Research On Sound Signal Recognition Algorithm Based On Few Shot Learning

In recent years,with the rapid improvement of computer processing ability,deep learning has flourished.Currently,deep learning has been widely applied in various fields such as underwater target recognition,biometric identification,autonomous driving,medical diagnosis and more.However,the prerequisite for these applications is the need for a large amount of annotated data to train neural networks.In some fields,the data requirements are extremely strict and even difficult to obtain,especially when the target is extremely rare.In addition,some data involving privacy cannot be easily obtained,and even if there is a large amount of data,labeling requires a lot of human and material resources.Therefore,scholars have proposed Few-Shot Learning,aiming to enable machines to learn like humans and achieve generalized learning,thus alleviating the pressure of the aforementioned problems.This paper will introduce relevant research on Few-Shot Learning in the field of sound signals,with the following content:(1)Due to the scarcity of annotated sound signal data,overfitting of models is often a problem.To alleviate this issue,this paper proposes a feature generation network that expands the diversity of features at the feature level.Furthermore,the generated feature vectors are subjected to non-parametric classification to make the generated features more classifiable.In addition,the network corrects the Gaussian distribution parameters of noise sampling using parameter correction algorithms,making the sampled data closer to the true feature distribution and providing the network with more prior knowledge.The experimental results show that compared to the current advanced few-shot learning algorithms based on data augmentation,this method achieves excellent performance on the clean and noise-free VCTK sound dataset.Moreover,it also achieves good results on the noisy Voxceleb 1 dataset.The model’s good generalization is demonstrated by its performance on different language datasets such as THCH-30,which did not result in a decrease in model performance.(2)Metric learning is one of the most common ways to solve the few-shot learning problem.Traditional features only consider marginal distribution information,while ignoring joint distribution information.The Brownian distance covariance proposed by scholars considers the latter.This paper improves on it by replacing the mean prototype with a weighted prototype to make it closer to the true distribution prototype.In addition,this paper proposes a new metric function that better measures the distance between features in the feature space.Experimental results show that compared with advanced metric learning-based few-shot learning,the proposed method achieves improvement in experiments on the noise-free dataset VCTK,the noisy background dataset Voxceleb 1,and the THCH-30 dataset.Moreover,the model exhibits good noise resistance in the experiments.