Bird repelling audio synthesis method and device based on few-shot concept learning

Abstract

The application provides a bird repelling audio synthesis method and device based on few-shot concept learning, which comprises the following steps: obtaining bird repelling audio samples corresponding to a target bird song concept, generating a bird repelling audio few-shot dataset, assigning a concept placeholder and constructing a text template to form a few-shot concept training set; inputting the few-shot concept training set into a pre-trained audio diffusion generation model to generate training latent variables and a conditional vector; updating the embedding vector corresponding to the concept placeholder under the condition that the model parameters of the frozen model except the embedding vector corresponding to the concept placeholder are updated to obtain a target concept embedding; and performing conditional guided inverse diffusion denoising based on the target concept embedding to generate bird repelling audio corresponding to the target bird song concept. The technical scheme of the application realizes few-shot personalized synthesis of bird repelling audio, improves the target voiceprint consistency and resource expansion efficiency.

Description

Technical Field

[0001] This invention belongs to the field of bird control technology for power transmission lines, specifically relating to a method and device for synthesizing bird-repelling audio based on few-sample concept learning. Background Technology

[0002] In power system operation and maintenance scenarios, areas surrounding transmission line towers, substation structures, distribution line crossarms, insulator strings, and other power equipment are susceptible to damage from bird activity, such as perching, nesting, foraging, and excretion. This can lead to problems like line short circuits, insulation flashovers, equipment corrosion, and power outages. To mitigate the impact of bird damage on the safe operation of the power grid, bird deterrence devices are typically deployed in power systems to drive birds away from target areas using acoustic methods. In acoustic bird deterrence scenarios, different target bird species respond differently to specific predator sounds, warning sounds, deterrent sounds, and sounds of fear. Therefore, developing bird deterrence audio with corresponding acoustic signatures for specific bird species is crucial for improving bird deterrence effectiveness.

[0003] In existing technologies, bird-repelling audio typically uses fixed audio from a pre-set audio library, manually recorded sounds of bird predators, warning sounds, or other highly stimulating sounds as playback content, or directly generates audio samples using a general audio generation model. In the field of audio generation, conditional diffusion models, latent variable coding models, and text-based conditional guidance models have been used for speech, music, and other sound content generation tasks. By inputting reference audio, text prompts, or conditional features, corresponding audio content can be generated or converted.

[0004] However, existing technologies still suffer from insufficient adaptability in personalized bird-repelling audio synthesis scenarios. First, some solutions rely on large-scale training samples and sufficient target audio data. For aggressive, warning, threatening, or frightened calls of specific target birds, obtaining a large number of high-quality samples in real-world scenarios is often difficult. This makes it challenging for existing general-purpose audio generation models to accurately learn the fine-grained semantics and voiceprint features of specific bird calls under limited sample conditions. Second, when faced with a small number of target bird audio samples, existing general-purpose audio generation methods are prone to issues such as inconsistencies between the generated results and the target bird call concepts, insufficient retention of core voiceprint features, and low distinguishability between different target bird species. This makes it difficult to meet the needs of targeted bird repelling in power system scenarios. Third, existing bird-repelling audio resources are usually based on fixed audio files. Expanding to include new bird species or new call types often requires collecting, organizing, and producing an entire set of audio resources again. This results in high costs and low update efficiency for expanding the bird-repelling sound source library, making it difficult to adapt to the diverse target bird species and significantly changing scenario requirements of the open environment of power systems.

[0005] Therefore, there is an urgent need for a bird-repelling audio synthesis method based on few-sample concept learning to improve the learning ability of specific target bird call concepts under few-sample conditions, target voiceprint consistency, and the efficiency of expanding bird-repelling audio resources. Summary of the Invention

[0006] This invention provides a method for synthesizing bird-repelling audio based on few-sample concept learning. The method constructs a bird-repelling audio synthesis scheme that includes a few-sample dataset of bird-repelling audio, concept placeholder text templates, a pre-trained audio diffusion generation model, concept embedding learning, and conditionally guided inverse diffusion generation. Based on freezing the main parameters of the pre-trained audio diffusion generation model, only the embedding vectors corresponding to the concept placeholders are optimized. Based on the learned target concept embedding, the bird-repelling audio corresponding to the target bird call concept is generated. This solves the problems in existing bird-repelling audio generation processes, such as difficulty in learning specific target bird call concepts under few-sample conditions, difficulty in maintaining fine-grained features of target voiceprints, and difficulty in quickly expanding personalized bird-repelling audio resources.

[0007] In a first aspect, the present invention provides a method for synthesizing bird-scare audio based on few-shot concept learning, the method comprising the following steps: Obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. The few-sample concept training set is input into a pre-trained audio diffusion generation model to generate training latent variables and a conditional vector containing the concept placeholders; Freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder, predict the noise residual based on the training latent variables and the conditional vector, update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise, and obtain the target concept embedding. Based on the text generation prompt containing concept placeholders and the target concept embedding, the generation conditions for generating the target bird song concept are generated. Based on the generation conditions, condition-guided inverse diffusion denoising is performed to generate the target latent variable. Based on the target latent variables, generate bird-repelling audio corresponding to the target bird call concept.

[0008] Using the above scheme, the present invention provides a method for synthesizing bird-repelling audio based on few-shot concept learning. This method generates a few-shot dataset of bird-repelling audio based on bird-repelling audio samples corresponding to the target bird-calling concept, and constructs a text template containing concept placeholders to achieve few-shot training representation of the target bird-calling concept. By inputting the few-shot concept training set into a pre-trained audio diffusion generation model, and updating only the embedding vectors corresponding to the concept placeholders while freezing the main parameters of the pre-trained audio diffusion generation model, it achieves few-shot learning of fine-grained voiceprint features of the target bird-calling concept. Furthermore, by generating the generation conditions of the target bird-calling concept based on the target concept embedding and performing condition-guided inverse diffusion denoising, it achieves the synthesized output of bird-repelling audio corresponding to the target bird-calling concept, significantly improving the generation accuracy, voiceprint consistency, and expansion flexibility of specific target bird-calling concepts under few-shot conditions.

[0009] In some embodiments of the present invention, generating a small sample dataset of bird-repelling audio based on the bird-repelling audio samples includes: The bird-repelling audio samples are subjected to target voiceprint detection. Audio samples with incomplete target calls, background noise exceeding a preset threshold, multiple species voiceprints mixed, or inconsistent species labeling are removed to obtain effective bird-repelling audio samples. The effective bird-repelling audio samples are processed by unifying the sampling rate, unifying the sound channels, normalizing the loudness, pruning silence, and fixing the segment length to generate the bird-repelling audio small sample dataset. The number of audio samples in the bird-repelling audio small sample dataset is 3 to 20.

[0010] In some embodiments of the present invention, the construction of a text template including the concept placeholder includes: Obtain the species tag, call type tag, and bird deterrence scene tag corresponding to the target bird call concept; The text template is generated by combining the concept placeholder with at least one of the species tag, the call type tag, and the bird deterrence scene tag; wherein, different target bird call concepts correspond to different concept placeholders, and the target bird call concepts include the target bird's aggressive call concept, warning call concept, deterrent call concept, or fear call concept.

[0011] In some embodiments of the present invention, the pre-trained audio diffusion generation model includes an audio time-frequency representation conversion module, a latent variable encoding module, a text conditional encoding module, a diffusion denoising module, and an audio decoding module; The audio time-frequency representation conversion module converts the audio samples in the few-sample concept training set into audio time-frequency representations; The latent variable encoding module encodes the audio time-frequency representation to generate the training latent variables; The text condition encoding module encodes the text template to generate a condition vector containing the concept placeholders; The diffusion denoising module predicts the noise residual based on the training latent variables and the conditional vector; The audio decoding module decodes the target latent variable to generate the bird-repelling audio.

[0012] In some embodiments of the present invention, the step of predicting the noise residual based on the training latent variables and the conditional vector includes: Sample diffusion time steps from a preset set of diffusion time steps; Based on the noise variance parameter and signal preservation parameter corresponding to the diffusion time step, Gaussian noise is injected into the training latent variables to generate noisy latent variables. The noisy latent variable, the diffusion time step, and the conditional vector are input into the diffusion denoising module of the pre-trained audio diffusion generation model to predict the noise residual.

[0013] In some embodiments of the present invention, updating the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise to obtain the target concept embedding includes: A concept embedding learning loss is constructed based on the difference between the noise residual and the injected noise; The embedding vectors corresponding to the concept placeholders are updated by backpropagation based on the concept embedding learning loss. During the update process, the model parameters in the pre-trained audio diffusion generation model, except for the embedding vector corresponding to the concept placeholder, remain unchanged; The updated concept placeholder corresponding to the embedding vector is determined as the target concept embedding.

[0014] In some embodiments of the present invention, the diffusion denoising module of the pre-trained audio diffusion generation model includes a cross-attention unit; The cross-attention unit maps the current latent variable features to a query matrix, and the condition vector to a key matrix and a value matrix; Attention features are calculated based on the query matrix, the key matrix, and the value matrix; The attention features are injected into the latent variable denoising process of the diffusion denoising module to generate denoised features corresponding to the target bird song concept.

[0015] In some embodiments of the present invention, the step of performing condition-guided inverse diffusion denoising based on the generation conditions to generate target latent variables includes: Initial latent variables are sampled from a standard Gaussian distribution; At each inverse diffusion time step, the current latent variable, the current inverse diffusion time step, and the generation conditions are input into the diffusion denoising module to generate the current noise prediction result; Based on the current noise prediction result, the current latent variable is denoised and updated to generate the latent variable corresponding to the next inverse diffusion time step; The latent variable obtained at the end of the reverse diffusion is determined as the target latent variable.

[0016] In some embodiments of the present invention, after generating the bird-repelling audio corresponding to the target bird call concept based on the target latent variable, the method further includes: The bird-repelling audio is subjected to loudness correction, beginning and end smoothing, duration adjustment and artifact removal to generate candidate bird-repelling audio; Based on acoustic feature distance, species identification results, or text-audio matching scores, the candidate bird-scare audio is subjected to target voiceprint consistency verification. Candidate bird-repelling audios that pass the target voiceprint consistency verification are archived according to species category, call type, bird-repelling scene tag, and playback parameters to generate a personalized bird-repelling audio resource library.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: by constructing a few-sample dataset of bird-repelling audio and building a text template containing concept placeholders, the target bird call concept can form a trainable concept expression under limited sample conditions; by inputting the few-sample concept training set into a pre-trained audio diffusion generation model, and updating only the embedding vectors corresponding to the concept placeholders while freezing the main parameters of the model, the fine-grained voiceprint features of the target bird call concept can be stably injected into the generation model; furthermore, by generating the generation conditions of the target bird call concept based on the target concept embedding and performing condition-guided inverse diffusion denoising, personalized bird-repelling audio synthesis of target bird call concepts such as aggressive calls, warning calls, deterrent calls, or fright calls is achieved, significantly improving the voiceprint consistency, generation controllability, and resource expansion efficiency of bird-repelling audio.

[0018] A second aspect of the present invention provides a bird-scare audio synthesis system based on few-shot concept learning, comprising: The training set construction module is used to obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. The latent variable generation module is used to input the few-sample concept training set into the pre-trained audio diffusion generation model to generate training latent variables and a condition vector containing the concept placeholders. The concept learning module is used to freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder, predict the noise residual based on the training latent variables and the conditional vector, and update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise to obtain the target concept embedding. The condition generation module is used to generate conditions for the target bird song concept based on the text containing concept placeholders and the target concept embedding, and to perform condition-guided inverse diffusion denoising based on the generation conditions to generate target latent variables. An audio synthesis module is used to generate bird-repelling audio corresponding to the target bird call concept based on the target latent variable.

[0019] A third aspect of the present invention provides a bird-repelling audio synthesis device based on few-shot concept learning, characterized in that the device includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, and when the computer instructions are executed, the device implements the bird-repelling audio synthesis method based on few-shot concept learning.

[0020] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0021] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] In the attached diagram: Figure 1 This is a flowchart illustrating a bird-scare audio synthesis method based on few-shot concept learning, provided as an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a bird-scare audio synthesis system based on few-shot concept learning, provided as an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0027] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0028] Figure 1 This is a flowchart illustrating a bird-scare audio synthesis method based on few-shot concept learning, provided by an embodiment of the present invention.

[0029] The method in this embodiment is applied to bird control scenarios in power systems. These scenarios include areas surrounding transmission line towers, substation structures, distribution line crossarms, insulator strings, and other power equipment areas where acoustic methods are needed to reduce the risk of bird damage. The bird-repelling audio can be the sounds of birds' natural enemies, calls of target birds of prey, warning calls, deterrent calls, frightened calls, or other acoustic signals that can stimulate and drive away target birds. This invention addresses the problem of personalized audio synthesis of the voiceprints of target birds of prey or other deterrent birds in bird control scenarios, focusing on solving the problems that existing general-purpose audio diffusion models struggle to accurately learn the fine-grained semantics of specific bird calls under limited sample conditions and are difficult to rapidly expand the bird-repelling sound source library.

[0030] This embodiment aims to address the challenges in personalized bird-repelling audio synthesis, such as the difficulty in stably learning specific target bird call concepts under few-sample conditions, the inaccurate preservation of fine-grained target voiceprint features, and the high cost of adding new bird call concepts. It achieves controllable synthesized output of bird-repelling audio corresponding to target bird call concepts by combining concept placeholders, a few-sample concept training set, and conditionally guided inverse diffusion generation. The technical approach of this invention is as follows: constructing a few-sample dataset of target bird-repelling audio; constructing an audio diffusion generation model including textual conditional encoding and latent variable encoding; performing few-sample concept embedding learning based on concept placeholders; performing conditionally guided inverse diffusion generation based on the learned concept embeddings; and forming an incrementally expandable resource library.

[0031] Example 1, as Figure 1 As shown, the present invention provides a method for synthesizing bird-scare audio based on few-shot concept learning, the method comprising the following steps: S1. Obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. S2. Input the few-sample concept training set into the pre-trained audio diffusion generation model to generate training latent variables and a conditional vector containing the concept placeholders; S3. Freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder. Predict the noise residual based on the training latent variables and the conditional vector. Update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise to obtain the target concept embedding. S4. Generate a target bird song concept based on the text prompt containing concept placeholders and the target concept embedding, and perform condition-guided inverse diffusion denoising based on the generation conditions to generate target latent variables. S5. Generate bird-repelling audio corresponding to the target bird call concept based on the target latent variables.

[0032] Using the above scheme, the present invention provides a method for synthesizing bird-repelling audio based on few-shot concept learning. This method generates a few-shot dataset of bird-repelling audio based on bird-repelling audio samples corresponding to the target bird-calling concept, and constructs a text template containing concept placeholders to achieve few-shot training representation of the target bird-calling concept. By inputting the few-shot concept training set into a pre-trained audio diffusion generation model, and updating only the embedding vectors corresponding to the concept placeholders while freezing the main parameters of the model, it achieves few-shot learning of fine-grained voiceprint features of the target bird-calling concept. Furthermore, by generating the generation conditions of the target bird-calling concept based on the target concept embedding and performing condition-guided inverse diffusion denoising, it achieves the synthesized output of bird-repelling audio corresponding to the target bird-calling concept, significantly improving the generation accuracy, target voiceprint consistency, and subsequent concept expansion efficiency of the bird-repelling audio under few-shot conditions. In some embodiments of the present invention… In some embodiments of the present invention, generating a small sample dataset of bird-repelling audio based on the bird-repelling audio samples includes: The bird-repelling audio samples are subjected to target voiceprint detection. Audio samples with incomplete target calls, background noise exceeding a preset threshold, multiple species voiceprints mixed, or inconsistent species labeling are removed to obtain effective bird-repelling audio samples. The effective bird-repelling audio samples are processed by unifying the sampling rate, unifying the sound channels, normalizing the loudness, pruning silence, and fixing the segment length to generate the bird-repelling audio small sample dataset. The number of audio samples in the bird-repelling audio small sample dataset is 3 to 20.

[0033] Specifically, in this embodiment, step S1 includes the following sub-steps: S111. Collect representative bird-repelling audio samples of the target birds. The target birds may be birds of prey or other birds with a deterrent effect. The bird-repelling audio samples may be collected based on the aggressive calls, warning calls, deterrent calls, or fright calls of the target species. To reflect the characteristics of learning with few samples, the number of samples for a single target concept is preferably 3 to 20. In a preferred embodiment, approximately 10 high-quality samples may be used for concept learning.

[0034] S112. The original bird-scare audio samples are screened for validity, removing samples with excessive background noise, incomplete target calls, significant multi-species overlap, severe recording distortion, or inconsistent target species labeling. Only candidate segments with clear voiceprint features and well-defined event boundaries are retained. Optionally, a bird sound recognition model is introduced to perform secondary species consistency verification on the candidate samples to improve the purity of the training data.

[0035] S113. Perform uniform preprocessing on the valid samples, unifying the retained samples to a preset sampling rate, a uniform number of channels, and a uniform loudness range, and perform silence trimming, valid signal extraction, and segment length fixing to obtain training audio with a consistent format. In one embodiment, the valid signal interval can be trimmed and aligned into a fixed-length audio segment of 10 seconds to meet the requirements of subsequent batch processing and stable training.

[0036] In some embodiments of the present invention, the construction of a text template including the concept placeholder includes: Obtain the species tag, call type tag, and bird deterrence scene tag corresponding to the target bird call concept; The text template is generated by combining the concept placeholder with at least one of the species tag, the call type tag, and the bird deterrence scene tag; wherein different target bird call concepts correspond to different concept placeholders, and the target bird call concepts include the target bird's aggressive call concept, warning call concept, deterrent call concept, or fear call concept. S114. Establish corresponding species tags, call type tags, and bird-repelling scene tags for each target bird call concept, and construct a text template with concept placeholders. The text template can use the format "a sound of..." In the form of "", These are placeholders for concepts to be learned. You can also add descriptive words such as scenario, intensity, and duration as needed to form a richer set of condition templates.

[0037] S115. Organize the processed audio samples and corresponding text templates into a few-sample concept training set. Each target concept corresponds to an independent set of data pairs, which are used for subsequent embedding vector optimization. For multiple target species or multiple call types, multiple concept training sets can be built in parallel using mutually independent placeholders to achieve rapid expansion of the subsequent bird deterrent sound source library.

[0038] In some embodiments of the present invention, the pre-trained audio diffusion generation model includes an audio time-frequency representation conversion module, a latent variable encoding module, a text conditional encoding module, a diffusion denoising module, and an audio decoding module; The audio time-frequency representation conversion module converts the audio samples in the few-sample concept training set into audio time-frequency representations; The latent variable encoding module encodes the audio time-frequency representation to generate the training latent variables; The text condition encoding module encodes the text template to generate a condition vector containing the concept placeholders; The diffusion denoising module predicts the noise residual based on the training latent variables and the conditional vector; The audio decoding module decodes the target latent variable to generate the bird-repelling audio.

[0039] Specifically, in this embodiment, step S2 includes the following sub-steps: S211. Establish an audio time-frequency representation conversion module to convert the input waveform into a two-dimensional acoustic representation suitable for diffusion modeling, preferably a Mel-spectral representation. This step is used to map the original time-domain audio to a time-frequency domain space that is more suitable for latent variable coding and diffusion denoising modeling.

[0040] S212. Establish a latent variable encoding module, using the encoder part of a pre-trained variational autoencoder to compress the time-frequency representation into a low-dimensional latent variable representation. The latent diffusion model first uses a pre-trained VAE encoder. audio signal The mel spectrum is mapped to latent variables. :

[0041] Latent variables are used to preserve the compact acoustic representations required for subsequent concept learning.

[0042] S213. Establish a text conditional encoding module, using a pre-trained text encoder to encode text templates containing concept placeholders into conditional vectors. Let the text template constructor be... The text encoder is Then the condition vector can be expressed as:

[0043] Map a text template containing concept placeholders to a semantic condition vector corresponding to the target bird song concept.

[0044] S214. Establish a conditional diffusion denoising module, constructing a diffusion denoising network based on U-Net, used to perform stepwise reconstruction from noisy latent variables to target latent variables in the latent variable space. Let the total number of diffusion steps be... In the In each time step, the noise variance is The signal retention coefficient is Their cumulative product is:

[0045] For noisy latent variables at any time step, they can be directly sampled using reparameterization: It can perform latent variable noise addition and denoising training based on noise variance parameters and signal preservation parameters.

[0046] S215. Establish a cross-attention injection and audio decoding module. Introduce a cross-attention mechanism into the diffusion denoising network to inject the text conditional vector into the latent variable denoising process. The general form of cross-attention can be expressed as:

[0047] Among them, the query matrix Obtained by projecting the current latent variable features, the key matrix K and value matrix V are derived from the conditional vector c(S). The latent variables are mapped to obtain semantic guidance for specific bird calls. After denoising, a decoder is used to restore the latent variables to time-domain waveforms, resulting in the final audio output.

[0048] S216. Define target concept placeholders and initialize embedding vectors. Define a dedicated placeholder S for the target bird song concept to be learned. And assign a learnable embedding vector v to it in the vocabulary of the text encoder. The embedding vector can be initialized randomly, with neighboring semantic words, or with predefined acoustic descriptors to improve subsequent convergence stability.

[0049] In some embodiments of the present invention, the step of predicting the noise residual based on the training latent variables and the conditional vector includes: Sample diffusion time steps from a preset set of diffusion time steps; Based on the noise variance parameter and signal preservation parameter corresponding to the diffusion time step, Gaussian noise is injected into the training latent variables to generate noisy latent variables. The noisy latent variable, the diffusion time step, and the conditional vector are input into the diffusion denoising module of the pre-trained audio diffusion generation model to predict the noise residual.

[0050] Further, updating the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise to obtain the target concept embedding includes: A concept embedding learning loss is constructed based on the difference between the noise residual and the injected noise; The embedding vectors corresponding to the concept placeholders are updated by backpropagation based on the concept embedding learning loss. During the update process, the model parameters in the pre-trained audio diffusion generation model, except for the embedding vector corresponding to the concept placeholder, remain unchanged; The updated concept placeholder corresponding embedding vector is determined as the target concept embedding. Specifically, in this embodiment, step S3 includes the following sub-steps: S312. Freeze the main parameters of the audio diffusion generation model. During the training phase, keep the main parameters of the latent variable encoding module, text conditional encoding module, diffusion denoising module, and audio decoding module fixed, and only allow the embedding vectors v corresponding to concept placeholders to be frozen. It participates in gradient updates. This strategy avoids the catastrophic forgetting and overfitting problems that may be caused by full parameter fine-tuning, retaining the model's original general audio generation capabilities while learning specific target bird calls using very few samples.

[0051] S313. Constructing Noisy Latent Variable Training Samples. After mapping a small number of training audio samples to training latent variables via the encoder, random sampling of diffusion time step t is performed during training, and Gaussian noise is injected into the training latent variables to generate noisy latent variables. This step is used to simulate the forward perturbation process in the training of the diffusion model, ensuring consistency between concept embedding learning and the diffusion denoising task.

[0052] S314. Calculate the concept embedding learning loss. This involves handling noisy latent variables. Time step and includes placeholder S The condition vector c(S) A common input diffusion denoising network is used to predict its noise residuals. The concept learning loss adopts the reconstruction error form of the diffusion model, which can be expressed as:

[0053] Used to characterize the difference between the predicted noise residual and the injected noise.

[0054] S315. Update only the concept placeholder embedding vector. Perform backpropagation based on the loss function, and update the embedding vector v corresponding to the concept placeholder. The iterative update can be described in the following direction:

[0055] Where η is the learning rate. In one implementation, the AdamW optimizer can be used, with momentum parameters β1=0.9, β2=0.999, and a weight decay coefficient of 10. 2 The numerical stability constant is 10. -8 The batch size per GPU can be set to 1 (meaning one GPU processes only one training sample at a time), and an equivalent batch size of 4 is formed through 4-step gradient accumulation; the base learning rate can be set to 5×10. 4 The warm-up steps can be set to 500, and the total training steps can be set to 3000.

[0056] In some embodiments of the present invention, the diffusion denoising module of the pre-trained audio diffusion generation model includes a cross-attention unit. The cross-attention unit maps the current latent variable features to a query matrix, and the condition vector to a key matrix and a value matrix; Attention features are calculated based on the query matrix, the key matrix, and the value matrix; The attention features are injected into the latent variable denoising process of the diffusion denoising module to generate denoised features corresponding to the target bird song concept.

[0057] In some embodiments of the present invention, the step of performing condition-guided inverse diffusion denoising based on the generation conditions to generate target latent variables includes: Initial latent variables are sampled from a standard Gaussian distribution; At each inverse diffusion time step, the current latent variable, the current inverse diffusion time step, and the generation conditions are input into the diffusion denoising module to generate the current noise prediction result; Based on the current noise prediction result, the current latent variable is denoised and updated to generate the latent variable corresponding to the next inverse diffusion time step; The latent variable obtained at the end of the reverse diffusion is determined as the target latent variable.

[0058] Specifically, in this embodiment, step S4 includes the following sub-steps: S411. Construct text generation prompts containing concept placeholders. During the inference phase, the input contains concept placeholders S. The text prompts form the conditions for generating the corresponding target bird call concept. Style control semantics such as "sharp alarmcall", "aggressive raptor call", and "short warning burst" can be added to the template to create more controllable and diverse bird-scare audio output.

[0059] S412. Initialize and generate latent variables. Sample initial latent variables from a standard Gaussian distribution as the starting point for the backdiffusion process. These initial latent variables do not contain explicit semantic information; the target bird call features are introduced step-by-step through the conditional vector and the backdiffusion process.

[0060] S413. Perform condition-guided stepwise backdiffusion denoising. In each backdiffusion time step, the current latent variable, time step encoding, and conditional vector are input into the backdiffusion denoising network to predict the noise components that should be removed in the current step, and gradually restore the latent variable representation of the target bird song concept. Through a cross-attention mechanism, the embedding vector corresponding to the concept placeholder continuously guides the generation direction throughout the backdiffusion process, thereby making the generated result acoustically approximate the target concept.

[0061] S414, Decoding yields personalized bird-repelling audio. The latent variables obtained after reverse diffusion are... The input is decoded and restored to a time-domain waveform signal, forming a personalized bird-scare audio output. While maintaining the consistency of the target bird's call concept, this audio can exhibit different durations, rhythms, and local acoustic details, which are used for subsequent audio decoding output.

[0062] In some embodiments of the present invention, after generating the bird-repelling audio corresponding to the target bird call concept based on the target latent variable, the method further includes: performing loudness correction, beginning and end smoothing, duration trimming, and artifact removal on the bird-repelling audio to generate candidate bird-repelling audio; performing target voiceprint consistency verification on the candidate bird-repelling audio based on acoustic feature distance, species identification results, or text audio matching scores; and archiving the candidate bird-repelling audio that has passed the target voiceprint consistency verification according to species category, call type, bird-repelling scene label, and playback parameters to generate a personalized bird-repelling audio resource library.

[0063] Specifically, in this embodiment, step S5 includes the following sub-steps: S511. Input the target latent variable obtained after inverse diffusion into the decoder to restore it as a time-domain waveform signal, forming a personalized bird-repelling audio output. This bird-repelling audio can exhibit different durations, rhythms, and local acoustic detail changes while maintaining the consistency of the target bird call concept.

[0064] S512. Perform loudness correction, beginning and end smoothing, duration trimming and basic artifact cleaning on the generated audio, remove abnormal samples with obvious breaks, metallic feel or non-target species pseudo-features, and retain candidate results suitable for device playback and engineering deployment.

[0065] S513. Perform target voiceprint consistency verification on the candidate results. Verify whether the generated audio retains the core voiceprint features of the target bird's song through methods such as acoustic feature distance, species identification accuracy, text-audio matching score, and human subjective listening. In one implementation, indicators such as BirdNET species identification accuracy, CLAP Score, CLAP-FD, and JSD can be used to assist in the evaluation.

[0066] S514. Generate a diverse set of bird-repelling audio samples. While maintaining conceptual consistency, generate multiple bird-repelling audio samples with different random seeds, different cue word templates, and different target durations to alleviate the auditory adaptation problems of birds caused by traditional fixed-play audio.

[0067] S515. Establish a personalized bird-repelling audio resource library. The selected synthesis results are archived according to species category, call type, bird-repelling scene tag, and playback parameters to form a continuously expandable personalized bird-repelling audio resource library for subsequent use in scenarios such as substations, farmland, and airports. When it is necessary to add new species or new call types, steps S1 to S4 are reused to reallocate independent placeholders for the new concepts and train their embedding vectors. Since different concepts correspond only to their respective embedding vectors, the training process for new concepts will not degrade the generation quality of existing concepts, thus supporting low-cost incremental expansion of the bird-repelling audio resource library.

[0068] After model training is completed, the pre-trained audio diffusion generation model is used to synthesize bird-repelling audio for the audio to be processed. Specifically, the same or corresponding text template construction and condition generation process as the training phase is performed on the audio to be processed to form a text generation prompt containing concept placeholders; based on the text generation prompt and the corresponding target concept embedding, the generation conditions for the target bird call concept are generated; the initial latent variables sampled from the standard Gaussian distribution are input into the trained diffusion denoising module, and the latent variable representation corresponding to the target bird call concept is gradually recovered during the inverse diffusion process; then, the target latent variables are decoded into time-domain waveforms by the audio decoding module, and the bird-repelling audio corresponding to the target bird call concept is output.

[0069] In one implementation, the bird-repelling audio can be a time-domain waveform or a playable audio file. The bird-repelling audio can be written into the audio library of a power system bird-repelling device, or sent to acoustic bird-repelling terminals deployed near transmission line towers, substation structures, distribution line crossarms, insulator strings, or other power facilities for playback. In this way, the generated bird-repelling audio not only retains the core acoustic signature of the target bird's call concept, but also generates playback content with differentiated duration, rhythm, and local detail variations according to different scenarios, thereby reducing the risk of birds adapting to fixed stimulus audio.

[0070] Through the above methods, this embodiment constructs a few-sample dataset of bird-repelling audio for the target bird call concept and establishes a mapping relationship between concept placeholders and text templates, enabling the target bird call concept to form a trainable concept expression under few-sample conditions. By freezing the main parameters of the pre-trained audio diffusion generation model, only the embedding vector corresponding to the concept placeholder is optimized, so that few-sample concept learning can accurately inject the fine-grained acoustic features of the target bird call concept while retaining the model's general generation capabilities. Furthermore, by performing inverse diffusion denoising and audio decoding output guided by the learned target concept embedding, target concepts such as aggressive calls, warning calls, deterrent calls, or fright calls can be generated and written into a personalized bird-repelling audio resource library, which is more suitable for acoustic bird-repelling applications in complex outdoor scenarios of power systems.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires bird-repelling audio samples corresponding to a target bird song concept, constructs a few-sample dataset of bird-repelling audio, assigns concept placeholders to the target bird song concept, and constructs a text template containing the concept placeholders, enabling the target bird song concept to form a trainable concept representation under limited sample conditions. Furthermore, by organizing the bird-repelling audio samples and text templates into a few-sample concept training set, a stable correspondence between the voiceprint information and text conditional information of the target bird song concept is achieved, reducing the dependence of traditional bird-repelling audio generation methods on large-scale labeled samples and improving the modeling ability of specific target bird song concepts in few-sample scenarios.

[0072] This invention inputs a few-shot concept training set into a pre-trained audio diffusion generation model to generate training latent variables and conditional vectors containing concept placeholders. By freezing all model parameters in the pre-trained audio diffusion generation model except for the embedding vectors corresponding to the concept placeholders, the embedding vectors corresponding to the concept placeholders are updated only based on the error between the noise residual and the injected noise. This allows fine-grained voiceprint features of the target bird call concept to be stably injected into the model while preserving its original general audio generation capabilities. Furthermore, through diffusion time-step sampling, noisy latent variable construction, and concept embedding learning loss constraints, the target concept embedding achieves accurate learning of target voiceprint features such as aggressive calls, warning calls, deterrent calls, or fright calls. This reduces overfitting and catastrophic forgetting problems caused by full-parameter fine-tuning, and improves the stability and generation consistency of few-shot concept learning.

[0073] This invention generates target bird call concepts by using text-based prompts containing concept placeholders and embedding target concepts to generate conditions for target bird call concepts. Based on these conditions, it performs condition-guided inverse diffusion denoising to generate target latent variables. Then, based on these latent variables, it generates bird-repelling audio corresponding to the target bird call concept, achieving controllable synthesis output of bird-repelling audio corresponding to a specific target bird call concept. Furthermore, by performing loudness correction, beginning and end smoothing, duration trimming, artifact removal, and target voiceprint consistency verification on the bird-repelling audio, and archiving the verified candidate bird-repelling audio according to species category, call type, bird-repelling scene tag, and playback parameters, it generates a personalized bird-repelling audio resource library. This significantly improves the target voiceprint consistency, playback adaptability, and resource expansion efficiency of bird-repelling audio, making it more suitable for personalized acoustic bird-repelling applications in complex outdoor scenarios of power systems.

[0074] Figure 2 This is a schematic diagram of a bird-scare audio synthesis system based on few-sample concept learning, provided by an embodiment of the present invention.

[0075] Example 2, as Figure 2 As shown, the present invention also provides a bird-scare audio synthesis system based on few-shot concept learning, comprising: a training set construction module S11, a latent variable generation module S12, a concept learning module S13, a condition generation module S14, and an audio synthesis module S15.

[0076] The training set construction module S11 is used to obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. The latent variable generation module S12 is used to input the few-sample concept training set into the pre-trained audio diffusion generation model to generate training latent variables and a condition vector containing the concept placeholders. The concept learning module S13 is used to freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder, predict the noise residual based on the training latent variables and the conditional vector, update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise, and obtain the target concept embedding. The condition generation module S14 is used to generate conditions for the target bird song concept based on the text containing concept placeholders and the target concept embedding, and to perform condition-guided inverse diffusion denoising based on the generation conditions to generate target latent variables. The audio synthesis module S15 is used to generate bird-repelling audio corresponding to the target bird call concept based on the target latent variable.

[0077] Example 3: The present invention also provides a bird-repelling audio synthesis device based on few-shot concept learning. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions. When the computer instructions are executed, the device implements the bird-repelling audio synthesis method based on few-shot concept learning.

[0078] Example 4, as Figure 3 As shown, the present invention also provides an electronic device 100 for implementing a bird-scare audio synthesis method based on few-sample concept learning.

[0079] The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0080] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the bird-repelling audio synthesis method based on few-shot concept learning as described in the first aspect of the present invention by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0081] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0082] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0083] The memory 101 in the electronic device 100 stores multiple instructions to implement a bird-scare audio synthesis method based on few-shot concept learning, and the processor 102 can execute multiple instructions to achieve the following: Obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. The few-sample concept training set is input into a pre-trained audio diffusion generation model to generate training latent variables and a conditional vector containing the concept placeholders; Freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder, predict the noise residual based on the training latent variables and the conditional vector, update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise, and obtain the target concept embedding. Based on the text generation prompt containing concept placeholders and the target concept embedding, the generation conditions for generating the target bird song concept are generated. Based on the generation conditions, condition-guided inverse diffusion denoising is performed to generate the target latent variable. Based on the target latent variables, generate bird-repelling audio corresponding to the target bird call concept.

[0084] Example 5: If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0085] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0086] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] 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.

[0088] 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.

[0089] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for synthesizing bird-scare audio based on few-shot concept learning, characterized in that, The method includes the following steps: Obtain bird-repelling audio samples corresponding to the target bird song concept, generate a bird-repelling audio few-sample dataset based on the bird-repelling audio samples, assign concept placeholders to the target bird song concept, construct a text template containing the concept placeholders, and generate a few-sample concept training set based on the bird-repelling audio few-sample dataset and the text template. The few-sample concept training set is input into a pre-trained audio diffusion generation model to generate training latent variables and a conditional vector containing the concept placeholders; Freeze the model parameters in the pre-trained audio diffusion generation model except for the embedding vector corresponding to the concept placeholder, predict the noise residual based on the training latent variables and the conditional vector, update the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise, and obtain the target concept embedding. Based on the text generation prompt containing concept placeholders and the target concept embedding, the generation conditions for generating the target bird song concept are generated. Based on the generation conditions, condition-guided inverse diffusion denoising is performed to generate the target latent variable. Based on the target latent variables, generate bird-repelling audio corresponding to the target bird call concept.

2. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The generation of a small sample dataset of bird-repelling audio based on the bird-repelling audio samples includes: The bird-repelling audio samples are subjected to target voiceprint detection. Audio samples with incomplete target calls, background noise exceeding a preset threshold, multiple species voiceprints mixed, or inconsistent species labeling are removed to obtain effective bird-repelling audio samples. The effective bird-repelling audio samples are processed by unifying the sampling rate, unifying the sound channels, normalizing the loudness, pruning silence, and fixing the segment length to generate the bird-repelling audio small sample dataset. The number of audio samples in the bird-repelling audio small sample dataset is 3 to 20.

3. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The construction of a text template containing the concept placeholders includes: Obtain the species tag, call type tag, and bird deterrence scene tag corresponding to the target bird call concept; The text template is generated by combining the concept placeholder with at least one of the species tag, the call type tag, and the bird deterrence scene tag; wherein, different target bird call concepts correspond to different concept placeholders, and the target bird call concepts include the target bird's aggressive call concept, warning call concept, deterrent call concept, or fear call concept.

4. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The pre-trained audio diffusion generation model includes an audio time-frequency representation conversion module, a latent variable encoding module, a text conditional encoding module, a diffusion denoising module, and an audio decoding module. The audio time-frequency representation conversion module converts the audio samples in the few-sample concept training set into audio time-frequency representations; The latent variable encoding module encodes the audio time-frequency representation to generate the training latent variables; The text condition encoding module encodes the text template to generate a condition vector containing the concept placeholders; The diffusion denoising module predicts the noise residual based on the training latent variables and the conditional vector; The audio decoding module decodes the target latent variable to generate the bird-repelling audio.

5. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The prediction of noise residuals based on the training latent variables and the conditional vector includes: Sample diffusion time steps from a preset set of diffusion time steps; Based on the noise variance parameter and signal preservation parameter corresponding to the diffusion time step, Gaussian noise is injected into the training latent variables to generate noisy latent variables. The noisy latent variable, the diffusion time step, and the conditional vector are input into the diffusion denoising module of the pre-trained audio diffusion generation model to predict the noise residual.

6. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The step of updating the embedding vector corresponding to the concept placeholder based on the error between the noise residual and the injected noise to obtain the target concept embedding includes: A concept embedding learning loss is constructed based on the difference between the noise residual and the injected noise; The embedding vectors corresponding to the concept placeholders are updated by backpropagation based on the concept embedding learning loss. During the update process, the model parameters in the pre-trained audio diffusion generation model, except for the embedding vector corresponding to the concept placeholder, remain unchanged; The updated concept placeholder corresponding to the embedding vector is determined as the target concept embedding.

7. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The diffusion denoising module of the pre-trained audio diffusion generation model includes a cross-attention unit; The cross-attention unit maps the current latent variable features to a query matrix, and the condition vector to a key matrix and a value matrix; Attention features are calculated based on the query matrix, the key matrix, and the value matrix; The attention features are injected into the latent variable denoising process of the diffusion denoising module to generate denoised features corresponding to the target bird song concept.

8. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, The step of performing condition-guided inverse diffusion denoising based on the generated conditions to generate target latent variables includes: Initial latent variables are sampled from a standard Gaussian distribution; At each inverse diffusion time step, the current latent variable, the current inverse diffusion time step, and the generation conditions are input into the diffusion denoising module to generate the current noise prediction result; Based on the current noise prediction result, the current latent variable is denoised and updated to generate the latent variable corresponding to the next inverse diffusion time step; The latent variable obtained at the end of the reverse diffusion is determined as the target latent variable.

9. The bird-scare audio synthesis method based on few-shot concept learning according to claim 1, characterized in that, After generating the bird-repelling audio corresponding to the target bird call concept based on the target latent variable, the method further includes: The bird-repelling audio is subjected to loudness correction, beginning and end smoothing, duration adjustment and artifact removal to generate candidate bird-repelling audio; Based on acoustic feature distance, species identification results, or text-audio matching scores, the candidate bird-scare audio is subjected to target voiceprint consistency verification. Candidate bird-repelling audios that pass the target voiceprint consistency verification are archived according to species category, call type, bird-repelling scene tag, and playback parameters to generate a personalized bird-repelling audio resource library.

10. A bird-scare audio synthesis device based on few-shot concept learning, characterized in that, The apparatus includes a computer device, which includes a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed, the apparatus implements the bird-scare audio synthesis method based on few-shot concept learning as described in any one of claims 1 to 9.

Images