An acoustic event detection method based on multi-scale spatial feature and coordinate attention fusion
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2026-03-23
- Publication Date
- 2026-06-09
AI Technical Summary
Existing sound event detection methods suffer from limited network receptive fields, making it difficult to capture multi-scale time-span features and accurately focus on target sounds and suppress noise in complex time-frequency spaces, resulting in insufficient detection accuracy and robustness.
A sound event detection method that integrates multi-scale spatial features and coordinate attention is adopted. The method uses a multi-scale coordinate attention module to perform deep feature extraction and temporal modeling of audio time-frequency features, and combines bidirectional gated loop unit to extract context information, thereby achieving accurate capture and noise suppression of multi-scale sound events.
It significantly improves the accuracy and robustness of sound event detection, can accurately locate target sounds in complex environments, enhances the ability to capture features across multiple time spans, and suppresses noise interference.
Smart Images
Figure CN122177148A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of acoustic event detection technology, specifically relating to an acoustic event detection method based on the fusion of multi-scale spatial features and coordinate attention. Background Technology
[0002] Sound event detection (SED) is an important research direction in the field of speech signal processing. Its goal is to identify the category of target sound events in a continuous audio stream and accurately determine their occurrence and end times, thereby precisely locating the events. With the rapid development of artificial intelligence and IoT technologies, systems such as smart homes, smart security, elderly care, and industrial equipment monitoring require real-time identification of key sound events during operation.
[0003] With the rapid development of deep learning technology, various neural network structures have been widely applied in the field of sound feature processing and have achieved remarkable results, such as CRNN (Convolutional Recurrent Neural Network) and Transformer. However, audio data in real-world environments often exhibits complex background noise, overlapping sound sources, and significant differences in sound duration (e.g., short impact sounds versus long alarm sounds). Common neural network models, when extracting audio time-frequency features, typically employ fixed-size convolutional receptive fields, lacking the ability to capture information across multiple time scales. Furthermore, they cannot accurately focus on discriminative target features in complex time-frequency spaces and are easily affected by environmental noise, leading to false alarms. Summary of the Invention
[0004] The purpose of this invention is to propose a sound event detection method based on the fusion of multi-scale spatial features and coordinate attention, in order to solve the problems in existing sound event detection methods, such as the limited receptive field of the network making it difficult to capture multi-scale time-span features and the inability to accurately focus on target sounds and suppress noise in complex time-frequency spaces, thereby improving the accuracy and robustness of sound event detection.
[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:
[0006] This invention provides a sound event detection method based on the fusion of multi-scale spatial features and coordinate attention, comprising the following seven steps:
[0007] Step 1: Obtain the raw audio signal to be detected;
[0008] Step 2: Preprocess the audio signal to be detected to obtain the audio time-frequency characteristics;
[0009] Step 3: Perform data augmentation on the preprocessed audio time-frequency features;
[0010] Step 4: Extract features from the time-frequency features of the augmented audio data, including initial spatial feature extraction and multi-scale coordinate attention depth feature extraction, to obtain depth features;
[0011] Step 5: Extract contextual information and perform temporal modeling on the extracted deep features to obtain temporal features;
[0012] Step 6: Input the temporal features into the classifier to obtain the frame-level and segment-level sound event detection results of the audio signal to be detected;
[0013] Step 7: Post-process and decode the frame-level and segment-level sound event detection results to obtain the sound event detection results to be detected.
[0014] The preprocessing of the audio signal to be detected to obtain audio time-frequency features includes:
[0015] The audio signal to be detected is converted to mono and length aligned;
[0016] The length-aligned signal is subjected to extreme value normalization.
[0017] The normalized signal is subjected to a short-time Fourier transform and mapped to a Mel filter bank. The logarithm is then taken to obtain a two-dimensional log-Mel spectrogram as the audio time-frequency feature.
[0018] The data augmentation of the preprocessed audio time-frequency features includes:
[0019] The audio time-frequency features after data preprocessing are randomly shifted along the time axis according to a Gaussian distribution to obtain the first enhancement result;
[0020] The time-frequency features and their labels of the preprocessed audio data in the same batch are randomly linearly mixed using mixing parameters that follow a beta distribution to obtain the second enhancement result;
[0021] The frequency bands are randomly divided in the frequency dimension, and a dynamic filter composed of logarithmic random gain multipliers is applied to perturb the frequency response, resulting in the third enhancement result.
[0022] Select the time or frequency window to be masked based on the mask, and mask the selected window with continuous feature frames to obtain the fourth enhancement result;
[0023] The fourth enhancement result is used as the result of data augmentation of audio time-frequency features.
[0024] The feature extraction of the time-frequency features of the augmented audio data includes:
[0025] The augmented audio time-frequency features are input into the CNN module for preliminary feature extraction to obtain initial features;
[0026] The initial features are input into the multi-scale coordinate attention module for in-depth feature extraction, resulting in in-depth feature extraction results.
[0027] The step of inputting the augmented audio time-frequency features into the CNN module for preliminary feature extraction to obtain initial features includes:
[0028] The preprocessed audio time-frequency features are input into a two-dimensional convolutional layer for feature extraction.
[0029] The output of the two-dimensional convolutional layer is input into the normalization layer for normalization processing;
[0030] The normalized result is input into the GLU gated linear unit for nonlinear activation processing;
[0031] The result after nonlinear activation is input into the average pooling layer for pooling downsampling to obtain the dimensionality reduction result.
[0032] The above steps are iterated over multiple times, and the output of the last average pooling layer is used as the initial feature.
[0033] The process of inputting initial features into a multi-scale coordinate attention module for in-depth feature extraction to obtain in-depth feature extraction results includes the following ten steps:
[0034] Step 1: Input the initial features into multiple parallel feature extraction branches. Each branch includes a 1×1 convolution branch, multiple dilated convolution branches that set the dilation rate only in the time dimension, and a global average pooling branch to obtain local and global features under different receptive fields.
[0035] Step 2: Perform bilinear interpolation upsampling on the output of the global average pooling branch, and then concatenate the features of all branches along the channel dimension to obtain the concatenated features;
[0036] Step 3: Input the spliced features into the feature fusion layer for dimensionality reduction to obtain multi-scale fused features;
[0037] Step 4: Perform one-dimensional adaptive average pooling on the multi-scale fused features along the frequency dimension and the time dimension respectively to obtain horizontal aggregated features and vertical aggregated features;
[0038] Step 5: Concatenate the horizontal and vertical aggregated features in the spatial dimension to obtain the intermediate feature;
[0039] Step Six: Input the intermediate features into the first convolutional layer and perform joint feature encoding with a nonlinear activation function;
[0040] Step 7: Divide the encoded intermediate features into horizontal and vertical tensors along the original concatenation dimension;
[0041] Step 8: Input the horizontal tensor and the vertical tensor into the second convolutional layer and the third convolutional layer, respectively;
[0042] Step 9: Input the outputs of the second and third convolutional layers into the Sigmoid activation function to obtain the attention weights in the time dimension and the frequency dimension.
[0043] Step 10: Multiply the multi-scale fusion features element-wise with the time dimension attention weights and frequency dimension attention weights to obtain the final in-depth feature extraction result.
[0044] The coordinate attention unit in the multi-scale coordinate attention module is used to perform biaxial time-frequency decoupling and attention recalibration on the input multi-scale fused features. The calculation formula is as follows:
[0045]
[0046]
[0047]
[0048] in, To further analyze the feature extraction results, let c represent the c-th feature channel, h be the height of the time dimension, and w be the width of the frequency dimension. This represents the horizontal aggregated feature at time height h obtained after pooling along the frequency dimension. This represents the vertical aggregated feature along the frequency width w obtained after pooling along the time dimension, where W is the total width dimension of the frequency dimension and H is the total height dimension of the time dimension. and For the feature element values participating in the one-dimensional adaptive average pooling calculation, The fused multi-scale feature values, For the time dimension attention weight corresponding to the c-th channel, The frequency dimension attention weight corresponds to the c-th channel.
[0049] The step of extracting contextual information and performing temporal modeling on the extracted deep spatial features to obtain temporal features, and inputting the temporal features into a classifier, includes:
[0050] The frequency dimension of the extracted deep spatial features is folded and compressed, and then input into the BiGRU module for contextual temporal modeling to obtain the first branch features.
[0051] The output of the BiGRU module is input into the strongly labeled classification fully connected layer and the attention weight fully connected layer, respectively.
[0052] The output of the strong label classification fully connected layer contains the strong label prediction result containing the prediction probability for each time frame;
[0053] The attention-weighted fully connected layer outputs the attention weight for each time frame after being activated by Softmax. The attention weight is then used to perform weighted summation and pooling on the strong label prediction results to obtain the weak label prediction results that include the prediction probability of the entire audio segment.
[0054] The present invention has the following beneficial effects:
[0055] (1) Introducing a multi-scale coordinate attention module into the model significantly enhances the model’s ability to capture multi-scale sound events, enabling the network to simultaneously and accurately capture the microscopic acoustic texture of short transient sounds (such as the sound of breaking glass) and the macroscopic contextual information of long continuous sounds (such as the roar of an engine and the sound of an alarm). On the other hand, it achieves accurate decoupling and directional focusing of time-frequency features, thereby achieving accurate localization of the target sound in complex time-frequency feature maps.
[0056] (2) Introducing a multi-scale coordinate attention module into the model can significantly improve the accuracy and robustness of sound event detection. Attached Figure Description
[0057] Figure 1 This is a flowchart of the acoustic event detection method in an embodiment of the present invention;
[0058] Figure 2 This is a structural diagram of the neural network model in an embodiment of the present invention;
[0059] Figure 3 This is a structural diagram of the multi-scale coordinate attention module in an embodiment of the present invention; Detailed Implementation
[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.
[0061] This embodiment proposes a sound event detection method based on the fusion of multi-scale spatial features and coordinate attention, such as... Figure 1 As shown, it includes:
[0062] Step S1: Obtain the original audio signal to be detected;
[0063] Step S2: Preprocess the audio signal to be detected to obtain audio time-frequency features;
[0064] Step S3: Perform data augmentation on the preprocessed audio time-frequency features;
[0065] Step S4: Extract features from the time-frequency features of the augmented audio data, including initial spatial feature extraction and multi-scale coordinate attention depth feature extraction, to obtain depth features;
[0066] Step S5: Extract contextual information and perform temporal modeling on the extracted deep features to obtain temporal features;
[0067] Step S6: Input the temporal features into the classifier to obtain the frame-level and segment-level sound event detection results of the audio signal to be detected;
[0068] Step S7: Post-process and decode the frame-level and segment-level sound event detection results to obtain the sound event detection results to be detected.
[0069] In a preferred embodiment of the present invention, in step S1, the raw audio signal to be detected is acquired. The dataset of the raw audio signal to be detected is the Indoor Sound Event Detection Dataset (DESED), which is currently the most mainstream dataset.
[0070] As a preferred embodiment of the present invention, in step S2, the process of preprocessing the audio signal to be detected to obtain audio time-frequency features specifically includes: performing mono conversion and length alignment on the audio signal to be detected; performing extreme value normalization on the length-aligned audio signal; using maximum normalization to map the audio waveform amplitude to a preset range; on this basis, performing a 2048-point Fast Fourier Transform (FFT) on the normalized audio signal to obtain spectral information; then using a Mel filter bank with a preset dimension to filter and map the spectrum; and taking the logarithm of the filtering result to obtain a two-dimensional time-frequency log-Mel spectrogram, which serves as the input data for the subsequent neural network model.
[0071] In a preferred embodiment of the present invention, step S3, the process of data augmentation of the preprocessed audio time-frequency features, specifically includes: First, randomly shifting the preprocessed audio time-frequency features along the time axis according to a Gaussian distribution to obtain new training data; Second, performing linear interpolation weighted fusion on two randomly arranged independent audio time-frequency features within the same data batch using a beta-distributed mixing parameter, and performing synchronous restricted linear interpolation on the event labels corresponding to the two features to obtain new training data; Third, randomly dividing frequency bands along the frequency dimension and applying a dynamic filter composed of logarithmic-level random gain multipliers to dynamically perturb the frequency response characteristics to obtain new training data; Finally, selecting the time or frequency window to be masked according to the mask, masking the selected window with continuous feature frames to obtain the final training data, and using the output of the above hybrid augmentation strategy as the final data-augmented audio time-frequency features. Using data augmentation methods can overcome the problem of insufficient training data and enhance the generalization ability of the model.
[0072] As a preferred embodiment of the present invention, such as Figure 2 As shown, in step S4, feature extraction of the augmented audio time-frequency features includes initial spatial feature extraction and multi-scale coordinate attention depth feature extraction. The process of finally obtaining the depth features specifically includes: the input augmented audio time-frequency features first enter the CNN module, sequentially passing through a two-dimensional convolutional layer, a normalization layer, a GLU-gated linear unit for nonlinear activation processing, and an average pooling layer for pooling downsampling processing. After multiple iterative cycles, the initial features are output. Using a multi-layer convolutional neural network (CNN) to extract shallow acoustic features can suppress unimportant background noise features, resulting in high-quality semantic features. Subsequently, the initial features are input to the multi-scale coordinate attention module. The initial features are processed by this module to finally obtain multi-scale fused features.
[0073] As a preferred embodiment of the present invention, such as Figure 3 As shown, the process of the multi-scale coordinate attention module specifically includes: firstly, inputting the initial features into multiple parallel feature extraction branches, the branches including a 1×1 convolution branch, multiple dilated convolution branches with dilation rates set only in the time dimension (dilation rates of 6, 12, 18), and a global average pooling branch for capturing global context information, to obtain local and global features under different receptive fields respectively.
[0074] The b-th dilated convolutional branch sets an increasing dilation rate only in the time dimension, while maintaining local connectivity in the frequency dimension. Its output in the feature map... Points on The calculation formula is as follows:
[0075]
[0076] in, This represents the output feature value of the b-th dilated convolution branch at time t and frequency f. This represents the weight value corresponding to the convolution kernel, X is the initial input feature, and u and v are the sliding indices of the convolution kernel in the time and frequency dimensions, respectively. The dilation rate (6, 12, 18) is set for the b-th dilated convolution branch in the time dimension. This is a bias term.
[0077] The output of the global average pooling branch is then upsampled using bilinear interpolation to restore it to the same spatial size as the original features, and the output features of all parallel branches are concatenated along the channel dimension. Finally, the concatenated features are input into a 1×1 feature fusion layer for channel dimensionality reduction to obtain the multi-scale fused feature X.
[0078] Next, a dual-axis time-frequency decoupling mechanism is used to generate directional attention weights for the multi-scale fused feature X. In the frequency dimension, one-dimensional adaptive average pooling is used to aggregate along the frequency width W, extracting horizontally aggregated features containing key activation time points. The formula for its calculation is:
[0079]
[0080] In the time dimension, one-dimensional adaptive average pooling is used to aggregate along the time height H to extract vertical aggregated features containing information about the distribution of the target acoustic bands. The formula for its calculation is:
[0081]
[0082] The horizontal and vertical aggregated features are then concatenated spatially to obtain intermediate features. These intermediate features are then input into a first convolutional layer and jointly encoded with a non-linear activation function to capture the cross-dimensional interaction between time and frequency dimensions. The encoded intermediate features are then split back into horizontal and vertical tensors along the original concatenation dimension. These horizontal and vertical tensors are then input into independent second and third convolutional layers, respectively, and mapped to the (0,1) interval via a sigmoid activation function, yielding the time-dimensional attention weights corresponding to the c-th channel. Attention weights in the frequency dimension .
[0083] Finally, the multi-scale fusion feature X is combined with the temporal dimension attention weight. Frequency dimension attention weight The formula for element-wise multiplication is as follows:
[0084]
[0085] in, The output is the in-depth feature extraction result. Through the multi-scale coordinate attention module described above, the network can assign higher response weights to the time-frequency region containing the target sound event and strongly suppress silent segments and background noise.
[0086] As a preferred embodiment of the present invention, in steps S5 and S6, the process of performing temporal modeling and classification on the extracted deep features specifically includes: folding and compressing the frequency dimension of the extracted deep spatial features, and inputting it into a bidirectional gated recurrent unit (BiGRU) module for contextual temporal modeling to obtain the first branch features; inputting the output of the BiGRU module into a strong label classification fully connected layer and an attention weight fully connected layer respectively; the strong label classification fully connected layer outputs a strong label prediction result containing the prediction probability of each time frame; the attention weight fully connected layer outputs the attention weight of each time frame after Softmax activation, and uses the attention weight to perform weighted sum pooling on the strong label prediction result to obtain the frame-level and segment-level sound event detection results of the audio signal to be detected.
[0087] As a preferred embodiment of the present invention, in step S7, the process of post-processing and decoding the frame-level and segment-level sound event detection results to obtain the detection results of the sound event to be detected specifically includes: post-processing and decoding the frame-level and segment-level predicted probabilities output by the network to finally output the sound event category that occurred in the audio signal to be detected and its corresponding start time and end time.
[0088] To verify the effectiveness of this invention, this embodiment uses the Polyphonic Sound Event Detection Score (PSDS) as the evaluation metric. The PSDS index is derived from the normalized area under the ROC curve enclosed by a series of coordinates. A higher PSDS value indicates better model performance, greater robustness to label subjectivity, and better insight into data bias and classification stability across sound classes. This invention calculates PSDS1 (focusing on the precise temporal localization ability of sound events) and PSDS2 (focusing on the ability to identify sound categories) to compare different model structures. To reduce the influence of random factors, each experiment is run 20 times using different random seeds, and the optimal result is taken as the final result. The experimental results for different model structures are shown in Table 1.
[0089] Table 1: PSDS1 and PSDS2 with different model structures Model Structure PSDS1 PSDS2 CRNN 0.353 0.553 CRNN-CABM 0.4 0.585 CRNN-SE 0.413 0.607 This example 0.421 0.638
[0090] As can be seen from Table 1, the model structure in this invention has a high accuracy rate, which proves the effectiveness of this invention.
[0091] Matters not covered in this invention are common knowledge.
[0092] All the above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for detecting acoustic events based on the fusion of multi-scale spatial features and coordinate attention, characterized in that, Includes the following steps: Step 1: Obtain the raw audio signal to be detected; Step 2: Preprocess the audio signal to be detected to obtain the audio time-frequency characteristics; Step 3: Perform data augmentation on the preprocessed audio time-frequency features; Step 4: Extract features from the time-frequency features of the augmented audio data, including initial spatial feature extraction and multi-scale coordinate attention depth feature extraction, to obtain depth features; Step 5: Extract contextual information and perform temporal modeling on the extracted deep features to obtain temporal features; Step 6: Input the temporal features into the classifier to obtain the frame-level and segment-level sound event detection results of the audio signal to be detected; Step 7: Post-process and decode the frame-level and segment-level sound event detection results to obtain the sound event detection results to be detected.
2. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 1, characterized in that, The audio signal to be detected is preprocessed to obtain audio time-frequency features, including: The audio signal to be detected is converted to mono and length aligned; The length-aligned signal is subjected to extreme value normalization. The normalized signal is subjected to a short-time Fourier transform and mapped to a Mel filter bank. The logarithm is then taken to obtain a two-dimensional log-Mel spectrogram as the audio time-frequency feature.
3. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 1, characterized in that, The aforementioned data augmentation of the preprocessed audio time-frequency features includes: The audio time-frequency features after data preprocessing are randomly shifted along the time axis according to a Gaussian distribution to obtain the first enhancement result; The time-frequency features and their labels of the preprocessed audio data in the same batch are randomly linearly mixed using mixing parameters that follow a beta distribution to obtain the second enhancement result; The frequency bands are randomly divided in the frequency dimension, and a dynamic filter composed of logarithmic random gain multipliers is applied to perturb the frequency response, resulting in the third enhancement result. Select the time or frequency window to be masked based on the mask, and mask the selected window with continuous feature frames to obtain the fourth enhancement result; The fourth enhancement result is used as the result of data augmentation of audio time-frequency features.
4. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 1, characterized in that, The feature extraction of the time-frequency features of the augmented audio data includes: The augmented audio time-frequency features are input into the CNN module for preliminary feature extraction to obtain initial features; The initial features are input into the multi-scale coordinate attention module for in-depth feature extraction, resulting in in-depth feature extraction results.
5. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 4, characterized in that, The step of inputting the augmented audio time-frequency features into the CNN module for preliminary feature extraction to obtain initial features includes: The preprocessed audio time-frequency features are input into a two-dimensional convolutional layer for feature extraction. The output of the two-dimensional convolutional layer is input into the normalization layer for normalization processing; The normalized result is input into the GLU gated linear unit for nonlinear activation processing; The result after nonlinear activation is input into the average pooling layer for pooling downsampling to obtain the dimensionality reduction result. The above steps are iterated over multiple times, and the output of the last average pooling layer is used as the initial feature.
6. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 4, characterized in that, The process of inputting the initial features into a multi-scale coordinate attention module for in-depth feature extraction to obtain the in-depth feature extraction results includes the following steps: Step 1: Input the initial features into multiple parallel feature extraction branches. Each branch includes a 1×1 convolution branch, multiple dilated convolution branches that set the dilation rate only in the time dimension, and a global average pooling branch to obtain local and global features under different receptive fields. Step 2: Perform bilinear interpolation upsampling on the output of the global average pooling branch, and then concatenate the features of all branches along the channel dimension to obtain the concatenated features; Step 3: Input the spliced features into the feature fusion layer for dimensionality reduction to obtain multi-scale fused features; Step 4: Perform one-dimensional adaptive average pooling on the multi-scale fused features along the frequency dimension and the time dimension respectively to obtain horizontal aggregated features and vertical aggregated features; Step 5: Concatenate the horizontal and vertical aggregated features in the spatial dimension to obtain the intermediate feature; Step Six: Input the intermediate features into the first convolutional layer and perform joint feature encoding with a nonlinear activation function; Step 7: Divide the encoded intermediate features into horizontal and vertical tensors along the original concatenation dimension; Step 8: Input the horizontal tensor and the vertical tensor into the second convolutional layer and the third convolutional layer, respectively; Step 9: Input the outputs of the second and third convolutional layers into the Sigmoid activation function to obtain the attention weights in the time dimension and the frequency dimension. Step 10: Multiply the multi-scale fusion features element-wise with the time dimension attention weights and frequency dimension attention weights to obtain the final in-depth feature extraction result.
7. The acoustic event detection method based on multi-scale spatial features and coordinate attention fusion according to claim 6, characterized in that, The coordinate attention unit in the multi-scale coordinate attention module is used to perform biaxial time-frequency decoupling and attention recalibration on the input multi-scale fused features. The calculation formula is as follows: in, To further analyze the feature extraction results, let c represent the c-th feature channel, h be the height of the time dimension, and w be the width of the frequency dimension. This represents the horizontal aggregated feature at time height h obtained after pooling along the frequency dimension. This represents the vertical aggregated feature along the frequency width w obtained after pooling along the time dimension, where W is the total width dimension of the frequency dimension and H is the total height dimension of the time dimension. and For the feature element values participating in the one-dimensional adaptive average pooling calculation, The fused multi-scale feature values, For the time dimension attention weight corresponding to the c-th channel, The frequency dimension attention weight corresponds to the c-th channel.
8. The acoustic event detection method based on multi-scale spatial feature and coordinate attention fusion according to claim 1, characterized in that, The step of extracting contextual information and performing temporal modeling on the extracted deep spatial features to obtain temporal features, and inputting the temporal features into a classifier, includes: The frequency dimension of the extracted deep spatial features is folded and compressed, and then input into the BiGRU module for contextual temporal modeling to obtain the first branch features. The output of the BiGRU module is input into the strongly labeled classification fully connected layer and the attention weight fully connected layer, respectively. The output of the strong label classification fully connected layer contains the strong label prediction result containing the prediction probability for each time frame; The attention-weighted fully connected layer outputs the attention weight for each time frame after being activated by Softmax. The attention weight is then used to perform weighted summation and pooling on the strong label prediction results to obtain the weak label prediction results that include the prediction probability of the entire audio segment.