An acoustic signal periodicity feature extraction method, a network model training method, a storage medium, and an apparatus

Abstract

This invention relates to a method for extracting periodic features from sound signals, a network model training method, a storage medium, and a device, belonging to the field of sound event detection technology. It addresses the problems of high difficulty in perceiving overlapping acoustic events and poor decoupling effects in current acoustic detection processes. For a sound signal i, it first maps it to a low-dimensional space through two different linear layers to obtain p and g. Then, it performs augmented convolution operations on p and g respectively to obtain pi. conv and g conv For p and g, Fourier features are obtained by feature encoding based on Fourier basis functions and gating mechanisms. For p conv and g conv The Fourier features are obtained in the same way, and the Hadamard product is used to obtain the periodic feature representation. During the training process of the corresponding model, the reconstruction error is calculated for each of the two features and the original signal i, and their L2 norms are calculated. The two L2 norms are then summed to obtain the Fourier series regularization term, which is used for model training.

Description

Technical Field

[0001] This invention belongs to the field of sound event detection technology, and relates to a method for extracting periodic features of sound signals and a network model training method, as well as corresponding storage media and devices. Background Technology

[0002] Sound is ubiquitous in the human world. It is not only a vital medium for communication and expression but also a key signal carrying information within the environment. Humans can quickly identify and understand various sound events through hearing, such as breaking glass, sirens, footsteps, or birdsong, thereby perceiving the state and changes of their surroundings. Simulating this process, enabling computers to "understand" environmental sounds like humans, is a crucial goal of sound perception and understanding research. Among these, Sound Event Detection (SED), as one of the core tasks in the field of computer hearing, aims to automatically identify and label different types of sound events and their occurrence times from continuous audio streams.

[0003] Early sound event detection methods primarily relied on manually designed acoustic features and traditional machine learning algorithms. While these methods performed well in controlled environments, their performance was often limited in noisy, multi-source real-world soundscapes. With the rise of deep learning, researchers have gradually shifted towards end-to-end modeling methods based on neural networks. Convolutional Neural Networks (CNNs) effectively capture local time-frequency features of sound, while Recurrent Neural Networks (RNNs) and their variants (such as LSTM and GRU) excel at modeling temporal dependencies in sound. Furthermore, models incorporating attention mechanisms or Transformer structures further enhance their ability to capture key acoustic patterns, enabling accurate identification and localization of multiple sound events in complex scenes.

[0004] In practical tasks, the annotation cost of audio data is high, especially the difficulty in obtaining precise temporal boundary annotations. To address this, researchers have proposed weakly supervised learning and semi-supervised learning methods. These methods effectively alleviate the data scarcity problem by relying only on segment-level or partially labeled data, combined with unlabeled audio for training. In recent years, strategies such as the teacher-student framework, consistency regularization, and pseudo-labeling have been widely applied to SED tasks, significantly improving the robustness and generalization performance of the models.

[0005] Sound event detection technology (SED) has shown broad application prospects in many fields. It can be used for abnormal sound alarms in smart homes, event recognition and soundscape analysis in urban security monitoring, animal call recognition in ecological monitoring, and multimedia retrieval and automatic annotation. SED technology is also playing an increasingly important role in areas such as automotive driver assistance systems, medical health monitoring, and industrial equipment fault detection. Summary of the Invention

[0006] This invention aims to address the problems of high difficulty in perceiving overlapping acoustic events and poor decoupling effects in current acoustic detection processes.

[0007] A method for extracting periodic features of a sound signal, comprising:

[0008] For the sound signal i, it is first mapped to a low-dimensional space through two different linear layers to obtain p and g. Then, augmented convolution operations are performed on p and g respectively to obtain... and ;against and Fourier features are obtained by feature encoding based on Fourier basis functions and gating mechanisms. ,against and Fourier features were obtained using the same method. ,right and The periodicity is represented by the Hadamard product.

[0009] Furthermore, the formulas for the expanded convolution operation during the expanded convolution operation on p and g are as follows:

[0010] (3)

[0011] Where k is the index of the k-th weight of the convolution kernel. This represents the weight of the k-th convolutional kernel; This represents the time position of the input sequence x. The values ​​at x are set as p and g, which are linearly mapped to the lower-dimensional space, respectively, and d is the expansion rate. The stride indicating "skipping".

[0012] Furthermore, targeting and Fourier features are obtained by feature encoding based on Fourier basis functions and gating mechanisms. The process includes:

[0013] (8)

[0014] Here, gate represents the gate coefficient in the gate control mechanism.

[0015] Furthermore, The derivation process of the form includes:

[0016] Fourier series expansion based on sound signal x , where f is the index of different frequencies. The angular frequency corresponding to the fundamental frequency. , , These are the Fourier coefficients; assuming , , It is obtained through automatic learning via network training, then we have ,in and To obtain weights for automatic learning during network training, then... This represents a feature map based on Fourier basis functions, expressed as a linear transformation of the sound signal x. , They represent , ,get By introducing a dynamic gating mechanism, we can obtain... The final representation of .

[0017] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method for extracting periodic features of a sound signal.

[0018] A device for extracting periodic features of a sound signal, the device comprising a processor and a memory, the memory storing at least one instruction, the at least one instruction being loaded and executed by the processor to implement the method for extracting periodic features of a sound signal.

[0019] A network model training method including a sound signal periodicity feature extraction step, wherein in the weakly supervised learning training process of the network model containing the sound signal periodicity feature extraction method, a Fourier series regularization term is used in the weakly supervised learning training process, and the construction of the Fourier series regularization term includes the following steps:

[0020] against and The reconstruction error is calculated by comparing it with the original signal i, and denoted as . and Calculate their L2 norms separately, and use the sum of the two L2 norms to obtain the regularized term of the Fourier series.

[0021] Furthermore, the regularization term of the Fourier series , It represents the 2-norm.

[0022] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the network model training method including a periodic feature extraction step of sound signals.

[0023] A network model training device including a periodic feature extraction step for sound signals is disclosed. The device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the network model training method including the periodic feature extraction step for sound signals.

[0024] Beneficial effects:

[0025] This invention first selects a frame from the preprocessed speech signal sequence and projects it onto two different low-dimensional representation spaces, thereby suppressing redundant information while preserving the main features. Based on the low-dimensional representation after linear mapping, expanded convolution is used to process the signal. Expanded convolution introduces spacing between convolution kernel elements, allowing the kernel to have a larger receptive field without increasing the number of parameters, thus better capturing long-term dependent regularities and periodic patterns in the audio signal. To characterize the audio signal at the frequency domain, Fourier transforms are performed on both the signal mapped to the low-dimensional space and the signal after expanded convolution. This step reveals the energy distribution of the signal across frequency components, providing a foundation for capturing periodic features. The two signals obtained from the Fourier transform are then multiplied by a Hadamard product. This operation fuses the features of the original signal and the expanded convolution result at the frequency domain, strengthening their correlation at the same frequency components, thereby obtaining a more explicit representation of periodic features. This periodicity can highlight the differences in frequency and time structure between different sound sources in overlapping acoustic events, thereby providing a stable structured prior for the subsequent perception and decoupling of sound sources. This reduces the difficulty of perceiving and decoupling overlapping acoustic events and improves the accuracy of perception and decoupling. The method of this invention makes the perception and processing of overlapping acoustic events relatively easy.

[0026] To further constrain the model's consistency in the time and frequency domains, the reconstruction errors of the two Fourier-transformed audio signals and the original signals were calculated separately, and the sum of their L2 norms was obtained to derive the Fourier series regularization term. This regularization term, when combined with the student-teacher framework, can enhance the stability of the teacher model's pseudo-labels, enabling the student model to obtain more reliable and consistent guidance in complex overlapping scenarios. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of a method for extracting periodic features from sound signals.

[0028] Figure 2 A comparison chart of the experimental results of the model corresponding to the method described in this invention and the baseline results;

[0029] Figure 3 The corresponding model effect diagram of the method described in this invention. Detailed Implementation Specific implementation method one:

[0031] This embodiment is a method for extracting periodic features of sound signals, including the following steps:

[0032] S1. Select one frame of the preprocessed audio signal, denoted as... ,in Let p represent the amplitude value of the m-th sampling point, and M represent the number of sampling points. The m-th sampling point is mapped to the low-dimensional space through two different linear layers, denoted as p and g respectively.

[0033] S2. Perform extended convolution operations on p and g respectively to obtain a concise and regular audio signal. and ;

[0034] S3, Regarding the phase parameter tensor With the corresponding magnitude tensor Feature encoding is performed based on Fourier basis functions to obtain two Fourier features, denoted as . and ;

[0035] S4. Select the features of the two audio signals after Fourier transform. and The periodicity is represented by the Hadamard product. Specific Implementation Method Two:

[0037] This embodiment is a computer storage medium that stores at least one instruction, which is loaded and executed by a processor to implement the method for extracting periodic features of a sound signal.

[0038] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0040] This embodiment is a device for extracting periodic features of sound signals. The device includes a processor and a memory. It should be understood that this includes any device including a processor and a memory described in this invention. The device may also include other units or modules that perform display, interaction, processing, control, and other functions through signals or instructions.

[0041] The memory stores at least one instruction, which is loaded and executed by the processor to implement the method for extracting periodic features of a sound signal.

[0042] Those skilled in the art will understand that at least one stored instruction constitutes a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0043] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0047] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations. Specific implementation method four:

[0049] This embodiment is a network model training method that includes a sound signal periodicity feature extraction step. Its main feature is that during the weakly supervised learning training process of a network model containing the sound signal periodicity feature extraction method described in Specific Embodiment 1, a Fourier series regularization term is used in the weakly supervised learning training process. The construction of the Fourier series regularization term includes the following steps:

[0050] against and The reconstruction error is calculated by comparing it with the original signal i, and denoted as . and Calculate their L2 norms separately, and then use the sum of the two L2 norms to obtain the regularized term of the Fourier series. . Specific implementation method five:

[0052] This embodiment is a computer storage medium that stores at least one instruction. The at least one instruction is loaded and executed by a processor to implement the network model training method that includes a periodic feature extraction step for sound signals.

[0053] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method six:

[0055] This embodiment is a network model training device that includes a periodic feature extraction step for sound signals. The device includes a processor and a memory. It should be understood that this includes any device described in this invention that includes a processor and a memory. The device may also include other units or modules that perform display, interaction, processing, control, and other functions through signals or instructions.

[0056] The memory stores at least one instruction, which is loaded and executed by the processor to implement the network model training method that includes a periodic feature extraction step for sound signals.

[0057] Those skilled in the art will understand that at least one stored instruction constitutes a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0058] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0059] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0060] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0061] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0062] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0063] In this embodiment of the invention, a preprocessed audio signal frame is selected as the feature vector during the model training phase. Two different linear layers are used to map the audio signal to a low-dimensional space. Then, extended convolution is applied to the signal to increase the receptive field and capture long-term patterns in the audio signal. Fourier transforms are performed on the original audio signal and the audio features after extended convolution to capture the periodic features of the audio signal in the frequency domain. The two audio signals after Fourier transform are multiplied by a Hadamard product to obtain representations of the periodic features of the audio signal at different frequency scales. Then, the reconstruction error of the original signal is calculated for the two audio signal features after Fourier transform, and the sum of their L2 norms is obtained to obtain the Fourier series regularization term.

[0064] Example:

[0065] In this embodiment of the invention, to address the problem of mutual interference of features from different sound sources and high overlap of time-frequency structures in overlapping acoustic events, which makes accurate model identification difficult, the proposed method for extracting periodic features of sound signals based on Fourier convolution units can perform structured decomposition of mixed signals at both the time and frequency domains. This embodiment of the invention uses the DESED dataset, which contains various sounds from everyday home environments. By applying the algorithm proposed in this invention to this dataset, effective perception of periodic information in sound signals is achieved, as detailed below. Figure 1 As shown, the representation of periodic features includes the following steps:

[0066] S1. Select one frame of the preprocessed audio signal, denoted as... ,in Let p represent the amplitude value of the m-th sampling point, and M represent the number of sampling points. The m-th sampling point is mapped to the low-dimensional space through two different linear layers, denoted as p and g respectively.

[0067] First, a frame, denoted as i, is selected from the preprocessed speech signal sequence and mapped to a low-dimensional space through two different linear layers, yielding p and g. These two linear layers have different weight parameters, allowing the input signal to be projected into two different low-dimensional representation spaces, thereby suppressing redundant information while preserving the main features. The formulas for the linear layer mapping are shown below:

[0068] (1)

[0069] (2)

[0070] In the formula The input signal is M, the input dimension is M, and the output dimension is output_dim. This is the weight matrix. The subscripts p and g of W and b represent the bias term, respectively, indicating two different linear layers.

[0071] S2. Perform expanded convolution operations on p and g respectively to obtain a concise and regular audio signal. and ;

[0072] Based on the low-dimensional representation after linear mapping, one-dimensional augmented convolutions are applied to p and g respectively. Augmented convolutions introduce spacing between kernel elements, allowing the kernel to have a larger receptive field without increasing the number of parameters. This enables better capture of long-term dependent regularities and periodic patterns in audio signals. The resulting output signal is simpler and more temporally structured than the original signal. The formula for augmented convolution is as follows:

[0073] (3)

[0074] Where k is the index of the k-th weight of the convolution kernel. This represents the weight of the k-th convolutional kernel; This represents the time position of the input sequence x. The value at point d is the expansion rate. The stride representing "skipping" is set with an expansion rate d=2, the kernel size K is set to 3, and the receptive field location index of the one-dimensional expanded kernel is used. Let x be a linear mapping to p and g in the lower-dimensional space, respectively.

[0075] S3, Regarding the phase parameter tensor With the corresponding magnitude tensor Feature encoding is performed based on Fourier basis functions to obtain two Fourier features, denoted as . and ;

[0076] For signals The Fourier transform of can be approximated by Fourier series expansion:

[0077] (4)

[0078] in, Used to represent sequence o or The single value at the m-th position, the sequence o in the derivation process. This is used to represent two sequences to illustrate the derivation process; m is the discrete sampling point of the signal in the time domain, F represents the number of terms in the Fourier series expansion of the signal, and f is the index of different frequencies. c is the angular frequency corresponding to the fundamental frequency. , The Fourier coefficients can be obtained by definite integrals, while and This can be used to calculate the spectrum of the Fourier transform;

[0079] The Fourier series expansion is discussed below. The Fourier series expansion of the sound signal x can be fitted by a shallow network, and its derivation process is as follows:

[0080] (5)

[0081] in, This represents an eigenmap based on Fourier basis functions, where c represents the Fourier coefficients. and Both indicate that the elements within the parentheses are subjected to cosine or sine operations, with T as a superscript indicating transpose; Fourier coefficients c, , Instead of calculating through definite integrals, the parameters are automatically learned through network training. In this case, c represents the learnable network parameters. and This can be expressed as a function of the signal x, through two learnable network parameters. , Mapped to obtain, The linear projection weights of the cosine terms are represented. Represents the linear projection weights of the sine term; It was also replaced with learnable network parameters. , The linear projection weights determine the angular frequency corresponding to the f-th frequency term; after splicing... , , They can be combined into weight matrices respectively. , and Among the above parameters, and It can be further merged into , This represents the learnable frequency domain weight matrix used to model the Fourier coefficients, while The angular frequency weights represent the fundamental frequency; the weight matrices corresponding to the Fourier coefficients and angular frequencies represent two different types of periodic information. In the calculation process of formula (5), the two types of weight matrices are sequentially applied to the feature representation in a series structure. Gradient analysis shows that the series structure leads to strong gradient coupling between the two, that is, the gradient of any one type of weight will be directly affected by the other type of parameter. To alleviate this coupling and improve their modeling independence, the series mode is changed to the parallel mode, so that the two types of weights can learn their respective frequency domain and time domain information, as shown in the following formula:

[0082] (6)

[0083] in, This represents the activation function; similarly, and It also functions as an activation function;

[0084] Using the above derivation process, it is possible to obtain the sound signal vector. and as well as and Approximate calculations are performed using Fourier series expansion. express , express The formula is shown below:

[0085] (7)

[0086] The above formula simply places periodic and linear features side by side without a mechanism to dynamically adjust their weights. Therefore, a dynamic gating mechanism is introduced. The learnable gating coefficients (gate) are represented by activation functions, and their Fourier feature mapping formula is shown below:

[0087] (8)

[0088] This structure contains only phase parameters. Enter the Fourier basis functions (cos, sin), amplitude parameters Non-orthogonal amplitude features are preserved through a linear path. In a one-dimensional dilated convolutional structure, the input receptive field position corresponding to the k-th convolutional kernel can be denoted as... Because dilated convolutions sample at predetermined dilation rates along the time axis, different kernel positions... This corresponds to different local context ranges. To enable the model to explicitly encode this location-related local context information, the phase and magnitude parameters after dilated convolution, i.e. and This is followed by Fourier feature mapping. Therefore, the receptive field location... The Fourier representation of is defined as: ,in and These are the original phase parameter p and magnitude parameter g after dilated convolution, and the values ​​are indexed from the corresponding receptive field. Local features extracted at the point. Obtained through convolution operations. and In fact, the receptive field location of the convolution kernel has been integrated. The included local context, therefore and Equivalent; using the same method to indicate That is: when correspond hour, correspond , The corresponding representation is as follows ;when correspond hour, correspond , The corresponding representation is as follows .

[0089] S4. Select the features of the two audio signals after Fourier transform. and The periodicity is represented by the Hadamard product.

[0090] The sound signal o (actually means) , ()) and receptive field location index of expanded convolution kernel The local features are subjected to Fourier mapping, and the time-domain convolution can be represented as a frequency-domain dot product using the convolution theorem. This Hadamard product can be simulated by a two-branch network, denoted as: Since the amplitude of the Fourier transform at each frequency is equal to the sum of the squares of the cosine and sine components, the periodicity of the Fourier convolution unit can be expressed as:

[0091] (9)

[0092] in, The f-th element of the periodic feature corresponds to the outer Fourier transform of the convolution. The energy of the f-th frequency; the subscripts o and k represent the receptive field indexes of the sound signal o and the expanded convolution kernel, respectively. The parameters; periodic features can be used to combine and fuse with features in a weakly supervised learning framework to jointly participate in the training of student and teacher models.

[0093] This periodicity can highlight the differences in frequency and time structure of different sound sources in overlapping acoustic events, thereby providing a stable structured prior for the perception and decoupling of subsequent sound sources, thus reducing the difficulty of perception and decoupling of overlapping acoustic events and improving the accuracy of perception and decoupling.

[0094] To further constrain the consistency of the model in the time and frequency domains, the two audio signals after Fourier transform were calculated separately. and The reconstruction error of the original signal i is calculated separately, and the sum of the second norms of the error vectors is obtained to obtain the Fourier series regularization term. The regularization term is then used in the training process of weakly supervised learning, which can effectively improve the training effect of the model and thus improve the model accuracy.

[0095] The construction of the Fourier series regularization term includes the following steps:

[0096] Select two audio signals after Fourier transform and The reconstruction error is calculated by comparing it with the original signal i, and denoted as . and ;

[0097] To further constrain the consistency of the model in the time and frequency domains, the reconstruction error between the Fourier-transformed audio signal and the original signal i is calculated, and the formula is shown below:

[0098] (10)

[0099] Here, the subscript x takes the value 0 or k, corresponding to 0 and the expanded convolutional kernel, respectively. The parameters are used to obtain the reconstruction error, which is denoted as follows: and .

[0100] The above process is actually and The reconstruction error is calculated by comparing the original signal i with the original signal i, and the results are obtained. and .

[0101] Select vector and Calculate their L2 norms separately, and use the sum of the two L2 norms to obtain the regularized term of the Fourier series.

[0102] Finally, for the error vector and Calculate its L2 norm to measure the magnitude of the error; sum the L2 norms of both to obtain the final Fourier series regularization term; the Fourier series regularization term is calculated as follows:

[0103] (11)

[0104] in, It represents the 2-norm.

[0105] The Fourier series regularization term can be added to the objective function of the student-teacher model as part of the objective function, participating in the training process of weakly supervised learning. When combined with the student-teacher framework, this regularization term can also enhance the stability of the pseudo-labels of the teacher model, enabling the student model to obtain more reliable and consistent guidance in complex overlapping scenarios; thus effectively improving the training effect of the model and enhancing its accuracy.

[0106] This invention compares the acoustic event detection task dataset from the 2024 Acoustic Scene and Event Detection and Classification Challenge (DCASE 2024) with baseline methods. The dataset consists of real recordings from everyday home environments and has been used in previous challenges. The baseline method employs a student-teacher model; both models have identical structures, using a multi-task acoustic model with CNN and BiGRU architectures, capable of handling strong, weak, and soft labels simultaneously, and improving robustness by ignoring missing labels. The evaluation metrics utilize the official evaluation script, where psds1_eval represents the overall PSDS score calculated under parameters such as the maximum false positive rate, and is the primary final performance metric. psds1_sed_scores represents the score obtained by directly calculating the frame-level prediction probabilities output by the model using the PSDS formula, used to analyze the quality of the model's predictions. The comparison results between this invention and the baseline are as follows: Figure 2 As shown.

[0107] Experimental results show that, compared with the original baseline method, the psds1_eval and psds1_sed_scores of the method in this invention are improved by 0.01 and 0.013, respectively. The experimental results demonstrate that the proposed method for extracting periodic features of sound signals can effectively improve the performance of sound event detection.

[0108] The final result is as follows Figure 3 As shown in the figure, the evaluation metrics demonstrate that the model can effectively perceive the periodicity of sound signals.

[0109] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the scope of this invention.

[0110] The above examples of this invention are merely illustrative of the computational model and process of this invention, and are not intended to limit the implementation of this invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of this invention are still within the scope of protection of this invention.

Claims

1. A method for extracting periodic features from sound signals, characterized in that, include: For the sound signal i, it is first mapped to a low-dimensional space through two different linear layers to obtain p and g. Then, augmented convolution operations are performed on p and g respectively to obtain... and ;against and Fourier features are obtained by feature encoding based on Fourier basis functions and gating mechanisms. ,use to indicate : （8） Where gate represents the gate coefficient in the gate control mechanism; against and Fourier features were obtained using the same method. ,right and The periodicity is represented by the Hadamard product. The derivation process of the form includes: Fourier series expansion based on sound signal x , where f is the index of different frequencies, The angular frequency corresponding to the fundamental frequency. , , These are the Fourier coefficients; assuming , , It is obtained through automatic learning via network training, then we have ,in and To obtain weights for automatic learning during network training, then... This represents a feature map based on Fourier basis functions, expressed as a linear transformation of the sound signal x. , They represent , ,get By introducing a dynamic gating mechanism, we can obtain... The final representation; This represents the activation function.

2. The method for extracting periodic features of a sound signal according to claim 1, characterized in that, The formulas for the expanded convolution operation during the expansion convolution operation on p and g are as follows: （3） Where k is the index of the nth weight of the convolution kernel. This represents the weight of the nth convolutional kernel; This represents the time position of the input sequence x. The values ​​at x are set as p and g, which are linearly mapped to the lower-dimensional space, respectively, and d is the expansion rate. The stride indicating "skipping".

3. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the method for extracting periodic features of a sound signal as described in any one of claims 1 to 2.

4. A device for extracting periodic features of sound signals, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the method for extracting periodic features of a sound signal as described in any one of claims 1 to 2.

5. A network model training method including a periodic feature extraction step for sound signals, characterized in that, In the weakly supervised learning training process of a network model containing the sound signal periodicity feature extraction method according to any one of claims 1 to 2, a Fourier series regularization term is used in the weakly supervised learning training process, and the construction of the Fourier series regularization term includes the following steps: against and The reconstruction error is calculated by comparing it with the original signal i, and denoted as . and Calculate their L2 norms separately, and use the sum of the two L2 norms to obtain the regularized term of the Fourier series.

6. A network model training method including a sound signal periodicity feature extraction step according to claim 5, characterized in that, Fourier series regularization term , It represents the 2-norm.

7. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the network model training method including the extraction of periodic features of sound signals as described in any one of claims 5 to 6.

8. A network model training device including a step for extracting periodic features of sound signals, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the network model training method including a periodic feature extraction step of sound signals as described in any one of claims 5 to 6.

Images