency range of 600-700Hz, which is a biomarker that helps to distinguish patients with depression;

[0053] (2) The present invention develops a multi-scale spectral attention module to recalibrate channel feature responses. MEB uses dilated convolution to explore frequency features of different scales on each filter. Meanwhile, FAB captures the dynamic behavior of the filter through multi-scale convolution. By stacking multiple MSSA layers, the model can extract key information from the fused multi-scale features and guide the filter to learn meaningful frequency sub-bands, thereby improving its sensitivity to depressive features.

[0054] (3) The present invention contributes to the research on depression detection by collecting an audio corpus based on neutral reading. In addition, the DALF framework is comprehensively evaluated on public datasets such as the Pain Analysis Interview Corpus, the Wizard of Oz, and the NRAC dataset. It not only outperforms the state-of-the-art methods in the SDD task, but also can automatically identify biomarkers related to depression. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the present invention;

[0056] Figure 2 It is a structural block diagram of the learning module of the present invention;

[0057] Figure 3 This is a structural block diagram of the multi-scale spectral attention module of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] See also Figure 1-Figure 3 The present invention provides a technical solution: a signal processing detection method based on an attention-guided learnable time-domain filter group, comprising the following steps:

[0060] S1. Establish a joint learning framework of learnable time-domain filter banks with attention guidance to directly process the original speech signal while exploring key frequency regions in an end-to-end manner, wherein the joint learning framework includes a learning module, a multi-scale spectral attention learning module, and a speech classification module;

[0061] S2. generating task-related spectral features through the learning module and generating biologically meaningful acoustic features by using a learnable time-domain filter;

[0062] S3. guiding the learnable time-domain filter to retain useful frequency sub-bands through the multi-scale spectral attention module;

[0063] S4. Establishing a classification module, through which the learned spectral features are used as input to generate segment-level label outputs;

[0064] S5. Collect a dataset based on an audio corpus of neutral readings and evaluate the performance of learnable time-domain filter banks on the audio corpus and the public DAIC-woz dataset.

[0065] In this embodiment, preferably, the learning module includes a learnable time domain filtering unit, an adaptive scale map pooling unit and a parameterized nonlinear transformation unit, the learnable time domain filtering unit is used to perform time domain filtering on the original signal through a learnable time domain filter, the adaptive scale map pooling unit is used to downsample the output generated by the time domain filter using an adaptive scale map pool, and the parameterized nonlinear transformation unit is used to activate the output of the time domain filter group using a parameterized nonlinear transformation.

[0066] In this embodiment, preferably, the specific steps of S2 include:

[0067] S201. Training a set of learnable time domain filters, performing time domain filtering on the original signal through the learnable time domain filters, to obtain the time-frequency features of the original speech signal, and to capture the frequency components related to depression in the speech;

[0068] S202. downsampling the output generated by the time domain filter using an adaptive scaling pool;

[0069] S203. Use parameterized nonlinear transformation to activate the output of the learning filter group, compress the dynamic range of the learning features while maintaining the local mode, and obtain the learned filter group features.

[0070] In this embodiment, preferably, the output of the time domain filter is a convolution impulse response φn∈Cκ, using the input segment waveform x ij ∈R1×l ij , including the following equation:

[0071] X ij [n,:]=x ij *φn,n=1,...,N;

[0072] In the formula, * represents the convolution operation, n is the filter, and X ij ∈R, N×V is a two-dimensional time-frequency feature, where V = l ij -κ+1, κ is the length of the filter;

[0073] The impulse response φn is defined as the multiplication of a Gaussian kernel and a complex sine kernel, including the following formula:

[0074] n=1,...,N;

[0075] Where t is the time step, σ n is the standard deviation, c n is the center frequency; N = 64, κ = 401, cn∈[0,0.5];

[0076] More specifically, the nth filter By a real and a fictional Partial composition;

[0077]

[0078] and Keep the filter as follows:

[0079]

[0080]

[0081] In this embodiment, preferably, the multi-scale spectral attention module includes a multi-scale feature extraction block and a frequency-aware attention block. The multi-scale feature extraction block is used to extract information under different receptive fields and obtain multi-scale features along the axis of the time domain filter. The frequency-aware attention block is used to obtain channel attention through convolutions at different time scales and assign different weights to the filters to extract discriminative acoustic features.

[0082] In this embodiment, preferably, the multi-scale feature extraction block further includes: extracting features at the following locations using an unfolded convolution method:

[0083] Output characteristics of filter banks of various sizes Applied to each channel separately, it allows to extract multi-scale contextual information related to depression in each sub-band Yu and Koltun. Given the nth sequence X n ∈R1×M filter function and a kernel ξ,0,...,κ1, the d-dilated convolution is defined as:

[0084]

[0085] Where d is the dilation factor, t is the time step, κ is the kernel length, and (td·μ) refers to the past time step;

[0086] K dilated units are stacked in the multi-scale spectral attention module, each of which has a 2K-1 receptive field. Specifically, each dilated unit contains a dilated convolution operation, an activation function, and two 1×1 convolutions. By stacking dilated units, the model allows the receptive field to be increased. In addition, the dilated unit has two outputs, one output Rk is passed through subsequent units along the backbone network, and the other is a skip connection output Slovak, allowing the block to explicitly capture multi-scale features at multiple hierarchical levels. The formula for these two outputs is written as:

[0087]

[0088]

[0089] in represents the dilated convolution weight at layer k, relu(·) is the rectified linear unit (ReLU), and is a trainable 1×1 convolution weight.

[0090] In this embodiment, preferably, the frequency-aware attention block is used to simulate the importance of different channels and capture the dynamic behavior of the filter. Specifically, different learnable time-domain filters are regarded as different channels and assigned different attention weights. First, an element-level addition strategy is used to fuse multi-scale features to enhance basic features and weaken irrelevant features, including the following formula:

[0091]

[0092] in, To fuzzify the features, three parallel 1D convolutional attention modules (CAMs) with kernels of size 3, 5, and 7 are used to effectively capture the important features of X at different time scales. Then, a fully connected (FC) layer and a softmax function are used to learn the important weights aa t From the time scale characteristics of the connection;

[0093]

[0094] Where T represents transposition, a n represents the attention weight corresponding to the nth filter,

[0095] Finally, we get the weighted Use the learned weight aa t , which is written as:

[0096]

[0097] Among them, · is the product operation of Hadamard, so the frequency-aware attention block will prioritize the frequency sub-bands that have more significant impact on the SDD task. By stacking several MSSA layers, the learnable time-domain filter is guided to capture key information from the fused multi-scale features and improve the sensitivity of its depressive features.

[0098] In this embodiment, preferably, the classification module includes several residual blocks, each residual block consists of three one-dimensional convolutional layers with kernel sizes of 7, 5 and 3, respectively. Batch normalization and ReLU function are performed after each convolutional layer, the block output features are flattened and input into the FC layer, and the sigmoid function is used to map the output to the probability of belonging to each class. The learned spectral features are used as input to generate the output of segment-level labels, and the topic-level labels are determined by the majority vote of the part-level predictions.

[0099] In this embodiment, preferably, the dataset collected in S5 based on the neutral reading audio corpus includes participant recruitment, clinical observer scoring, self-assessment and speech data collection, which is used to systematically investigate the differences in speech differences between individuals diagnosed with depression and a normal control group.

[0100] In this embodiment, preferably, the performance of the audio corpus data set is evaluated in S5 by using PHQ-9 and HAMD17 scales to comprehensively verify the audio corpus data set.

[0101] Experimental data analysis:

[0102] The audio corpus dataset includes 155 depressions and 110 nc samples; in addition, the severity of depression of the participants was assessed using two scales: Hamilton Depression Rating Scale 17 Items (HAMD17) and Patient's Health Status;

[0103] Table 1 NRAC and corresponding depression severity

[0104]

[0105] The PHQ-9 is an observer-rated scale used by clinicians, while the PHQ-9 is a self-rated scale completed by the participants themselves. Based on these two scales, the participants were divided into two groups:

[0106] PHQ-9 group: Participants in this group rated their depression severity using the PHQ-9, with voice-recorded scores ranging from 0 to 27, with scores on the scale corresponding to the following levels of depression severity: normal control group (0-4), mild depression (5-9), moderate depression (10-19), and severe depression (≥20);

[0107] HAMD-17 group: Participants in this group rated their depression severity using HAMD17. The voice recordings were marked with a score from 0 to 52. The severity level based on the scale score was:

[0108] Normal control group (0-7), mild depression (8-16), moderate depression (17-23), severe depression (≥24);

[0109] During the data collection phase, all participants were placed in a peaceful and closed environment;

[0110] Clear instructions were provided to the participants, emphasizing the importance of remaining in a natural and relaxed state throughout the recording process. To capture the audio, a standardized recording pen was used to capture the audio. The participants were asked to read a specific text, "Let life be as beautiful as the flowers in summer". The speech length ranged from 2 to 4 min. The speech files were recorded at a sampling rate of 44, 1 kHz, PCM, at 16 bits.

[0111] Table 1 summarizes the basic information of the NRAC dataset, and Appendix B provides demographic details. All participants were informed of the study according to the regulations set by the Medical Ethics Committee of Nanjing Brain Hospital.

[0112] and provided written informed consent, signed by themselves and their legal guardians. The study received official approval number 2021-KY108-01 from MECNJBH.

[0113] 2. DAIC-woz: Since the eight-item Patient Health Questionnaire (PHQ-8) has been shown to be as effective as the PHQ-9 for screening depression (Shin, Lee, Han, Yin, & Han, 2019), the Distress Analysis Interview Corpus The Wizard of Oz (DAIC-woz) (Gratch et al., 2019) was also used. In the study, DAICwoz is a public English dataset that includes a series of clinical interviews to diagnose depression, anxiety, post-traumatic stress disorder, and other mental disorders. The length of each audio file ranges from 7-33min, with a fixed sampling rate of 16000Hz. Each participant received a PHQ-8 score based on their responses to the questionnaire. Each recording was labeled with a PHQ-8 score and a PHQ-8 binary label, where the binary score was determined by a threshold of 10 to divide the patients into depression and NCs groups. As shown in Table 2, the train set and development set of DAIC-woz include 10 7 samples and 35 samples, respectively;

[0114] Table 2

[0115]

[0116] Parameter settings: The hyperparameter settings of each module are shown in Table 3; for optimization, the Adam optimizer with a weight decay of 0.001 is used, the batch size is set to 32, and the initial learning rate is 0.0005.

[0117] Evaluation Metrics: For the DAIC-woz dataset, the same random sampling procedure as for deep networks (Ma, Yang, Chen, Huang, & Wang, 2016) was followed, and the results were reported based on the development set. For each task in NRAC, all datasets were split into 5 times to retain the percentage of samples in each class. In addition, 10% of the training data was selected as a validation dataset for hyperparameter tuning. To evaluate the performance of the model, accuracy, recall, f1 score, and precision were used as the selected evaluation metrics. The mean and standard deviation of all results were reported, and a paired t-test at the 5% significance level was used to explore whether there were statistical differences between the DALF metric and other methods.

[0118] Table 3

[0119] Hyperparameter settings for the model, where length represents the kernel size.

[0120]

[0121]

[0122] Performance evaluation on the audio corpus dataset:

[0123] To intuitively explore the performance of DALF on SDD tasks, we first conducted a comprehensive validation on the NRAC dataset using the PHQ-9 and HAMD17 scales. Specifically, we cast depression detection as a classification task. The depression severity classification (SC) task aims to explore discriminative patterns through speech. In addition, the diagnosis-based classification (DC) task focuses on detecting speech differences between NCs and patients.

[0124] Table 4 reports the results of DALF, and there are some observations worth highlighting:

[0125] 1. DALF performed well in the PHQ-9 group and HAMD17 group in the classification of depression and NCs, with F1 scores of 87.3% and 81.7%, respectively;

[0126] 2. When comparing the classification results for differentiating NCs from patients with mild, moderate, and severe symptoms in the PHQ-9 group, it is noteworthy that the classification performance was consistently good regardless of the severity of depression. These results suggest that patients with different symptoms exhibited significant differences in speech compared to NCs. In addition, the analysis of both groups showed that the model performance improved significantly with symptom severity, indicating that the acoustic features associated with depression were more pronounced in severe cases.

[0127] 3. When observing the depression severity classification results, including the classification task of NCs. For each severity category and the classification task of different severity categories, we can find that the classification performance of mild, moderate, and severe symptoms is poor, especially when the symptoms are close. This is because depressive speech often shows common features and similar expressions, resulting in weak differences between different symptom severities.

[0128] 4. Considering the weak performance of the HAMD17 group in the classification of mild and NCs, it can be inferred that the PHQ9 scale provides a fair reflection of depressive speech. In other words, the change in language is evident regardless of the score. In addition, the weak performance in the HAMD17 group was found. The PHQ-9 scale, which is based on the actual feelings of the patient, is very suitable for initial screening.

[0129] Table 4 Experimental results on NRAC, boldfaced

[0130]

[0131] On one hand, the proposed method is compared with several state-of-the-art methods using the public DAIC-woz dataset. In particular, a comprehensive comparison of the DALF model with traditional methods and deep learning methods is performed. On the other hand, several comparable methods are tested on the PHQ-9 group of the NRAC dataset. For all the compared methods, the same settings as the original paper are adopted to ensure that they are competitive in the comparison. All Mel-spectral features (Nfb=64, Nmfcc=40) are performed on a 16kHz sampler using a 25ms window length and a 10ms hop length.

[0132] Traditional Methods: Two common ML classifiers were selected as baselines: Support Vector Machine (SVM) and Random Forest (RF). We then used openSMILE (Eyben et al.) to extract acoustic features, including low-level MFCs, comparisons and eGeMAPS, and input all concatenated feature representations into the model.

[0133] Deep Learning Model: We compare our model with the following models: DepAudioNet1 (Ma et al. 2016), which is an SDD framework that uses a 1D CNN and LSTM to capture mid-term and long-term dependencies to produce a more comprehensive speech representation.

[0134] DEPA (Zhang et al., 2021) is a self-supervised audio embedding pre-trained model for depression detection, in which a self-supervised encoder-decoder model is trained to predict and reconstruct the central spectrogram.

[0135] Mfcc-LSTM (Rejaibi et al.) is a continuous LSTM-based model that utilizes MFCC features to detect depression and assess its severity.

[0136] CNN-ae (Sardari, Nakisa, Rastgoo, & Eklund, 2022) is an end-to-end depression detection model that uses a CNN autoencoder to learn highly relevant and discriminative features from raw audio signals.

[0137] En-CNN (Vázquez-Romero and Gallardo-Antolin, 2020) uses an ensemble learning strategy to integrate a voting mechanism with the 1D-CNN architecture to improve the performance of depression classification.

[0138] Vladgrup2 (Shen et al.) used NetVLAD to generate same-dimensional speech embeddings from extracted Mel spectrograms and used Gate Recurrent Units (GRU) to automatically summarize representations of depressive speech.

[0139] ConvBiLSTM3 (Wei et al.) achieves end-to-end depression estimation by leveraging a layered CNN and Bi-LSTM structure to capture both short-term and long-term temporal and spectral features.

[0140] Table 5 shows that our model consistently achieves better performance than the baseline. In particular, we have the following observations

[0141] Table 5

[0142] Comparison with other methods on different datasets, where RS represents the original speech signal. The best indicator is in bold, and the t indicator p value < 0 is statistically significant.

[0143]

[0144] DALF achieves superior performance on DAIC-woz compared to the state-of-the-art. Moreover, compared to traditional approaches such as handcrafted features combined with RF and SVM, DALF observes significant improvements of 21.4% and 38.4% in F1, respectively. Similarly, our DALF outperforms other deep learning methods that use only audio modality data. For example, DALF achieves an average improvement of 10.20% in F1 compared to CNN-based models (DepAudioNet, CNN-AE, and En-CNN). Furthermore, comparisons are made with RNN-based methods (DEPA, Mfcc-LSTM, Vlad-GRU, and ConvbiLSTM). DALF achieves an average F1 im-12.7%. These results highlight the effectiveness of the DALF framework.

[0145] 2. On the NRAC dataset, as shown in Table 5, our model consistently achieves the best results in terms of accuracy, F1, and other metrics compared to other state-of-the-art models. This improvement shows that the DALF framework significantly outperforms the state-of-the-art methods. This is due to the fact that the DFBL module captures more relevant spectral features associated with depression in high-dimensional signals.

[0146] 3. The classification results obtained by these models on NRAC are generally better than those on the DAIC-woz dataset. This can be attributed to the fact that the DAICwoz dataset is collected based on conversational scenarios with fewer training samples and more unbalanced class distribution. In contrast, the NRAC dataset is collected using neutral reading for speech, with lower heterogeneity in speech content and more balanced samples. Therefore, we believe that more clinically valuable findings can be obtained from the NRAC dataset.

[0147] Principles and advantages of the present invention:

[0148] The present invention proposes a learning module for depression features, which adaptively constructs unbiased filter group output features through a learnable method. Different from traditional millimeter spectrum features, the learnable time domain filter group can automatically modulate the corresponding frequency region related to the classification task. By analyzing the cumulative frequency response of the learnable time domain filter, the frequency region of interest can be quantified to enhance the interpretability of the model. In addition, the learning module can easily replace the existing Mel spectrum features in the deep SDD framework, and the center frequency of the learnable time domain filter is concentrated in the specific resonance peak range of Chinese vowels, especially in the frequency range of 600-700Hz, which is a biomarker that helps to distinguish patients with depression. By developing a multi-scale The spectral attention module can recalibrate the channel feature response, MEB uses dilated convolution to explore the frequency features of different scales on each filter, and FAB captures the dynamic behavior of the filter through multi-scale convolution. By superimposing multiple MSSA layers, the model can extract key information from the fused multi-scale features and guide the filter to learn meaningful frequency sub-bands, thereby improving its sensitivity to depressive features; by collecting an audio corpus based on neutral reading, it is helpful for the research on depression detection. In addition, the DALF framework is comprehensively evaluated on the public datasets Pain Analysis Interview Corpus, The Wizard of Oz, and NRAC datasets, which not only outperforms the most advanced methods in the SDD task, but also can automatically identify biomarkers related to depression.

[0149] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A signal processing detection method based on attention-guided learnable time-domain filter banks, characterized in that: The following steps are involved: S1. Establish a joint learning framework of learnable time-domain filter banks with attention guidance to directly process the original speech signal while exploring key frequency regions in an end-to-end manner, wherein the joint learning framework includes a learning module, a multi-scale spectral attention learning module, and a speech classification module; S2. generating task-related spectral features through the learning module and generating biologically meaningful acoustic features by using a learnable time-domain filter; S3. guiding the learnable time-domain filter to retain useful frequency sub-bands through the multi-scale spectral attention module; S4. Establishing a classification module, through which the learned spectral features are used as input to generate segment-level label outputs; S5. Collect a dataset based on an audio corpus of neutral readings and evaluate the performance of learnable time-domain filter banks on the audio corpus and the public DAIC-woz dataset.

2. According to claim 1, a signal processing detection method based on attention-guided learnable time-domain filter bank is characterized in that: The learning module includes a learnable time domain filtering unit, an adaptive scale map pooling unit and a parameterized nonlinear transformation unit. The learnable time domain filtering unit is used to perform time domain filtering on the original signal through a learnable time domain filter, the adaptive scale map pooling unit is used to downsample the output generated by the time domain filter using an adaptive scale map pool, and the parameterized nonlinear transformation unit is used to activate the output of the time domain filter group using a parameterized nonlinear transformation.

3. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 2, characterized in that: The specific steps of S2 include: S201. Training a set of learnable time domain filters, performing time domain filtering on the original signal through the learnable time domain filters, to obtain the time-frequency features of the original speech signal, and to capture the frequency components related to depression in the speech; S202. downsampling the output generated by the time domain filter using an adaptive scaling pool; S203. Use parameterized nonlinear transformation to activate the output of the learning filter group, compress the dynamic range of the learning features while maintaining the local mode, and obtain the learned filter group features.

4. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 3, characterized in that: The output of the time domain filter is obtained by convolving the impulse response φn∈Cκ with the input segment waveform x ij ∈R1×l ij , including the following equation: X ij [n,:]=x ij *φn,n=1,...,N; In the formula, * represents the convolution operation, n is the filter, and X ij ∈R, N×V is a two-dimensional time-frequency feature, where V = l ij -κ+1, κ is the length of the filter; The impulse response φn is defined as the multiplication of a Gaussian kernel and a complex sine kernel, including the following formula: Where t is the time step, σ n is the standard deviation, c n is the center frequency; N = 64, κ = 401, cn∈[0,0.5]; More specifically, the nth filter By a real and a fictional Partial composition; and Keep the filter as follows:

5. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 1, characterized in that: The multi-scale spectral attention module includes a multi-scale feature extraction block and a frequency-aware attention block. The multi-scale feature extraction block is used to extract information under different receptive fields and obtain multi-scale features along the axis of the time domain filter. The frequency-aware attention block is used to obtain channel attention through convolutions at different time scales and assign different weights to the filters to extract discriminative acoustic features.

6. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 5, characterized in that: The multi-scale feature extraction block further includes: extracting features at the following locations using an unfolded convolution method: Output characteristics of filter banks of various sizes Applied to each channel separately, it allows to extract multi-scale contextual information related to depression in each sub-band Yu and Koltun. Given the nth sequence X n ∈R1×M filter function and a kernel ξ,0,...,κ1, the d-dilated convolution is defined as: Where d is the dilation factor, t is the time step, κ is the kernel length, and (td·μ) refers to the past time step; K dilated units are stacked in the multi-scale spectral attention module, each of which has a 2K-1 receptive field. Specifically, each dilated unit contains a dilated convolution operation, an activation function, and two 1×1 convolutions. By stacking dilated units, the model allows the receptive field to be increased. In addition, the dilated unit has two outputs, one output Rk is passed through subsequent units along the backbone network, and the other is a skip connection output Slovak, allowing the block to explicitly capture multi-scale features at multiple hierarchical levels. The formula for these two outputs is written as: in represents the dilated convolution weight at layer k, relu(·) is the rectified linear unit (ReLU), and is a trainable 1×1 convolution weight.

7. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 6, characterized in that: The frequency-aware attention block is used to simulate the importance of different channels and capture the dynamic behavior of the filter. Specifically, different learnable time-domain filters are regarded as different channels and assigned different attention weights. First, an element-level addition strategy is used to fuse multi-scale features to enhance basic features and weaken irrelevant features. Includes the following formulas: in, To fuzzify the features, three parallel 1D convolutional attention modules (CAMs) with kernels of size 3, 5, and 7 are used to effectively capture the important features of X at different time scales. Then, a fully connected (FC) layer and a softmax function are used to learn the important weights aa t From the time scale characteristics of the connection; Where T represents transposition, a n represents the attention weight corresponding to the nth filter, Finally, we get the weighted Use the learned weight aa t , which is written as: Among them, · is the product operation of Hadamard, so the frequency-aware attention block will prioritize the frequency sub-bands that have more significant impact on the SDD task. By stacking several MSSA layers, the learnable time-domain filter is guided to capture key information from the fused multi-scale features and improve the sensitivity of its depressive features.

8. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 1, characterized in that: The classification module contains several residual blocks, each of which consists of three one-dimensional convolutional layers with kernel sizes of 7, 5 and 3, respectively. Batch normalization and ReLU function are performed after each convolutional layer, the block output features are flattened and input into the FC layer, and the sigmoid function is used to map the output to the probability of belonging to each class. The learned spectral features are used as input to generate the output of segment-level labels, and the topic-level labels are determined by the majority vote of the part-level predictions.

9. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 1, characterized in that: The dataset collected in S5, which is a neutral reading-based audio corpus, includes participant recruitment, clinical observer ratings, self-assessment, and speech data collection, and is used to systematically investigate differences in speech differences between individuals diagnosed with depression and normal controls.

10. The signal processing detection method based on attention-guided learnable time-domain filter bank according to claim 1, characterized in that: The performance of the audio corpus dataset is evaluated in S5, and the audio corpus dataset is comprehensively validated using the PHQ-9 and HAMD17 scales.

Images